Early detection of heart failure using trend similarity metrics combined with hybrid deep learning based on real-world health monitoring data

Heart failure (HF) is a global health crisis, with over 18 million annual deaths attributed to cardiovascular diseases. Early detection is critical to enable timely lifestyle interventions and targeted pharmacotherapy, yet existing prediction methods often fail to integrate high-dimensional multi-source data or capture dynamic temporal patterns in patient health records. Here, we present a cross-disciplinary framework that combines trend similarity analysis (rooted in Haar wavelet decomposition) and hybrid deep learning to address these limitations. Leveraging real-world data from the HeartCarer telemonitoring system-encompassing 2976 patients with HF, 18 203 family members and 295 801 healthy individuals (2018-2024)-we integrate physiological sensor data (e.g. heart rate, blood pressure) and electronic health records (EHRs, e.g. disease history, lifestyle factors) through structured/categorical feature fusion. Our trend similarity metric quantifies temporal consistency in health data sequences, while the deep learning model (LogitBoost meta-classifier with leaky rectified linear unit (LReLU) activation) performs binary HF risk classification. The framework achieves 98.9% prediction accuracy, outperforming traditional machine learning methods (e.g. support vector machine (SVM), random forest: 70-90%) and our prior work (90%, 98.5%). This approach bridges computational science and clinical practice, providing a scalable tool for early HF warning that is validated in real-world clinical settings.