# Causal deep learning to personalize medicine: Which intensive care patients with sepsis will benefit from corticosteroid therapy?

## TL;DR

This study uses deep learning to identify which ICU patients with sepsis are likely to benefit from corticosteroid treatment based on their initial condition.

## Contribution

A novel causal deep learning approach is applied to personalize corticosteroid therapy for sepsis patients.

## Key findings

- Patients with severe metabolic acidosis and impaired circulation are likely to benefit from corticosteroids.
- The model showed internal and external validation AUROCs of 0.79 and 0.71, respectively.
- Physicians' treatment decisions often diverged from the model's predictions, suggesting room for improvement in clinical assessment.

## Abstract

Sepsis, defined as life-threatening organ dysfunction due to dysregulated host response to an infection, often requiring intensive care treatment. There is a strong rationale for the administration of corticosteroids for immunomodulation; however, clinical trials are inconclusive, which may be attributable to heterogeneity in therapeutic effects between individual patients. Leveraging deep learning within a causality framework, we aimed to identify for which intensive care patients with sepsis corticosteroids lead to improved survival.

We trained the treatment agnostic representation network (TARNet) to estimate the reduction in predicted probability of 28-day mortality following initiation of corticosteroid treatment of intensive care patients with sepsis. We used the freely available and public AmsterdamUMCdb ICU database for causal model development, considering 19 predictor variables from the first 24 h of admission, and validated the model with Medical Information Mart for Intensive Care (MIMIC-IV) version 2.2 data. A cut-off of 10% reduction in predicted probability of mortality was used to classify treatment responders.

According to the Sepsis-3 criteria, a total of 2920 admissions in AmsterdamUMCdb were eligible. Of these, 1378 were assigned to the intervention group and 1542 to the control group. Internal validation of predictions of the observed outcomes showed an area under the receiver operating characteristic curve (AUROC) of 0.79, while external validation yielded an AUROC of 0.71. Covariate balance of the TARNet model latent representation, as measured by the Wasserstein distance, was 3.6 × 10⁻⁷ for the internal data set and 4.2 × 10⁻⁷ for the external data set. Based on the estimated reduction of predicted mortality, a distinction was made between treatment responders (n=245), non-responders (n=2098), and those predicted to be harmed by corticosteroid treatment (n=577).

Corticosteroid treatment responders were those with severe metabolic acidosis and impaired circulation, whereas patients who were less ill based on these parameters were more likely to have increased mortality rates by corticosteroid treatment. There was also a notable discrepancy between the model’s suggestions and the physicians’ treatment that was carried out, implying improvements in the clinical assessment of patients with sepsis are necessary. Given recent years have not yielded new treatments for sepsis, computational clinical decision-support systems are worth exploring.