💊 AEGIS — Medication Adherence Risk Explorer P2 · static

Score a CSV of patients

Upload a CSV with the 11 model columns: AGE, ANNUALCONTRIBUTION, ANNUALCLAIMAMOUNT, UNITSTOTAL, GENDER_M, SCHEMETYPE_MEDIUM, SCHEMETYPE_PREMIUM, DIAGNOSIS_HYPERTENSION, COVERTYPE_STANDARD, COMORBIDITY_NO_COMORBIDITY, COMPLICATIONDEVELOPMENT_NO_COMPLICATION (one-hot columns 0/1). Scored with the model selected above.

How the 5 candidate models compared during development

All 5 models are shown for transparency, even though only Gradient Boosting, Random Forest, and Extra Trees are offered for live prediction.

What drives the prediction (Random Forest feature importance)

Total medication units and annual claim amount dominate, with age a distant third — the same pattern reported in the original case study.

What this app is (and isn't)

This is a testing wrapper, not a validated clinical or underwriting tool. The point of this phase of AEGIS is to see how the chosen model behaves on real inputs in something simple and interactive.

No trained model or cleaning script survived from the original project run. This app's models were retrained from scratch on the recovered "Final Prepared Dataset" (24,084 patients, 11 features). The retrained Random Forest's feature importances match the original project's to several decimals, and Logistic Regression reproduces the exact collapse-to-majority-class behaviour reported in the original case study — both strong signals the rebuild is faithful.

Data source: Kanyongo, Moyo, Ezugwu & Fonou Dombeu, "Dataset for analysing medication adherence among diabetes and hypertension patients," Mendeley Data, DOI 10.17632/zkp7sbbx64.2 (CC0).

Target framing: the model predicts the probability a patient is non-adherent (the higher-risk class for care management), not the probability of adherence.

Static edition — runs entirely in your browser; no server. Predictions are bit-exact reproductions of the original scikit-learn models (verified to machine epsilon).