Externally validated AI-driven multimodal prediction of progression from Stage B (asymptomatic LV dysfunction) to overt heart failure: incremental prognostic value of GLS and NT-proBNP
Y Qianyi, M Juneja, R Khan, H Noor AIn, H Ramteke, P Aiswarya, I Thomas jollyAbstract
Background
Stage B heart failure (asymptomatic LV dysfunction) represents a critical transition phase preceding overt HF, yet risk stratification remains inadequate. Conventional clinical models lack precision. Advanced imaging (GLS) and biomarkers (NT-proBNP) may detect subclinical progression earlier.
Purpose
To develop and externally validate an AI-driven multimodal prediction model integrating clinical variables, echocardiographic strain (GLS), and NT-proBNP to identify asymptomatic Stage B heart failure patients at high risk of progression to overt Stage C HF within 3–5 years.
Methods
We conducted a retrospective cohort study of 1,000 asymptomatic Stage B HF patients with baseline clinical data, NT-proBNP, and echocardiography including GLS. Progression to Stage C HF over 3.5 years was assessed. An AI-driven gradient boosting model was developed, internally cross-validated, and externally validated against a standard clinical model.
Results
A total of 1,000 patients with Stage B heart failure (asymptomatic LV dysfunction; LVEF 40–50% and/or abnormal GLS) were prospectively analysed and followed for a median of 3.5 years. Mean age was 64±11 years, and 42% were female. Overall, 120 patients progressed to Stage C HF (12%; 95% CI 10.0–14.2%). At baseline, progressors were older (71±9 vs. 63±12 years, p<0.001) and more frequently had diabetes (45% vs. 26%, p<0.001) and hypertension (85% vs. 72%, p=0.003). Biomarker and echocardiographic differences were marked: NT-proBNP was higher (median 480 vs. 125 pg/mL, p<0.001), GLS more impaired (−12.8±1.8% vs. −15.8±2.0%, p<0.001), LVEF lower (42±3% vs. 47±4%, p<0.001), E/e′ ratio higher (14.1±3.2 vs. 9.8±2.5, p<0.001), and LV mass index greater (112±22 vs. 95±16 g/m², p<0.001). In multivariable Cox regression, independent predictors of progression included GLS >−13% (HR 2.92; 95% CI 2.12–4.02), NT-proBNP >300 pg/mL (HR 3.38; 95% CI 2.47–4.63), diabetes (HR 1.81; 95% CI 1.29–2.54), and LV mass index per 10 g/m² (HR 1.24; 95% CI 1.10–1.39) (all p<0.001). The clinical model (age, diabetes, hypertension) demonstrated moderate discrimination (AUC 0.68; 95% CI 0.63–0.73), sensitivity 62% (95% CI 53–71%), specificity 64% (95% CI 61–67%), recall 0.62, and F1-score 0.34. The multimodal model integrating GLS and NT-proBNP significantly improved performance (AUC 0.84; 95% CI 0.80–0.88; p<0.001), with sensitivity 78% (95% CI 70–85%), specificity 81% (95% CI 78–84%), recall 0.78, and F1-score 0.56. Net reclassification improvement was 0.29 (95% CI 0.18–0.40). These findings demonstrate substantial incremental prognostic value of combined strain and biomarker assessment in identifying the Stage B “transition zone.”
Conclusion
An externally validated AI-driven multimodal model integrating GLS and NT-proBNP significantly improves prediction of progression from Stage B to overt heart failure, enabling targeted preventive intervention.