2F-03 - Benefits of Machine Learning Methods for Identifying Clinical Features and Patient Characteristics Associated with Treatment Outcomes in Autoimmune Conditions

Background: To provide high-value healthcare, providers seek information regarding patients who may benefit from a new therapy. Machine learning (ML) methods emerged to complement traditional regression modeling for identifying outcome-related features.

Objectives: To demonstrate how ML methods complement regression approaches in elucidating patient, clinical and treatment history features associated with treatment outcomes in autoimmune conditions.

Methods: The prospective, non-interventional CorEvitas® Psoriasis (PsO) and Rheumatoid Arthritis (RA) Registries collect data on adults diagnosed with PsO or RA, respectively. Baseline and follow-up data are collected from patients and providers during routine clinical visits using Registry questionnaires.

In RA, change in disease activity was assessed using continuous outcome CDAI (Clinical Disease Activity Index). Linear regression models identified clinical features associated with change in CDAI, using backward elimination and Akaike Information Criterion. ML methods including Least Absolute Shrinkage and Selection Operator (LASSO) regression with stability selection identified clinical features associated with therapy benefits and decision trees (DTs) identified combinations of clinical and patient features that may benefit from target therapy.

In PsO, logistic regression, using a Bayesian Information Criterion with stepwise elimination, identified features associated with switching biologic agents. ML methods including Random Forest (RF) ranked importance of variables, with accumulated local effect (ALE) and SHapley Additive exPlanations (SHAP) plots describing how covariates influenced predictive ability. Unobtainable from logistic regression models, DTs identified rare combinations of clinical and patient features likely to lead to therapy switch.

Results: In RA, linear regression identified baseline CDAI, comorbid conditions, demographic factors and patient-reported outcomes associated with change in CDAI. LASSO identified the strongest predictors including baseline CDAI whereas DT identified combinations of features (e.g. baseline CDAI and history of comorbidities). In PsO, logistic regression identified change in Body Surface Area (BSA), baseline BSA, additional treatment, concomitant conditions and quality of life as important variables associated with switching. DT identified specific combinations of these.

Conclusions: While traditional modeling identified similar features, ML methods (LASSO, DT and RF) offered additional insights beyond these models, including relative importance of clinical features and patient characteristics to predict treatment outcomes in autoimmune conditions.