AI is Smart. Is it Wise? Quantifying the Effect of Patient-Choice (β) on Physical Outcomes

By analyzing SPORT trial data, this study quantifies a significant contribution of patient choice (beta) to physical outcomes that is distinct from biological treatment effects (alpha), thereby highlighting a critical limitation of current AI systems that rely on population-level evidence and cannot capture individual preference-driven decision-making.

The Gist

The Big Question: Is AI Smart, but Not Wise?

Imagine you are trying to decide between two routes to get to a new city.

• Route A (Surgery): Fast, but bumpy and risky.
• Route B (Non-surgery): Slower, but smoother and safer.

A super-smart GPS (Artificial Intelligence) looks at millions of past trips. It tells you: "Statistically, both routes get you to the destination in about the same amount of time. The difference is tiny."

The GPS is Smart. It knows the data. But is it Wise? No. Because it doesn't know you. It doesn't know if you hate bumpy roads, if you are in a huge rush, or if you are terrified of getting a flat tire.

This paper asks: Does the driver's personal choice actually change the outcome of the trip? And can AI figure that out?

---

The Experiment: The "Spine" Study

The researchers looked at a famous medical study called SPORT. It compared back surgery vs. non-surgical treatment for a slipped disc.

• The Twist: The medical data showed that surgery and non-surgery were basically equal in effectiveness.
• The Chaos: Despite being randomly assigned to a group, nearly half of the patients (42–45%) ignored the assignment and switched to the treatment they wanted. They "crossed over."

Usually, scientists get annoyed by this. They say, "The patients broke the rules! The data is messy."
This paper says: "No, that's not noise. That's a signal!" The fact that people switched shows they had strong feelings about what was right for them.

The Two Ingredients: Alpha (α) and Beta (β)

The authors created a simple formula to explain what happens in the body and the mind.

1. Alpha (α) = The Biological "Magic"

This is the actual medical power of the treatment.

• Analogy: This is the engine of the car. Does the surgery fix the back? Does the pill stop the pain?
• The Finding: In this study, the "engine" of surgery and the "engine" of non-surgery were almost identical. The medical difference was tiny.

2. Beta (β) = The Patient's "Choice Power"

This is the effect of the patient choosing their path.

• Analogy: This is the driver's confidence. If you choose to drive the bumpy road because you believe in it, you might drive better, handle the bumps better, and feel less stressed. If you feel forced into a path you hate, you might feel worse, even if the road is technically fine.
• The Finding: The researchers calculated that Beta (Choice) had a huge, measurable impact on how much pain patients felt and how well they could move.

The "Aha!" Moment

Here is the math in plain English:

• The difference between the two treatments (Alpha) was 0.65 points. (Basically, zero difference).
• The power of the patient's choice (Beta) was 7.4 points.

The Conclusion: When the medical treatments are equal, the patient's choice becomes the most important factor in whether they get better.

Why AI Gets Stuck Here

The paper argues that current AI (like Large Language Models) is Alpha-Biased and Beta-Blind.

• How AI works: It reads millions of medical papers. Those papers only talk about Alpha (the biological results). They say, "Surgery works 50% of the time; Non-surgery works 50% of the time."
• What AI misses: AI cannot read a patient's mind. It cannot know that this specific patient hates surgery, or that that specific patient is willing to take a risk to get back to work faster.
• The Result: AI is Smart (it knows the stats) but not Wise (it can't weigh personal values). It might tell a patient, "It doesn't matter which you pick," but for the patient, it matters everything.

The "Informed Choice" vs. "Informed Consent"

The paper makes a distinction between two types of "knowing":

1. Informed Consent: "I understand the risks and I agree to do what the doctor says." (Good for when there is one clear best option).
2. Informed Choice: "I understand the risks, I know my own values, and I am picking the path that fits my life." (Required when options are equal).

The authors argue that in medicine, we need to move from just getting consent to facilitating Choice.

The Takeaway

1. Choice is Medicine: When you choose your treatment, that act of choosing actually helps you heal. It's not just a feeling; it's a physical factor.
2. AI has a Blind Spot: AI is great at summarizing medical facts, but it is terrible at understanding human values. It cannot replace the conversation between a doctor and a patient.
3. The Future: We need AI that doesn't just say "Here are the stats," but helps doctors ask, "What matters most to you?"

In short: AI can tell you which car is faster, but only you can decide which car you want to drive. And sometimes, driving the car you want makes the ride smoother than the one that is technically "faster."

Technical Summary

1. Problem Statement

The paper addresses a critical gap in the application of Artificial Intelligence (AI), specifically Large Language Models (LLMs), to clinical decision-making.

• The Limitation of Current AI: LLMs are trained on population-level evidence (clinical trials, meta-analyses) which primarily quantifies biological treatment effects (denoted as $\alpha$). These models excel in "preference-neutral" conditions where one treatment is statistically superior.
• The "Preference-Sensitive" Gap: In many clinical scenarios (e.g., lumbar disc herniation), treatments yield comparable outcomes ($\Delta\alpha \approx 0$). In these cases, the optimal decision depends on individual patient values, risk tolerance, and life priorities.
• The Unquantified Variable: The authors argue that "patient choice" acts as a distinct causal driver of physical outcomes, denoted as $\beta$. Current AI systems are "blind" to $\beta$ because their training data lacks the individual utility functions required to model value-based trade-offs. The paper seeks to empirically quantify $\beta$ and demonstrate its magnitude relative to biological treatment effects.

2. Methodology

The study is a secondary analysis of the Spine Patient Outcomes Research Trial (SPORT), a landmark study comparing surgical discectomy vs. nonoperative treatment for lumbar disc herniation. The SPORT dataset is ideal for this analysis due to its high rate of "crossover" (42–45% of randomized patients switched treatments within 3 months), which the authors reframe not as noise, but as a signal of strong patient preference.

Analytical Framework:

• Data Sources: Published aggregate results from the randomized cohort ($n=501$) and the observational cohort ($n=743$).
• Primary Outcomes: SF-36 Bodily Pain (BP), SF-36 Physical Function (PF), and Oswestry Disability Index (ODI) measured at 3, 12, and 24 months.
• Causal Inference Strategy:
  • Instrumental Variable (IV) Analysis: Used randomized treatment assignment as an instrument to estimate unbiased treatment effects ($\alpha$) while accounting for non-compliance (crossover).
  • Two-Stage Least Squares (2SLS): Implemented to model treatment received as a function of assignment, followed by modeling outcomes as a function of predicted treatment.
  • Complier Average Causal Effect (CACE): Calculated to estimate the effect specifically for those who adhered to their assigned treatment.
• Decomposition of Effects:
  • $\alpha$ (Biological Effect): Estimated via ITT (Intention-to-Treat) and IV/CACE methods.
  • $\beta$ (Patient Choice Effect): Defined as the residual difference between the Observational Contrast ($\Delta_{OBS}$, where patients self-select) and the Causal Contrast ($\Delta_{ITT}$ or $\Delta_{IV}$).
  • Four Estimators for $\beta$:
    1. $\beta_{Arithmetic} = \Delta_{OBS} - \Delta_{ITT}$
    2. $\beta_{CACE1} = \Delta_{OBS} - (\Delta_{ITT} / \pi)$ (where $\pi$ is compliance rate)
    3. $\beta_{CACE2} = \Delta_{OBS} - \Delta_{CACE}$ (2SLS based)
    4. $\beta_{IV} = \Delta_{OBS} - \Delta_{IV}$
• Robustness Checks: E-values were calculated to assess sensitivity to unmeasured confounding, and bootstrap resampling (10,000 iterations) was used for confidence intervals.

3. Key Results

The analysis yielded three primary quantitative findings:

A. Quantification of Patient Choice ($\beta$)

• Patient choice has a measurable and significant causal effect on physical outcomes.
• The aggregate patient-choice effect ($\beta_G$) was calculated at 7.4 points (SD 3.4) on a 0–100 scale.
• This effect is statistically significant ($Z \approx 3.6$, $P < 0.001$) across all outcomes and time points.
• The magnitude of $\beta$ exceeds the Minimal Clinically Important Difference (MCID) for SF-36 and ODI scales, indicating clinical relevance.

B. Comparison of $\alpha$ vs. $\beta$

• Aggregate Treatment Effect ($\alpha_G$): The overall biological benefit of treatment was 15.7 points.
• Net Treatment Difference ($\Delta\alpha$): The difference between surgical and nonoperative biological effects was negligible at 0.65 points.
• Dominance of Choice: In the context of clinical equipoise (where $\Delta\alpha \approx 0$), the patient-choice effect ($\beta_G = 7.4$) is 11 times larger than the net difference in biological efficacy.

C. Validation of Crossover as Signal

• Observational contrasts (self-selected groups) overestimated treatment effects by approximately 67–71% compared to causal estimates (IV/CACE).
• This inflation is attributed to $\beta$: patients who chose a treatment derived additional benefit from the alignment of that treatment with their values, a factor invisible to standard ITT analysis.

4. Key Contributions

1. Empirical Quantification of $\beta$: The study moves beyond theoretical arguments for "values-based medicine" by providing a causal-inference framework to measure the specific contribution of patient choice to physical outcomes.
2. Reframing Crossover: It challenges the traditional view of crossover in RCTs as a protocol violation, proposing instead that high crossover rates are evidence of strong, outcome-shaping patient preferences.
3. Defining the "AI Blind Spot": The paper establishes that current LLMs are $\alpha$-biased (trained on biological efficacy) and $\beta$-blind (incapable of encoding individual utility functions). It argues that in preference-sensitive conditions, AI recommendations may be suboptimal or misleading because they ignore the dominant driver of outcomes ($\beta$).
4. Distinction between Informed Consent and Informed Choice: The authors clarify that "informed consent" (autonomy in preference-neutral settings) is insufficient for "preference-sensitive" conditions, which require "informed choice" (deliberative alignment of values and risks).

5. Significance and Implications

• For Clinical Practice: When biological differences between treatments are small, the "wisdom" of a decision lies in the patient's value system, not just the clinical data. Clinicians must prioritize preference elicitation over algorithmic prediction in these scenarios.
• For AI Development:
  • Architectural Limit: Current token-based LLMs cannot access real-time, context-dependent utility functions. They cannot simulate the "lived reasoning" required for $\beta$.
  • Future Directions: The authors propose "hybrid systems" that fuse evidence synthesis ($\alpha$) with structured preference-elicitation tools. They also call for a redesign of clinical trials to prospectively estimate $\beta$ and report it alongside $\alpha$.
• For Policy and Ethics: The findings suggest that regulatory frameworks and reimbursement models for AI in healthcare must account for the "wisdom gap." Deploying AI in preference-sensitive settings without mechanisms for value alignment risks misrepresenting efficacy and undermining shared decision-making.

Conclusion:
The paper concludes that while AI is "smart" (capable of processing vast evidence for $\alpha$), it is not yet "wise" (unable to navigate the value-laden $\beta$). In conditions of clinical equipoise, patient choice is not merely a confounder but a therapeutic driver, and ignoring it renders AI-driven decision support fundamentally incomplete.