CARE-HEART: AI-driven prediction in heart failure for timely advance care planning

Background End stage heart failure (HF) places a significant burden on patients and caregivers. Advance care planning (ACP) allows clinicians and patients to discuss their preferences in treatment approach. However, unpredictable HF progression causes prognostic uncertainty, which is a barrier to timely implementation of ACP. Purpose In this study we developed machine learning models to predict mortality and HF (re)hospitalisation at 3, 6, and 12 months in patients with HF at a large tertiary hospital in the Netherlands. Methods Separate prognostic models were developed for discharge after hospitalisation and for outpatient visit. Each encounter was treated as an inclusion event, with weights inversely proportional to the number of inclusion events per patient to ensure equal contribution. Candidate features included demographics, vital signs, conditions, medication, laboratory values, and procedures, resulting in 490 candidate features. For some features, short-term (≤3 years before encounter) and long-term (>3 years) variants were included. Two algorithms were compared: ℓ1-regularised Logistic Regression (Lasso) and Random Forests (RF). Models were evaluated for discrimination (AUC) and calibration (Brier Score), and benchmarked against the MAGGIC risk score for 1-year mortality. Results We included 77,597 patient encounters (15,732 inpatient; 61,865 outpatient). Median age was 71 [IQR: 61-78] years and 39.2% was female. Both Lasso and RF significantly outperform the MAGGIC benchmark (AUC: 0.706 [95%-CI: 0.683, 0.728]), achieving AUCs of 0.815 [0.796, 0.833] and 0.811 [0.793, 0.830] in the inpatient setting, respectively. Similar performance is achieved in the outpatient setting. Performance does not differ significantly between Lasso and RF across settings and outcomes (p > 0.10), and remains consistent across time horizons. Given its greater interpretability, we focussed on Lasso. The AUC for 1-year HF (re)hospitalisation is 0.705 [0.686, 0.724] inpatient and 0.741 [0.731, 0.752] outpatient. Brier Scores are 0.121 (mortality) and 0.208 (HF (re)hospitalisation) inpatient, and 0.090 and 0.152 outpatient (Figure 1). Top predictors for mortality are age, urea, and sodium. For HF (re)hospitalisation, encounter characteristics such as prior HF hospitalisation or unplanned visits are most important (Figure 2). Conclusion The machine learning models significantly outperform the MAGGIC benchmark with promising calibration and discrimination. The models are deployable as clinical decision support at any inpatient or outpatient encounter and the comparable performance across time horizons enables clinicians to personalise ACP planning for each individual patient. This is an early indication that artificial intelligence can be leveraged to reduce prognostic uncertainty in HF progression and support clinicians in timely ACP initiation.