Determining the Number of Clusters in Clinical Response of TMS Treatment using Hyperdimensional Computing

• Published in: Journal of signal processing systems Vol. 96; no. 8; pp. 509 - 523
• Main Authors: Ge, Lulu, McInnes, Aaron N., Widge, Alik S., Parhi, Keshab K.
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.09.2024; Springer Nature B.V
• Subjects: Algorithms; Circuits and Systems; Clustering; Computation; Computer Imaging; Electrical Engineering; Engineering; Harnesses; Image Processing and Computer Vision; Pattern Recognition; Pattern Recognition and Graphics; Signal,Image and Speech Processing; Transcranial magnetic stimulation; Vision; Transcranial magnetic stimulation (TMS); Latent class mixed modeling (LCMM) and clustering; Hyperdimensional computing (HDC); Vector symbolic architecture (VSA); Major depressive disorder (MDD)
• ISSN: 1939-8018; 1939-8115
• DOI: 10.1007/s11265-024-01921-y

This paper addresses clustering of clinical response of subjects with major depressive disorder (MDD) after they are treated with transcranial magnetic stimulation (TMS). Specifically, we present an approach to determine the number of clusters using hyperdimensional computing (HDC). In the clinic, MDD patients are required to receive weeks of repetitive TMS (rTMS). To achieve response rates from TMS that are comparable to or better than medication, we need to understand the TMS mechanism and its effect on cognitive control. This paper investigates the variability of rTMS treatment across patients, which is essentially a clustering problem. We propose an algorithm based on HDC to estimate the number of clusters for a group of clinical responses. Our algorithm harnesses an existing HDC-based clustering approach–HDCluster. Experimental results show that: i) . For n = 27 patients, the original clinical trajectories typically exhibit four clusters, while the baseline-corrected trajectories tend to have three clusters. A similar finding is obtained for a larger dataset when n = 176 . ii) . Though the number of clusters revealed by the proposed HDC algorithm is different from that determined by the traditional latent class mixed modeling (LCMM), both algorithms share a similar clustering pattern.