To develop and validate a transformer-based model (TRisk) for predicting all-cause mortality in heart failure (HF) patients using electronic health records (EHR), addressing the limitations of current models.
Key Findings:
TRisk achieved a higher Concordance-index (C-index: 0.845) compared to MAGGIC-EHR (C-index: 0.728), indicating superior predictive performance.
TRisk demonstrated better discrimination across key subgroups, providing a more nuanced risk stratification.
Both models were well calibrated, but TRisk showed a more complex predictive distribution, indicating its ability to capture a wider range of risk.
Interpretation:
The TRisk model outperforms existing models in predicting mortality in HF patients, addressing limitations in current risk assessment approaches by utilizing comprehensive EHR data, which may lead to improved clinical outcomes.
Limitations:
The study is limited to specific cohorts from the UK and USA, which may affect generalizability to other populations.
Potential biases in EHR data and missing values could influence model performance and its applicability in diverse clinical settings.
Conclusion:
TRisk represents a significant advancement in mortality prediction for HF patients, emphasizing the need for robust models that incorporate multi-factorial risk profiles to enhance clinical decision-making.
by Shishir Rao, Nouman Ahmed, Gholamreza Salimi-Khorshidi, Christopher Yau, Huimin Su, Nathalie Conrad, Folkert W. Asselbergs, Mark Woodward, Rod Jackson, John GF Cleland, Kazem Rahimi
