Predictive healthcare ML: from cohort definition to operational action

One discipline across several prediction problems

Readmission, length of stay, emergency-department utilization, quality gaps, and transportation anomalies have different labels and operating rhythms. They share a harder requirement: the score has to arrive with enough context for a real team to act responsibly.

This case study is a public-safe synthesis of multiple engagements. It shows the modeling and delivery discipline without collapsing distinct clients, cohorts, or outcomes into one fictional product.

My role

I worked across cohort definition, data-quality rules, feature engineering, baseline and challenger models, calibration, explainability, evaluation slices, deployment, monitoring, and the analytics surfaces used by operations and care teams.

The product question was always paired with the statistical one: who will use the score, what action can follow, what capacity exists, and what failure is more costly?

Cohorts and labels before algorithms

A technically clean model can still be wrong for the workflow if the index event, observation window, prediction horizon, exclusions, or outcome definition drift from the operational decision. I treated those definitions as versioned contracts.

Claims, ADT events, demographics, diagnoses, medications, prior utilization, quality signals, and social-risk features were joined with explicit provenance and point-in-time boundaries.

• Prevent post-outcome leakage
• Keep training and scoring definitions aligned
• Measure missingness and freshness by source
• Document proxy and access concerns

Baselines, challengers, and calibration

Interpretable baselines established whether additional complexity earned its operational cost. Logistic models, rules, and simple risk scores were compared with boosted trees and task-specific challengers.

AUROC alone was never the release decision. Precision-recall behavior, Brier score, calibration curves, threshold tradeoffs, and slice-level stability mattered because teams consume ranked probabilities under limited capacity.

Explainability as a review aid

Feature attribution helped debug data and communicate why a record moved in the ranking. It did not establish causality or justify an intervention by itself.

Useful reviewer surfaces paired risk bands with current evidence, temporal context, missingness, and the action boundary. Sensitive or weak proxies required explicit review rather than a prettier explanation chart.

From score to operating system

Batch and API delivery paths were designed around the receiving workflow: refresh cadence, available capacity, escalation rules, feedback capture, and the difference between informational and action-triggering outputs.

Monitoring separated service health, data drift, score distribution, calibration, slice behavior, downstream action, and outcome. A healthy endpoint was not treated as a healthy model.

What failed or underperformed

Leakage from post-event fields, changes in coding practice, sparse subgroups, shifting utilization patterns, and threshold choices disconnected from operational capacity all created failure modes that a single aggregate metric would hide.

Where a model did not beat an interpretable baseline or could not be connected to a responsible action, the right outcome was to simplify, keep it in shadow mode, or stop.

Documented outcomes

In the transportation FWA workflow, explainable anomaly review supported an 18% reduction in waste. Other models supported care-management prioritization, utilization insight, quality work, and the reusable analytics platform described in the adjacent case study.

These outcomes belong to their documented workstreams. This page does not imply clinical validation, universal performance, or that a model alone produced a program result.

What I would improve next

I would formalize dataset and model cards in the release path, expand temporal and subgroup validation, pair calibration monitoring with action-rate monitoring, and require a shadow-mode checkpoint before every material cohort or feature change.