An Expansion for Automated Cardiac Anomaly Detection Frameworks with Multimodal Missing Data Imputation

Monitoring of the cardiovascular system is vital to detect cardiac arrhythmia which is a leading cause of death in the United States. The most popular cardiac monitoring modalities are Electrocardiography (ECG) and Photoplethysmography (PPG). Utilization of the critical features extracted from these signals allows an automated pipeline for early detection. However, recorded data is often incomplete due to equipment sampling issues or loss of contact. Completion of missing values through imputation permits these datasets to be examined for arrhythmia through machine learning. Because a comprehensive investigation of prominent imputation algorithms has not been conducted within this context before, we explored several algorithms to accurately predict the missing feature values extracted from both modalities. Our workflow consists of preprocessing, feature extraction, and imputation algorithms to estimate these clinically important values. We collected 2700 minutes of ECG and PPG data from 36 patients within the PhysioNet MIMIC-III database. Approximately 150 features were extracted for both ECG and PPG prior to the imputation. We compared seven imputation algorithms (Zero, Mean, Hot Deck, Linear Regression, KNN, Random Forest and Neural Networks) using MSE and R
²
as the evaluation metrics. KNN, Random Forest and Linear Regression gave the highest R
²
values of 0.904, 0.908, 0.898 respectively, for a subset of the features. Overall, KNN was the most consistent in providing the highest R
²
values and the lowest MSE values across all the features tested. Our investigation provides an expanded preliminary platform to conduct real-time arrhythmia detection in hospitals or remotely via wearable devices.