Deep Learning Model Predicts Sepsis Patients Likely to Benefit from Steroid Treatment

Even with modern intensive care, only about 60–70% of patients with septic shock survive the first 30 days. Corticosteroids remain one of the most hotly debated treatments — they help some patients but may harm others, and traditional clinical trials aren’t built to untangle that kind of individual variability. Now, researchers have used deep learning to identify patients with sepsis who are most likely to respond to corticosteroids, improving treatment precision.

A research team from the Department of Intensive Care Medicine, Amsterdam UMC, Netherlands) turned to causal deep learning to pinpoint which patients truly benefit. They built a predictive model using TARNet, focusing on 28-day mortality, and trained it on 2,920 Sepsis-3 patients in the AmsterdamUMCdb database. Each profile included 19 commonly collected clinical variables such as lactate, pH, and the PaO₂/FiO₂ ratio.

To test how well the model held up elsewhere, they validated it on the large American MIMIC-IV v2.2 dataset of 30,639 patients. In findings published in the Journal of Intensive Medicine, the model showed a strong performance: AUROC was 0.79 internally and 0.71 in the external cohort, with Brier scores of 0.14 in both sets. Calibration remained robust, and TARNet achieved near-perfect covariate balance, significantly outperforming conventional propensity score matching.

Researchers then grouped patients based on a clinically meaningful 10% shift in predicted 28-day mortality. This created three subgroups: responders (245 patients), non-responders (2,098 patients), and those likely to be harmed (577 patients). Individuals with severe metabolic acidosis and circulatory dysfunction, characterized by low pH, low bicarbonate, elevated lactate, and higher creatinine, were the most likely to benefit from corticosteroids. These findings align with current understanding of sepsis physiology.

By learning shared representations before estimating each patient’s potential outcome, TARNet could cleanly separate benefit from no benefit or harm, while its excellent covariate balancing reduced confounding. The dual-database approach also added weight to the findings, with the model performing consistently even though the MIMIC-IV population was less severely ill. Importantly, all 19 input variables are routinely collected worldwide, and the response threshold mirrors real-world decision-making, making the results highly practical for ICUs.

“Our study addresses a critical gap in current practice. Causal deep learning enables the estimation of individualized treatment effects, overcoming the limitations of traditional population level studies,” said Dr. Ameet Jagesar. “This is the first application of the TARNet model in the context of corticosteroid therapy for sepsis.”

Related Links:
Amsterdam UMC