Clinical Scorecard: A Transformer-Based Model for Predicting All-Cause Mortality in Heart Failure Patients: Insights from a Multi-Cohort Analysis
At a Glance
Category
Detail
Condition
Heart failure (HF)
Key Mechanisms
Complex, multifactorial risk including cardiovascular function, comorbidities, and interventions; use of electronic health records (EHR) and AI transformer models to capture longitudinal patient data
Target Population
Patients aged 40-90 years diagnosed with heart failure in primary or secondary care
Care Setting
Primary and secondary care settings using routine electronic health records
Key Highlights
TRisk, a transformer-based AI model, predicts 36-month all-cause mortality in HF patients with superior discrimination (C-index 0.845) compared to adapted MAGGIC-EHR model (C-index 0.728).
TRisk incorporates comprehensive longitudinal EHR data including diagnoses, medications, and procedures without imputing missing values, enabling nuanced risk stratification.
TRisk demonstrated better calibration, higher net clinical benefit across risk thresholds, and generalizability validated in UK and US cohorts.
Guideline-Based Recommendations
Diagnosis
Utilize comprehensive EHR data capturing diagnoses, medications, procedures, and test results for risk assessment in heart failure patients.
Consider models that integrate multimorbidity and interventions beyond traditional cardiovascular markers.
Management
Apply robust risk prediction models like TRisk to inform medium-term prognosis and guide clinical decision-making and interventions.
Use risk stratification to facilitate effective communication with patients and audit quality of care.
Monitoring & Follow-up
Monitor patient outcomes using validated AI models that incorporate longitudinal health data for dynamic risk assessment.
Regularly update risk predictions as new EHR data become available to capture evolving patient status.
Risks
Be aware that traditional models relying on limited cardiovascular markers may underestimate risk due to omission of multimorbidity.
Ensure transparency in model performance metrics and validation across diverse populations to mitigate bias.
Patient & Prescribing Data
Heart failure patients aged 40-90 years with diverse comorbidities and medication histories
Approximately one-third of patients had diabetes, myocardial infarction history, or beta-blocker prescriptions at baseline; TRisk model incorporates medication codes to enhance prognostic accuracy.
Clinical Best Practices
Incorporate AI-based transformer models like TRisk that utilize comprehensive, timestamped EHR data for mortality risk prediction in HF.
Validate prognostic models externally across different geographic populations to ensure generalizability.
Use models that do not require imputation of missing data to maintain data integrity and improve prediction accuracy.
Employ decision curve analysis to evaluate net clinical benefit of risk prediction models across relevant thresholds.
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
