Table 3.

AI in glucocorticoid therapy for sepsis.

Title	First author	Objects and number of studies	Models	Corticosteroid regimen	Observations	Reference
Use of IFNγ/IL10 Ratio for Stratification of Hydrocortisone Therapy in Patients With Septic Shock	König	83 patients from the Berlin cohort, a subset of the CORTICUS study.	machine learning model, not specified	Patients received 200 mg/day hydrocortisone for 5 days followed by a tapering dose until day 11.	A low IFNγ/IL10 ratio indicated increased survival in the hydrocortisone group, whereas a high ratio suggested better survival in the placebo group.	[76]
Assessment of Machine Learning to Estimate the Individual Treatment Effect of Corticosteroids in Septic Shock	Pirracchio	2548 adult patients with severe infectious shock from 4 randomized controlled trials	Super Learner, with integrated machine learning algorithms.	1. IV hydrocortisone 50 mg/6 h for 5-7 days, tapered or not. 2. Same dosage plus oral flutriasone 50 mcg daily, no tapering.	Corticosteroids (hydrocortisone or hydrocortisone plus flutriasone) reduced 90-day death risk (RR 0.89, P = .004).	[77]
Machine learning and murine models explain failures of clinical sepsis trials	Stolarski	119 adult inbred wild-type ICR mice divided into two sepsis models: abdominal infection (CLP) and pneumonia (PNA)	Machine learning algorithms using Lasso regression and 10-fold cross-validation	Mice received hydrocortisone, vitamin C, and thiamine (HAT).	HAT benefited mice with abdominal infections but not those with pneumonia.	[78]
Learning Optimal Treatment Strategies for Sepsis Using Offline Reinforcement Learning in Continuous Space	Wang	6660 ICU patients meeting sepsis-3 criteria	Offline deep reinforcement learning based on continuous state-action space (including history capture models, generative models, perturbation models and Q-networks)	Use of hydrocortisone or not	Mortality was lowest when the dose matched the AI's decision.	[79]
Development of a Reinforcement Learning Algorithm to Optimize Corticosteroid Therapy in Critically Ill Patients with Sepsis	Bologheanu	3051 patients with sepsis admitted to the intensive care unit (ICU) were screened according to Sepsis-3 criteria	Reinforcement Learning Algorithms Based on Markov Decision Processes and Temporal Differences in Actor-Critic Methods	Systemic glucocorticoid doses were converted to hydrocortisone equivalents and categorized: 0 mg, 1-100 mg, 101-200 mg, 201-300 mg, and > 300 mg.	Mortality decreased when following the algorithm's steroid recommendations.	[80]
Machine-learning-derived sepsis bundle of care	Kalimouttou	42,735 adults with sepsis or septic shock	LASSO regression machine learning model	Antimicrobials, balanced crystalloid, insulin therapy, corticosteroids, pressor and sodium bicarbonate therapy	In the validation cohort and modeling cohort, the full treatment group had a lower 28-day mortality rate compared to the incomplete treatment group (15.5 % vs 37.8 % and 11.3 % vs 15.1 %).	[81]