ArticleBMJ health & care informatics2026
Mechanistic interpretability of reinforcement learning in Medicaid care coordination.
Article in BMJ health & care informatics, 2026. The graph could read no effect estimate from its abstract, so it casts no vote on the map. Not yet cited in PubMed.
What it found
Each row is one number read from the abstract, on the scale the paper reported it, with its interval. Left of the dashed line favours the treatment, right favours the comparator. Under each row is the sentence it came from. New to these charts? A ten-minute tutorial.
The abstract states no effect estimate the extractor could read, or names no intervention and outcome on the map, so this paper lights no cell and moves no belief. It is still indexed, cited and linked below.
The trial behind it
Trials whose registry record cites this paper, or whose number appears in the abstract. A trial that started after this paper was published is citing it as background, not reporting it.
Neither the registry nor the abstract names a trial number. If this is a trial report, that itself is worth knowing.
Who cites it
0 citing papers in PubMed.
No citing paper in PubMed yet.
Corrections and comments
PubMed lists nothing against this paper. Absence here is not a guarantee, only a check that was made.
Authors and funding
7 authors.
Funding
No grant is acknowledged in the PubMed record.
Abstract
objectiveTo expose reasoning pathways of a reinforcement learning policy for Medicaid care coordination, develop an error taxonomy and implement fairness-aware guardrails.
designRetrospective interpretability audit using attention analysis, Shapley explanations, sparse autoencoder feature discovery and blinded clinician adjudication.
settingMedicaid care coordination programmes in Washington, Virginia and Ohio (July 2023-June 2025).
participants250 000 intervention decisions; 200 divergent cases reviewed by five clinicians.
main outcome measuresCalibrated harm prediction; algorithmic clearance and residual harm rates; error taxonomy frequencies; subgroup fairness metrics.
resultsThe conformal model achieved area under the receiver operating characteristic curve of 0.80 (95% CI 0.78 to 0.82), clearing 89.5% (95% CI 88.9% to 90.1%) of decisions with 1.22% (95% CI 1.14% to 1.30%) residual harm versus 6.67% (95% CI 6.02% to 7.32%) for flagged decisions. Sparse autoencoders identified seven reasoning motifs linking social determinants to clinical cascades. The error taxonomy revealed premise errors (48%, 95% CI 41% to 55%), calibration failures (27%, 95% CI 21% to 33%) and contextual blind spots (25%, 95% CI 19% to 31%). Divergence was higher for telehealth visits (11.2%) and behavioural health patients (10.7% vs 6.9%, p<0.001). Fairness optimisation reduced race-group disparity by 37% (95% CI 22% to 48%) and sex-group disparity by 28% (95% CI 14% to 39%). Reviewers rated 23% (95% CI 17% to 29%) of overridden recommendations as well-matched, confirming appropriate human oversight.
conclusionsMechanistic interpretability transforms opaque algorithmic assistance into auditable decision support, providing a governance scaffold for clinical artificial intelligence deployment.
Indexed as
Identifiers
What OpenQuestion holds
Registered trials
Read under generation 80e0d062 · epoch 390. Bibliography from PubMed, PubMed Central and OpenAlex; grants from NIH RePORTER; trial links from ClinicalTrials.gov; estimates, votes and beliefs from the OpenQuestion graph.