ML-HF Risk Stratifier (Gradient Boosting model)

ML-HF Risk Stratifier — Explanation and Clinical Context The ML-HF Risk Stratifier presented here is a prototype web tool that approximates the behavior of a gradient-boosting model (CatBoost) for predicting a composite 30-day outcome after a heart failure (HF) emergency department visit or hospitalization: HF-related ED visit, HF hospital readmission, or all-cause death. The underlying published work (Fine et al., 2024) trained CatBoost on >50,000 patients from a large administrative dataset and demonstrated superior discrimination compared with logistic regression for both 30-day and 1-year endpoints (CatBoost AUC-ROC ≈ 0.74 at 30 days for the composite endpoint; logistic regression ≈ 0.62). The model used deep feature synthesis to expand administrative data into engineered predictors and identified a set of top features contributing to risk.

The prototype scoring algorithm in this page uses those published feature importance signals to create an interpretable weighted score mapped to a predicted probability via a logistic transform. This approximation is useful for web demonstration, educational purposes, and for local research; it is NOT a replacement for the original trained CatBoost artifact and should not be used alone for clinical decisions. How to interpret: The output is a predicted probability of the composite 30-day event (HF ED visit, HF readmission, or death).

Risk categories are intentionally conservative: Very low <5%, Low 5–15%, Moderate 15–30%, High >30%. Clinicians should treat this as supportive information only and combine model output with individual clinical assessment and pathway protocols. Limitations: 1) The original CatBoost model was trained on administrative data from Alberta (2002–2016) and performance / calibration may differ in other health systems or time periods.

2) The original model’s trained parameters are not public in this page — exact replication requires access to the original training artifacts or re-training on local data. 3) This implementation approximates gradient-boosting behavior via a weighted logistic mapping derived from feature importance ranks for transparency; it cannot capture complex non-linear interactions present in a fully trained ensemble. External validation and prospective calibration are required before any clinical use.