ML-HF Risk Stratifier (Gradient Boosting model)

Clinical calculation support · Verify results before clinical use

Post-acute heart failure encounter review

Enter encounter characteristics to review the available variables. The published CatBoost model is not embedded in this page.

Research prototype · no validated prediction
Important model limitation. The published study's performance cannot be reproduced from feature rankings alone. The original trained model and its complete feature-engineering pipeline are not supplied here; therefore this page intentionally does not generate a numerical probability or risk category. This prevents an unvalidated estimate from being mistaken for a clinically calibrated prediction.
Enter 0 for an ED encounter without an inpatient stay, if applicable.
Optional. Leave blank if unavailable.
Optional. Leave blank if unavailable.
Reset form

About the published model

Fine and colleagues reported a machine-learning approach for predicting 30-day and 1-year outcomes after an HF-related emergency department visit or hospital admission. The study used administrative health data from Alberta, Canada, covering 2002–2016, and included 50,630 patients with 93,552 HF emergency visits or hospitalizations.

The primary composite endpoint included HF-related emergency department visits, HF hospital readmission, or all-cause death. In the holdout validation cohort, the 30-day AUROC was 0.7416 for CatBoost and 0.6225 for logistic regression.

Why this page does not estimate risk

  • Published feature importance is not a set of model coefficients and cannot be converted directly into calibrated probabilities.
  • The original pipeline used engineered features from multiple administrative data sources; the small set of bedside inputs on this page is not an exact reproduction.
  • The arbitrary weights and risk thresholds in the previous prototype were not reported parameters from the publication and have no demonstrated calibration. They have therefore been removed.
  • Clinical use would require access to the validated model and its preprocessing pipeline, followed by local external validation and calibration.

Reference

Fine NM, Kalmady SV, Sun W, Greiner R, Howlett JG, White JA, et al.Machine Learning For Risk Prediction After Heart Failure Emergency Department Visit or Hospital Admission Using Administrative Health Data. PLOS Digital Health. 2024;3(10):e0000636.https://doi.org/10.1371/journal.pdig.0000636PMID: 39453878
Clinical disclaimer: This page is a research and educational interface, not a validated clinical prediction instrument. It does not calculate an individual probability of readmission, emergency department return, or death. Do not use it as the sole basis for triage, discharge, follow-up, or treatment decisions. Use clinical assessment and locally approved pathways.