DESIGN OF AN IMPROVED MODEL FOR CARDIOVASCULAR DISEASE DETECTION USING DEEP CANONICAL CORRELATION ANALYSIS AND BIOINSPIRED OPTIMIZATION

• Published in: Journal of mechanics of continua and mathematical sciences Vol. 20; no. 6; pp. 126 - 141
• Main Authors: Sahoo, Prakash Chandra, Pattanayak, Binod Kumar, Mohanty, Rajani Kanta, Mohanty, Ayasa Kanta
• Format: Journal Article
• Language: English
• Published: Institute of Mechanics of Continua and Mathematical Sciences 01.06.2025
• Subjects: cardiovascular disease; correlation analysis; deep canonical genetic algorithm; explainable ai; multimodal data fusion; scenarios
• ISSN: 0973-8975; 2454-7190
• DOI: 10.26782/jmcms.2025.06.00008

Cardio Vascular Diseases (CVDs) are one of the most prevalent causes of death in the world and require an appropriate early detection method that could satisfactorily integrate diversified patient data available in today's healthcare. Traditional diagnosis is often based on single-modality data, either ECG or imaging, which seldom can unfold the complex and multi-faceted nature of cardiovascular conditions. Moreover, these models have incomplete interpretation and optimization issues, which do not suit their application in a clinical setting. On this, we propose a novel framework for the detection of cardiovascular diseases and presiding analysis through multimodal data fusion, optimized neural networks, and explainable AI techniques. Our approach begins with Deep Canonical Correlation Analysis (DCCA), which fuses multiple modalities of data such as ECG time series, medical imaging, electronic health records, and genetic data into a unified latent representation that represents correlated information across these heterogeneous sources. This will not only enhance the prediction accuracy but also retain modality-specific unique aspects, thus going beyond traditional models. We will go one step beyond this by using a Genetic Algorithm in combination with the Neuro-evolution of Augmenting Topologies for optimization not only for neural network architecture and hyperparameters but also for going into the process. This bioinspired methodology makes dynamic adjustments in the complexity of a model, substantially reducing error rates. To ensure interpretability in our predictions, we will finally integrate Shapley Additive explanations (SHAP) into the multimodal fusion network. SHAP values provide a clear, quantitative measure of the contribution of each feature and modality to the model predictions, most significantly corresponding to a priori known clinical risk factors that offer critical insights for healthcare professionals. Impact: we have more than halved error rates by 15%, reached an Area Under the Curve of 0.92, and demonstrated a very strong correlation with expert-annotated risk scores of r = 0.87 using SHAP values. This framework has set a new standard in the CVD prediction area by putting together cutting-edge AI techniques and practical, interpretable healthcare applications.