Within a Heartbeat: A Machine Learning Approach Using VCG for Heart Failure Assessment in Sinus Rhythm.

Background/Objectives: Heart failure (HF) is associated with structural changes such as ventricular dilation, hypertrophy and fibrosis that alter conduction pathways and leave characteristic traces in the electrical activity of each heartbeat. Individuals diagnosed with HF exhibit a range of clinical manifestations and require diverse therapeutic interventions. Nevertheless, an accurate and timely diagnosis must precede any therapeutic approach. Patients with HF and a high ejection fraction (EF) frequently remain undiagnosed, as preliminary medical assessments cannot effectively distinguish them from asymptomatic individuals. Methods: In this work, vectorcardiography (VCG) data (277 samples) were used from the PTB and MUSIC databases, with a primary endpoint of identifying individuals with a high likelihood of HFpEF (EF ≥ 50%) compared to healthy subjects to help mitigate overlooked cases and lay the groundwork for patient-centered care. The second endpoint was to screen whether an individual was healthy or had an HF subtype (HFrEF or HFpEF), making it relevant as a potential triage support in clinical practice. To achieve these objectives, heart-rate variability (HRV) features and continuous wavelet transform (CWT) features for each QRS complex and ST-T segment were extracted to train traditional machine learning (ML) models: support vector machine (SVM), random forest (RF), and eXtreme gradient boosting (XGBoost). Results: Across 100 random train/test splits, the first endpoint achieved Recall = 0.88 and AUROC = 0.98. For the second endpoint, aggregated across three internal test sets, the classification of healthy individuals, HFrEF and HFpEF resulted in Recall = 0.88, AUROC = 0.99, Recall = 0.76, AUROC = 0.87, Recall = 0.60 and AUROC = 0.79, respectively. Conclusions: These findings suggest that ECG/VCG-based machine learning approaches may contribute to HF subtype stratification and have the potential to broaden access to precision diagnosis and personalized medicine. As healthy and HF patients come from different databases and the study lacks external validation, findings should be interpreted with care.