Multimodal Heart Disease Risk Assessment: A Machine Learning Practice Integrating Lifestyle and Clinical Data

This project (multimodal-heart-risk-ml) was developed by Durga200422, aiming to improve the accuracy of heart disease risk prediction by integrating BRFSS lifestyle survey data and Cardio clinical indicators. The project uses a LinearSVM baseline model and a Stacking ensemble strategy, combined with SHAP explainable AI technology and Streamlit interactive applications, to provide a solution for medical AI that balances performance and transparency.

Heart disease is one of the major global health threats. Traditional risk assessment relies on a single data source (clinical or lifestyle), but heart health is influenced by multiple factors such as physiology and behavior. This project addresses the limitations of single data sources through multimodal data fusion to build a more comprehensive risk assessment model.

BRFSS Lifestyle Data: A large-scale survey led by the U.S. Centers for Disease Control and Prevention, covering factors such as smoking and diet. It needs to address issues like categorical data, missing values, and correlation; Cardio Clinical Data: Includes physiological indicators such as blood pressure and cholesterol, which are accurate but have high collection costs; Integration Value: For example, the synergistic risk effect of high blood pressure + smoking—models can learn such interaction patterns to improve accuracy.

LinearSVM Baseline: Suitable for high-dimensional data, with strong generalization ability and easily interpretable decision boundaries; Stacking Ensemble: After training multiple base learners, use a meta-learner to combine predictions, capturing the complementarity of different models; Optimization: Hyperparameter tuning to balance sensitivity (avoiding missed diagnoses) and specificity (avoiding misdiagnoses).

Necessity: Black-box models are difficult for doctors to accept; it is necessary to understand decision logic to avoid bias; SHAP Analysis: Quantify the contribution of individual features to predictions, answering 'why is this patient at high risk'; Permutation Importance: Evaluate the global importance of features, providing references for public health policies; Combination of Local and Global: Meet the different needs of doctors (single case) and researchers (overall patterns).

The project provides an interactive web application based on Streamlit, with features including: data input forms (lifestyle + clinical data), real-time risk prediction, personalized explanations (main factors affecting the score), and a visualization dashboard (population risk distribution trends), achieving an end-to-end user-friendly experience.

Value: Assist preventive medicine, early identification of high-risk groups to take intervention measures; Limitations: Model generalization is affected by population differences; predictions are based on correlation rather than causation; Future: Introduce genetic/wearable data, deep learning architectures, and continuous learning mechanisms to optimize the model.