Interpretable and Intuitive Machine Learning Approaches for Predicting Disability Progression in Relapsing-Remitting Multiple Sclerosis Based on Clinical and Gray Matter Atrophy Indicators

• Published in: Academic radiology Vol. 31; no. 7; p. 2910
• Main Authors: Yan, Zichun, Shi, Zhuowei, Zhu, Qiyuan, Feng, Jinzhou, Liu, Yaou, Li, Yuxin, Zhou, Fuqing, Zhuo, Zhizheng, Ding, Shuang, Wang, Xiaohua, Yin, Feiyue, Tang, Yang, Lin, Bing, Li, Yongmei
• Format: Journal Article
• Language: English
• Published: United States 01.07.2024
• Subjects: Adult; Atrophy; Disability Evaluation; Disease Progression; Female; Gray Matter - diagnostic imaging; Gray Matter - pathology; Humans; Machine Learning; Magnetic Resonance Imaging - methods; Male; Middle Aged; Multiple Sclerosis, Relapsing-Remitting - diagnostic imaging; Multiple Sclerosis, Relapsing-Remitting - pathology; Retrospective Studies; Multiple sclerosis; Gray matter; Disease progression; Machine learning
• ISSN: 1878-4046
• DOI: 10.1016/j.acra.2024.01.032

To investigate whether clinical and gray matter (GM) atrophy indicators can predict disability in relapsing-remitting multiple sclerosis (RRMS) and to enhance the interpretability and intuitiveness of a predictive machine learning model. 145 and 50 RRMS patients with structural MRI and at least 1-year follow-up Expanded Disability Status Scale (EDSS) results were retrospectively enrolled and placed in the discovery and external test cohorts, respectively. Six clinical and radiomics feature-based machine learning classifiers were trained and tested to predict disability progression in the discovery cohort and validated in the external test set. Partial dependence plot (PDP) analysis and a Shiny web application were conducted to enhance the interpretability and intuitiveness. In the discovery cohort, 98 patients had disability stability, and 47 patients were classified as having disability progression. In the external test set, 35 patients were disability stable, and 15 patients had disability progression. Models trained with both clinical and radiomics features (area under the curve (AUC), 0.725-0.950) outperformed those trained with clinical (AUC, 0.600-0.740) or radiomics features only (AUC, 0.615-0.945). Among clinical+ radiomics feature models, the logistic regression (LR) classifier-based model performed best, with an AUC of 0.950. Only the radiomics feature-only models were applied in the external test set due to the data collection problem and showed fair performance, with AUCs ranging from 0.617 to 0.753. PDP analysis showed that female patients and those with lower volume, surface area, and symbol digit modalities test (SDMT) scores; greater mean curvature and age; and no disease modifying therapy (DMT) had increased probabilities of disease progression. Finally, a Shiny web application (https://lauralin1104.shinyapps.io/LRshiny/) was developed to calculate the risk of disability progression. Interpretable and intuitive machine learning approaches based on clinical and GM atrophy indicators can help physicians predict disability progression in RRMS patients for clinical decision-making and patient management.