Key result
Machine learning models predict in-hospital heart failure mortality with an AUROC of ~0.98.
Why the study?
Heart failure is a major public health challenge in Portugal with high mortality and healthcare costs, motivating the development and validation of artificial intelligence models to predict in-hospital mortality using integrated primary and hospital care data.
Do machine learning models (Neural Networks, LightGBM, XGBoost) accurately predict in-hospital mortality in adult heart failure patients using integrated primary and hospital care data?
Cohort (n=7,462)
No
Do machine learning models (Neural Networks, LightGBM, XGBoost) accurately predict in-hospital mortality in adult heart failure patients using integrated primary and hospital care data?
Machine learning models (LightGBM and XGBoost) using routinely collected primary and hospital care data in Portugal effectively predicted in-hospital heart failure mortality with AUROC up to 0.98, while revealing significant care integration gaps.
Answers are grounded in this paper and cite their sources.
Gaps in primary care HF documentation may compromise care continuity; leaves open the reliability of integrated records for AI mortality prediction.
Background Heart failure (HF) represents a major public health challenge in Portugal, associated with high mortality and healthcare costs. This study aimed to develop and validate artificial intelligence models to predict in-hospital mortality among HF patients at the Baixo Alentejo Local Health Unit, using integrated primary and hospital care data from 2017-2023. Methods In this retrospective observational study, we analysed 7,462 adult HF patients using data from primary care records and hospital admissions. Models were developed using various machine learning approaches (Neural Networks, LightGBM, XGBoost) incorporating clinical, sociodemographic, and deprivation index data. Performance was evaluated using AUROC, AUPRC, and other metrics, with 10-fold cross-validation. Results The cohort, 54.9% female, median age 81, showed 5-year survival of 68.5%. Of 5,796 hospital admissions, 1,135 resulted in death. Only 46.6% of hospitalized patients had HF documented in primary care. The best-performing models (LightGBM and XGBoost) achieved AUROC values up to 0.98, with age, recent clinical monitoring, NT-proBNP levels, and comorbidities as key predictors. Different imputation strategies and SMOTE balancing were evaluated, with Reference and Median imputation methods showing superior performance. Conclusions This pioneering study in Portugal demonstrates the potential of machine learning to predict HF mortality using routinely collected healthcare data. The robust performance of our models suggests their potential utility in clinical settings, while highlighting opportunities for improving care integration between primary and hospital services. Further validation in other Portuguese healthcare organizations is necessary. Key messages • This study demonstrates that machine learning models can effectively predict in-hospital mortality for HF patients using routinely collected Portuguese healthcare data. • The finding that only 46.6% of hospitalised patients had HF documented in primary care highlights important care integration gaps.
No takes yet. Share an insight, caveat, or question.
Diniz et al. (2026) conducted a cohort in Heart failure (n=7,462). Machine learning models (Neural Networks, LightGBM, XGBoost) was evaluated on In-hospital mortality prediction (AUROC up to 0.98). Machine learning models (LightGBM and XGBoost) predicted in-hospital mortality in heart failure patients with an AUROC of up to 0.98 using routinely collected healthcare data.
Synapse has enriched 10 related papers. Consider them for comparative context: