Personalized anomaly detection using contrastive learning technique of real time physiological signal datas

Physiological signals obtained from humans are crucial in many fields, especially in healthcare. Analyzing information such as a person's heart rate, body temperature, blood pressure, electrodermal activity, etc., can provide many benefits such as diagnosing heart diseases, determining stress levels, monitoring exercise performance, and more. With the advancement of technology, smartwatches and wristbands that make physiological signals more easily obtainable have become part of our lives. This allows individuals to track their health status while exercising or in their daily lives using these developed technologies. One of the factors affecting human health alongside nutrition, physical activity, sleep patterns, genetic factors, is stress. Stress, which has become a widespread problem today, also has serious effects on our health. Furthermore, the effects of stress can also be observed on various physiological signals in the body. As part of a thesis study, the aim is to use a contrastive learning model to determine whether a person is stressed by examining the effects of stress on physiological signals. Anomaly refers to events or situations that differ from what is normally expected, are unusual, or attention-grabbing. Therefore, within the scope of the study, a person being stressed is considered an anomaly. contrastive learning is a deep learning method that emerged to be used on unlabeled data. Anomaly detection in time series data often poses a problem due to the lack of labeling. In order to solve this problem, the aim is to detect anomalies using a contrastive learning model with time series data of blood volume pulse, electrodermal activity, and body temperature recorded from children with different conditions (obstetric brachial plexus injury, dyslexia, intellectual disability, and typically developing). Within the study, first anomaly detection is made by using all data, an then personalized anomaly detections are detected. As a result, it is observed that personalized anomaly detection achieves more successful results than general anomaly detection.