Interoperability, Data Governance, and the Analytics Ask
A regional accountable care organization (ACO) wants to build a predictive model that flags patients at high risk of 30-day readmission so care managers can intervene. Data come from three sources: the ACO's primary care EHR, two independent hospital EHRs with different vendors, and a health information exchange (HIE) feed that is incomplete because one large hospital system opts out of sharing behavioral health data. The compliance officer flags that behavioral health data carries heightened confidentiality protections under state law and 42 CFR Part 2 for substance use disorder records. The analytics team proposes using all available fields, including behavioral health diagnoses, to maximize model accuracy. The ACO's attributed population is 41,000; the model would be deployed to prioritize roughly 2,000 outreach calls per quarter.
1. What data-governance constraints must shape the model design, and how should the team proceed?
Read the worked answer
The team must reconcile model performance against legal and ethical constraints on data use:
- 42 CFR Part 2 (SUD records) requires specific patient consent for disclosure and re-disclosure limits; SUD data generally cannot be freely used in a shared analytics pipeline without consent. State behavioral health confidentiality laws may be stricter still.
- Opt-out fragmentation. The HIE gap means the model's inputs are systematically missing for patients seen at the opting-out system - a missing-not-at-random problem that can bias predictions and disadvantage an entire subpopulation.
- Minimum necessary and purpose limitation. Using behavioral health fields to prioritize outreach may exceed the permitted purpose absent consent.
How to proceed:
- Conduct a data-use inventory and legal mapping of each field against Part 2, state law, and HIPAA minimum-necessary before modeling.
- Exclude or de-identify restricted behavioral health fields from the shared model; if clinically valuable, pursue an explicit consent pathway rather than assume access.
- Document and quantify the missingness from the opt-out system and test whether the model performs equitably across patients with and without HIE coverage.
- Obtain IRB/ethics and compliance sign-off and build an audit trail.
The tested judgment: accuracy is not the only objective; lawful, unbiased, purpose-limited data use governs the design.
2. How should the ACO evaluate whether the model is fair and effective before scaling outreach?
Read the worked answer
Before scaling to 2,000 calls/quarter, the ACO should run a fairness and effectiveness evaluation:
- Performance by subgroup: measure sensitivity, positive predictive value, and calibration across race, language, insurance, disability, and by data-source (patients with full vs. partial HIE coverage). A model that predicts well overall but poorly for the opt-out population would systematically miss those patients.
- Bias review: examine whether behavioral health exclusion or missingness proxies for protected characteristics and skews who gets called.
- Effectiveness test: run a pilot with a comparison group (e.g., randomized or stepped-wedge) to confirm that flagged patients actually benefit from outreach - a high-risk flag is useless if the intervention doesn't change outcomes.
- Capacity alignment: 2,000 calls/quarter against a 41,000 population means the model must rank, not just classify; validate that the top-ranked patients are genuinely actionable and that care-management capacity matches.
- Monitoring plan: define drift, override, and equity metrics to review quarterly post-deployment.
The tested judgment: validate fairness and real-world benefit, not just AUC, before committing scarce outreach capacity.