Latypov Timur H, Yakubov Rose, Jörgens Daniel, Tsai Pascale, Srisaikaew Patcharaporn, Hung Peter Shih-Ping, Walker Matthew R, Tawfik Marina, Mikulis David, Rudzicz Frank, Hodaie Mojgan
Krembil Research Institute, University Health Network, Toronto, Ontario, Canada, M5T 0S8.
Institute of Medical Science, Temerty Faculty of Medicine, University of Toronto, Toronto, Ontario, Canada, M5S 3K3.
Brain Commun. 2025 Jun 17;7(3):fcaf178. doi: 10.1093/braincomms/fcaf178. eCollection 2025.
Chronic pain remains a challenge for clinicians, with limited individualized predictive tools that can aid with diagnosis, disease course, or prediction of treatment outcomes. We hypothesized that a comprehensive analysis, encompassing a patient's complete pain-related clinical data, medical history and brain imaging, can identify key contributors linked to surgical outcomes and stratify specific outcome categories for trigeminal neuralgia (TN)-chronic facial pain syndrome. Using supervised and unsupervised machine learning approaches, we analysed data from 102 subjects with classical TN. Pre-surgical clinical data were processed through unsupervised learning to delineate key clinical contributors of TN outcome stratification and their correlation with surgical response. Concurrently, we applied supervised learning to pre-surgical T1-weighted brain magnetic resonance imaging. Clinical data analysis uncovered pain and non-pain-related measures-including pain frequency, degree of medication relief, pain character, presence of diabetes and cancer history-as the most significant in forecasting surgical outcome. Analysis revealed strong correlation of pre-surgical clinical data with surgical response duration ( = 0.5, < 0.00001). Imaging data analysis used a support vector machine classification model with high recall for subjects who would be either long-term responders or non-responders 0.79 and 0.86 with the area under the receiver operating characteristic curve (AUC) of 0.86 and 0.84, respectively. The average multiclass accuracy in predicting the duration of surgical response categories was 78% (AUC 0.8). Together, these results show that TN surgical outcome categories are distinguishable, and surgical outcome can be stratified based on combined clinical and brain imaging data available prior to surgical treatment. We suggest a novel perspective on different strata of chronic pain disorders, each with structural imaging, clinical correlates and specific surgical outcomes.
慢性疼痛仍然是临床医生面临的一项挑战，目前能够辅助诊断、疾病进程或治疗结果预测的个体化预测工具有限。我们假设，综合分析患者完整的疼痛相关临床数据、病史和脑部成像，可以识别与手术结果相关的关键因素，并对三叉神经痛（TN）-慢性面部疼痛综合征的特定结果类别进行分层。我们使用监督式和非监督式机器学习方法，分析了102例典型TN患者的数据。术前临床数据通过非监督式学习进行处理，以确定TN结果分层的关键临床因素及其与手术反应的相关性。同时，我们将监督式学习应用于术前T1加权脑磁共振成像。临床数据分析发现，疼痛和非疼痛相关指标——包括疼痛频率、药物缓解程度、疼痛特征、糖尿病史和癌症史——在预测手术结果方面最为显著。分析显示，术前临床数据与手术反应持续时间有很强的相关性（= 0.5，< 0.00001）。成像数据分析使用了支持向量机分类模型，对于长期反应者或无反应者的召回率很高，分别为0.79和0.86，受试者工作特征曲线（AUC）下面积分别为0.86和0.84。预测手术反应类别持续时间的平均多类准确率为78%（AUC 0.8）。这些结果共同表明，TN手术结果类别是可区分的，并且可以根据手术治疗前可用的临床和脑成像数据组合对手术结果进行分层。我们提出了一个关于慢性疼痛疾病不同分层的新观点，每层都有结构成像、临床相关性和特定的手术结果。
