Apppublic

erikabear95/transitionrx-dashboard

sourceHugging Facemitupdated 1mo agoView on Hugging Face

0likes

App README

TransitionRx — Medication Complexity & Care-Transition Readmission Risk

Decision-support tool for hospital pharmacy and care-management leadership.

Project question: For a hospital facing CMS Hospital Readmissions Reduction Program (HRRP) penalties, can discharge-level data already on hand — medication count, regimen churn, and care-transition quality — identify which diabetes patients are at elevated risk of a preventable 30-day readmission, so limited pharmacist time is targeted where it matters?

What's in the app

Tab	Contents
📖 Start here	Business problem, the decision supported, how to use the tool
📊 Risk dashboard	Cohort risk distribution, capacity-vs-captured-risk curve, ranked worklist
👤 Patient brief	Single-patient scoring with SHAP driver explanation
💬 Guidance Q&A	RAG over CMS/AHRQ discharge-planning guidance, with citations
💰 Cost / value	Scenario table, net-value sensitivity, resource impact
⚠️ Limitations	Data gaps, performance ceiling, fairness caution, appropriate use

Every headline figure is driven by sidebar assumptions the user controls — pharmacist capacity, cost per readmission, cost per intervention, and assumed intervention effectiveness.

Configuration

Set these under Settings → Variables and secrets:

Name	Type	Required	Purpose
`HF_MODEL_REPO`	Variable	If artifacts aren't committed	Model repo, e.g. `username/transitionrx-readmission-xgboost`
`OPENAI_API_KEY`	Secret	No	Enables generated RAG answers. Without it the Q&A tab runs in retrieval-only mode

Files this Space expects

Committed directly, or downloaded from HF_MODEL_REPO:

—transitionrx_xgb_model.pkl — tuned XGBoost classifier
—transitionrx_calibrated_model.pkl — isotonic-calibrated version
—transitionrx_metadata.json — metrics, threshold, assumptions
—transitionrx_feature_columns.csv — required feature order
—sample_cohort.csv — demo cohort (optional; synthetic data used if absent)
—transitionrx_faiss.index + transitionrx_chunks.parquet — RAG index (optional)

Model

Tuned XGBoost on the UCI Diabetes 130-US Hospitals dataset (1999–2008). Cohort excludes expired/hospice discharges and keeps the first encounter per patient. Target is readmitted == '<30'.

ROC-AUC is in the mid-0.60s — the realistic ceiling for administrative billing data and consistent with published work on this dataset. This ranks patients for a workflow decision. It is not a clinical diagnostic.

Important limitations

—Training data is 1999–2008; practice patterns have changed
—No vitals, lab trends, social determinants, adherence, or clinical notes
—Subgroup fairness across race, age, sex, and payer has not been audited
—The model predicts readmission risk, not preventability or responsiveness to intervention
—The intervention-effectiveness assumption comes from literature, not measurement here

Educational course project. Not a medical device.