Predictive Clinical Decision Support System Using Machine Learning and Imaging Biomarkers in Patients With Neurostimulation Therapy: A Pilot Study

• Published in: Pain physician 2021-12, Vol.24 (8), p.E1279-E1290
• Main Authors: De Andres, Jose, Ten-Esteve, Amadeo, Harutyunyan, Anushik, Romero-Garcia, Carolina S, Fabregat-Cid, Gustavo, Asensio-Samper, Juan Marcos, Alberich-Bayarri, Angel, Marti-Bonmati, Luis
• Format: Article
• Language: English
• Subjects: Biomarkers; Chronic pain; Decision support systems; Decision Support Systems, Clinical; Discriminant analysis; Electric stimulation therapy; Humans; Insular Cortex; Machine Learning; Magnetic Resonance Imaging; Neuroimaging; Pilot Projects; Prospective Studies; Spinal cord
• ISSN: 1533-3159; 2150-1149

Chronic pain is correlated with alterations in brain structure and function. The selection process for the ideal candidate for spinal cord stimulation (SCS) therapy is based on functional variables analysis and pain evaluation scores. In addition to the difficulties involved in the initial selection of patients and the predictive analysis of the trial phase, the large rate of explants is one of the most important concerns in the analysis of the suitability of implanted candidates. To investigate the usefulness of imaging biomarkers, functional connectivity (FC) and volumetry of the whole brain in patients with Failed back surgery syndrome (FBSS) and to create a clinical patient-based decision support system (CDSS) combining neuroimaging and clinical data for predicting the effectiveness of neurostimulation therapy after a trial phase. A prospective, consecutive, observational, single center study. The Multidisciplinary Pain Management Department of the General University Hospital in Valencia, Spain. A prospective, consecutive, and observational single-center study. Using Resting-state functional magnetic resonance imaging (rs-fMRI) and Region of interest (ROI) to ROI analysis, we compared the functional connectivity between regions to detect differences in FC and volume changes. Basal magnetic resonance images were obtained in a 1.5T system and clinical variables were collected twice, at the basal condition and at 6-months post-SCS implant. We also conducted a seed-to-voxel analysis with 9 items as seed-areas characterizing the functional connectivity networks. A decreased in 10 units in the Pain Detect Questionnaire (PD-Q) score was established to define the subgroup of Responders Group (R-G) to neurostimulation therapy. The clinical variables collected and the imaging biomarkers obtained (FC and volumes) were tested on a set of 6 machine learning approaches in an effort to find the best classifier system for predicting the effectiveness of the neurostimulator. Twenty-four patients were analyzed and only seven were classified in the R-G. Volumetric differences were found in the left putamen, F = 34.06, P = 0.02. Four pairwise brain areas showed statistical differences in the rs-fMRI including the right insular cortex. Linear Discriminant Analysis showed the best performance for building the CDSS combining clinical variables and significant imaging biomarkers, the prediction increased diagnostic accuracy in the R-G patients from 29% in current practice to