OBJECTIVES:

This review provides an overview of existing ML methods for estimating individualized or heterogeneous treatment effect.

METHODS:

The paper begins with a brief overview of the potential to use CI for personalised medicine and its application in CEA for HTA. This is followed by a review and taxonomy of existing ML methods to estimate subgroup-specific and individualised treatment effects (ITE) from real-world data under specific types of outcome variables, treatment exposure and data structure. The paper helps practitioners identify the most appropriate method to use, depending on: the available data (cross-sectional or longitudinal); the outcome of interest (continuous, binary or time-to-event (TTE)); whether the method can handle observed or unobserved confounders; if the method explicitly quantifies measure(s) of uncertainty; the software (R, Python or Stata) used to implement it. By contrasting the taxonomy against the information required to conduct CEA for HTA, the paper highlights the gaps that ML methods developers need to address for ML to become integral part of the next-generation toolbox used in HTA.

RESULTS: There is extensive literature on ML methods for ITE estimation, although not all produce estimates consistent within a CI framework. Most of the methods can handle confounding at baseline, but cannot accommodate time-varying and hidden confounding. Those ML methods that estimate ITE in longitudinal settings and account for time-varying confounding, have been developed for use with continuous outcomes. Only one ML method can estimate ITE for TTE outcomes while accounting for time-varying confounders. Most methods produce point estimates using non parametric estimation and do not formally quantify uncertainty around their predictions.

CONCLUSIONS:

More work is required to further develop and integrate CI and ML methods for the analysis of real-world data to inform treatment and funding decisions.