AB0205 A NOVEL METHOD FOR PREDICTING 1-YEAR RETENTION OF ABATACEPT USING MACHINE LEARNING TECHNIQUES: DIRECTIONALITY AND IMPORTANCE OF PREDICTORS

Abstract

Background:In the ACTION (NCT02109666) study, multivariable Cox proportional hazards regression models showed that the predictors of 1-year retention to abatacept treatment were: patient global pain assessment, country, reason for stopping last biologic, number of prior biologic treatments, abatacept monotherapy, RF/anti-cyclic citrullinated peptide (CCP) status, previous neoplasms, psychiatric disorders and cardiac disorders.1 Machine learning techniques, using the gradient-boosting model, subsequently identified additional predictors of abatacept retention in patients with moderate-to-severe RA enrolled in ACTION; however, the analysis did not show the directionality of the predictors.2Objectives:To improve the clinical interpretability of the machine learning model in terms of directionality and the importance of each variable in predicting retention.Methods:Previous analyses using the gradient-boosting model to identify predictors of abatacept retention at 1 year in the ACTION study have been described.2 This analysis used SHapley Additive exPlanations (SHAP), a mathematical framework, to show how a particular predictor value influences prediction in the context of all other predictors. Higher SHAP values indicate a higher likelihood of retention. The contribution of every variable in the model’s prediction (with the exception of country variables) was computed for each data point to capture individual variable impact. This enabled interpretation for level of importance and directionality at a patient level.Results:Using data from 2350 patients enrolled in ACTION (May 2008 to December 2013), the mean retention rate at 1 year was 59.3% (n=1393). Overall variable importance is shown in Figure 1. After removal of country variables, the top five baseline predictors of retention were: no previous corticosteroid use, ACR functional class II, ≥2 prior biologic treatments prior to abatacept initiation, abatacept monotherapy and HAQ-DI. In terms of directionality, no previous corticosteroid use, ≥2 prior biologic treatments prior to abatacept initiation, abatacept monotherapy and a higher HAQ-DI score at baseline were associated with a lower likelihood of retention; ACR functional class II was associated with a higher likelihood of retention.Conclusion:The gradient-boosting model previously identified predictors of abatacept retention from ACTION;2 the addition of SHAP in this analysis has provided information on the importance and directionality of those predictors. The most important predictor of abatacept retention was no previous corticosteroid use, which was associated with lower retention. The models and predictors identified could be further refined by using additional datasets from clinical trials. Machine learning offers an innovative and complementary approach to biostatistics and could be used to identify treatment response predictors at an individual patient level, leading to a more personalised treatment approach.