A Hybrid Machine Learning Algorithm for Predicting Resting Motor Thresholds in Patients With Schizophrenia and Healthy Individuals Undergoing Transcranial Magnetic Stimulation

• Published in: IEEE transactions on magnetics Vol. 61; no. 9; pp. 1 - 6
• Main Authors: Saxena, Yash R., Lewis, Connor J., Sabbir Alam, Muhammad, Atulasimha, Jayasimha, Mehta, Urvakhsh M., Hadimani, Ravi L.
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.09.2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Artificial neural network (ANN); Artificial neural networks; Biological neural networks; Brain modeling; Charge coupled devices; Depolarization; Electric fields; Errors; Functional magnetic resonance imaging; functional magnetic resonance imaging (fMRI); Grey matter; hybrid neural network; Long short term memory; long short-term memory (LSTM); Machine learning; Magnetic resonance imaging; Motors; Neural networks; neuromodulation; Parameters; Schizophrenia; Stimulators; Testing; Training; Transcranial magnetic stimulation; transcranial magnetic stimulation (TMS)
• ISSN: 0018-9464; 1941-0069
• DOI: 10.1109/TMAG.2025.3554122

Due to the complex and varied neuroanatomy and functional states of brains, it is difficult to predict the resting motor threshold (RMT) needed as a dose parameter for treatment with transcranial magnetic stimulation (TMS). Our prior publications have shown that anatomical parameters, such as coil-to-cortex distance (CCD), gray matter volume (GMV), depolarized GMV (DGMV), and maximum electric field (E-field) value, neuroanatomy, and connectivity derived from functional magnetic resonance imaging (fMRI) are all associated with RMT. For 54 subjects with schizophrenia and 43 healthy subjects, fMRI blood oxygen-level detection (BOLD) in 25 brain regions was turned into time series and fed into a long short-term memory (LSTM) model. The outputs of the LSTM are concatenated with the schizophrenia status, CCD, GMV, percentage of gray matter voxels depolarized over 50 V/m (DGMV50) and 100 V/M (DGMV100), and maximum E-field value and then fed into an artificial neural network (ANN) that predicted the RMT. The training and testing mean absolute errors (MAEs) are 0.1176 and 0.0845, respectively, corresponding to the errors of 3.6456% and 2.6195% of the maximum stimulator output (%MSO) in the predicted RMT values. Our novel hybrid LSTM-ANN neural network can be used as a pretreatment procedure to reduce the number of trials needed to measure RMT for patients and increase patient comfort and confidence in the procedure administered.