Using Causal Inference and Machine Learning Methods to Predict Cognitive Behavioral Treatment Response

点击翻译按钮获取中文摘要

PROJECT SUMMARY/ABSTRACT Many patients with anxiety and fear disorders (AFDs) report minimal benefits when treated with an evidence- based psychotherapy (e.g., cognitive-behavioral therapy [CBT]) or pharmacotherapy (e.g., antidepressants; benzodiazepines). Conversely, some AFD patients are likely to benefit from virtually any treatment (e.g., supportive therapy). Using data collected in randomized controlled trials (RCTs), there has been limited progress determining how to use pre-treatment characteristics to match AFD patients to the treatment that is most likely to provide benefit. As a result, NIMH has forwarded Strategic Objectives focused on identifying treatment moderators and developing tools that predict differential treatment response. The broad goal of this proposed secondary data analysis is to apply causal inference and machine learning methods to prospective observational data to predict differential treatment response among patients with AFDs. The sample (n = 1,528) is from a longstanding NIMH-funded study of AFD patients who received: (a) CBT with concurrent pharmacotherapy, (b) CBT without pharmacotherapy, or (c) treatment as usual (TAU). Targeted maximum likelihood estimation (a causal inference method) and super learning (an ensemble machine learning method) will be used to accomplish the proposed Aims. Aim 1 will estimate the (overall) average effects of the three treatment types. Aim 2 will estimate “optimal treatment rules” to determine if differential treatment response can be meaningfully predicted based on a patient's multidimensional profile of pre-treatment symptoms. Aim 3 will estimate “optimal treatment rules” to determine if differential treatment response can be meaningfully predicted using all available pre-treatment covariates. The proposed study is highly innovative and could significantly impact the growing literature focused on predicting differential treatment effects and personalizing treatment for patients with AFDs. Although treatment effects estimates obtained using observational data and causal inference methods (i.e., adjusted for nonrandom treatment selection) are similar to those estimated in RCTs, this would be the first study to apply such methods to AFD patient data. This study will also be the first to use ensemble machine learning to develop composite moderators for AFDs (i.e., optimal treatment rules). In comparison, prior attempts to develop composite moderators have relied on less flexible model-building procedures prone to overfitting and unable to capture complex predictor-outcome associations (e.g., interactions among predictors; nonlinear associations). In achieving the proposed Aims, the current study would be a catalyst for future research using causal inference and machine learning to study predictors of differential treatment response among AFD patients. Results will be used to justify future research aimed at expanding and validating the models in larger observational samples and pragmatic RCTs (e.g., optimal treatment rules for specific medications/doses; timing of pharmacotherapy relative to CBT; second-wave CBT versus acceptance-based CBT).