transitionrx-readmission-xgboost

TransitionRx - 30-Day Readmission Risk (Diabetes Encounters)

Gradient-boosted classifier that ranks diabetes inpatient discharges by risk of a 30-day readmission, so limited pharmacist and case-manager time can be targeted at the patients most likely to bounce back.

Intended use

Use: ranking discharges to prioritize pharmacist-led medication reconciliation and post-discharge follow-up under a fixed staffing capacity.

Not for: clinical diagnosis, treatment decisions, denying care, or any individual-level determination made without a clinician. The model predicts readmission risk, not preventability and not responsiveness to intervention.

Data

UCI Diabetes 130-US Hospitals, 1999-2008.

Cohort construction:

• Expired and hospice discharges removed (disposition IDs 11, 13, 14, 19, 20, 21) - these patients cannot be readmitted.
• First encounter per patient only - prevents outcome leakage across the train/test split.
• Target: readmitted == '<30'. Prevalence approximately n/a.

Features

116 features engineered from medication columns, prior utilization, discharge disposition, and diagnosis groupings. Headline engineered features:

• complexity_index - diabetes drug count + 2 x regimen changes
• regimen_change_score - dose escalations + de-escalations during the stay
• prior_inpatient_band, prior_emergency_band - prior-year utilization
• home_no_services, transferred_facility - care-transition quality proxies

Performance (held-out test set)

ROC-AUC in the mid-0.60s is the realistic ceiling for this dataset and consistent with published work on it. Administrative billing data contains no vital signs, lab trends, social determinants, medication adherence, or discharge-summary text. This is a ranking tool for allocating a scarce resource, not a clinical diagnostic.

Threshold is a business decision

The default 0.5 cutoff is not used. The operating threshold of 0.57 corresponds to reviewing roughly the top 20% of discharges - a pharmacist capacity constraint, not a statistical optimum. Re-tune it to your own capacity.

Files

Use the calibrated model for anything involving expected cost. scale_pos_weight distorts raw probabilities upward; they rank well but are not literal probabilities.

Usage

import joblib, pandas as pd
from huggingface_hub import hf_hub_download

model = joblib.load(hf_hub_download("erikabear95/transitionrx-readmission-xgboost", "transitionrx_xgb_model.pkl"))
cols = pd.read_csv(hf_hub_download("erikabear95/transitionrx-readmission-xgboost", "transitionrx_feature_columns.csv"))["feature"]

X_new = X_new.reindex(columns=cols, fill_value=0)   # column order must match exactly
risk = model.predict_proba(X_new)[:, 1]

Limitations

• Data is 1999-2008; coding practice and care patterns have changed.
• No vitals, lab trends, social determinants, adherence, or clinical notes.
• Trained on 130 US hospitals; case mix at your site will differ.
• Subgroup performance across race, age, and payer was not audited. Do this before any real deployment.
• The intervention-effectiveness assumption used in the value model is drawn from literature, not measured here. A prospective pilot is required to validate it.