JMIR Publications

ABSTRACT

Background:

The Cox proportional hazards (CPH) model is a common choice for analyzing time to treatment interruptions in patients on antiretroviral therapy (ART), valued for its straightforward interpretability and flexibility in handling time-dependent covariates. Machine learning (ML) models have increasingly been adapted for handling temporal data, with added advantages of handling complex, non-linear relationships, large datasets, and provide clear practical interpretations.

Objective:

This study aims to compare the predictive performance of the traditional CPH model and ML models in predicting treatment interruptions among patients on ART, while also providing both global and individual-level explanations to support personalized, data-driven interventions for improving treatment retention.

Methods:

Using data from 621,115 patients who started ART between 2017 and 2023, in Kenya, we compared the performance of the CPH with 6 different ML models – Gradient Boosting Machine, Extreme Gradient Boosting, Regularized Generalized Linear Models (Ridge, Lasso and Elastic-Net) and Recursive Partitioning – in predicting first and multiple treatment interruptions. Explainable surrogate technique (model-agnostic) was applied to interpret the best-performing model's predictions globally, using variable importance and partial dependence profiles (PDP), and at individual-level, using break-down additive (BD), Shapley additive explanations (SHAP), and ceteris paribus (CP).

Results:

Recursive partitioning (RP) model achieved the best performance with a predictive concordance index score (C-Index) of 0.81 for first treatment interruptions and 0.89 for multiple interruptions, outperforming the CPH model, which scored 0.78 and 0.87 for the same scenarios, respectively. RP’s performance can be attributed to its ability to model non-linear relationships and automatically detect complex interactions. The global model-agnostic explanations aligned closely with the interpretations offered by hazard ratios in the CPH model, while offering additional insights into the impact of specific features on the model's predictions. The BD and SHAP explainers demonstrated how different variables contribute to the predicted risk at the individual patient level. The CP profiles further explored the time-varying model, to illustrate how changes in a patient’s covariates over time could impact their predicted risk of treatment interruption.

Conclusions:

In conclusion, our results highlight the superior predictive performance of ML models and their ability to provide patient-specific risk predictions and insights that can support targeted interventions to reduce treatment interruptions in antiretroviral therapy care.