Clinical feasibility of motor hotspot localization based on electroencephalography using convolutional neural networks in stroke

• Published in: Journal of neuroengineering and rehabilitation Vol. 22; no. 1; pp. 193 - 17
• Main Authors: Choi, Ga-Young, Seo, Jeong-Kweon, Kim, Kyoung Tae, Chang, Won Kee, Yoon, Sung Whan, Paik, Nam-Jong, Kim, Won-Seok, Hwang, Han-Jeong
• Format: Journal Article
• Language: English
• Published: London BioMed Central 26.09.2025; BioMed Central Ltd; BMC
• Subjects: Adult; Aged; Algorithms; Biomedical and Life Sciences; Biomedical Engineering and Bioengineering; Biomedicine; Care and treatment; Convolutional Neural Networks; Deep Learning; Electroencephalography; Electroencephalography - methods; Evoked potentials (Electrophysiology); Feasibility Studies; Female; Hand - physiopathology; Health aspects; Humans; Machine learning; Male; Measurement; Methods; Middle Aged; Motor cortex; Motor Cortex - physiopathology; Motor hotspot; Neural networks; Neural Networks, Computer; Neurology; Neuromodulation; Neurorehabilitation; Neurosciences; Rehabilitation Medicine; Stroke; Stroke - physiopathology; Stroke patients; Stroke Rehabilitation - methods; Transcranial Magnetic Stimulation - methods; South Korea; Deep learning; Stroke; Electroencephalography; Neurorehabilitation; Neuromodulation; Motor hotspot
• ISSN: 1743-0003; 1743-0003
• DOI: 10.1186/s12984-025-01736-3

Background Although transcranial magnetic stimulation (TMS) is the optimal tool for identifying individual motor hotspots–specific regions of the brain that are essential for controlling voluntary muscle movements–it involves a cumbersome procedure that requires patients to visit the hospital regularly and relies on expert judgment. To address this, we propose an advanced electroencephalography (EEG)-based motor hotspot identification algorithm using a deep-learning and assess its clinical feasibility and benefits by applying it to EEGs for stroke patients, considering the noticeable variations in EEG patterns between stroke patients and healthy controls. Methods Motor hotspot locations were estimated using a two-dimensional convolutional neural network (CNN) model. We utilized various types of input data, depending on the five processing levels, the five types of input data, depending on the processing levels, to assess the signal processing capability of our proposed deep-learning model using EEGs of thirty healthy subjects measured during a simple hand movement task. Furthermore, we applied our proposed deep-learning algorithm to the hand-movement-related EEGs of twenty-nine stroke patients. Results The mean error distance between the motor hotspot locations identified by TMS and our approach for healthy subjects was 0.35 ± 0.04 mm when utilizing power spectral density (PSD) features. The mean error distance was 2.27 ± 0.27 mm for healthy subjects and 1.64 ± 0.14 mm for stroke patients, when using raw data without any feature engineering. Our proposed motor hotspot identification algorithm showed robustness concerning the number of electrodes; the mean error distance was 2.34 ± 0.19 mm when using only 9 channels around the motor area for healthy subjects, and 1.77 ± 0.15 mm using only 5 channels around the motor area for stroke patients. Conclusion We demonstrate that our EEG-based deep-learning approach can effectively identify individual motor hotspots, and the clinical feasibility of our algorithm by successfully applying the proposed approach to stroke patients. It can be used as an alternative to TMS for identifying motor hotspots, potentially enhancing the effectiveness of rehabilitation strategies.