Machine Learning Estimates Treatment Effect of Corticosteroids in Septic Shock

Close up of artificial intelligence brain with binary number code.
The researchers’ objective was to determine whether machine learning-derived estimated individual corticosteroid therapy effect yields better results than treat all or treat no one strategies in adult patients with septic shock.

A study published in JAMA Open Network found evidence for a positive net benefit from an individualized treatment strategy (ITE) deciding which septic shock patients to treat with corticosteroids. Furthermore, the treatment strategy was based on the estimated individual treatment effect as derived from machine learning.

Individual patient data from 4 trials on steroid supplementation in adults with septic shock were used as training cohort to model the ITE using a machine learning approach. In total, 2548 participants with a mean age of 66 (IQR, 55-76) years of whom 1656 (65.0%) were men, were included.

Among this cohort, 515 patients received only hydrocortisone, 1009 received a combination of hydrocortisone plus fludrocortisone, and 1024 received either no treatment or placebo. On day 1, the median Simplified Acute Physiology (SAPS II) and Sepsis-related Organ Failure Assessment (SOFA) scores were 55 (IQR, 42-69) and 11 (IQR, 9-13), respectively.

Corticosteroids, either hydrocortisone or hydrocortisone plus fludrocortisone, decreased risk of 90-day mortality compared with placebo or usual care; the crude pooled relative risk (RR) of 0.89 (95% CI, 0.83-0.96; P =.004) and absolute risk reduction (ARR) of 5.11% (95% CI, 1.50-8.72). When based on SAPS II, the mean estimated probability of death in the overall sample was 55.0% (95% CI, 53.8-56.1). The area under the curve (AUC) for SAPS II was 0.64 (DeLong 95% CI, 0.62-0.67).

Based on the optimal individual model, mean estimated probability of death was 47.7% (95% CI, 46.8-47.8) in the overall sample. The estimated ARR using the optimal individual model was 2.90% (95% CI, 2.79-3.01) and the AUC based on a validation cohort of 75 patients was 0.77 (95% CI, 0.59-0.92).

The net benefit of treating all patients with hydrocortisone or hydrocortisone with fludrocortisone was positive for any number willing to treat (NWT) greater than 25.

With an NWT of 25, the net benefit for the treat all with hydrocortisone strategy was 0.01; -0.01 for the treat all with hydrocortisone and fludrocortisone; 0.06 for treat by SAPS I; and 0.31 for the optimal individual strategy. For a smaller NWT, the net benefit of SAPS II and the optimal individual model converge to 0, but the individual model was consistently superior.

The results may not be generalizable because the data were constrained by the inclusion criteria of the studies used, said investigators. Also, although SAPS II was used as an alternative to the optimal individual model, it was developed to estimate hospital mortality and not the primary outcome of 90-day mortality.

Investigators added that in order to refine the ITE, an appropriate NWT accounting for frequency and severity of adverse effects is needed, and the decision tree generated to illustrate the use of the ITE should be interpreted with caution.

Based on these findings, investigators concluded, “that an individualized estimation-based treatment strategy to decide which patients with septic shock to treat with corticosteroids and which corticosteroid regimen to administer yielded positive net benefit regardless of potential corticosteroid-associated adverse effects.” However, this result needs to be further validated in prospective studies, said investigators.


Pirracchio R, Hubbard A, Sprung CL, Chevret S, Annane D; Rapid Recognition of Corticosteroid Resistant or Sensitive Sepsis (RECORDS) Collaborators. Assessment of machine learning to estimate the individual treatment effect of corticosteroids in septic shock. JAMA Netw Open. 2020;3(12):e2029050. doi:10.1001/jamanetworkopen.2020.29050