Background/Purpose: In the ACTION study (NCT02109666), previous multivariable Cox proportional-hazards regression models showed that predictors of 1-year retention to intravenous abatacept treatment included: patient global pain assessment, country, reason for stopping last biologic, number of prior biologic treatments, abatacept monotherapy, RF/anti-CCP status, previous neoplasms, psychiatric disorders, and cardiac disorders.¹ Machine learning techniques, using the gradient-boosting model, identified additional predictors of abatacept retention in patients with moderate-to-severe RA enrolled in ACTION; however, the analysis did not show whether the variables predicted reduced or increased retention (directionality).² The objective of this analysis was to assess the clinical importance and directionality of each patient demographic or disease characteristic for predicting retention.

Methods: The gradient-boosting model was used to identify predictors of abatacept retention at 1 year in the ACTION study.² Retention was defined as treatment for > 365 days or ≤ 365 days in patients who achieved remission or major clinical response (using EULAR response criteria, based on DAS28). This analysis expanded on the previous model, adding SHapley Additive exPlanations (SHAP), a mathematical framework, to show how important each characteristic was for predicting abatacept retention. Higher SHAP values indicate a higher likelihood of retention. Every characteristic’s contribution in the model’s prediction (except country) was computed for each data point to capture the impact of each individual characteristic. 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 abatacept retention rate at 1 year was 59.3% (n = 1393). Figure 1 shows how important each characteristic was for predicting retention. Lower treatment retention was predicted by no previous corticosteroid use, ≥ 2 prior biologic treatments prior to abatacept initiation, abatacept monotherapy, and a higher HAQ-disability index score at baseline, whereas higher treatment retention was predicted by ACR functional class II (vs functional classes I, III, and IV).

Conclusion: Adding SHAP to the gradient-boosting model allowed for a more clinically relevant analysis to be performed, showing how the identified variables impacted retention, either positively or negatively. The most important baseline characteristics that were predictive of abatacept retention at 1 year were no previous corticosteroid use, which was associated with lower patient retention in the study, and ACR functional class II, which was associated with higher retention. 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 personalized treatment approach.

References:
1. Alten R, et al. RMD Open 2017;3:e000538.
2. Alten R, et al. Presented at the virtual ACR Convergence 2020; November 5–9, 2020. Poster 1745.

Medical writing: Claire Line, PhD (Caudex), funded by Bristol Myers Squibb

« Back to ACR Convergence 2021

ACR Meeting Abstracts - https://acrabstracts.org/abstract/prediction-of-1-year-intravenous-abatacept-retention-in-patients-with-ra-using-novel-machine-learning-techniques-directionality-and-importance-of-predictors/