Transformers for all-cause mortality prognostication in patients with heart failure: validation of risk modelling and risk factor discoveries across international electronic health record datasets

Abstract Background Accurately predicting clinical outcomes in patients with heart failure (HF) can help inform patient management and service audits. Current risk scores rely on specialised tests (e.g., ejection fraction), but offer modest discrimination in large part as they fail to capture the evolving multi-factorial risk profiles of HF patients. Purpose Firstly, to develop and validate the Transformer-based Risk assessment survival model (TRisk) model, an artificial intelligence approach, for prediction of 36-month mortality in patients with HF using longitudinal data from routine UK electronic health records (EHR). Secondly, to employ post-hoc explainability methods to identify novel or under-appreciated mortality risk factors and better understand evolving patient profiles throughout the HF disease course. Methods A cohort of 403,534 HF patients aged 40 to 90 years was identified from 1,418 English general practices (1,063 for derivation and the remaining 355 for external validation). TRisk analyses temporal patient journeys—incorporating diagnoses, medications, and procedures documented up to the point of assessment—to generate personalised risk assessments for clinical decision support. TRisk was compared with the Meta-Analysis Global Group in Chronic HF model adapted for use with routine linked EHR (named as MAGGIC-EHR) 1. A second validation of TRisk using transfer learning was conducted on 21,767 HF patients from USA hospital admission data. Lastly, explainability analysis using the integrated gradients method examined which features TRisk prioritised for prognostication in each cohort. Results In the UK cohort, with median follow-up of 9 months (interquartile interval: 2, 29), the well-calibrated TRisk demonstrated a concordance index (C-index) of 0.845 (95% confidence interval: 0.841, 0.849), outperforming MAGGIC-EHR with a C-index of 0.728 (0.723, 0.733) for predicting 36-month all-cause mortality (Figure 1). Adapting TRisk to USA data using transfer learning yielded a C-index of 0.802 (0.789, 0.816) for predicting all-cause mortality. Explainability analyses showed TRisk captured established risk factors while also identifying several under-appreciated factors, notably cancers and hepatic failure, that were consistently important across both cohorts (Figure 2A). Additionally, the prognostic utility remained strong for cancers occurring even a decade before baseline, while other factors weakened with time (Figure 2B panel 3). Conclusions TRisk demonstrated well-calibrated, accurate prediction of mortality across both UK and USA healthcare settings using routinely collected EHR. Explainability analyses identified several risk factors not included in previous expert-driven models, underscoring the value of tracking longitudinal and evolving health profiles. By capturing complex patient journeys through routine EHR, the model would better stratify HF patients and guide management.Figure 1:Model discrimination (UK data) Figure 2:Risk factor discovery