The International Conference on Acoustics, Speech and Signal Processing (ICASSP) is the world’s largest and most comprehensive technical conference on signal processing and its applications. The e-Prevention: Person Identification and Relapse Detection from Continuous Recordings of Biosignals challenge was hosted at the ICASSP 2023, which aimed to discover the distinctive behavioral patterns and relapse courses of patients with psychiatric disorders from the long-term continuous biosignal recordings of smartwatches.

Samsung R&D Institute China - Beijing (SRC-B) participated in the “Person Identification” track, which identifies a smartwatch’s wearer using daily sleep information and continuous measurements from accelerometers, gyroscopes, and heart rate monitors. The competition attracted many followers and participants from the industry and academia. SRC-B ranked 1st and the excellent results achieved by SRC-B in the e-Prevention challenge represented a breakthrough for SRC-B in the fields of wearable devices and health.

SRC-B’s team members

Currently, user demand for monitoring physical conditions using smart watches or wristbands has inspired manufacturers to develop wearable device features continuously. However, users’ daily behavior habits and physiological characteristics are different, and the signals collected by wearable devices are individually biased. This problem affects the accuracy and robustness of advanced features, such as disease monitoring and mental state detection. By identifying the wearer using physiological signals, their unique behavioral patterns can be better explored to provide more natural and accurate personalized services.

During wearer identification, valid physiological signal data collected by wearable devices may be short, and these signals have multiple missing and abnormal values. Thus, the valid data was divided into numerous fine-grained fragments to solve this problem and entered into a 1D-CNN (Convolutional Neural Networks), and the user ID of a day is predicted through the voting of fragment results. Based on this framework, multiple base networks with different input lengths or differing signal channels are trained, and these base networks are aggregated for ensemble learning. The recognition accuracy of the ensemble model has been significantly improved, providing a basis for the active personalized services of wearable devices.