Do machine learning methods lead to similar individualized treatment rules? A comparison study on real data.

Identifying patients who benefit from a treatment is a key aspect of personalized medicine, which allows the development of individualized treatment rules (ITRs). Many machine learning methods have been proposed to create such rules. However, to what extent the methods lead to similar ITRs, that is, recommending the same treatment for the same individuals is unclear. In this work, we compared 22 of the most common approaches in two randomized control trials. Two classes of methods can be distinguished. The first class of methods relies on predicting individualized treatment effects from which an ITR is derived by recommending the treatment evaluated to the individuals with a predicted benefit. In the second class, methods directly estimate the ITR without estimating individualized treatment effects. For each trial, the performance of ITRs was assessed by various metrics, and the pairwise agreement between all ITRs was also calculated. Results showed that the ITRs obtained via the different methods generally had considerable disagreements regarding the patients to be treated. A better concordance was found among akin methods. Overall, when evaluating the performance of ITRs in a validation sample, all methods produced ITRs with limited performance, suggesting a high potential for optimism. For non-parametric methods, this optimism was likely due to overfitting. The different methods do not lead to similar ITRs and are therefore not interchangeable. The choice of the method strongly influences for which patients a certain treatment is recommended, drawing some concerns about their practical use.

Estimating individualized treatment effects using an individual participant data meta-analysis.

BACKGROUND: One key aspect of personalized medicine is to identify individuals who benefit from an intervention. Some approaches have been developed to estimate individualized treatment effects (ITE) with a single randomized control trial (RCT) or observational data, but they are often underpowered for the ITE estimation. Using individual participant data meta-analyses (IPD-MA) might solve this problem. Few studies have investigated how to develop risk prediction models with IPD-MA, and it remains unclear how to combine those methods with approaches used for ITE estimation. In this article, we compared different approaches using both simulated and real data with binary and time-to-event outcomes to estimate the individualized treatment effects from an IPD-MA in a one-stage approach. METHODS: We compared five one-stage models: naive model (NA), random intercept (RI), stratified intercept (SI), rank-1 (R1), and fully stratified (FS), built with two different strategies, the S-learner and the T-learner constructed with a Monte Carlo simulation study in which we explored different scenarios with a binary or a time-to-event outcome. To evaluate the performance of the models, we used the c-statistic for benefit, the calibration of predictions, and the mean squared error. The different models were also used on the INDANA IPD-MA, comparing an anti-hypertensive treatment to no treatment or placebo ( N = 40 237 , 836 events). RESULTS: Simulation results showed that using the S-learner led to better ITE estimation performances for both binary and time-to-event outcomes. None of the risk models stand out and had significantly better results. For the INDANA dataset with a binary outcome, the naive and the random intercept models had the best performances. CONCLUSIONS: For the choice of the strategy, using interactions with treatment (the S-learner) is preferable. For the choice of the method, no approach is better than the other.