Causal machine learning for predicting treatment outcomes

TL;DR

The benefits of causal ML (relative to traditional statistical or ML approaches) are discussed and the key components and steps are outlined and recommended, offering practical guidance for appropriate clinical use.

Abstract

Causal machine learning (ML) offers flexible, data-driven methods for predicting treatment outcomes including efficacy and toxicity, thereby supporting the assessment and safety of drugs. A key benefit of causal ML is that it allows for estimating individualized treatment effects, so that clinical decision-making can be personalized to individual patient profiles. Causal ML can be used in combination with both clinical trial data and real-world data, such as clinical registries and electronic health records, but caution is needed to avoid biased or incorrect predictions. In this Perspective, we discuss the benefits of causal ML (relative to traditional statistical or ML approaches) and outline the key components and steps. Finally, we provide recommendations for the reliable use of causal ML and effective translation into the clinic.

Figures (3)

• Figure 1: Causal ML for predicting treatment outcomes.a, Different from traditional ML, causal ML aims at (i) estimating the treatment effect or (ii) predicting the patient outcomes themselves due to treatments. b, Causal ML is challenging due to the fundamental problem of causal inference in that not all potential outcomes can be observed and are thus missing in the data. c, Treatment effect heterogeneity.
• Figure 2: Workflow for causal ML in medicine.
• Figure 3: Formalizing tasks for causal ML.a, A causal graph must be assumed such as the example here. The arrows ($\rightarrow$) indicate the causal relationships between different variables. Note that the causal graph allows for possible unobserved variables (not confounders) that are correlated with treatment and confounders, or correlated with confounders and outcome. b, The research question defines what causal quantity is of interest, that is, the so-called estimand. The estimand can vary by the effect heterogeneity (average vs. individualized) and treatment type (binary vs. continuous). Here, $Y(a)$ is the potential outcome for treatment $a$. c, The research question determines the problem setup, which comes with different assumptions that must be made to ensure identifiability and thus reliable inferences.