Prediction of cardiovascular complications using multi-modal data

This study aims to predict the risk of cardiovascular complications in elderly patients after open-heart surgery based on multimodal data. I used single-lead electrocardiogram (ECG) signals and tri-axial accelerometer data from 80 patients, recorded through a chest-worn heart rate monitor during various physical tests, including the veloergometry test, six-minute walk test, stair climbing test, time up and go (TUG) test, and gait analysis on a treadmill. To ensure data reliability and analytical accuracy, I conducted multi-level preprocessing of the ECG and accelerometer data, including resampling, filtering, outlier detection and removal, and data segmentation. In data analysis, I initially employed machine learning algorithms, including decision trees, KNN, SVM, and random forests, to classify the types of physical activities (e.g., stair climbing, walking, and cycling) performed by the patients. Subsequently, I used a random forest model to predict patients' heart function status under different activity contexts, adopting the New York Heart Association (NYHA) Functional Classification as an evaluation standard. For each prediction task, I compared the results using only clinical features (e.g., height, age, weight, postoperative days, type of surgery), only sensor features (ECG and accelerometer data), and a fusion of both in a multimodal approach. The results show that using only clinical information or sensor data alone is insufficient for accurate predictions, while multimodal data fusion significantly improves model performance. By analyzing the effectiveness of multimodal data in predicting cardiovascular complications, I verified the varying impact of different activities and features on the prediction outcomes (e.g., veloergometry test is more predictive than stair climbing) and found that integrating results from multiple activities provides a more comprehensive risk assessment. Through feature importance analysis, I further identified the key features with the greatest influence on prediction results, offering valuable clinical insights for early diagnosis and risk management of cardiovascular diseases.