By Preet Mehta
Every year, health plans send nurses to the homes of members who will not be members much longer. A Healthy Home Visit generates a paid claim, which feeds the risk-adjustment machinery, which eventually moves state payments — but only if the member is still eligible when the state takes its snapshot. Worse, the visit’s value decays: every subsequent claim eats away at what the visit can still explain. A visit to someone who disenrolled the next month is money spent for nothing. Across our Medicaid population, that waste ran to roughly $300 million. So we built a model to answer one question before scheduling a single visit: will this member still be here in three months?
Medicaid churn is an eligibility problem wearing a clinical disguise
Medicaid makes this harder than it looks. Unlike Medicare’s single national model, Medicaid is a patchwork of state programmes with their own eligibility rules — and eligibility churn, not member choice, drives most departures. People lose coverage because paperwork lapsed or income shifted, not because they picked a competitor.
We pulled thirteen data domains — enrollment and eligibility, visit dispositions, claims, diagnoses, pharmacy, primary-care relationships, demographics, risk scores, episodes — into a member-level table anchored to monthly prediction dates. Every feature was strictly point-in-time: nothing the model saw could leak information from after the anchor date. Seven thousand candidate features went in; about three thousand survived. Databricks AutoML handled the model search; the design discipline was ours.
Choose which mistake you can afford
We optimised for recall — 90%, with the false-negative rate under 10% — and accepted 60% precision in exchange. That was a business decision, not a statistical one. A missed high-risk member means a wasted visit; a false alarm means a care coordinator spends a few minutes reviewing a name. The confusion matrix is a budget document. Accuracy was 75% here, and it told us almost nothing about whether the model was useful. An AUC of 0.80 confirmed the ranking worked. What mattered was the trade-off, chosen deliberately and defended in plain language.
Eligibility, not illness, predicts who leaves
The model’s strongest signals were not clinical at all. Enrollment-stability patterns — coverage gaps, recertification timing, eligibility-category changes — dwarfed diagnoses and utilisation. Termination reason codes let us separate eligibility-driven churn, which dominates Medicaid, from voluntary switching. It was a humbling insight for a team that started by reaching for clinical features: the business already knew members churn, but it took the model to show that the churn signal lived in administrative data, not medical data. That changed where we invest in features — and where care teams focus retention effort.
Fairness as a design constraint
We excluded sensitive demographic features entirely and put the model through bias evaluation before it touched production. It passed — strong performance without protected attributes. Responsible ML works best as a constraint at design time, not an audit after the fact. If your model needs someone’s race or postcode to predict whether they’ll keep their coverage, the problem is your features, not your fairness metric.
The model runs in production on a quarterly cycle. High-risk lists go to the home-visit nurse company, and visits for those members are suppressed — the knock that would have been wasted never happens. Savings to date are roughly $300 million, and the methodology ports to any payer wrestling with eligibility churn.
A note on the ethics of not knocking, because it deserves a direct answer. Let’s be honest about what these visits are: payer-initiated and scheduled for risk adjustment — but a nurse in the home still does real clinical work, reviewing medications and spotting fall hazards. No member loses a doctor’s appointment, a prescription, or any covered benefit; while enrolled, all of that continues unchanged. The harder question is what happens after the nurse leaves.
A visit that surfaces a problem creates an obligation to follow through, and follow-through needs a relationship the plan will not have with a member who disenrolls next month. Sending nurses where the plan can act on what they find — and not where it can’t — is not rationing care. It is refusing to start care it cannot finish.
The broader lesson for analytics leaders: the highest-ROI model is rarely the cleverest one — it is the one aimed at the largest pile of preventable waste. Before asking what your next model should predict, ask where your organisation is already spending on people who will not be around to benefit. That list is your roadmap.
Outreach targeting deserves the same rigour as clinical prediction. Plans that predict disenrollment first stop paying for empty door knocks; the rest keep scheduling visits to members who have already left. The bottom line is unglamorous and worth repeating: know who will still be there before you go. In Medicaid, that single discipline is worth hundreds of millions.
– Preet Mehta is Manager, Healthcare Analytics at EXL Service in New York City, where he leads machine-learning and data-platform work for US healthcare payers