104-OR: Identifying Painful Diabetic Neuropathy Subtypes from Resting-State Functional Magnetic Resonance Brain Imaging—A Novel Unsupervised Machine Learning Approach

• Published in: Diabetes (New York, N.Y.) Vol. 72; no. Supplement_1; p. 1
• Main Authors: SELVARAJAH, DINESH, TEH, KEVIN, ANANDHANARAYANAN, APARNA, SLOAN, GORDON P., TESFAYE, SOLOMON
• Format: Journal Article
• Language: English
• Published: New York American Diabetes Association 20.06.2023
• Subjects: Brain mapping; Cerebral hemispheres; Cortex (parietal); Cortex (somatosensory); Diabetes; Diabetes mellitus; Diabetic neuropathy; Functional magnetic resonance imaging; Learning algorithms; Machine learning; Magnetic resonance imaging; Neural networks; Neuroimaging; Pain perception; Patients
• ISSN: 0012-1797; 1939-327X
• DOI: 10.2337/db23-104-OR

The use of neuroimaging methods to develop biomarkers in painful DN have yielded promising results but have yet to demonstrate clear clinical utility. Our aim was to delineate the biological heterogeneity in painful DN by first defining the functional connectivity subtypes in patients then assessing the clinical significance with respect to prediction of treatment outcomes. Methods: 82 painful-DN subjects and 36 HV underwent resting-state functional MR imaging (RS-fMRI). Painful-DN subjects were divided into training (n=48) and testing (n=34) datasets. Clinical pain subtypes were identified by analysing, via unsupervised, machine learning (k-means algorithm), the RS-fMRI functional connectivity of key somatosensory/nociceptive brain regions. Patients in our testing dataset were divided into responders (VAS<4; n=23) and non-responders (VAS>4; n=11). Results: RS-fMRI functional connectivity defines two clinically relevant painful-DN subtypes. The two subtypes were characterised by strong functional connectivity differences in the postcentral (p<0.001, 95%CI=0.22:0.65) and precentral (p=0.002, 95%CI=0.15:0.59) parietal cortex functional connectivity. The interhemispheric connectivity between homologous regions was also notable. RS-fMRI functional connectivity differences in the default mode network (p=0.002;95%CI=-0.35:-0.08) was also identified. The performance of our unsupervised k-means model to predict treatment response was accuracy=0.77, recall=0.83, precision=0.83 and an AUC=0.73. Conclusion: We have identified neurobiological markers of painful DN subtypes from brain functional connectivity using an unsupervised data-driven approach. The subtype whose functional connectivity differed most from HV failed to respond to neuropathic pain treatment. External validation on a large independent dataset is now required to confirm future clinical utility.