Multimodal prediction of incident heart failure using clinical data, polygenic risk scores, and electrocardiogram-based artificial intelligence
A Shahid, S Kany, S Friedman, E Altonen, A Henry, R Rathod, S Koyama, K Miyazawa, R T Lumbers, J W Cunningham, N Mars, M Maddah, J E Ho, P T Ellinor, S KhurshidAbstract
Background
Heart failure (HF) is a progressive syndrome associated with high morbidity and mortality. The presence of clinically overt HF is a marker of worsening prognosis, highlighting opportunities for upstream prevention targeting at-risk individuals. Several methods to predict future HF exist, including clinical risk factor scores, polygenic risk scores (PRS), and electrocardiogram-based artificial intelligence (ECG-AI) models. Although individual methods demonstrate moderate discrimination of HF risk, the comparative utility of an integrated approach remains unknown.
Methods
We estimated the risk of incident HF using three previously validated approaches: the PREVENT-HF clinical risk score, a contemporary HF PRS, and an ECG-AI model for future HF (ECG2HF). HF events were identified using ICD codes and a validated electronic health record-based natural language processing (NLP) model. We assessed discrimination using time-dependent AUROC in two biobank cohorts external to the development of each risk component: the UK Biobank (UKB) and the Mass General Brigham Biobank (MGBB). Cox models incorporating two- and three-component (Tri-HF) combinations were developed in UKB and assessed in both datasets (Figure 1).
Results
Models were evaluated in UKB (N = 51,237, HF events = 388, age 64.7 ± 7.5 years, 51.6% women) and MGBB (N = 14,800, HF events = 492, age 55.7 ± 12.0 years, 53.8% women). In UKB, Tri-HF outperformed all single-component models with respect to HF discrimination (Tri-HF AUROC 0.811 [95% CI 0.789-0.834]; PREVENT-HF 0.788 [0.767-0.809]; PRS 0.544 [0.514-0.574]; ECG2HF 0.725 [0.696-0.754]). Performance was similar in MGBB (Tri-HF 0.811 [0.789-0.832]; PREVENT-HF 0.746 [0.723-0.770]; PRS 0.545 [0.518-0.571]; ECG2HF 0.786 [0.764-0.808]). Of note, the two-component model comprising PREVENT-HF + ECG2HF performed nearly identically to Tri-HF (UKB: AUROC 0.811 [0.787-0.834]; MGBB: 0.810 [0.790-0.831]). When stratified by Tri-HF, individuals in the highest tertile of risk showed higher cumulative incidence of HF (UKB: high 2.88% [95% CI 2.54-3.22%], intermediate 0.56% [0.41-0.70], low 0.25% [0.15-0.34]; MGBB: high 11.36% [10.17-12.53], intermediate 2.96% [2.37-3.54], low 0.76% [0.47-1.05]). When assessing HF incidence across model components, individuals in the highest tertile of risk for all three components showed increasing cumulative HF incidence (UKB: 4.72% [3.65-5.79]; MGBB: 16.36% [13.11-19.49]) versus those classified as high-risk by only two, one, or zero (UKB: 0.30% [0.20-0.41]; MGBB: 0.99% [0.66-1.32]) models (Figure 2).
Conclusions
An integrated HF prediction model offers favorable predictive utility compared to clinical factors, PRS, or ECG-AI prediction alone. Yet, population-level HF risk prediction using clinical factors and ECG-AI is not meaningfully improved by the addition of HF PRS. Prospective evaluation of preventive interventions focused on individuals at high predicted risk of HF is warranted.Figure 1Figure 2