A Machine Learning Algorithm to Define Disease Severity in Chronic Inflammatory Demyelinating Polyneuropathy

OBJECTIVES: We aim to define health states representing different levels of disease severity to inform a cost-effectiveness model of efgartigimod in Chronic Inflammatory Demyelinating Polyneuropathy (CIDP). To this scope, we developed a machine learning algorithm that classifies CIDP severity from the adjusted Inflammatory Neuropathy Cause and Treatment (aINCAT) disability score, a measure of functional limitation of arms and legs.

METHODS: The algorithm was developed on data from ADHERE, a Phase 2 trial that included 322 CIDP patients. The algorithm aggregated different aINCAT scores in clusters. The EuroQoL 5 Dimensions (EQ-5D) data based on UK preference weights was used as a measure of disease severity. Starting from eleven clusters corresponding to each aINCAT score, each iteration of the algorithm tested possible aggregations of the existing clusters and selected the aggregation associated with the lowest increase in within-cluster dissimilarity across the sample. Only clusters of adjacent aINCAT scores were considered for aggregation to limit the number of possible combinations. Dissimilarity was assessed using the sum of squares residuals, and the optimal number of clusters was selected using the Gap statistic. We ran the algorithm on 1000 bootstrap data samples to assess which cluster sets were most frequently selected as optimal and assessed their quality using ANOVA.

RESULTS: The dataset included 1057 observations. The two cluster sets most frequently selected were: aINCAT 0, aINCAT 1-2, aINCAT 3, aINCAT 4, aINCAT 5-6, aINCAT 7-8, aINCAT 9-10 (Cluster Set A, 8.1% of the simulations); and aINCAT 0-1, aINCAT 2-3, aINCAT 4, aINCAT 5-6, aINCAT 7-8, aINCAT 9-10 (Cluster Set B, 7.2% of the simulations). P-values from ANOVA were significant (<0.001) for both Cluster Sets A and B.

CONCLUSIONS: Results show that both Cluster Sets may be appropriate to represent different levels of disease severity. Further validation from clinical experts is needed to confirm this result.