Combining clinical data, polygenic risk scores, and electrocardiogram-based artificial intelligence for multimodal prediction of incident heart failure

S. Kany (Hamburg)¹, A. Shahid (Cambridge)², S. Friedman (Cambridge)², E. Altonen (Helsinki)³, A. Henry (Darlinghurst)⁴, R. Rathod (Cambridge)², S. Koyama (Cambridge)², T. Lumbers (London)⁵, J. Cunningham (Cambridge)⁶, C. Magnussen (Hamburg)¹, M. Maddah (Cambridge)², J. Ho (Cambridge)², P. Ellinor (Boston)⁷, S. Khurshid (Cambridge)^2
1Universitäres Herz- und Gefäßzentrum Klinik für Kardiologie Hamburg, Deutschland; ²Cambridge, USA; ³Helsinki, Finnland; ⁴Darlinghurst, Australien; ⁵London, Großbritannien; ⁶Broad Institute of MIT and Harvard Cardiovascular Disease Initiative Cambridge, USA; ⁷Institute for Technology Assessment Massachusetts General Hospital, Harvard Medical School Boston, USA

Background Heart failure (HF) remains a progressive syndrome with substantial morbidity and mortality, making the upstream identification of at-risk individuals a priority for preventive cardiology. While individual tools, such as clinical risk scores, polygenic risk scores (PRS), and emerging electrocardiogram-based artificial intelligence (ECG-AI), can predict incident HF, the comparative value of integrating these modalities is unknown. We sought to determine whether a multimodal approach improves HF risk stratification.

Methods We estimated incident HF risk using three validated tools: the PREVENT-HF clinical score, a contemporary HF PRS, and an ECG-AI model for future HF (ECG2HF). We identified HF events using ICD-9 and ICD-10 codes, along with a validated electronic health record-based natural language processing (NLP) model. Cox proportional hazards models combined these components into two-component and three-component (Tri-HF) risk estimators in the UK Biobank (UKB). We subsequently evaluated discrimination using time-dependent area under the receiver operating characteristic curve (AUROC) and average precision in the UKB and an external validation cohort, the Mass General Brigham Biobank (MGBB).

Results Models were evaluated in the 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 the UKB, Tri-HF outperformed all single-component models for 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 consistent in the 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]). Notably, a two-component model combining PREVENT-HF and ECG2HF performed nearly identically to the Tri-HF model (AUROC in UKB: 0.811 [0.787-0.834]; MGBB: 0.810 [0.790-0.831]). Individuals in the highest Tri-HF risk tertile demonstrated a higher cumulative incidence of HF (UKB: high risk 2.88% [95% CI 2.54-3.22], intermediate risk 0.56% [0.41-0.70%], low risk 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]). Furthermore, individuals categorized in the highest risk tertile across all three separate components exhibited a markedly increased cumulative HF incidence (UKB: 4.72% [3.65-5.79%], MGBB: 16.36% [13.11-19.49%]) compared to those flagged as high-risk by only two, one, or zero models (UKB: 0.30% [0.20-0.41%], MGBB: 0.99% [0.66-1.32%]).

Conclusions Integrating clinical risk factors with ECG-AI provides superior prediction of incident HF compared to isolated risk modalities. However, the addition of genomic data via PRS does not meaningfully improve population-level risk discrimination beyond clinical and electrocardiographic assessment alone. These findings support the prospective evaluation of scalable, multimodal clinical and ECG-AI tools to identify high-risk patients for targeted preventive interventions.


Figure 1: Study Overview


Figure 2: Cumulative risk of incident HF stratified by integrated HF risk estimation