Counterfactual inference framework for joint estimation of medical costs, treatment effect and time-to-event

Abstract

Optimal treatment selection is extremely crucial in emergency situations such as for a patient admitted in ICU. However, the chosen medical treatment may not necessarily be a financially favorable choice. In some of the developing countries like India, there is a lack of a good public health insurance system and not everyone can afford private health insurance. Therefore, balancing the trade-off between medical costs and best treatment choice is important not just from patient point of view but also for settings where free healthcare services are provided. In order to make the best decision, it is essential for clinicians to know about the treatment which will lead to shorter duration of inpatient hospital stay as well as the associated medical costs with each potential treatment choice. Various authors have tried to predict duration of stay or medical costs of inpatient ICU stays or the treatment effect with time-to-event as the outcome variable. However, to the best of our knowledge, there is no research work that proposes joint estimation of medical cost and duration of stay in a hospital taking into consideration individual treatment effect. Our research work addresses this issue, and provides two novel frameworksMedCI andMedSCI, that not only predict time-to-stay and associated medical costs for a given treatment and counterfactual treatment choice but also return individual treatment effect for both the outcomes. Our work is a mixture of Survival Analysis, Causal Inference and Deep Learning. The results are obtained on a semisynthetic and synthetic dataset for MedCI while MedSCI is evaluated on a synthetic dataset.