📄 medicine

Measuring and reducing intervention misalignment in next best action selection for care management: a within-patient natural experiment

This study quantifies a 27% intervention misalignment in care management for rising-risk Medicaid patients, primarily driven by the under-addressing of social needs, and demonstrates that the PEARL policy, which leverages within-patient natural experiments and fairness constraints, can reduce this misalignment to 2.0% while avoiding the severe performance degradation seen in fairness-agnostic variants.

Original authors: Sanjay Basu, Parth Sheth, Sadiq Patel

Published 2026-07-10

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Sanjay Basu, Parth Sheth, Sadiq Patel

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a care manager, a superhero in scrubs, tasked with helping patients who are juggling multiple problems at once: high blood pressure, anxiety, a lack of food, and no way to get to the doctor. You only have 15 minutes per patient. You can't fix everything at once. You have to pick one thing to tackle first. This is the "Next Best Action" problem.

For years, the system has been like a vending machine that only dispenses the same snack to everyone. If a patient walks in, the computer says, "Here, take a behavioral health referral!" or "Here, take a care access coordination ticket!" It doesn't really look at whether that specific snack is the one the patient actually needs most to get better.

The authors of this paper, Sanjay Basu, Parth Sheth, and Sadiq Patel, decided to measure how often this "vending machine" gets it wrong. They call this mistake intervention misalignment. It's the percentage of patients who get an action that isn't the one expected to give them the biggest health boost.

The Big Reveal: The Vending Machine is Broken

The team looked at 34,971 rising-risk Medicaid patients across three states between 2023 and 2025. They used a clever trick called a "within-patient natural experiment." Think of it like this: because the care program rolled out to different groups of people at slightly different times (due to logistics, not because the patients were sicker or healthier), the researchers could compare the same person's health before they got help versus after they got help. This let them figure out what the patient actually needed without relying on guesswork.

Here is the shocking part: Under the current system, 27.0% of patients (that's more than 1 in 4) were given the wrong "next best action."

But the real story is what was wrong. The current system was obsessed with two things: "care access coordination" and "behavioral health referrals." Together, these made up 80.3% of all actions taken.

Meanwhile, the patients' actual needs were screaming for something else. The data showed that patients would have benefited most from social needs like transportation, food security, housing, and financial benefits.

The current system picked transportation as the top priority only 1.9% of the time, even though the patients' own data suggested it should have been the top choice 5.1% of the time. That's a 25.5-fold gap.
For food security, the system chose it 1.9% of the time, but patients needed it 5.2% of the time. That's a 17.3-fold gap.
For housing, the gap was 12.5-fold.

The paper explicitly argues against the idea that this is just a simple "poor people need more help" gradient. The misalignment was high across all income levels and races, showing that the system is systematically ignoring social needs for everyone, not just the most deprived.

The Solution: PEARL

The researchers didn't just point out the problem; they built a new "vending machine" called PEARL (Policy Evolution through Aligned Retrospective Learning).

PEARL is like a super-smart coach that learns from the "natural experiments" of the past. It has three secret ingredients:

The Within-Patient Signal: It learns from the specific history of each patient (what worked for them before).
Group-Stratified Fairness: It makes sure it doesn't accidentally favor one group of people over another.
A Fourteen-Expert Router: It has a team of 14 different "experts" (one for each type of help, like hypertension or housing) who vote on the best move.

When they tested PEARL, it was a game-changer.

It reduced intervention misalignment from 27.0% down to just 2.0%.
That is a 92.6% reduction toward the perfect possible score.
The paper notes that the "within-patient signal" alone did most of the heavy lifting, dropping misalignment to 4.4%. However, if you take away the "fairness" part, the system actually gets worse than before (jumping to 38.7–42.4%). This proves that fairness isn't just a nice-to-have; it's essential for the system to work at all.

What They Ruled Out

The paper is very careful about what it doesn't claim.

It's not a magic bullet for everyone: They tested 13 different policies. Some, like "Conservative Q-learning" (a type of AI), improved the overall number of hospital visits but didn't fix the "wrong action" problem. It reduced the event rate but kept the misalignment high at 32.0%.
It's not just about being "nice": Simple copying of past decisions (behavioral cloning) without the fairness tweak made things worse.
The Camden Coalition Trial: The authors re-analyzed a famous past trial (the Camden Coalition) where everyone got the exact same intense help. They found that in that trial, 90.9% of patients got the "wrong" action because they all got the same thing. In a simulation, they showed that if PEARL had been used instead, the readmission rate could have been 15.1 percentage points lower. But the authors are clear: this is a simulation based on their model, not a proof that the Camden trial would have definitely succeeded.

How Sure Are We?

The authors are quite confident in their numbers, but they are also honest about the limits.

They measured the misalignment using a robust method called "doubly robust counterfactual modeling."
They ran 20 different sensitivity analyses (changing the rules slightly to see if the result holds up), and PEARL stayed at 2.0% in almost all of them.
They calculated an "E-value" of 11.92. This is a fancy way of saying: "For our results to be wrong, there would have to be a hidden factor so powerful it predicts both the action and the outcome 11.92 times better than any factor we actually measured." That's huge, so they are pretty sure the result is real.
However, they admit that the "within-patient natural experiment" measures the effect of getting care vs. not getting care, not necessarily the head-to-head battle of "Housing vs. Food" in a perfect lab setting. They are measuring the quality of the choice, not running a new clinical trial to prove the choice works in real-time (though they are planning a new trial to do exactly that).

The Bottom Line

The paper concludes that "intervention misalignment" is a real, measurable thing that we can fix. The current system is ignoring social needs like food and housing by a massive margin, even when those are the things patients need most. By using a new method (PEARL) that listens to the patient's own history and treats everyone fairly, we can fix this mismatch.

The authors are now preparing a new, large-scale experiment across 24 care teams to see if this new way of choosing actions actually reduces hospital visits in the real world. Until then, the data suggests that if we want to help patients, we need to stop guessing and start listening to the signals that tell us what they actually need first.

Technical Summary: Measuring and Reducing Intervention Misalignment in Next Best Action Selection

Problem Definition
Care management programs for Medicaid beneficiaries with multiple chronic conditions and social needs face a "next best action" problem. Due to operational constraints (high patient-to-staff ratios, short encounter durations, and evidence-based guidelines recommending single-goal interventions), care managers must prioritize one intervention type per encounter from a set including chronic disease management, behavioral health, medication adherence, and social needs navigation. Current programs largely rely on risk scores to identify who needs care but lack mechanisms to determine which specific action yields the largest expected health gain for a specific patient. This leads to "intervention misalignment," defined as the fraction of patients whose assigned next best action is not the action expected to produce the largest health gain. The authors hypothesize that uniform or heuristic-based assignment dilutes average treatment effects, contributing to inconsistent results in randomized trials (e.g., the Camden Coalition trial) and high rates of potentially preventable acute care events.

Methodology
The study employs a within-patient natural experiment design using retrospective data from 34,971 rising-risk Medicaid patients across three states (Washington, Virginia, Ohio) from 2023 to 2025.

Natural Experiment Construction: The authors leverage staggered enrollment timing into care management programs. Because enrollment timing is driven by administrative logistics (data-feed connectivity, team training) rather than patient health status, it creates exogenous variation. Patients serve as their own controls, comparing a pre-enrollment period (no care management) with a post-enrollment period (care management).
Preference Pairs: From 39,317 assessed patients, 1,707 eligible "within-patient preference pairs" were constructed. A primary pair consists of a patient with a favorable 90-day acute care outcome post-enrollment and an unfavorable outcome pre-enrollment. Secondary pairs include weak-positive outcomes and intent-to-treat windows based on Medicaid eligibility churn.
Intervention Misalignment Metric: Misalignment is defined as the fraction of patients where the assigned action ( $A_i$ ) is not the action ( $a$ ) expected to minimize the 90-day acute care event probability by at least a threshold $\epsilon = 0.02$ . The expected outcomes are estimated using Doubly Robust (DR) counterfactual modeling (Augmented Inverse Probability Weighting, AIPW) across 14 candidate intervention types.
Policy Evaluation: Thirteen pre-registered routing policies were evaluated on a held-out test set (N=6,995) using Doubly Robust Off-Policy Evaluation (DR-OPE).
- PEARL (Policy Evolution through Aligned Retrospective Learning): The primary proposed policy combines three components: (1) a within-patient signal derived from preference pairs, (2) group-stratified fairness constraints to prevent demographic disparities, and (3) a fourteen-expert router (one expert per intervention type).
- Comparators: Included behavioral cloning, observational Direct Preference Optimization (DPO), Conservative Q-learning (CQL), causal forests, risk scores (LACE, HOSPITAL), and Decision Transformers.
Validation: Six pre-specified falsification tests (T1–T6) were conducted to validate the natural experiment, including covariate balance checks, placebo outcomes, and tests for treatment effect heterogeneity.

Key Results

Baseline Misalignment: Under current behavioral routing, intervention misalignment was 27.0% (95% CI, 25.9–28.1%). This indicates that over one in four patients received a suboptimal intervention.
Disproportionate Under-prioritization of Social Needs: Current routing heavily favored care access coordination (52.5%) and behavioral health (27.8%), assigning social needs (transport, food, housing, financial benefits) to only 1.9% of patients. In contrast, within-patient preference pairs indicated social needs should be the next best action in 20.9% of cases. This represents a 12.5- to 25.5-fold gap in prioritization.
PEARL Performance: The PEARL policy reduced intervention misalignment to 2.0% (a reduction of 25.0 percentage points, $p < 0.0001$ ), achieving 92.6% of the maximum possible reduction toward a theoretical oracle. It also achieved the lowest DR-OPE (0.027) among deployable policies.
Component Ablation:
- The within-patient signal was necessary; observational DPO without it only reduced misalignment to 20.4%.
- Behavioral cloning on within-patient preferences alone reduced misalignment to 4.4%, suggesting the signal itself is the primary driver of improvement.
- Group-stratified fairness was critical; PEARL variants without fairness constraints increased misalignment to 38.7–42.4%, performing worse than the baseline.
Camden Coalition Reanalysis: In a matched subgroup of 242 patients resembling the Camden Coalition trial criteria, the uniform protocol produced 90.9% misalignment. The authors' outcome model simulated a 15.1 percentage point reduction in 30-day readmissions if PEARL's personalized selection had been used instead.

Significance and Claims
The paper claims to establish intervention misalignment as a measurable and modifiable property of care management selection policies. The study demonstrates that:

Current routing mechanisms systematically misalign interventions, particularly by under-prioritizing social needs despite strong within-patient evidence for their necessity.
Reducing misalignment requires a specific combination of causal signals (within-patient natural experiments), fairness constraints, and expert routing architectures.
High misalignment is a plausible mechanism for the null results observed in previous uniform-intervention trials (like Camden Coalition), as uniform protocols applied to heterogeneous populations inherently produce high misalignment.
The proposed PEARL framework is deployable (coverage relative to behavioral support set is 1.00) and significantly outperforms existing risk-score and reinforcement learning baselines in off-policy evaluation.

The authors note limitations, including that the natural experiment identifies the effect of care management receipt rather than specific head-to-head action comparisons, and that external validation in non-Medicaid populations is pending. They conclude by announcing a planned stepped-wedge cluster-randomized trial to prospectively validate these findings.