A Two-Layer Ensemble Method for Detecting Epileptic Seizures Using a Self-Annotation Bracelet With Motor Sensors

• Published in: IEEE transactions on instrumentation and measurement Vol. 71; pp. 1 - 13
• Main Authors: Dong, Chunjiao, Ye, Tianchun, Long, Xi, Aarts, Ronald M., van Dijk, Johannes P., Shang, Chunheng, Liao, Xiwen, Chen, Wei, Lai, Wanlin, Chen, Lei, Wang, Yunfeng
• Format: Journal Article
• Language: English
• Published: New York IEEE 2022; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accelerometer (ACM); Algorithms; Annotations; Brain modeling; Convulsions & seizures; Data collection; Data models; Datasets; epilepsy; gyroscope; Hospitals; Injury prevention; Jewelry; Labelling; Machine learning; Monitoring; Motion sensors; motor sensor; Night; Random forests; Segments; seizure detection; Seizures; Sensitivity; Sensors; signal processing; Support vector machines; wearable device; Wearable technology
• ISSN: 0018-9456; 1557-9662
• DOI: 10.1109/TIM.2022.3173270

Using monitoring devices could help avoid injuries and even death. Currently, wearable sensors such as motion sensors and other sensors are used to detect when the patient is having a seizure and alarm their caregivers. However, the development phase of these devices requires labor-intensive work on labeling the collected data, resulting in difficulties in developing wearable monitoring devices. Thus, a more automated auxiliary method of labeling seizure data and a wearable device to detect seizures for daily monitoring use are necessary. We collected data from epileptics outside the hospital with our proposed bracelet. The subjects were asked to press the mark button after they had seizures. We also presented an automatically extraction and annotation of moving segments (EAMS) algorithm to exclude nonmoving segments. Then, we used a two-layer ensemble model (TLEM) using machine learning methods to classify seizures and non-seizure moving segments, which was designed to deal with imbalanced dataset. Then, we build two individual TLEM models separately for the overall (all day and night) seizure detection case and the night seizure detection case, owing to different imbalance of these datasets. The EAMS algorithm exclude 93.9% raw inactive data. The TLEM model achieved 76.84% sensitivity (SEN) and 97.28% accuracy (ACU) for the overall case and achieved 94.57% SEN and 91.37% ACU for the night case. These results indicate that this bracelet can capture seizures efficiently, and our proposed TLEM has higher SEN and ACU than single-layer machine learning models.