MSN-04 // DEPLOYED

HOSPITAL READMISSION

90% recall on readmissions, from a Gradient Boosting model tuned for the costly miss.

25K Records90% RecallROC-AUC 0.66Threshold Tuned

BRIEFING

Readmission risk scoring on 25,000 anonymized records. Three models were compared on ROC-AUC, Logistic Regression, Random Forest, and Gradient Boosting, and the Gradient Boosting champion (ROC-AUC 0.66) was threshold-tuned to catch 90% of readmissions. In a clinical setting a missed high-risk patient costs far more than a false alarm, so recall is the right target. The model's feature importances surfaced prior inpatient visits, medication count, and lab procedures as the dominant drivers, pointing discharge planning at the right patients.

ROLE / METHOD / OUTCOME

Role
Solo analysis.
Method
Three-model comparison ranked by ROC-AUC, champion threshold-tuned for recall.
Outcome
90% of readmissions caught before discharge (ROC-AUC 0.66).

STACK: Python · scikit-learn · Gradient Boosting

MORE MISSIONS

Eight more case files — ML, BI, and supply-chain strategy.

VIEW ALL MISSIONS