A proof-of-principle study of tractography-based machine learning for predicting transcranial magnetic stimulation motor responsiveness.
Authors
Affiliations (4)
Affiliations (4)
- Department of Physics, Faculty of Philosophy, Sciences and Letters of Ribeirão Preto (FFCLRP), University of São Paulo (USP), Av. Bandeirantes, 3900, Monte Alegre, 14040-901, Ribeirão Preto, SP, Brazil.
- Department of Neuroscience and Biomedical Engineering, Aalto University School of Science, Rakentajanaukio 2, 02150, Espoo, Finland. Electronic address: [email protected].
- Department of Physics, Faculty of Philosophy, Sciences and Letters of Ribeirão Preto (FFCLRP), University of São Paulo (USP), Av. Bandeirantes, 3900, Monte Alegre, 14040-901, Ribeirão Preto, SP, Brazil; Department of Medical Imaging, Hematology and Clinical Oncology, Ribeirão Preto Medical School (FMRP), University of São Paulo (USP) and Ribeirão Preto Medical School Hospital das Clínicas, Av. Bandeirantes, 3900, 14049-900, Ribeirão Preto, SP, Brazil.
- Department of Computing and Mathematics, Faculty of Philosophy, Sciences and Letters of Ribeirão Preto (FFCLRP), University of São Paulo (USP), Av. Bandeirantes, 3900, Monte Alegre, 14040-901, Ribeirão Preto, SP, Brazil.
Abstract
Identifying the brain stimulation target is fundamental for transcranial magnetic stimulation (TMS). Currently, this process is time-consuming and heavily dependent on the operator's expertise. Here, we present a proof-of-principle study evaluating a deep learning-based approach that leverages structural brain connectivity to support stimulation site prediction and real-time cortical excitability mapping during neuronavigation. We introduce a tractography-guided method for TMS hotspot identification based on subject-specific structural connectivity. Diffusion MRI-derived tractography features are combined with machine learning to predict, within individual subjects, cortical regions with high stimulation responsiveness. This approach enables data-driven, individualized target selection. Cortical prediction of responsive cortical areas were tested by comparing real motor mapping data with prediction from machine learning algorithms. Tractography-derived connectivity features and distal myographic responses were used to train four neural network models across five subjects. Inputs were varied using tractography alone, coil coordinates alone, or hybrid combinations. In most subjects, hybrid models integrating coil coordinates and fiber information showed numerically similar F1 scores and accuracies to unimodal models, with small descriptive performance differences between architectures. Substantial inter-subject variability limits the ability to draw definite conclusions regarding model accuracies, but multimodal models showed greater spatial congruence between experimentally derived motor maps and predicted regions of highest responsiveness in the tested subject. This approach may be extended to other brain regions and functional domains beyond the motor cortex. Despite promising feasibility, substantial inter-subject variability highlights the need for larger studies and more robust training data.