Development of a Personalized Anomaly Detection Model to Detect Motion Artifacts Over PPG Data Using Catch22 Features

As remote health monitoring grows, it's crucial to distinguish high-quality biomedical signals from low-quality ones. Identifying and mitigating motion artifacts (MAs) is essential for accurate data from wearable devices. Methods: In this work, a high-performing subset of time-series features denoted as catch22 (22 CAnonical Time-series CHaracteristics) was used to detect the presence of MAs in photoplethysmogram (PPG) data acquired from the brachial and digital artery of 31 healthy subjects. Three unsupervised algorithms were employed along with catch22 to detect MAs within the dataset, these were: One-Class Support Vector Machine, Isolation Forest, and Local Outlier Factor. Results: Aggregated precision, recall, and F1-score were computed per each method to assess the detection performances according to a variety of features and anomalies' distribution. These metrics resulted respectively equal to 0.5, 0.64, and 0.55 for OC- SVM, 0.91, 0.94, and 0.92 for IF, and 0.74, 0.75, and 0.74 for LOF. Conclusion: Experimental findings illustrate that by employing the catch22 feature subset, it is viable to discern the presence of MAs in beat-to-beat pulse waveforms without recurring to prior knowledge or data-driven PPG features.