Multimodal machine learning framework for cardiovascular risk stratification in adult obesity

Background and objectives

Cardiovascular diseases account for approximately 80% of all deaths caused by known medical conditions, making them the leading cause of mortality worldwide. The present study investigates the use of electrocardiogram (ECG) non-linear features and different topological medical features (heart rate, anthropometry, blood, glucose, and lipid profile, and heart rate variability) to discriminate between different Framingham Cardiovascular Risk Scale status groups in adult obesity using machine learning.

Methods

We conducted a cross-sectional study between November 2023 and May 2024 in Fortaleza, Ceará, Brazil. Based on the Framingham Cardiovascular Risk Scale, patients were categorized into three cardiovascular risk groups: Low (22 participants), Moderate (14 participants), and High (17 participants). From ECG signals at two different positions (ECG_Down and ECG_UP), 27 non-linear features were extracted using multi-band analysis. Additionally, 42 medical features provided by physicians were included. From a pool of 19 machine learning classifiers, models were trained and tested within a nested leave-one-out cross-validation procedure using information solely from ECG, solely from medical features, and combining both (multimodal), respectively, to distinguish between Low vs. Moderate, Low vs. High, Moderate vs. High, and All vs. All.

Results

The multimodal model presented the best results for every comparison group, reaching (1) 88.89% Accuracy and 0.8831 area under the curve (AUC) for Low vs. Moderate; (2) 97.44% Accuracy and 0.9706 AUC for Low vs. High; (3) 93.55% Accuracy and an AUC of 0.9412 for Moderate vs. High; (4) 86.79% Accuracy and 0.9346 AUC for All vs. All.

Conclusions

In a cohort of 53 patients, we extracted 27 non-linear ECG features from two positions and 42 physician-curated clinical features and, using nested LOOCV with analysis of variance F-value selection, trained models to discriminate Framingham risk strata. Across Low vs. Moderate, Low vs. High, Moderate vs. High, and All vs. All tasks, the multimodal model consistently outperformed ECG-only and medical features-only models, achieving 86–97% Acc with AUCs up to 0.97. ECG-derived non-linear features, especially from the ECG_D position, were the principal drivers of discrimination, while medical features provided complementary gains, indicating the proposed multimodal approach is a promising tool to support clinical triage.

Full text:

https://www.xiahepublishing.com/2472-0712/ERHM-2025-00037

The study was recently published in the Exploratory Research and Hypothesis in Medicine.

Exploratory Research and Hypothesis in Medicine (ERHM) publishes original exploratory research articles and state-of-the-art reviews that focus on novel findings and the most recent scientific advances that support new hypotheses in medicine. The journal accepts a wide range of topics, including innovative diagnostic and therapeutic modalities as well as insightful theories related to the practice of medicine. The exploratory research published in ERHM does not necessarily need to be comprehensive and conclusive, but the study design must be solid, the methodologies must be reliable, the results must be true, and the hypothesis must be rational and justifiable with evidence.

Follow us on X: @xiahepublishing

Follow us on LinkedIn: Xia & He Publishing Inc.