Data-driven subtypes reveal hidden diversity in chronic inflammatory demyelinating polyneuropathy

Chronic inflammatory demyelinating polyneuropathy, or CIDP, has long been treated as a single disease with a handful of variants, but a large retrospective analysis of German patient records suggests the condition is far more fragmented than current diagnostic labels imply. Researchers screened 425 patients diagnosed with typical CIDP or CIDP variants at ten German university hospitals between 2010 and 2024, and used an unbiased machine-learning clustering approach to sort 325 of them into a discovery cohort. Instead of the familiar split into typical and atypical forms, the data resolved into four distinct subgroups, each with its own demographic profile, electrophysiological signature, comorbidity burden and disease trajectory. Two of the four clusters behaved as red-flag phenotypes, marked by high disability and, in one case, relentless progression over three years of follow-up.

The study, published in the Journal of Neurology, was designed to overcome the limitations of earlier attempts at CIDP subtyping, which the authors note were constrained by small sample sizes, restricted variable sets or a lack of external validation. Rather than asking clinicians to assign patients to predefined categories, the team harmonized demographic, clinical, electrophysiological, laboratory, comorbidity and treatment variables and applied hierarchical clustering based on Gower distance, a metric suited to mixed data types. The resulting tree pointed to four clusters, which were then confirmed using partitioning around medoids with a fixed random seed for reproducibility. A separate validation cohort of 100 patients, drawn by center-stratified random sampling, showed sufficient agreement with the discovery solution, with an accuracy of 67 percent and a Cohen’s kappa above 0.5.

The first cluster was the most striking departure from textbook expectations. It comprised 51 patients, 75 percent of them women, with the highest leg disability scores at diagnosis among all groups, a mean INCAT leg score of 2 points. Weakness, ataxia and areflexia were frequent, yet neuropathic pain was the least common of any cluster. Diagnostically, more than a third of tested patients carried antinuclear antibodies, and autoimmune comorbidities such as hypothyroidism, present in 31 percent, and rheumatoid arthritis, present in 11 percent, far exceeded general population estimates. Because CIDP is usually more common in men, this female-dominated, autoimmunity-enriched, high-disability group stands out as a potential signature of distinct immune mechanisms, and the authors flagged it as one of the two red-flag subtypes.

Cluster 2, the largest with 120 patients, looked much closer to the conventional picture of a stable, treatment-responsive disease. It was 78 percent male, with the lowest INCAT arm and leg disability scores at diagnosis, the highest proportion of purely demyelinating patterns on nerve conduction studies, and significantly fewer comorbidities of every kind captured in the dataset. Over 36 months of follow-up, this was the only cluster whose total INCAT scores improved. The finding aligns with earlier reports that demyelinating electrophysiology predicts better treatment response, whereas early axonal loss has been associated with severity and refractoriness, and it suggests this subgroup carries the lowest overall disease burden.

Cluster 3, with 107 patients, was overwhelmingly male at 88 percent and defined by a combination of high symptom levels, axonal-demyelinating damage on neurography in 92 percent of cases, and a heavy cardiovascular comorbidity load. Arterial hypertension affected 69 percent of these patients, and nicotine and alcohol abuse were also common. Notably, this group received steroids most often, both overall and as induction therapy, yet reported the lowest subjective treatment response at just over half of patients. The authors caution that it remains unclear whether the hypertension was caused or worsened by corticosteroid treatment itself, a question they say warrants prospective investigation, but the association positions vascular and metabolic factors as possible contributors to disease trajectory in this subgroup.

Cluster 4 was the smallest, at 47 patients, and the most ominous. It consisted almost entirely of multifocal CIDP, with no typical CIDP cases, and 85 percent of patients were men. Arm disability at diagnosis was the highest of any cluster, ataxia and areflexia were frequent, and axonal-demyelinating damage appeared in 93 percent of tested neurographies. Over the 36-month observation window, this cluster showed progressive deterioration in INCAT scores, making it the second red-flag subgroup. The authors note that previous work has suggested multifocal CIDP may involve a predominantly cellular immune response and disruption of the blood-nerve barrier at sites of conduction block, raising the possibility that multifocality is not merely a distributional variant but a biologically distinct form of the disease that demands closer monitoring and earlier therapeutic adaptation.

Beyond describing the clusters, the team tested whether the classification could be reproduced by a simpler clinical rule. A pruned classification and regression tree, trained on 80 percent of the discovery cohort and evaluated on an independent 20 percent test set, achieved 81 percent overall accuracy, with 84 percent sensitivity and 93 percent specificity. The decision pathways were clinically intuitive: cluster 1 was reached through the absence of multifocal disease combined with areflexia and female sex, cluster 3 through areflexia, male sex and an axonal-demyelinating pattern, and cluster 4 through multifocal CIDP alone, which served as the sole criterion leading to that subgroup. Receiver operating characteristic analyses showed that typical versus multifocal CIDP discriminated cluster 4 particularly well, with area under the curve values of at least 0.8.

Treatment patterns differed markedly across the subgroups even though treatment variables were deliberately excluded from the clustering procedure. Steroid use was least common in clusters 2 and 4, at 56 and 49 percent respectively, while 73 percent of cluster 3 patients received steroids. Intravenous immunoglobulins dominated maintenance therapy in all clusters, but were used least in cluster 3. Objective treatment response, measured against the 2021 European Academy of Neurology and Peripheral Nerve Society criteria, did not differ significantly between clusters, although subjective response ranged from 85 percent in cluster 4 down to 51 percent in cluster 3, a gap that may reflect the differing comorbidity and symptom profiles of these groups.

The findings arrive at a moment when the field is actively searching for mechanism-based, individualized therapies, since up to half of CIDP patients show no objective response to initial treatment with corticosteroids, intravenous immunoglobulins or plasma exchange. The 2021 EAN/PNS guidelines updated diagnostic criteria but, as the authors point out, the conventional classification does not reliably predict severity, course or treatment responsiveness. By showing that a purely data-driven partition of real-world patient data yields subgroups with meaningfully different trajectories, the study supports the idea that CIDP comprises several biologically distinct disease processes rather than one entity with variable expression. The authors also connect their results to earlier unsupervised clustering work by Chang and colleagues, noting that despite differences in cohort, variables and methodology, both studies converge on heterogeneity beyond established phenotypic categories.

Important limitations temper the conclusions. The retrospective design, choices about feature selection, scaling and mixed-data handling, and the possibility that prior treatment decisions indirectly shaped the clinical characteristics used for clustering may all have influenced cluster formation. Detailed information on treatment dosage and duration was not systematically available, and nodal and paranodal antibody testing, available for only 20 percent of patients with uniformly negative results, could not be incorporated. A significant association between cluster membership and recruiting center, with a moderate Cramér’s V of 0.48, points to differences in referral patterns, diagnostics and treatment practices across sites. The authors therefore describe the clusters as exploratory and emphasize that external validation in prospective cohorts will be essential before the framework can guide individualized monitoring or therapy, potentially aided by neurophysiological, ultrasound or digital assignment tools.

Subject of Research: Data-driven identification of clinical subtypes in chronic inflammatory demyelinating polyneuropathy

Article Title: Clinical heterogeneity in chronic inflammatory demyelinating polyneuropathy: data-driven subtypes challenge current classification

Article References: Quint, P., Kohle, F., Oeztuerk, M., Motte, J., Fisse, A.-L., Stascheit, F., Gilliam, E., De Lorenzo, A., Afzali, A. M., Fischer, K., Disse, P., Nelke, C., Mueller-Miny, L., Mausberg, A. K., Stettner, M., Pitarokoili, K., zu Hoerste, G. M., Brunkhorst, R., Pawlitzki, M., … Ruck, T. (2026). Clinical heterogeneity in chronic inflammatory demyelinating polyneuropathy: data-driven subtypes challenge current classification. Journal of Neurology, 273(11), Article 681. https://doi.org/10.1007/s00415-026-14221-1

Image Credits: AI Generated

DOI: 10.1007/s00415-026-14221-1

Keywords: CIDP, clustering, neurology, autoimmunity, demyelination, multifocal CIDP, disability, comorbidities, patient subtypes, electrophysiology, peripheral neuropathy, disease progression