Novel fusion-based time-frequency analysis for early prediction of sudden cardiac death from electrocardiogram signals

• Published in: Medical engineering & physics Vol. 141; p. 104370
• Main Authors: Karimulla, Shaik, Patra, Dipti
• Format: Journal Article
• Language: English
• Published: England Elsevier Ltd 01.07.2025
• Subjects: Average fusion; Death, Sudden, Cardiac; Deep Learning; ECG; Electrocardiography; Female; Humans; Male; Middle Aged; Scalogram; Signal Processing, Computer-Assisted; Spectrogram; Sudden cardiac death; Time Factors; Time-frequency analysis; Transfer learning; Time-frequency analysis; Scalogram; Average fusion; Transfer learning; ECG; Sudden cardiac death; Spectrogram
• ISSN: 1350-4533; 1873-4030; 1873-4030
• DOI: 10.1016/j.medengphy.2025.104370

•Introducing a novel methodology for early prediction of SCD using fusion-based time-frequency representation (TFR) techniques.•Transformation of pre-processed ECG signals into 2D plots illustrating frequency variations over time via diverse time-frequency analysis methods.•Utilization of spectrograms, scalogram images, and their fused images as input data for transfer learning models to enhance predictive accuracy.•Extending the SCD prediction time to as early as 60 minutes in advance makes it an adequate and timely alert system against SCD.•The predictive model becomes more comprehensive, inclusive, and applicable by considering more classes, including NSR, SCD, CAD, CHF, and AF. Sudden cardiac death (SCD) is one of the leading causes of global mortality, often occurring without warning and driven by complex cardiac dynamics. Despite significant advances in cardiovascular diagnostics, accurately predicting SCD at an early stage remains a critical challenge. This study proposes a novel fusion-based time-frequency (T-F) deep learning framework for the early prediction of SCD by classifying associated cardiac conditions. Electrocardiogram (ECG) signals were first denoised and segmented to isolate clinically relevant patterns. These signals were then transformed into two-dimensional T-F representations using spectrograms and scalograms, capturing complementary temporal and spectral information. An average fusion technique merged these representations, enriching T-F images with enhanced discriminatory power. The fused images were used to train deep learning (DL) models, and performance was evaluated using subject-wise data splits to assess generalizability across individuals. The proposed approach achieved a classification accuracy of 94.60 %, effectively identifying cardiac conditions associated with SCD one hour before its onset. This fusion-based framework shows strong potential for integration into real-time, automated diagnostic systems, enabling early warning, personalized monitoring, and timely intervention to reduce fatal outcomes.