Framing Clinical Intervention Selection as a Constrained Optimization Problem
A Formal Framework for Achieving Optimal Patient Outcomes in Physical Rehabilitation
by Brent Brookbush, DPT, PT, MS, CPT, HMS, IMT
Introduction
Problems with our industry: Intervention selection is often "modality driven," based on practitioner preference, and rationalized by any improvement in an outcome measure, often selected post hoc. Too often, this measure is a subjective assessment of current pain, which is not a reliable predictor of short-term or long-term outcomes. I know this sounds harsh, but an examination of educational content throughout the industry and social media posts will quickly highlight an obsession with promoting or demonizing interventions, with almost no reference to a systematic approach, assessment-driven decisions, the relative efficacy of an intervention, comparative research, or reliable, objective outcome measures.
Intervention selection, sometimes referred to as clinical decision-making, is a topic that receives insufficient attention in college and university curricula, as well as in professional continuing education. Often, it receives no dedicated instruction during coursework, resulting in vague rationales and messy heuristics that fall apart when used in practice. What’s most disappointing is that logic, set theory, decision theory, and information science have advanced significantly over the past 80 years, largely due to technology, while our professions have largely ignored or failed to integrate these advancements.
The thought experiment that inspired this article: Imagine placing every physical rehabilitation technique ever conceived in a pile on a table; every modality, manual technique, and exercise from every profession (PT, ATC, DC, OT, DO, etc.). Which would you choose? Most of us would want to select the best possible techniques, or the best possible combination of techniques, within our scope of practice. But what does “best possible” actually mean? Too often, “best” is defined by personal preference, familiarity, or a profession's perceived specialty. Worse, any intervention that produces a positive effect is often considered valid, with little consideration for whether it was the best possible choice. This defaults clinical reasoning to “it worked for me” rather than “this would result in the best possible outcome.”
There is a better alternative: The central thesis of this article is that “best” is a measurable and modelable quantity. It is not merely a matter of subjective opinion or professional experience. In formal terms, the best possible intervention, or combination of interventions, can be determined using the principle of expected value: the product of reliability (frequency of positive outcomes) and effect size (magnitude of outcome). This shifts intervention selection from guesswork to a formal optimization problem, constrained by session length and grounded in comparative research, probabilistic modeling, and adaptive decision-making.
Abstract
Framing Clinical Intervention Selection as a Constrained Optimization Problem: Toward an Adaptive, Personalized Model for Physical Rehabilitation
Clinical decision-making in physical rehabilitation involves selecting from a set of interventions under strict time and resource constraints. While individual treatment modalities may be subjectively and individually preferred or dismissed, there is no widely adopted systematic framework for optimizing the full set of possible interventions that could compose a clinical session. This paper introduces a novel decision-theoretic model that conceptualizes intervention selection as a convergence-capable, multi-objective knapsack problem. Grounded in axioms derived from expected value theory and optimization logic, the model assumes that outcomes are probabilistic, choices are relative, and the effectiveness of each intervention can be operationalized as an expected value (frequency × effect size).
The framework incorporates five key principles: (1) constrained resource allocation, (2) diminishing marginal utility of sequential interventions, (3) adaptive optimization through empirical feedback, (4) convergence to optimal or near-optimal solutions over repeated sessions, and (5) subgroup-specific personalization through ongoing assessment. This model departs from traditional treatment planning by integrating both real-time learning (as in reinforcement learning paradigms) and population-level comparative effectiveness data (peer-reviewed and published comparative research. By unifying concepts from health economics, machine learning, and clinical reasoning, it provides a formalized structure for achieving maximal patient outcomes within session limits.
No existing framework in medicine, rehabilitation, or performance training explicitly combines constrained intervention bundling, probabilistic modeling, dynamic adaptation, and subgroup-specific optimization. As such, this model offers a novel and testable foundation for clinical decision support, AI-driven care planning, and systems-level research in allied health disciplines.

Quick Summary
Primary Hypothesis
There exists an objectively measurable, optimal set of interventions that, on average, maximizes patient outcomes within a given treatment session. This optimal set can be identified by modeling intervention selection as a constrained optimization problem, one that prioritizes interventions by their expected value and accounts for diminishing returns, intervention synergy, and individual variability.
This hypothesis can be operationalized using the following methodology:
- Model Development from Comparative Research: Build an intervention selection model using comparative research to assign expected value estimates to intervention categories and individual techniques. These values should reflect the probability and magnitude of outcome improvement. Intervention categories should be reorganized and labeled to reflect shared mechanisms or outcome targets, facilitating model construction and synergy mapping.
- Subgroup-Specific Prioritization via Assessment: Use assessment findings to stratify patients into subgroups that consistently respond best to different intervention sets. These subgroup distinctions should guide adaptive reprioritization of interventions, improving model specificity and relevance to individual patients.
- Iterative Reassessment and Bayesian Updating: Employ a structured process of assessment, intervention, and reassessment to empirically update the expected value of each intervention for a given patient. This real-time feedback loop enables Bayesian refinement of the model, improving the precision of future intervention sets.
- Exploratory Allocation for Model Expansion: Allocate a small portion of each session to testing underutilized or emerging interventions. The goal is to empirically identify techniques that outperform their prior expected values, allowing the model to evolve and incorporate higher-yield strategies over time.
Axioms Defining the Formal Optimization of Intervention Selection
- Outcomes are probabilistic. The effect of an intervention on clinical outcomes is inherently probabilistic. Accordingly, clinical decisions should be based on the likelihood and magnitude of observable outcomes, rather than on theoretical mechanisms or the practitioner's intent.
- Choices are relative. Assuming all candidate interventions produce outcomes superior to doing nothing, intervention selection becomes a problem of relative effectiveness. The value of any intervention must be considered in comparison to available alternatives.
- “Best” is a measurable quantity. Effectiveness can be defined as expected value, calculated as the product of the probability of success (reliability) and the magnitude of effect (effect size). For each intervention, the expected value is determined by multiplying its estimated frequency of producing a favorable outcome by the size of the effect it produces.
- Intervention selection is constrained (zero-sum). The finite duration of a clinical session imposes constraints on the number or duration of interventions that can be delivered. Consequently, intervention selection becomes a zero-sum game: selecting one intervention often requires excluding another. This creates a need to maximize expected value within limited resources.
- Maximizing expected value maximizes expected outcomes. If the total expected outcome of a session is defined as the sum of the expected values of selected interventions, the optimal intervention set is the one that maximizes this total under session constraints.
- Expected values should be initialized using peer-reviewed, comparative research. Intervention selection should begin with population-level data derived from peer-reviewed, comparative studies. These data provide initial estimates of expected value by quantifying the average reliability and effect size of each intervention in relevant patient populations, and demonstrating its relative effectiveness compared to one or more additional interventions.
- Deliberate experimentation is necessary to avoid local maxima. To ensure that intervention selection does not plateau at suboptimal combinations, a portion of clinical sessions should be allocated to structured experimentation. Testing novel interventions or combinations allows for the discovery of higher-value alternatives and supports adaptive learning over time.
- Diminishing marginal utility governs the value of additional interventions. As more interventions are added to a treatment session, the additional benefit (marginal utility) provided by each successive intervention typically decreases. This principle supports the prioritization of high-value interventions early in the decision process and recognizes that total expected outcome gains may plateau or decline when too many interventions are applied, especially under time constraints.
- The system converges toward optimal intervention sets over time. Given a bounded set of candidate interventions, repeated sessions, and ongoing empirical feedback, an adaptive system will asymptotically converge to a set of intervention combinations that maximize expected value within session constraints. While multiple optimal solutions may exist, such a system will reliably select only from within this optimal set. With the addition of deterministic tie-breaking rules or selection heuristics, the system can converge on a consistent, unique intervention policy.
- Assessment enables subgroup-specific, multi-objective optimization. Assessment allows for the stratification of patients into subgroups, each of which may have distinct expected value functions for the same interventions. This transforms the problem from a single-objective knapsack model into a multi-objective optimization problem, where the goal is to select the intervention set that best fits the unique outcome priorities and response patterns of each subgroup.
Formal Optimization Model for Intervention Selection
Let:
- xi ∈ {0, 1}: Binary decision variable indicating whether intervention i is selected.
- di: Duration or resource cost of intervention i.
- pi: Probability of success for intervention i.
- mi: Magnitude of effect for intervention i.
- EVi: Expected value of intervention i, defined as EVi = pi ⋅ mi
- t: Total time or resource budget per session.
- n: Number of available interventions.
Expected Value Definition
EVi = pi ⋅ mi
- This expresses the effectiveness of an intervention as the product of its reliability and the size of its benefit, reflecting Axioms 1–3.
Zero-Sum Constraint
∑i=1n xi ⋅ di = t
- This ensures that the total duration of selected interventions does not exceed the session constraint, consistent with Axiom 4.
Objective Function: Maximize Expected Outcomes
Maximize ∑i=1n xi ⋅ EVi
- This is a classic 0/1 knapsack optimization model that maximizes the total expected value of selected interventions under constraints, reflecting Axiom 5.
Extension Concepts
Diminishing Marginal Utility: While not easily written symbolically without a utility function, you can model this using a concave utility function U(S), where S is the set of selected interventions. For instance:
- Additional interventions yield smaller incremental gains.
- Could be approximated using penalty functions or diminishing weights per added intervention.
Bayesian or Empirical Updating: Not symbolically shown here, but conceptually:
- At each session t, update pi and mi based on observed outcomes (e.g., via Bayesian posterior updating or empirical moving average).
- This implements adaptive learning (Axioms 6–9).
Multi-Objective and Subgroup Optimization: Let EVig be the expected value of intervention i for subgroup g. Then:
Maximize ∑i=1n xi ⋅ EVig
- This reflects Axiom 10: subgroup-specific expected values define multiple objective functions depending on the patient classification.

Section 1: There is A Best Approach
Outcomes are probabilistic, not deterministic.
Practitioners do not perform an intervention and know the effect on outcomes with certainty. The relationship between interventions and outcomes is inherently uncertain and best described as a probability of effect. Interventions are often taught deterministically, as if a specific technique will reliably produce a specific outcome. In reality, this is either an oversimplification of a complex interaction or a misunderstanding. All interventions yield outcomes that vary in both reliability (the likelihood of success) and effect size (the magnitude of benefit), the very components that define expected value in clinical decision-making.
Assertions of causation are hypotheses, and these only influence outcomes insofar as they influence practice. That is, the impact of a hypothesis on outcomes is not inherent to the idea itself, but depends on how it changes clinical behavior (e.g., intervention selection).
The phrase "correlation is not causation" is often repeated but rarely unpacked. Demonstrating causation is not as simple as this cliché implies. In media and casual discourse, it’s often suggested that correlation is a shortcut, and a little extra effort would uncover the “true” causal link. This is misleading. While correlations are relatively easy to demonstrate, a causal relationship is generally accepted only when a hypothesis consistently explains observed outcomes and enables more accurate predictions than alternative explanations.
It would be more accurate, although perhaps incomplete, to say that the convergence of multiple consistent correlations supports causation. It's also worth noting that attempts to “prove” causation often devolve into a slippery slope argument, where each answer only reveals deeper gaps in our understanding.
This complexity reinforces a core principle: the relationship between interventions and outcomes is fundamentally probabilistic. Even if we don’t always calculate exact probabilities, recognizing this uncertainty is essential. It demands that we shift away from dogmatic claims of effectiveness and instead adopt a framework grounded in expected value and constrained optimization. Clinical outcomes cannot be predicted with certainty; intervention selection must reflect that truth.
Critique Is Not Binary
The efficacy of an intervention does not compete with doing nothing; it competes with the efficacy of all other interventions available within the limited time of a clinical session. When a technique is critically evaluated by a professional, the decision is not binary—whether the technique should or should not be used—unless the technique has an expected value less than or equal to doing nothing. Techniques less effective than doing nothing are likely rare due to survivorship bias.
Survivorship bias suggests that the techniques still in use today represent only those that have "survived." Interventions that consistently worsen outcomes are unlikely to be used beyond a practitioner’s tolerance for experimentation and are unlikely to spread professionally, as clinicians tend not to share ineffective strategies. Of course, some techniques may also fail to persist due to complexity, external competition, or a lack of dissemination, but these issues will be addressed later in the section on experimentation.
The majority of techniques used in rehabilitation have likely demonstrated at least some benefit beyond placebo—even if limited to short-term symptom relief for certain patients. At minimum, more techniques have demonstrated efficacy in randomized controlled trials (RCTs) than can reasonably be applied within a single session. This implies that critical evaluation should focus on how effective a technique is relative to the other interventions that could be chosen instead.
Unsupported Default Position Fallacy
Another common reasoning error must be addressed: the unsupported default position fallacy. This fallacy is the belief that identifying a flaw in an opposing view automatically strengthens one’s position. In short: “Proving you wrong makes me right.” This is fallacious unless there are only two mutually exclusive options, and one of them must be correct. That is rarely, if ever, the case in physical medicine, and certainly not when selecting among interventions.
When more than two solutions are possible, or when multiple options may be simultaneously flawed, each position must be evaluated on its own merits. Demonstrating flaws in one approach does not excuse or validate the alternative; it still must be shown that the other approach is less flawed or has greater expected value.
This fallacy is particularly rampant on social media, where critics often dismiss interventions without recognizing that the techniques they endorse may suffer from equivalent or even greater shortcomings. A good example is seen when proponents of Pain Neuroscience Education (PNE) critique biomechanical or postural dysfunction models. The implication is that if biomechanical models are flawed, then PNE must be superior. However, randomized controlled trials comparing PNE to movement-impairment-based approaches have consistently shown PNE to be relatively ineffective (See PNE Research Review ) (1).
Models of movement impairment and posture require refinement, just as every scientific domain does. Medicine is an evolving field. Identifying flaws is essential to progress, but doing so does not imply that another model is automatically better. In summary, even if a description or implementation of an intervention is flawed, if it yields the highest expected value, it remains the optimal choice. The goal is to select the most effective intervention based on measurable outcomes. The goal is not to find a flawless intervention; such an option may not even exist.
“Best” is a Measurable Quantity
Effectiveness is a measurable quantity.
The primary hypothesis asserted by this paper is that there is an attainable best treatment approach that our professions should strive to achieve. This “best” approach may be defined objectively and mathematically, rather than subjectively, using reliable, outcome-based measures. Furthermore, the best outcomes may be achieved through the optimal selection of interventions. It is important to note that “best outcomes” refers to the best average outcomes across all patients.
The most influential contributors to average outcomes are likely the reliability and magnitude of an intervention’s effect. These can also be described as frequency (how often the intervention is effective) and value (the size of the effect). Their product is known as the expected value. The formula, expressed, is: Frequency (reliability) x Value (effect size) = Expected Value (effect on outcomes). Expected value, a concept commonly used in game theory and economics, helps address the challenge of comparing interventions that differ in both consistency and impact (2). For example, without this formula, it would be difficult to objectively compare a technique that is highly reliable but produces small effects to one that produces large effects but is rarely effective. By using the product of these two variables, we can more clearly evaluate interventions and prioritize those most likely to improve average outcomes.
Practitioners often fall prey to the availability heuristic, giving undue weight to memorable outcomes, such as dramatic successes or failures, while ignoring the frequency of those outcomes. This bias is particularly prevalent among professionals who are committed to a specific modality or method. For example, a technique that produces a remarkable result for one patient may only achieve this outcome 1 out of 10 times. Despite the strong anecdotal impression, its average effectiveness may be lower than that of more reliable alternatives. Note, however, that such a technique could become an excellent choice if assessment tools can identify which patients fall into the “1 out of 10” subgroup (this will be discussed further below).
A common example of this issue in physical rehabilitation is practitioners who become Graston-certified and begin treating a wide range of conditions with instrument-assisted soft tissue mobilization (IASTM). However, our systematic research review on IASTM suggests that it has lower efficacy than more specific manual techniques, such as ischemic compression, dry needling, joint mobilizations, and joint manipulations (3). This does not mean that IASTM should be excluded from practice. Rather, it implies that IASTM should be used strategically, often in conjunction with other manual techniques, to enhance mobility further when time allows (discussed below).
Why Can’t We Base Selections on the Intended Effect of an Intervention?
Outcomes depend on the variables we can modify, rather than on our understanding of how those variables affect outcomes (mechanism of effect). The Bradford Hill Criteria of Causality, published in 1965, documented this interesting logical distinction. The Bradford Hill Criteria outline several factors that strengthen a hypothesis of a causal relationship (4, 5). One criterion is a supportable causal hypothesis; however, Bradford Hill also notes that it is not necessary to know how a variable affects an outcome to know that it does, in fact, affect an outcome. That is, knowing how something works is not required to know that something works.
Some everyday examples of this logic include making a phone call without understanding telecommunications, reducing a headache with aspirin without knowing its pharmacodynamics, or reaping the benefits of resistance training without understanding exercise physiology. Similarly, in clinical practice, modifying a variable may improve an outcome regardless of whether the practitioner understands or correctly identifies the underlying mechanism of effect. The mechanism is a hypothesis about why a change occurs, while modification of variables focuses on what reliably produces better outcomes.
Further, a patient can benefit from the effects of an intervention even if both the patient and practitioner believe in an inaccurate causal explanation. For example, cervical manipulations may reduce pain even if both parties believe the benefit is due to the correction of a “subluxation.” In fact, research has demonstrated that patient expectations, practitioner preference, and false narratives do not influence the effectiveness of manual therapy interventions (see: False Narratives, Nocebo, and Negative Expectations do NOT affect Manual Therapy Outcomes: Research Confirmed ) (6).
It’s important to understand that hypotheses do not affect outcomes unless they affect behavior, specifically, intervention selection. Therefore, we must base our choices on measured outcomes, not on intent or proposed mechanisms. This is not to say that mechanisms of effect are irrelevant; their value lies in generating new hypotheses or helping us identify additional variables worth modifying. In summary, outcomes improve when we modify variables in a way that increases effectiveness. Hypothesized mechanisms are useful only if they lead us to discover or refine the way we modify those variables.
Professional Title Does Not Affect Outcomes
A professional designation is not a variable that directly affects patient outcomes. For example, we would not expect a chiropractor and a physical therapist to achieve vastly different results from performing the same joint mobilization technique. While it may be argued that one profession is more skilled in a particular intervention, this argument pertains to individual proficiency, not to the profession itself as a factor that modifies outcomes.
Even if we consider skill level as a meaningful variable, research suggests it is unlikely to significantly influence the average outcome of an intervention relative to other available interventions. This means that variations in skill may have a minimal impact on how interventions should be prioritized or how the optimal intervention set is selected. In short, the best approach is determined by the relative effectiveness of interventions, not by professional title.
This insight has significant implications for the future evolution of our field. “Scope wars” between professions are counterproductive. Every clinician treating a particular patient population should have access to the most effective interventions for that population. Scopes of practice should be informed by the optimal combinations of interventions required for optimal outcomes for a patient population, not by protectionist agendas advanced by professional organizations.
Furthermore, when professionals treat similar patient populations, they should use similar treatment strategies. All patients deserve access to the best approach, regardless of the provider’s credentials. Divisions between professions should instead reflect distinct patient populations with different needs. For example, orthopedic outpatient care and neurological rehabilitation likely require different intervention sets and, therefore, distinct areas of specialization. Considering these issues, outpatient-focused PTs, ATCs, and DCs should not function as three separate professions; however, it may be appropriate to distinguish between physical therapists who specialize in orthopedics versus those who specialize in neurological rehabilitation, or between those working in outpatient versus inpatient care settings.
Chance of Multiple Best Solutions
It is theoretically possible that more than one “best approach” exists; however, principles from Bayesian probability suggest that this scenario is exceedingly unlikely. If the best possible approach is defined as the combination of interventions with the highest total expected value, where expected value is the product of reliability and effect size, then two distinct intervention sets would have to produce statistically indistinguishable and maximally effective outcomes. This would require equivalence not only in the expected value of individual interventions but also in their cumulative and interactive effects. As the number of interventions in each set increases, the probability that two distinct combinations will result in the exact same best outcome decreases dramatically. This is analogous to the unlikely event of two players holding different, but equally winning, hands in a multi-player poker game.
A potential exception to this improbability would be the existence of threshold effects, points at which additional intervention no longer leads to further improvement. Threshold phenomena have been documented in resistance training research, such as findings that 5, 6, 7, and 8 sets per muscle group per session result in similar improvements in muscle growth (hypertrophy). However, comparable threshold effects have not been consistently demonstrated in physical medicine. Interventions in this domain often result in a wide range of outcomes, ranging from rapid symptom resolution to prolonged recovery. Even if two treatment plans produce similar clinical outcomes due to a shared threshold, other differentiating variables, such as patient discomfort, session duration, or the ability to incorporate it into a self-management routine, would likely affect their overall expected value.
In summary, while more than one treatment strategy can result in clinically meaningful results, the probability of two distinct approaches producing the same best outcome is vanishingly small. This reinforces the core premise: there is most likely a single best approach for any given patient population, defined by the combination of interventions with the highest cumulative expected value.
Intervention Selection is a Zero-Sum Optimization Problem
Intervention selection is a zero-sum game because the number of interventions selected will always be limited by the amount of time available in a session. A zero-sum game is a scenario in which one choice is selected at the expense of another. In decision theory, this means that the inclusion of one intervention necessarily excludes others. Although more than one technique may be implemented per session, we can think of a “set” as the total number of interventions that can reasonably be performed within the constraints of a session. Once that set is full, adding a new intervention requires removing one or more existing interventions. Similarly, if two systems of care are being compared, each set would replace the other because both cannot be implemented within the same time frame. If this were not true, there would be no need to compare interventions because an infinite number of interventions could be combined.
Research Supports That the “Best Intervention” Is a Set
Nearly all available research comparing a single intervention to combinations of interventions supports the notion that a combination of interventions is more effective than any single technique performed in isolation. For example, the combination of manual therapy and exercise consistently outperforms either alone. However, this does not imply that any additional technique improves outcomes. Obviously, relatively ineffective techniques may not add to the expected value of treatment. Techniques with similar effects may also result in no additional value. For instance, combining joint mobilizations and a joint manipulation for the same segment. Furthermore, the diminishing marginal utility of additional techniques for a particular clinical goal may suggest that adding techniques beyond a certain number is not worth the cost. These issues can be addressed by selecting techniques based on expected value, avoiding redundancy through the selection of techniques based on categories, and perhaps limiting the number of techniques intended for a specific clinical goal, all of which are discussed below.
More on Sets: Time, Optimization, and Efficiency
The number of interventions included in a full set is not absolutely fixed and should be optimized based on expected value. Parameters such as session length, frequency (e.g., sessions per week), and total number of sessions can be adjusted, with diminishing marginal utility predicting an upper limit to the total session time. In summary, research and clinical practice must consider the efficiency of sessions, which will be limited by marginal utility, and can be mathematically expressed as the outcome (expected value) achieved per unit of time. Note that additional considerations may include patient tolerance, financial resources, and dependence on third-party payers. Furthermore, it is worth noting that for most practices, once a session length is determined, it remains relatively fixed due to various business concerns that would render a variable session length unmanageable.
In this context, determining the optimal session length becomes a parameter that should be initially considered for adjustment. The length of a session should be designed to match the upper limit of time required for the majority of optimal sets for the clinic's patient population. As mentioned above, the upper limit would be set by diminishing marginal utility, ensuring that additional time results in meaningful improvements in outcomes per unit of time. If a 45-minute session is sufficient time for the gross majority of sessions to achieve the highest expected value, lengthening it to 60 minutes is likely not worth the additional value, and less time is likely to decrease efficiency and potentially increase the number of required sessions.
Just as the optimal intervention set should determine session length, it should also be used to determine the techniques within a professional's scope of practice. If the optimal set includes movement assessment, dry needling, joint manipulation, exercise instruction, kinesiology taping, and a home exercise program, then this should be the scope of a single provider who can perform these techniques during a single session. Fragmenting that set across multiple sessions with multiple providers reduces efficiency, increases time and cost (e.g., transportation, wait times, documentation), and may degrade outcomes by disrupting additive effects between techniques.
Implications for Scope of Practice and Education
This logic has profound implications for the evolution of our professions. Scopes of practice should be designed around the optimal combination of interventions for specific patient populations, not the protectionist agendas of professional organizations. Similarly, clinical education should match the set of interventions that yields the highest expected value for those populations. Professionals treating the same population should use the same best approach. For example, outpatient-focused PTs, ATCs, and DCs should not function as three separate professions. Instead, scopes of practice and training pathways should be reorganized to reflect population-based optimal treatment sets. A logical approach may be to consolidate similar professions into a single doctorate of physical medicine (e.g., combining PTs, DCs, ATCs, OTs, and acupuncturists), with population-based specialization introduced through elective tracks and clinical affiliations, similar to the model used in physician education. This reorganization must be dynamic and evidence-informed. As comparative outcomes research and real-world experimentation continue to identify better-performing combinations of interventions, scopes of practice, session structures, and education must be continuously updated to reflect the current best set.
Prioritizing Based on “Best Interventions” Will Result in “Best Outcomes”
When multiple interventions are considered, the expected outcome of a session is determined by summing the expected values of each intervention. Prioritizing interventions with the highest expected values (again, defined as the product of an intervention’s reliability and effect size) yields the most favorable outcomes. Because expected value is relative and intervention selection is constrained by limited session time, ordering interventions by expected value is essential to optimizing results.
Nobody Wants the Second-Best Option
Although individual patients may occasionally respond better to an intervention other than the one with the highest average expected value, no one would intentionally start with a second-best choice. This is best illustrated with a thought experiment: Imagine you have two techniques to choose from. One is the most effective for 70% of patients, while the other is best for 30%. Without additional information, no rational practitioner would start with the 30% option. There is no reason to assume a given patient is a “different responder” from the outset. And if the 70% intervention proves ineffective, the 30% option remains available for subsequent sessions.
This point becomes even more compelling as additional techniques are introduced. Consider a scenario with four possible interventions, with the probability of being the best choice for 55%, 30%, 10%, and 5% of patients, respectively. If a practitioner starts with the 5% and 10% options, the probability of selecting the best intervention is only 15%. Conversely, starting with the 55% and 30% options raises the probability of achieving the best outcome to 85%. Prioritizing higher expected value interventions ensures that most patients receive the most effective care in the fewest number of attempts.
Although research provides the most accurate data for determining relative efficacy, a degree of trial and error is necessary in practice. That is, assessment, intervention, and reassessment should be used to refine treatment based on the individual response. However, these efforts should be anchored in a rank-order approach informed by peer-reviewed comparative research. This strategy maximizes efficiency and ensures that each patient begins with the highest-probability interventions available.
The Problem with Current Practice: Random Sorting and Practitioner Preference
If it is not standard practice to prioritize interventions based on expected value, then the order of techniques performed is likely, at least in part, to be random. This logic is similar to expecting six dice to naturally arrange themselves from highest to lowest simply by rolling them, or even just half of them. Even if three dice had known values and could be ordered in advance, rolling the other three is unlikely to result in an optimal final arrangement. A similar situation occurs in clinical practice: some techniques may be correctly prioritized based on known higher expected values, but the remainder are ordered arbitrarily due to lack of knowledge, or perhaps lack of effort. Importantly, this lack of knowledge is often not due to a lack of research. Far more comparative outcome data exists than is currently being applied in practice (as discussed further below).
It may be argued that the random sorting problem is less concerning than the issue of practitioner preference. Many professionals choose interventions based on their allegiance to a particular modality or “school of thought,” or on their comfort with specific intervention types. For example, comparative research indicates that manual therapy is likely more effective than exercise during the acute phase of orthopedic pain, and that combining manual therapy with exercise yields better results than either treatment alone (see: Active vs. Passive... ) (7). Yet some clinicians refuse to perform manual therapy, citing allegiance to "active" approaches. As previously noted, some prioritize IASTM over specific manual techniques, despite research suggesting that specific manual techniques produce larger improvements in outcomes (3). Perhaps most obviously, therapeutic ultrasound (US) is often used in practice despite being demonstrably less effective than manual therapy or exercise techniques that already consume the majority of a session. Choosing to spend session time on US necessarily reduces the time available for more effective interventions (1).
It is, of course, possible that these lower expected value interventions will be the most effective choice for a specific patient. However, practitioners must be cautious not to fall victim to the availability heuristic, assuming that an intervention is broadly effective simply because of one or two memorable successes. The goal is not to find a rare match between a low-value intervention and a rare high-responder. The goal is to deliver the highest overall success rate across all patients. If one of these lower expected value interventions is, in fact, the best choice for a given patient, and the practitioner is prioritizing based on expected value, then the technique will eventually be tried. However, starting with the interventions that have the highest expected values ensures that the greatest number of patients receive the best outcomes within the fewest number of interventions.
What About Patient Preference?
Although many issues of patient preference can be addressed respectfully with education (e.g., “I understand that you enjoy manual therapy, but for this issue, this exercise is more effective”), some preferences warrant genuine consideration. For example, certain patients may express strong fear of joint manipulation due to false beliefs about high risk. While these beliefs may be inaccurate (see: “Risk of Adverse Events ”) (12), their psychological impact still matters. Research suggests that low to moderate stress does not significantly alter manipulation outcomes, but high levels of stress may reduce their effectiveness (13). This implies that if a patient is genuinely frightened of manipulation, it should not be performed. That said, patient preference does not alter the methodology for determining the optimal treatment plan. In this scenario, the next best intervention, likely a joint mobilization, should be selected to preserve the highest expected value possible under these circumstances. In short, patient preference may shape the final intervention selection, but it does not change the optimization process used to identify the best available strategy.
Issues with the Prioritization by Expected Value Approach
Determining the best combination of interventions should begin by prioritizing techniques with the highest expected values. However, this approach presents two potential challenges.
The first issue is the risk of redundancy when selecting techniques with high expected values. Research will likely demonstrate that the first- and second-ranked techniques belong to the same category and produce similar effects. For example, thoracic screw manipulation and thoracic pistol manipulation may both be highly effective for improving thoracic mobility. While both techniques may be independently effective, performing them sequentially is unlikely to provide a significant additive benefit compared to selecting one and following it with a different, yet still effective, intervention from another category, such as ischemic compression or a specific exercise. In other words, interventions that are too similar may result in diminishing marginal utility when used together.
The second issue is that some combinations of techniques may produce greater additive effects than others, even if they do not follow a strict rank order based on individual expected values. This accounts for the potential for synergistic relationships among interventions. For example, the top-ranked interventions for cervical pain might include joint manipulations, dry needling, IASTM, and manual release—all manual therapy techniques targeting mobility. However, the optimal combination for outcomes may instead include dry needling, joint manipulation, a home exercise program, and kinesiology taping, representing a balanced approach across the categories of mobility, stability, and self-management.
These issues can be addressed by improving the sorting, labeling, and modeling of intervention strategies. Many techniques are variations of a shared underlying approach. If interventions were properly categorized based on type and mechanism of effect, clinicians could then prioritize the most effective technique within each category. The optimal intervention set would comprise the top techniques from the most effective categories, rather than simply the top individual techniques. For example, in the thoracic manipulation scenario, selecting the most effective thoracic manipulation followed by the top technique from the next highest-value category is likely to yield better outcomes than using two similar manipulations in succession.
Identifying synergistic combinations across categories may be refined through clinical practice and experience. However, the relationships between assessment findings, intervention categories, and outcomes suggest a need for more structured modeling of intervention plans. This concept will be explored further in a subsequent section.
The Goal of Assessment: Decision Theory
From the perspective of decision theory, the goal of an assessment is to differentiate patient populations into subgroups that achieve optimal outcomes from different intervention plans. That is, if an assessment can be used to identify a subgroup of patients who respond better to an intervention that differs from the intervention with the highest expected value based on general population data (e.g., different responders), then interventions may be reprioritized to optimize the expected value for that subgroup. Further, this is also likely to increase the average expected outcome for the original group, whose averages may have been negatively affected by the lower-than-average results exhibited by the “different responders.” The better an assessment is at accurately identifying a subgroup, the more assessments that can be developed to identify additional subgroups, and the better the prioritization of interventions (or categories of interventions) based on expected value, the more likely it is that expected outcomes for all patients will increase. The upper limit to the division of subgroups would be constrained by the number of assessments that result in meaningful reprioritization of techniques based on expected value.
From an optimization perspective, assessments serve to modify the expected value function applied to each intervention within a constrained knapsack, enabling more personalized and more efficient maximization of outcome per unit of session time.
Referring to some of the examples above, if it is possible to use an assessment before treatment to identify the 30% of patients who exhibit a larger effect size from an intervention that does not generally result in the best outcomes (different responders), then that intervention may be prioritized higher for those patients. This reduces the number of patients who would have to wait for the technique to be attempted second. A common example of this idea is the assessment of a range of motion, in which the identification of hypomobility results in the practitioner performing mobility techniques, and the identification of hypermobility results in the practitioner performing stabilization techniques. The identification of accurate and relevant assessments may also address some of the issues with allegiance to modalities based on a few remarkable results (e.g., availability heuristic bias). If an assessment can differentiate the subgroup of remarkable responders, it may be possible to increase the reliability of the intervention several times, increasing the expected value of the technique, which would result in reprioritization of the technique for those remarkable responders.
Assessment Issues
Assessments should not be performed unless they lead to a reprioritization of interventions that improve outcomes. If an assessment does not influence intervention selection—by modifying expected values based on more specific or individualized data—it should not be included in the session. Assessments that do not change the intervention plan have no utility. Importantly, identifying "red flags" for referral is an exception, as this represents a fundamental change in the treatment plan. In this context, red flag assessments divide the patient population such that the optimal intervention still exists, but should not be performed by the current provider.
Clinicians should routinely ask:
- “What will I do if I get a positive result from this assessment?”
- “What will I do if I get a negative result from this assessment?”
If the answers to both questions are the same, the assessment is unlikely to affect intervention prioritization and should be discarded.
Additionally, short-term subjective measures of pain and symptoms (e.g., pain intensity scales, verbal reports, and patient history) tend to lack specificity. Although reducing symptoms is a core goal, the symptoms themselves are often poor predictors of which interventions will produce the best long-term outcomes. This is evident in the strong short-term effects of placebo on pain, despite having little to no effect on objective, functional outcomes. For example, lumbar effleurage may reduce pain intensity temporarily, yet have no lasting impact on motor control deficits or movement impairments that correlate with recovery.
The inherent imprecision of the sensory system further limits subjective measures. Even highly trained professionals cannot reliably identify underlying pathology based solely on symptom location or quality. For instance, distinguishing between a sore throat caused by post-nasal drip, a virus, strep throat, or cancer based solely on symptom reports is nearly impossible—yet each requires a different intervention. Similarly, a patient reporting pain resolution after treatment for low back pain sustained during a basketball collision may seem improved, but if they do not return to sport, the functional goal has not been achieved.
Ultimately, the value of an intervention should be based on its impact on outcomes, rather than the strength of the correlation between the targeted factor and the pathology. This distinction is often misunderstood. The magnitude of change that an intervention can produce on a correlated factor—and how that change translates into improved outcomes—is more important than the correlation itself. This mirrors the classic economic question: “Would you rather have 70% of $50 ($35) or 25% of $100 ($25)?”—illustrating how relying solely on percentages or correlations without considering absolute outcomes leads to suboptimal decisions.
In physical medicine, this fallacy is evident in the misapplication of research focused on pain-related factors, rather than studies directly comparing interventions. For example, while biopsychosocial factors like kinesiophobia are highly correlated with pain, interventions such as pain neuroscience education (PNE) have consistently shown weak effects on outcomes (see: PNE Research Review ) (1). Even if cognitive factors show stronger correlations with pain than biomechanical issues, the current tools available to treat cognitive contributors may be so ineffective that targeting moderately correlated—but highly modifiable—biomechanical impairments yields better overall outcomes.
Defining Which Outcome Measures We Should Address
The best approach is likely to consist of selecting interventions with the highest expected value based on their carry-over effects on treatable factors correlated with both short-term and long-term outcomes. Identifying reliable, accurate assessments for these treatable factors is essential, as they support the refinement of intervention selection within a session. For example, if a technique intended to improve range of motion (ROM) results in no measurable change on a goniometric assessment, the clinician can immediately pivot to an alternative technique. This reflects the “assess–address–reassess” strategy—an applied form of trial-and-error through a rank-ordered list of interventions prioritized by expected value.
Ideally, these assessments would be reliable, objective outcome measures for treatable factors most strongly correlated with desirable outcomes. However, the research base is incomplete. Not all treatable correlated factors have clearly defined assessments, and the existing literature disproportionately favors certain regions of the body. For instance, there is substantially more research guiding shoulder assessment than for the elbow. Likewise, some meaningful impairments lack direct measurement tools. For example, while excessive anterior scapular tipping has been associated with shoulder impingement syndrome (SIS) (8), there is no standardized goniometric method to quantify anterior or posterior tipping of the scapula. Despite these limitations, a considerable body of research remains underutilized in guiding clinical decision-making. For example, the Overhead Squat Assessment has demonstrated reliability in detecting signs of dysfunction (e.g., knee valgus), and reducing valgus has been shown to decrease knee pain and/or the risk of future injury.
Furthermore, short-term improvements do not always align with long-term outcomes. Although this issue is more pronounced with subjective measures, even objective short-term changes may not reflect the most effective long-term strategies. One way to mitigate this limitation is to evaluate the carry-over effect of an intervention—that is, the extent to which an in-session improvement is retained at the beginning of the following session. If intervention selection is guided not only by immediate effect size but also by the carry-over observed at the next visit, incongruence between short-term and long-term results can likely be minimized.
To illustrate the importance of carry-over, consider the example of cryotherapy. Ice can produce immediate pain relief, which may enhance short-term function. However, these effects are transient and unlikely to persist into the next session. In contrast, joint manipulation often yields measurable mobility improvements that are at least partially retained, especially when combined with a home exercise program. Although home exercise constitutes an additional intervention, it is unlikely that ice would enhance the carry-over of exercise in the same way. In summary, intervention selection should be based on expected value; however, it may be more accurate to calculate the expected value based on the measurable carryover observed at the start of the subsequent session.

Section 2: Further Refinements for Achieving the "Best Approach"
The Best Intervention Should Be Determined by Comparative Research Whenever Possible
To prioritize the “best interventions” with the greatest accuracy, it is necessary to base relative effectiveness on the most reliable data available. Unfortunately, there is rarely enough data to precisely determine the reliability and average effect size (i.e., expected value) of every possible intervention, especially when carry-over effects are considered. However, it may be assumed that average outcomes (expected value) are the product of reliability and effect size, including carry-over effects. The most accurate source for estimating these expected values remains peer-reviewed and published research (referred to throughout this article as “research”). Further, because our goal is to prioritize interventions based on relative efficacy, the research used must be comparative.
More on “Why Research?”
Although it is tempting to assert that only “high-quality” research should be considered, this position usually relies on the levels of evidence hierarchies or pyramids, which are based on oversimplified generalizations and flawed assumptions. Worse, relying too rigidly on these hierarchies often leads to the dismissal of a large portion of the available data - a type of unintentional cherry-picking that can hinder more accurate modeling of expected value. To illustrate the problem, ask five colleagues how levels of evidence are ranked. Most will reply “by quality,” but this term is subjective unless explicitly defined. "Quality" must be linked to an objective measure (e.g., error rate, reproducibility, or risk of bias). For example, someone may feel that a Chevrolet is higher quality than a Toyota, but without objective comparison metrics (e.g., mechanical failure rates), this remains an opinion.
Further, even if we attempted to quantify bias or error, a study comparing study designs would be necessary to assess the actual accuracy of conclusions by research type —a process known in meta-science as evidence synthesis or meta-research, which is still in development. The situation becomes even more complicated when hierarchies attempt to compare study types directly. If level-2 research is “better” than level-3, then by how much? Does one level-2 study outweigh five level-3 studies? Ten? These seemingly simple questions reveal the weak assumptions underlying most evidence hierarchies.
It is worth noting that evidence hierarchies, such as those from the Oxford Centre for Evidence-Based Medicine or GRADE, are heuristic tools, not absolute or universally valid ranking systems. Many methodologists share the views above, which are reflected in several common criticisms of these hierarchies:
- They often overlook the quality of studies within each level.
- They may overvalue randomized controlled trials (RCTs), even when RCTs are inappropriate for the research question (e.g., rare harms, implementation science).
- They may undervalue observational studies that provide meaningful insights into real-world applications.
Importantly, study design is not the same as methodological rigor. A randomized controlled trial may be well-designed or poorly executed. The same is true for observational or cohort studies. Additionally, different study designs are suited to different questions. RCTs may be optimal for testing the efficacy of acute interventions, but they are not well-suited to modeling longitudinal outcomes, cost-effectiveness, or rare harms. No single design universally outperforms others in all contexts.
We will explore these issues further in a forthcoming article focused on interpreting research for clinical decision-making. For now, a more pragmatic and logically defensible hierarchy is proposed, based on the number of controls present and the scale of evidence, rather than generalizations about study design types.
A Better Level of Evidence Hierarchy
- More research studies are generally better than one study.
- Research is better than a single case.
- Objective outcome measures in clinical practice are more reliable than expert opinion.
- Expert opinion is better than non-expert opinion.
- Non-expert opinion should not guide clinical decision-making.
This simplified hierarchy does not claim that all studies are equally valuable or that expert opinion should be ignored. Rather, it acknowledges that data derived from larger, more controlled, and replicated studies tend to be more reliable than single cases or anecdotes. Controls such as peer review, statistical analysis, blinding, and independent replication reduce error and bias (e.g., selection bias, availability heuristic, personal allegiance). However, since these traits apply across multiple study designs, we cannot rank RCTs above observational studies without additional context about how well these controls were applied.
Why This Matters for Intervention Modeling
In the context of our knapsack optimization model, expected value is the critical variable for determining which interventions should be included in a time-limited session. When available, comparative research provides the most precise estimates of the relative expected value of various interventions. Prioritizing interventions without this data results in reliance on clinical experience, personal bias, and/or adherence to convention, all of which are vulnerable to various biases, fallacies, and errors.
By aggregating and integrating all comparative research, the accuracy of expected value estimates will continue to increase. This strengthens the decision-making model, increases the likelihood of delivering the most effective intervention plan, and enables better personalization through assessment-driven subgrouping. In short, the better the evidence base, the more accurate and efficient the optimization process becomes.
Research Must Be Comparative
The simplified hierarchy above emphasizes that research (the disciplined application of the scientific method, statistical analysis, and expert peer review) is our best tool for minimizing bias and error and likely results in the most accurate data currently available. Further, for the purposes of prioritizing interventions, the research must be comparative, directly evaluating the relative efficacy of at least two interventions.
Comparative research may include both experimental and observational designs, with or without control groups. While randomized controlled trials (RCTs) benefit from increased internal validity, they are not strictly necessary for our purposes. For the purposes of comparing the relative efficacy of interventions, interventions must be compared. If previous research has demonstrated that both interventions are generally effective, a control may not be necessary to obtain data useful for improving intervention selection. Note that non-comparative studies, including studies that compare an intervention to a randomized control group (no treatment or baseline), are likely less useful. Furthermore, separate studies should not be directly compared to one another due to the indeterminable number of confounding variables that may influence outcomes across studies. The problem with comparing studies may be partly be demonstrated by the Reproducibility Project , which has failed to demonstrate similar outcomes when replicating studies. Some of the issues with replication have included confounding variables that could not be controlled or were not considered prior to study replication.
This principle also applies to how we interpret meta-analyses (MAs) in our field. A meta-analysis, at its core, is an average of averages. It aggregates effect sizes across studies, but it can also amplify certain sources of error and bias. Unfortunately, serious interpretive errors are common in our field, particularly when drawing conclusions from meta-analyses (MAs) that include studies of varying quality or questionable comparability. For example, if an MA includes nine studies, of which seven demonstrate significantly better outcomes from one intervention over another, two show non-significant differences, but one of those two shows a clear trend similar to the others. Performing a meta-analysis on these studies only adds the potential of a failure to refute the null hypothesis. So although the trend in the research is clear prior to performing the MA, all the MA can potentially due is increase confusion. In fact, if the MA fails to refute the null, something is likely problematic with the design, the research question, or there was insufficient data to overcome regression to the mean. The inappropriate elevation of MAs over concise, peer-reviewed, and published comparative research has led to a Nihilistic view of our field (nothing works/refutes the null), and is preventing progress toward more effective intervention selection.
Again, a more detailed discussion of these interpretive issues will be presented in a forthcoming article, which focuses specifically on research synthesis and statistical inference in physical medicine.
Simpler, But More Accurate, Interpretation
The issues discussed above, subjective quality ratings, flawed evidence hierarchies, and limitations in research design, imply that a simpler approach to research interpretation may actually yield more accurate conclusions. In summary, the highest “quality” information available (based on the greatest number of controls for reducing bias and error) is peer-reviewed research. To develop a prioritization of interventions based on relative efficacy, we must rely on comparative research.
Further, an objective and systematic method is needed to interpret the findings of multiple comparative studies. The approach recommended by the Brookbush Institute is known as vote counting. Notably, in our field, “votes” are rarely close—studies comparing two interventions often show clear trends favoring one over the other, or reveal that differences between interventions are negligible.
The following is the vote-counting rubric used by the Brookbush Institute for all systematic reviews and educational content:
- Comparison Rubric (The goal is to use available research to determine the most likely trend.)
- A is better than B in all studies → Choose A
- A is better than B in most studies, and additional studies show similar results between A and B → Choose A
- A is better than B in some studies, and most studies show similar results between A and B → Choose A (with reservations)
- Some studies show A is better, some show similar results, and some show B is better → Results are likely similar (unless there is a clear moderator variable such as age, sex, or injury status that explains the divergence)
- A and B show similar results in the gross majority of studies → Results are likely similar.
- Some studies favor A, others favor B → Unless the number of studies overwhelmingly supports one side, results are likely similar.
When Additional Information is Needed
Although research should serve as the primary source for determining the most effective intervention, not all interventions have been investigated in peer-reviewed and published studies, nor have all interventions been directly compared to other interventions. Research may permit some indirect comparisons; for example, research may suggest that joint manipulation results in better outcomes than instrument-assisted soft tissue mobilization (IASTM), and that IASTM outperforms ultrasound, leading to the assumption that joint manipulation is superior to ultrasound. While this inference may be reasonable in certain cases, it must be used cautiously. Interventions may have varying levels of efficacy for different patient populations, diagnoses, or impairments. Untested assumptions inherently limit the validity of indirect comparisons, and any conclusions drawn from them should be treated as provisional estimates until more direct comparisons are available.
Further, an absence of evidence is not evidence of absence. A lack of research does not imply that an intervention is ineffective; it only implies that its relative expected value cannot be estimated with research. In the context of a knapsack optimization model, this creates a gap: we are unable to assign a reliable expected value to that intervention relative to other interventions. In such cases, additional methods must be used to estimate the expected value and guide prioritization.
When research is unavailable, in-practice comparisons become essential. However, these comparisons should not be based on intuition or gut-level impressions of trends. As discussed above, professionals are prone to a wide range of cognitive biases (e.g., confirmation bias, availability heuristic, anchoring, and misinterpretation of nonlinear change). Instead, these comparisons should be based on reliable, objective outcome measures. Whenever possible, the effects of two interventions should be tested within the same patient to reduce inter-individual variability, and those comparisons should be repeated across multiple patients before drawing conclusions.
In this framework, the clinician becomes a data-generating agent, approximating experimental research methods in practice. When results consistently show that Intervention A outperforms Intervention B in terms of reliability, magnitude of effect, speed of improvement, or session-to-session carryover, these results should be documented and used to reprioritize interventions. This process reflects a Bayesian approach: prior beliefs (established from research or training) are updated with observed outcomes (in-practice comparisons), refining the expected value estimate over time.
In short, when research is unavailable, professionals should approximate a controlled experiment, use objective outcome measures, and apply systematic tracking to support informed decision-making. These observations can then be integrated back into the knapsack model, adjusting value estimates in a manner consistent with Bayesian updating, to improve the accuracy of intervention selection over time. As these field-derived estimates accumulate, they serve as iterative refinements to the optimization framework, ensuring that intervention selection improves both within and across sessions.
There Is More Research Available Than Is Currently Utilized
One critique of this article during early review was that, if a single “best” approach exists, the current body of research is insufficient to identify it. While I acknowledge that substantial gaps remain in the literature, I would confidently argue that far more research exists than is currently applied in practice. The Brookbush Institute was founded with the intent of developing the first comprehensively evidence-based education platform, with every course built from a systematic review of all relevant peer-reviewed research on that topic. Our goal is to achieve the highest level of accuracy, ensuring that every conclusion is supportable by the best evidence available. When this is combined with careful development of a systematic approach, the result should be unparalleled outcomes.
What has been especially striking is that, without exception, every topic we have reviewed reveals dozens, sometimes hundreds, of studies that were overlooked in previously published reviews and educational materials. These studies often address nuanced questions, contribute critical details, and clarify apparent contradictions. The persistent critique that there is “not enough research” overlooks the substantial gains in accuracy and effectiveness that could be achieved by fully leveraging the existing research.
Even if this process falls short of identifying a singular, universally superior approach, it represents significant progress toward that goal. The point is not perfection, but optimization; continual refinement through the comprehensive application of current evidence.
Experimentation is Necessary to Overcome Local Maxima
In the context of physical rehabilitation, a local maximum refers to a treatment approach that yields a relatively high level of effectiveness based on current knowledge and available evidence. This approach may appear optimal within the boundaries of what has already been studied and validated. However, it may not represent the global maximum, the absolute best intervention strategy that can be achieved. Discovering whether better outcomes are possible requires experimentation with new interventions or novel combinations of existing interventions. Only by exploring alternatives—including those not currently ranked among the highest expected values—can we determine whether a superior outcome lies beyond the current local maximum.
A note of caution regarding the barrage of recommendations shared on social media: while it is technically possible for a novice or non-professional to stumble upon a more effective intervention, such discoveries are unlikely in a mature scientific field. It is far more likely that individuals with limited training will repeat interventions already proven to be ineffective or prematurely abandon promising interventions due to missed changes in outcome measures. In contrast, experienced and well-educated professionals are more likely to recognize the nuanced limitations of current approaches, identify unexplored treatment variations, and develop plausible hypotheses worth testing. While experimentation inherently involves uncertainty, it is most likely to yield meaningful advances when guided by a systematic process informed by existing evidence.
As with all non-research-based comparisons, conclusions drawn from clinical experimentation should not rely on intuition or anecdotal impressions. Interventions should be compared using reliable, objective outcome measures, ideally within the same patient and across multiple patients, to minimize the impact of confounding variables. As stated previously, the clinician should strive to approximate an experiment in practice.
One pragmatic recommendation for doing so is what we call “giving yourself 5 minutes to suck.” If we assume a standard session lasts 60 minutes, five minutes may be reserved for experimenting with a new technique, while the remaining 55 minutes are dedicated to interventions with known, high expected values. This structure ensures that even if the new intervention is ineffective, it is unlikely to reduce the overall session efficacy. However, if the experimental technique shows promise and is tracked systematically, it may improve expected value estimates over time. The name reflects the humility and tolerance required to pursue professional growth, acknowledging that learning often begins with not being very good at something.
This recommendation parallels Nassim Nicholas Taleb’s barbell strategy, originally applied to financial portfolios. In Taleb’s framework, a portfolio is divided between conservative, low-risk investments (e.g., 95%) and speculative, high-risk opportunities (e.g., 5%) to balance stability and optionality. In physical rehabilitation, the barbell strategy can be mirrored by dedicating the majority of the session (e.g., 55 minutes) to validated, low-risk interventions, while allocating a small portion (e.g., 5 minutes) to experimental techniques. This minimizes the potential downside while allowing for high-reward learning opportunities.
Critically, these “5-minute experiments” provide empirical observations that can be used to update the expected value assigned to that intervention within the broader intervention selection model. As outcome data accumulate across patients, impairments, or conditions, the observed effect size, reliability, and carry-over can be re-estimated, refining the value parameter assigned to the intervention. This process aligns with Bayesian updating: the clinician begins with a prior (e.g., unknown or low expected value due to lack of evidence) and incrementally updates that estimate as new data are gathered. These refined values feed directly back into the knapsack optimization model, improving future prioritization decisions and accelerating convergence toward a more globally optimal intervention strategy.
Experimentation is Necessary to Overcome Local Maxima
In the context of physical rehabilitation, a local maximum refers to a treatment approach that yields a relatively high level of effectiveness based on current knowledge and available evidence. This approach may appear optimal within the boundaries of what has already been studied and validated. However, it may not represent the global maximum, the absolute best intervention strategy that can be achieved. Discovering whether better outcomes are possible requires experimentation with new interventions or novel combinations of existing interventions. Only by exploring alternatives—including those not currently ranked among the highest expected values—can we determine whether a superior outcome lies beyond the current local maximum.
A note of caution regarding the barrage of recommendations shared on social media: while it is technically possible for a novice or non-professional to stumble upon a more effective intervention, such discoveries are unlikely in a mature scientific field. It is far more likely that individuals with limited training will repeat interventions already proven to be ineffective or prematurely abandon promising interventions due to missed changes in outcome measures. In contrast, experienced and well-educated professionals are more likely to recognize the nuanced limitations of current approaches, identify unexplored treatment variations, and develop plausible hypotheses worth testing. While experimentation inherently involves uncertainty, it is most likely to yield meaningful advances when guided by a systematic process informed by existing evidence.
As with all non-research-based comparisons, conclusions drawn from clinical experimentation should not rely on intuition or anecdotal impressions. Interventions should be compared using reliable, objective outcome measures, ideally within the same patient and across multiple patients, to minimize the impact of confounding variables. As stated previously, the clinician should strive to approximate an experiment in practice.
One pragmatic recommendation for doing so is what we call “giving yourself 5 minutes to suck.” If we assume a standard session lasts 60 minutes, five minutes may be reserved for experimenting with a new technique, while the remaining 55 minutes are dedicated to interventions with known, high expected values. This structure ensures that even if the new intervention is ineffective, it is unlikely to reduce the overall session efficacy. However, if the experimental technique shows promise and is tracked systematically, it may improve expected value estimates over time. The name reflects the humility and tolerance required to pursue professional growth, acknowledging that learning often begins with not being very good at something.
This recommendation parallels Nassim Nicholas Taleb’s barbell strategy, originally applied to financial portfolios. In Taleb’s framework, a portfolio is divided between conservative, low-risk investments (e.g., 95%) and speculative, high-risk opportunities (e.g., 5%) to balance stability and optionality. In physical rehabilitation, the barbell strategy can be mirrored by dedicating the majority of the session (e.g., 55 minutes) to validated, low-risk interventions, while allocating a small portion (e.g., 5 minutes) to experimental techniques. This minimizes the potential downside while allowing for high-reward learning opportunities.
Critically, these “5-minute experiments” provide empirical observations that can be used to update the expected value assigned to that intervention within the broader intervention selection model. As outcome data accumulate—across patients, impairments, or conditions—the observed effect size, reliability, and carry-over can be re-estimated, refining the value parameter assigned to the intervention. This process aligns with Bayesian updating: the clinician begins with a prior (e.g., unknown or low expected value due to lack of evidence) and incrementally updates that estimate as new data are gathered. These refined values feed directly back into the knapsack optimization model, improving future prioritization decisions and accelerating convergence toward a more globally optimal intervention strategy.
Modeling Is Likely Necessary
In this context, modeling refers to the construction of a simplified, structured representation of a dynamic system to simulate, predict, and refine decision-making. Complex problems involving multiple interacting variables, such as those encountered in physical rehabilitation, often require modeling to ensure optimal outcomes. Models also assist in managing variable relationships, testing alternative configurations, and refining intervention strategies over time.
The assertions presented thus far suggest that modeling is not merely helpful but likely necessary for identifying the best approach. First, research has consistently demonstrated that combinations of interventions result in better outcomes than any single intervention. Second, improved labeling and categorization reduce the likelihood of selecting interventions with redundant effects—an essential step that requires formal sorting structures. Third, prioritizing interventions based on relative efficacy enables the assembly of complementary interventions across categories. However, synergistic effects between interventions, the influence of assessment results on intervention selection, and the need for individualized plans based on subgroup characteristics all add additional layers of complexity.
Assessments must be carefully chosen not only to measure change but to identify and differentiate patient subgroups that benefit most from specific intervention strategies. These assessments, in turn, should guide the reprioritization of interventions to improve expected value across individuals. Additionally, as discussed in previous sections, a small amount of session time should be reserved for experimentation, necessary for advancing beyond local maxima toward globally optimal strategies. These interrelated, multi-variable dynamics are precisely the conditions under which modeling becomes essential.
This direction is reflected in a statement developed during the Brookbush Institute’s systematic review and course development process:
“The future is likely the modeling of dysfunction by identifying impairments correlated with a symptom or symptom cluster, identifying reliable and accurate assessments to differentiate those clusters from other conditions (that would ideally be treated differently), and determining the combination of interventions that result in the best possible objective outcome measures (e.g., reliability, effect size), assuming those outcome measures are also correlated with both short- and long-term patient outcomes (e.g., pain, function, return to sport, etc.).”
Although this perspective predated the development of this article, it clearly attempted to define a model-driven pathway toward an objectively measurable “best possible approach.” This article refines that concept through the following principles:
- Use outcomes (expected value) from comparative research to build an intervention model that prioritizes intervention categories by relative efficacy and identifies the highest-performing intervention within each category. (Note: this process depends on improved labeling and sorting of intervention types.)
- Use assessments to identify meaningful patient subgroups that achieve optimal outcomes from different intervention combinations, allowing for subgroup-specific reprioritization.
- Apply a systematic methodology of assessment, intervention, and re-assessment in clinical practice to refine intervention choices for individual patients, following the rank order of expected value.
- Reserve a small portion of each session for experimental approaches to incrementally discover interventions with higher expected value than those currently in the model. These outcomes can then be used to update the value parameters and improve the model over time.
In sum, a formal model supports convergence toward an optimal intervention strategy through structured integration of empirical data, subgroup differentiation, in-practice testing, and ongoing refinement, consistently aligned with the principles of expected value, Bayesian updating, and constrained optimization.
When Modeling Is Not Helpful
If the goal of modeling is to improve outcomes, then the model must be deliberately constructed to predict the relationship between interventions and improvements in those outcomes. A valid model for intervention selection must be outcome-driven—built with the explicit intent of optimizing objective, measurable results. Models that focus solely on correlates of outcomes, or that aim primarily to explain causes rather than predict effect, are not suitable for this purpose.
For example, there is no such thing as a “reps model” for resistance training unless the desired outcome is to perform more repetitions. A model centered on repetitions might be helpful for specific performance goals but is not generalizable to broader training outcomes such as strength, power, hypertrophy, or endurance unless those variables are explicitly targeted.
Unfortunately, the term model is often misused in education and professional discourse. Many certifications and courses employ the word to describe loose collections of preferred interventions, conceptual frameworks, or personal philosophies. These so-called models often function as a means to dismiss disfavored techniques, justify favored interventions, or lend structure to an otherwise arbitrary mix of approaches—none of which serve the purpose of improving outcomes through expected value optimization. Without an intent to predict, compare, or prioritize based on objective outcomes, such frameworks cannot be considered true models in the sense used throughout this article.
A common example of a misapplied model in our field is the Biopsychosocial (BPS) model. While valuable in its intended domain, the BPS model was developed to describe the multifactorial nature of the pain experience. It is a model of explanation, not a model for outcome prediction. It does not provide a method for selecting interventions based on relative efficacy, nor does it offer a framework for prioritization or optimization. The problem lies not in the BPS model itself, but in its misapplication by professionals who attempt to use a causal explanation framework to guide treatment decisions, without regard to comparative outcomes.
In summary, not every grouping of concepts constitutes a model, and not every use of the word model implies a useful tool for clinical decision-making. To be useful in a knapsack optimization framework—or any framework aimed at improving outcomes—a model must (1) be predictive, (2) be testable against objective measures, and (3) support systematic prioritization based on expected value. Any model that fails to meet these criteria either misuses the term or serves a different purpose altogether.
Evidence of Constrained Optimization Models in Medicine, Rehabilitation, and Sports
Several publications in healthcare, behavioral science, and clinical decision-making have referred to components of a constrained optimization framework, including diminishing marginal returns, expected value modeling, and adaptive personalization. However, none of these publications appear to integrate these principles into a singular, unified structure as proposed in this article. In my opinion, and I believe this is well supported by the article, the best approach (global optimum) cannot be achieved without integrating all components.
Health Resource Allocation and Diminishing Returns
In public health planning, constrained optimization is frequently used to allocate limited resources across interventions to maximize population-level outcomes (e.g., QALYs or cases averted). Meyer-Rath et al. (2017) demonstrated this in an HIV intervention package optimizer that explicitly modeled diminishing marginal returns (14). Their findings showed that as intervention coverage increased, the incremental benefit decreased, and cost-effectiveness declined by up to fourfold in some cases. This work underscores the importance of considering resource saturation and diminishing utility when evaluating interventions. Note that these models of diminishing returns are typically static, population-based, and designed for budget optimization, rather than individual-level treatment planning or dynamic learning based on patient outcomes.
Multicomponent Intervention Frameworks
Frameworks like the Multiphase Optimization Strategy (MOST) and Sequential Multiple Assignment Randomized Trials (SMART) treat intervention design as an optimization problem. MOST, as described by Collins, Murphy, and Strecher (15), focuses on evaluating combinations of components under real-world constraints (e.g., limited session time or budget). Similarly, SMART designs enable the adaptive sequencing of interventions, allowing plans to be adjusted based on patient response (16). These frameworks recognize that interventions may not produce linear additive effects, and that some combinations may be redundant or less effective due to overlapping mechanisms. However, they optimize at the research design level, not dynamically during clinical care. As a result, they lack real-time updating based on observed outcomes and typically rely on population averages rather than individualized optimization.
Adaptive Treatment Strategies and Reinforcement Learning
Adaptive and personalized treatment models, particularly those grounded in reinforcement learning (RL) and sequential decision-making, are perhaps the closest analogue to the dynamic elements of the proposed framework. As summarized in Yu et al. (17), RL models frame intervention selection as an ongoing optimization problem, adjusting policies based on new data and observed patient responses. These models explicitly aim to maximize expected utility by updating probability distributions over time, incorporating uncertainty and heterogeneity in treatment effects. Lavori and Dawson (18) also highlighted dynamic treatment strategies in clinical trials, offering methods for refining interventions based on prior outcomes. However, most RL and dynamic sequencing models operate in a stepwise, sequential manner, choosing the next best action, rather than solving for a constrained portfolio of interventions within a fixed session. While they do accommodate personalization, probabilistic effectiveness, and adaptive learning, they often lack explicit modeling of intervention interactions, diminishing returns, or concurrent resource constraints.
Adaptive Systems in Rehabilitation Technology
In rehabilitation technology, early implementations of adaptive learning systems have begun to reflect key principles of the proposed model. For example, Grimm, Naros, and Gharabaghi (2016) developed a closed-loop VR rehabilitation system that modulates task difficulty in real-time using patient performance data (19). This mirrors the concept of adaptive reassessment and continuous refinement of expected value based on individual response. However, such systems typically focus on single-modality adaptation (e.g., exercise difficulty) rather than comprehensive intervention selection across categories. They do not typically include a formal mechanism for optimizing multiple concurrent interventions under session constraints or tracking synergistic versus redundant effects.
Synthesis and Novelty
Taken together, these models show that constrained optimization, diminishing marginal utility, and adaptive outcome-based refinement are not novel concepts in isolation. They are applied across health economics, behavioral intervention design, machine learning in clinical decision-making, and emerging rehabilitation technologies. However, no existing framework appears to combine all of the following:
- Selection of multiple interventions per session under a fixed time/resource constraint (i.e., a knapsack problem),
- Prioritization based on the relative expected value of intervention categories, refined through comparative research,
- Bayesian or empirical updating of effectiveness estimates over time,
- Synergy modeling to avoid redundant effects, and
- Individual-level adaptive decision-making using ongoing reassessment.
The integration of these elements into a unified system, particularly with the goal of maximizing session-level effectiveness in clinical or rehabilitation contexts, represents a novel contribution. The existing literature reflects key components, but the full model, as articulated in this article, appears to be original and has not been previously represented in the current medical, rehabilitation, or sports science literature.
A Sample of the Developing Models from the Brookbush Institute
The following is an applied example of the intervention selection model developed by the Brookbush Institute for Lower Extremity Dysfunction. This model was constructed using a systematic review of peer-reviewed research, including the identification of impairments reliably correlated with symptom clusters. The intent was to develop a decision-support framework that (1) reduces bias, (2) implies useful assessments, (3) supports predictable outcomes, and (4) aligns with a knapsack optimization model by maximizing expected value within session constraints.
The development of this model began with grouping symptoms and diagnoses that tend to co-occur or are consistently mentioned in research investigating the target segment. These clusters were organized to ensure practical applicability within the time constraints of a typical session, balancing comprehensiveness (e.g., the entire lower extremity and trunk) with feasibility (e.g., ankle-only treatment). A list of impairments was then derived from this clustering process, further refined to highlight impairments that can be reliably and objectively assessed, categorized by structure and dysfunction type (e.g., joint mobility, muscle activation), and aligned with interventions demonstrating the highest relative efficacy for that type of structure and dysfunction.
Importantly, assessment serves as the operational bridge between symptom clusters and intervention selection. A standardized set of reliable and objective assessments is used to identify the impairments present in a given patient. Once impairments are identified, the clinician selects the relevant techniques from the list of interventions, which are organized by category. Note, interventions have been pre-selected based on expected value. Further, the sequence of technique categories reflects synergistic effects and modifiability based on assessment findings. The model was iteratively refined through repeated application in practice and continued review of emerging literature.
Symptoms and Diagnoses Clusters That Result in Similar Dysfunction
- Ankle/Foot
- Medial tibial stress syndrome
- Pronation
- Ankle sprain
- Ankle instability
- Achilles tendinopathy
- Tibialis posterior tendinopathy
- Plantar fasciitis
- Knee
- Anterior cruciate ligament injury
- Functional valgus
- Knee pain (patellofemoral pain syndrome, jumper's knee )
- Lateral knee pain (iliotibial band syndrome, runner's knee)
- Knee osteoarthritis
- Knee effusion
- Proximal tibiofibular joint pathology
- Tibiofibular joint subluxation/dislocation
- Medial and lateral heel whip
- Hip
- Trigger points
- Abductor tendon tear
- Adductor Groin Strain
- Femoral acetabular impingement
- Hip Osteoarthritis
- Lumbosacral dysfunction
- Low Back Pain
- Sacroiliac joint pain
The Impairments Correlated with Lower Extremity Dysfunction
Correlated Changes in Overhead Squat Assessment (Sit-to-Stand Mechanics)
- Loss of dorsiflexion (inadequate forward translation of the knee, i.e. tibia on foot dorsiflexion)
- Feet Flatten (a.k.a. functional pes planus, pronation, eversion, calcaneus valgus, positive navicular drop test, etc.)
- Feet turn out (a.k.a. turn-out, heel flare, heel whip, etc.)
- Knees Bow In (a.k.a. functional knee valgus, medial knee displacement, hip adduction, etc.)
- Excessive Forward Lean (excessive hip flexion, forward trunk position, tibia/torso angle)
Correlated Changes in Muscle Activity and Length
Correlated Changes in Joint Mobility (Increased stiffness)
- Ankle : Inadequate posterior glide of the talus on the tibia
- Ankle : Inadequate posterior glide of the lateral malleolus on the tibia
- Ankle /Knee : Inadequate anterior glide of the fibular head on the tibia
- Knee : Inadequate anterior glide of the tibia on the femur (the lateral compartment may be more restricted)
- Hip : Inadequate posterior/inferior glide of the femur in the acetabulum
Correlated Changes in Fascia Mobility (Loss of Extensibility)
- Sacrotuberous ligament
- Iliotibial Band
- Crural Fascia
- Achilles Tendon
- Plantar Fascia
Altered Subsystem Recruitment Patterns
- Under-active (Integrate)
- Over-active (Release and Avoid)
Intervention Model (Prioritization of Categories Based on Relative Efficacy, Modifiable via Assessment Findings):
- Mobilize
- Release
- Mobilize
- Lengthen
- IASTM (optional)
- Activate
- Isolated activation
- Core integration
- Reactive activation
- Subsystem integration
Note that this case study demonstrates the use of objective, reliable assessments of factors correlated with lower extremity symptoms (symptom clusters) to select a group of interventions that have demonstrated the highest relative efficacy for addressing those factors. All of this was developed from all of the relevant peer-reviewed research available at the time of the systematic review for a particular variable and tested in practice.
Case Study:
Patient: 47yo, female, complains of shin pain. An avid long-distance runner since high school. History of knee and foot-related complaints (e.g. PFPS, Achilles tendinopathy, plantar fasciitis)
- Overhead Squat Assessment:
- Goniometry
- Dorsiflexion Goniometr y <15 - 20°
- Manual Muscle Testing MMT:
Sample Intervention (Ankle Dorsiflexion Restriction)
- Manual Release
- Mobilization or Manipulation
- Manual Lengthening
- Instrument Assisted Soft Tissue Mobilization
- Activation
- Reactive Activation:
- Integration
This case study illustrates how a model built on peer-reviewed evidence, organized around symptom clusters, impairment identification, and intervention prioritization, can yield structured and optimized treatment strategies. It exemplifies the translation of expected value modeling into clinical practice, where assessment identifies constraints and modulates the value parameters used in the intervention selection algorithm. This is how real-world decisions, driven by comparative data and refined through practice, move us closer to a universally optimal intervention framework.

Bibliography
Please forgive me for referencing myself. Note that the only references to my previously published materials are references to comprehensive research reviews of a topic, and these references include annotated bibliographies. Citing these articles was only intended to improve the readability of this article by reducing dozens or hundreds of citations to a few citations. We genuinely hope that you will critically review these references, and we humbly ask for feedback if you believe our conclusions are anything less than objective and conservative. Additionally, some may question our reference to Wikipedia; however, the pages referenced provide a more comprehensive summary of those topics than any other single reference (e.g., The Branford Hill Criteria).
- Brookbush, B. (2024) Pain neuroscience education (PNE) is relatively ineffective: Research confirmed. Brookbush Institute. https://brookbushinstitute.com/articles/pain-neuroscience-education-pne-is-relatively-ineffective-research-confirmed
- Wikipedia contributors. (n.d.). Expected value. Wikipedia. Retrieved September 24, 2024, from https://en.wikipedia.org/wiki/Expected_value
- Brookbush, B., Campione, J. (2024) Instrument-assisted soft tissue mobilization (IASTM): Comprehensive systematic research review [Online course]. Brookbush Institute. https://brookbushinstitute.com/courses/instrument-assisted-soft-tissue-mobilization-iastm-comprehensive-systematic-research-review
- Hill, A. B. (2015). The environment and disease: association or causation?. Journal of the Royal Society of Medicine, 108(1), 32-37.
- Wikipedia contributors. Bradford Hill criteria. Wikipedia. Retrieved September 24, 2024, from https://en.wikipedia.org/wiki/Bradford_Hill_criteria
- Brookbush, B. (2024). False narratives: Nocebo and negative expectations do not affect manual therapy outcomes. Brookbush Institute. https://brookbushinstitute.com/articles/false-narratives-nocebo-and-negative-expectations-do-not-affect-manual-therapy-outcomes
- Brookbush, B. (2024). Active versus passive: Is exercise more effective than manual therapy? Brookbush Institute. Retrieved September 24, 2024, from https://brookbushinstitute.com/articles/active-versus-passive-is-exercise-more-effective-than-manual-therapy
- Lawrence, R. L., Braman, J. P., LaPrade, R. F., & Ludewig, P. M. (2014). Comparison of 3-dimensional shoulder complex kinematics in individuals with and without shoulder pain, part 1: sternoclavicular, acromioclavicular, and scapulothoracic joints. journal of orthopaedic & sports physical therapy, 44(9), 636-645.
- Brookbush, B. (2015). Lower extremity dysfunction. Brookbush Institute. Retrieved September 26, 2024, from https://brookbushinstitute.com/courses/lower-extremity-dysfunction
- Burns, P. B., Rohrich, R. J., & Chung, K. C. (2011). The levels of evidence and their role in evidence-based medicine. Plastic and reconstructive surgery, 128(1), 305-310.
- Taleb, N. N. (2012). Antifragile: Things that gain from disorder. Random House.
- Brookbush, B. (2020). Joint mobilizations and manipulations: Risk of adverse events. Brookbush Institute. https://brookbushinstitute.com/courses/joint-mobilization-and-manipulation-risk-of-adverse-events
- Brookbush, B. (2024). False narratives, nocebo, and negative expectations do not affect manual therapy outcomes: Research confirmed. Brookbush Institute. https://brookbushinstitute.com/articles/false-narratives-nocebo-and-negative-expectations-do-not-affect-manual-therapy-outcomes
- Meyer‑Rath, G., Schnure, M., & De Bruin, M. (2017). Optimization of HIV intervention packages considering diminishing marginal returns. BMC Public Health, 17, Article 123. https://doi.org/10.1186/s12889-017-4128-0
- Collins, L. M., Murphy, S. A., & Strecher, V. (2007). The multiphase optimization strategy (MOST) and the sequential multiple assignment randomized trial (SMART): new methods for more potent eHealth interventions. American journal of preventive medicine, 32(5), S112-S118.
- Lei, H., Nahum-Shani, I., Lynch, K., Oslin, D., & Murphy, S. A. (2012). A" SMART" design for building individualized treatment sequences. Annual review of clinical psychology, 8(2012), 21-48.
- Yu, C., Liu, J., Nemati, S., & Yin, G. (2021). Reinforcement learning in healthcare: A survey. ACM Computing Surveys (CSUR), 55(1), 1-36.
- Lavori, P. W., & Dawson, R. (2014). Introduction to dynamic treatment strategies and sequential multiple assignment randomization. Clinical trials, 11(4), 393-399.
- Grimm, F., Naros, G., & Gharabaghi, A. (2016). Closed-loop task difficulty adaptation during virtual reality reach-to-grasp training assisted with an exoskeleton for stroke rehabilitation. Frontiers in neuroscience, 10, 183574.