International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 12 Issue: 10 | Oct 2025
p-ISSN: 2395-0072
www.irjet.net
Multimodal Deep Learning Algorithms for Predictive Modeling of Cardiovascular Disease Onset and Progression: A Tetramodal Architecture Integrating HER, Imaging, Genomics, and Wearables Data Dr. Emmanuel Ameh1, Dr. Alexander Paselk2 Dept. Machine Learning, Capitol Technology University, Maryland, USA Dept. Occupational Health and Safety, Capitol Technology University, Maryland, USA ---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Background: Cardiovascular diseases (CVDs)
in clinical care, risk prediction remains dominated by traditional models such as the Framingham Risk Score, which rely on a limited set of clinical variables (cholesterol, age, smoking status) and fail to leverage the breadth of modern patient data sources. These limitations contribute to misclassification of high-risk patients and insufficient monitoring of disease progression, particularly in those with subclinical presentations.
remain the leading global cause of mortality, accounting for over 18 million deaths annually. Traditional risk scores, such as the Framingham Risk Score, fail to fully capture the complex interactions between genetic, physiological, and lifestyle factors. Deep learning methods, particularly multimodal architectures, offer opportunities to integrate heterogeneous data sources for improved prediction of both CVD onset and progression. Methods: This retrospective, quantitative, cohort study developed and evaluated a tetramodal deep learning framework integrating electronic health records (EHR), imaging, genomic, and wearables data. Individual modality architectures included convolutional neural networks (CNNs) for imaging, transformer encoders for wearables, and fully connected networks for EHR and genomics. Early fusion integration was employed to construct the tetramodal architecture. Data preprocessing, feature extraction, and multimodal fusion strategies were optimized via ablation studies. Model performance was evaluated for binary classification (CVD_Presence) and regression (Time_to_Event) tasks using AU-ROC, F1-score, mean absolute error (MAE), and the Friedman chi-squared test with post-hoc analysis. Results: The tetramodal model significantly outperformed all single-, bi-, and tri-modal models for both CVD onset and progression prediction. Inclusion of wearables data improved predictive accuracy for progression tasks with statistical significance (p = 0.04). The architecture demonstrated superior sensitivity, specificity, and clinical applicability compared to unimodal baselines. Conclusions: This study presents a scalable, interpretable tetramodal deep learning framework capable of integrating heterogeneous clinical and real-time physiological data for enhanced CVD risk stratification. Future work will focus on real-world deployment, optimization for low-resource environments, and external validation on diverse patient populations.
Recent advances in deep learning (DL) enable the processing of high-dimensional, multimodal datasets, ranging from imaging and genomic sequences to continuous physiological monitoring from wearable devices. Multimodal deep learning, specifically tetramodal integration in this study, addresses the complexity of CVD pathophysiology by fusing heterogeneous data modalities for richer, more personalized risk assessment. The objective of this study was to design, implement, and evaluate a tetramodal deep learning architecture for predicting both CVD onset (binary classification: CVD_Presence) and CVD progression (regression: Time_to_Event). Despite its promise, the application of multimodal deep learning to CVD prediction remains limited. Persistent challenges include the interpretability of complex models in clinical settings, the computational cost of training and deployment, integration of longitudinal patient data, and variability in data formats and quality (Abbas & Daena, 2025; Nadella et al., 2023). Furthermore, the absence of standardized algorithms for multimodal integration and reliance on small-scale datasets reduce clinical translatability (Abbas & Daena, 2025; Esteva et al., 2021). Research on leveraging multimodal deep learning for predicting CVD progression, especially in patients with multiple comorbidities, remains scarce (Kasula, 2023; Soenksen et al., 2022; Vaid et al., 2023).
Key Words: Deep Learning Models, Cardiovascular Disease Prediction, Machine Learning Models, Multimodal Data Modeling, Tetramodal Data, Wearable Data
Addressing these gaps requires robust, generalizable multimodal deep learning algorithms capable of integrating heterogeneous data streams to predict both onset and progression of CVD. Accurate, timely predictions could facilitate earlier interventions, reduce complications, and improve patient outcomes (V. V. Paul & Masood, 2024; Terranova & Venkatakrishnan, 2024; Tian et al., 2024). The present study develops and evaluates such a framework,
1.INTRODUCTION Cardiovascular diseases (CVDs) constitute a persistent public health burden, causing approximately one-third of global deaths annually (WHO, 2024). Despite advancements
© 2025, IRJET
|
Impact Factor value: 8.315
|
ISO 9001:2008 Certified Journal
|
Page 854