Establishment of a new classification system for chronic inflammatory demyelinating polyneuropathy based on unsupervised machine learning

Introduction/Aims: A model for predicting responsiveness to immunotherapy in patients with chronic inflammatory demyelinating polyneuropathy (CIDP) has not been well established. We aimed to establish a new classifier for CIDP patients based on clinical characteristics, laboratory findings, and electrophysiological features. Methods: The clinical, laboratory, and electrophysiological features of 172 treatment-naive patients with CIDP between 2003 and 2019 were analyzed using an unsupervised hierarchical clustering. The identified pivotal features were used to establish simple classifications using a tree-based model. Results: Three clusters were identified: 1, n = 65; 2, n = 70; and 3, n = 37. Patients in Cluster 1 scored lower on the disability assessment score before treatment. More patients in Clusters 2 (90.0%) fulfilled demyelinating criteria than patients in Cluster 1 (30.8%, p < .001). Cluster 3 had more patients with chronic kidney disease (CKD) (27.0%) and hypoalbuminemia (3.40 g/dL) than did Cluster 2 (CKD: 0%, p < .001; hypoalbuminemia: 4.09 g/dL, p < .001). The responsiveness to pulse steroid therapy was higher in Cluster 2 (70.0%) than in Clusters 1 (31.8%; p = .043) and 3 (25.0%; p = .014). A tree-based model with four pivotal features classified patients in our cohort into new clusters with high accuracy (89.5%). Discussion: The established hierarchical clustering with the tree-based model identified key features contributing to differences in disease severity and response to pulse steroid therapy. This classification system could assist clinicians in the selection of treatments and could also help researchers by clustering patients for clinical treatment trials.