497: Machine Learning Neuroimaging Biomarkers to Predict Treatment Response in Neuropathic Pain

Topic: Clinical Science
Subtopic: Pain category (ICD-11), Chronic neuropathic pain, Artificial Intelligence/Machine Learning

Background and Aims: Chronic neuropathic pain is a prevalent and disabling condition characterised by marked inter-individual variability in treatment response and a paucity of reliable predictive markers. This heterogeneity hampers optimal treatment selection and contributes to prolonged trial-and-error prescribing and suboptimal outcomes. The present study applies machine learning techniques to multimodal neuroimaging data to identify objective biomarkers associated with treatment response. By delineating neurobiological signatures that predict therapeutic benefit, this work aims to advance precision medicine approaches in neuropathic pain and to support more informed, evidence-based clinical decision-making.

Methods: Multisite neuroimaging data were analysed from 76 patients with painful diabetic neuropathy (Sheffield and Oxford) and 56 with osteoarthritis (OpenNeuro). Responders were defined as NRS <4 or ≥30% reduction post-treatment; non-responders had NRS >4 without such reduction. Structural MRI and resting-state fMRI were processed to derive cortical, subcortical, and functional connectivity features. Supervised machine learning models were trained on Sheffield and Oxford data and externally validated on OpenNeuro. Harmonisation methods (NeuroComBat and domain adaptation) reduced inter-site variability. Models were optimised via Bayesian tuning and evaluated using MCC, AUROC, recall, and specificity.

Results: In the Sheffield cohort, the best-performing model achieved a balanced accuracy of 0.689, recall of 0.796, AUROC of 0.806, and MCC of 0.407. When the
Sheffield and Oxford datasets were combined, the best-performing model achieved a balanced accuracy of 0.693, recall of 0.75, specificity of 0.667, AUROC
of 0.705, and MCC of 0.382 when tested on the OpenNeuro cohort. Notably, the model’s ability to predict placebo responders was poor, with a balanced
accuracy of 0.502, recall of 0.548, specificity of 0.455, AUROC of 0.503, and MCC of 0.

Conclusions: This study demonstrates the potential of neuroimaging-based machine learning models as objective biomarkers for predicting treatment response in neuropathic pain. The integration of robust MRI harmonisation, secure multisite dataset curation, and optimised analytical pipelines supports the feasibility of a scalable multimodal AI framework for neuropathic pain research. External validation across independent sites and conditions showed fair and generalisable performance. Notably, the model predicted treatment response more accurately than placebo response, suggesting meaningful target engagement and clinical relevance. Future work should prioritise larger validation cohorts, enhanced feature selection, and improved model interpretability to facilitate clinical translation.

References: Jiayang Zhang , Lalu Muhammad Riza Rizky , Solomon Tesfaye , Jim Wild , Andrew Segerdahl , Douglas Steele , David Bennett , Lesley
Colvin , Haiping Lu , Dinesh Selvarajah
University Of Sheffield, School of Medicine and Population Health, Sheffield, United Kingdom, University Of Sheffield, School of Computer Science,
Sheffield, United Kingdom, Sheffield Teaching Hospitals NHS Foundation Trust, Diabetes Research Unit, Sheffield, United Kingdom, University OF Oxford,
Nuffield Department of Anaesthetics; Nuffield Department of Clinical Neurosciences, Oxford, United Kingdom, University Of Dundee, School of Medicine,
Dundee, United Kingdom, University Of Oxford, Nuffield Department of Anaesthetics; Nuffield Department of Clinical Neurosciences, Oxford, United
Kingdom

Presenting Author: Dr Dinesh Selvarajah
Submitting Author: Dinesh Selvarajah

Senior Author: Dr Dinesh Selvarajah

IASP 2026 Global Year theme of Neuropathic Pain: Yes

• Dr Kevin Teh: PhD - University Of Sheffield, School of Medicine and Population Health, Sheffield, United Kingdom, University Of Sheffield
• Dr Jiayang Zhang: PhD - School of Computer Science, University of Sheffield, United Kingdom,
• Dr Lalu Muhammad Riza Rizky: PhD - University Of Sheffield, School of Computer Science, Sheffield, United Kingdom
• Prof Solomon Tesfaye: FRCP, MD - Sheffield Teaching Hospitals NHS Foundation Trust, Diabetes Research Unit, Sheffield, United Kingdom
• Prof JIm WIld: PhD - Division of Clinical Medicine, School of Medicine and Population Health, Sheffield, United Kingdom,
• Dr Andrew Segerdahl: PhD - University OF Oxford, Nuffield Department of Anaesthetics; Nuffield Department of Clinical Neurosciences, Oxford, United Kingdom
• Prof Douglas Steele: PhD - University Of Dundee, School of Medicine, Dundee, United Kingdom
• Prof David Bennett: PhD - Nuffield Department of Anaesthetics; Nuffield Department of Clinical Neurosciences, Oxford, United Kingdom
• Prof Lesley Colvin: PhD - University Of Dundee, School of Medicine, Dundee, United Kingdom
• Prof Haiping Lu: PhD - University Of Sheffield, School of Computer Science, Sheffield, United Kingdom
• Dr Dinesh Selvarajah: MBChB, MRCP, MPH, PhD - Division of Clinical Medicine, University Of Sheffield, School of Medicine and Population Health, Sheffield, United Kingdom,