Predicting end-of-life in older women with heart failure: development and internal validation of clinically actionable prognostic models using routinely collected national data

Prognostic uncertainty in older women with heart failure often delays end-of-life discussions and leads to unnecessary interventions. This study aimed to develop, internally validate, and evaluate the clinical utility of prognostic models for identifying patients nearing the end-of-life phase using nationally representative data from the Australian Longitudinal Study on Women’s Health (ALSWH), linked with administrative health records. This prognostic study included older women from the 1921–26 ALSWH cohort who had a documented diagnosis of heart failure. Predictors were selected through a systematic review and expert input including age, smoking status, body mass index, breathing difficulties, need for regular assistance with daily activities, being confined to bed or a chair for most or all of the day, the number of medications supplied, and the number of hospital admissions. We developed, validated, and compared five models: a multivariable logistic regression (LR) model and four supervised machine learning (ML) models (Gradient Boosting, Extreme Gradient Boosting (XGBoost), Support Vector Machine, and Random Forest). Model performance was assessed using discrimination (area under the receiver operating characteristic curve [AUROC]) and calibration (calibration plot, slope, and intercept). Clinical utility was evaluated through decision curve analysis. The analysis included data from 1,630 older women with heart failure (mean baseline age: 72.5 ± 1.5 years). The LR model demonstrated good discrimination (AUROC: 0.740; 95% CI: 0.716–0.763) and excellent calibration (slope: 1.00; 95% CI: 0.87–1.13). Among the ML models, Gradient Boosting and XGBoost showed similar discrimination (AUROC: 0.733; 95% CI: 0.709–0.757), with good calibration for Gradient Boosting (slope: 1.04; 95% CI: 0.90–1.17) and slight miscalibration for XGBoost (slope: 0.88; 95% CI: 0.77–1.00). Random Forest had the lowest discrimination (AUROC: 0.720; 95% CI: 0.696–0.745) and the poorest calibration (slope: 0.55; 95% CI: 0.47–0.63). Decision curve analysis indicated a net clinical benefit across a broad range of threshold probabilities for both LR and ML models. Prognostic models based on routinely collected health data can effectively identify older women with heart failure who are nearing the end-of-life stage. These models, particularly LR and Gradient Boosting, show promise in supporting timely palliative care referrals and guiding personalised care planning in clinical practice.