Transformer-Based Model TRisk Improves Mortality Prediction in Heart Failure

Overview

The TRisk model, a Transformer-based AI leveraging comprehensive EHR data, significantly outperforms the adapted MAGGIC-EHR model in predicting 36-month all-cause mortality in heart failure patients. Validated on large UK and US cohorts, TRisk demonstrates superior discrimination, calibration, and clinical net benefit across diverse patient subgroups.

Background

Heart failure (HF) prognosis is challenging, especially for medium-term outcomes over years, which are critical for guiding interventions and care quality assessment. Existing risk models like MAGGIC rely on specialized tests and focus mainly on cardiovascular predictors, often missing the impact of multimorbidity that contributes to over 40% of mortality in HF. The availability of large-scale electronic health records (EHR) enables advanced AI models such as TRisk to capture complex, longitudinal patient data for improved risk stratification. TRisk processes extensive diagnosis, medication, and procedure codes from primary and secondary care, aiming to enhance mortality prediction accuracy and clinical utility.

Data Highlights

Key Findings

• TRisk achieved a significantly higher C-index (0.845) compared to MAGGIC-EHR (0.728) for 36-month mortality prediction.
• TRisk demonstrated better calibration with low integrated calibration index (ICI) and calibration curves closely aligned to reference.
• TRisk showed a bimodal risk prediction distribution, enabling nuanced stratification, unlike the unimodal distribution of MAGGIC-EHR.
• Decision curve analysis indicated TRisk provides greater net clinical benefit across relevant risk thresholds up to ~0.6.
• TRisk outperformed MAGGIC-EHR consistently across key patient subgroups with less deviation from overall cohort performance.
• TRisk processes extensive longitudinal EHR data including 2639 diagnosis codes, 633 medication codes, and 852 procedure codes without imputing missing values.

Clinical Implications

TRisk’s superior predictive performance and nuanced risk stratification can enhance medium-term mortality risk assessment in HF patients, facilitating timely clinical decision-making and personalized care planning. Its ability to integrate comprehensive, routinely collected EHR data supports scalable implementation across healthcare settings without reliance on specialized testing. Clinicians may leverage TRisk to improve prognostication accuracy and optimize resource allocation for patients at highest risk.

Conclusion

The Transformer-based TRisk model represents a significant advancement in heart failure mortality prediction by effectively utilizing large-scale EHR data to capture complex patient risk profiles. Its robust validation across UK and US cohorts underscores its potential for broad clinical adoption to improve patient outcomes.

References

1. Original Article -- A Transformer-Based Model for Predicting All-Cause Mortality in Heart Failure Patients