SurvITE: Learning Heterogeneous Treatment Effects from Time-to-Event Data

Abstract

We study the problem of inferring heterogeneous treatment effects from time-toevent data. While both the related problems of (i) estimating treatment effects for binary or continuous outcomes and (ii) predicting survival outcomes have been well studied in the recent machine learning literature, their combination -albeit of high practical relevance -has received considerably less attention. With the ultimate goal of reliably estimating the effects of treatments on instantaneous risk and survival probabilities, we focus on the problem of learning (discrete-time) treatment-specific conditional hazard functions. We find that unique challenges arise in this context due to a variety of covariate shift issues that go beyond a mere combination of wellstudied confounding and censoring biases. We theoretically analyse their effects by adapting recent generalization bounds from domain adaptation and treatment effect estimation to our setting and discuss implications for model design. We use the resulting insights to propose a novel deep learning method for treatment-specific hazard estimation based on balancing representations. We investigate performance across a range of experimental settings and empirically confirm that our method outperforms baselines by addressing covariate shifts from various sources.

Optimizing adjuvant treatment options for patients with glioblastoma

“…A total of 20,443 patients with complete BCSS records were included in this study, with a median follow-up time of 12 (6)(7)(8)(9)(10)(11)(12)(13)(14)(15)(16)(17)(18)(19)(20)(21) months and an overall BCSS mortality rate of 75.9% [95% confidence interval (CI): 75.3-76.5%]. The median age was 62 (54-70) years, and 40.6% of patients were male.…”

“…This is primarily due to the fact that a patient cannot simultaneously receive both treatments, and confounding variables are prevalent in observational studies ( 15 ). Benefitting from advances in machine learning (ML) and statistical theories, we can use balanced representation-based ( 16 ), tree-based ( 17 ), and conditional average treatment effect (CATE)-based ( 18 , 19 ) methods to counterfactually infer patients’ individual treatment effect (ITE) directly from observational data and thus attempt identify the relatively optimal treatment choice for specific individuals. With the development of deep learning (DL) and representation learning, novel techniques enable combining DL with survival models and learning balanced representations directly from the data to reason about unbiased counterfactual survival outcomes ( 20 ).…”

Individualized survival prediction and surgery recommendation for patients with glioblastoma

“…The individual treatment effect (ITE) can only be obtained by inferring from data ( 10 ). With the ideal way of including treatment as a covariate ( 11 ), although it is predictive, as the model will be biased from confounders if the treatment is not allocated randomly ( 12 ), it is not an unbiased estimate. Alternatives include conditional average treatment effect (CATE)- ( 13 ), matching- ( 14 ), and representation-based approaches ( 15 ).…”