Table 1. Summary of the models created for prediction. AUROC of only the best performing algorithm in each study have been included. The two values for studies in the size of dataset column indicate training set and test set where applicable.

AUROC: area under receiver operating curve; ARDS: acute respiratory distress syndrome; APACHE III: Acute Physiological and Chronic Health Evaluation III; NLP: natural language processing; ML, machine learning; ICU: intensive care unit; EHR: Electronic Health Record; GA: genetic algorithm; MIMIC III : Medical Information Mart for Intensive Care III; P/F: PaO₂/FiO₂

Author(s)	Dataset used	Size of dataset	AUROC	Conclusion of study
Brown et al. (2011) [17]	ARDS network trial	(1,800, 222)	0.71 vs. 0.73 (APACHE III)	Simple classification rule was developed that stratified patients according to hospital mortality which was comparable to widely used APACHE III
Afshar et al. (2018) [21]	Data from the 533 patients admitted to certain wards of a tertiary medical center	533	0.8	NLP and ML were used to build a computable phenotype of ARDS
Christie et al. (2019) [5]	Observational cohort data	1,494	0.84-0.89	Superlearner fits provide versatile means of helping clinicians integrate big data on severely injured patients into real-time, dynamic decision-making support
Ding et al. (2019) [19]	ICU data of patients admitted to five different centers in Beijing	296	0.82	A model for predicting ARDS was developed in Chinese patients which included 11 features
Zeiberg et al. (2019) [15]	Single-center EHR data	(1,621, 1,122)	0.81	Feasibility of ML models to risk stratify ARDS patients solely based on EHR data was demonstrated
Zhang (2019) [20]	Secondary analysis of two randomized controlled trials conducted across 44 hospitals	1,071	0.821 vs. 0.665 (APACHE III)	A model based on neural network using GA was developed which outperformed the conventional scoring system for predicting mortality in ARDS patients
Yang et al. (2020) [16]	MIMIC III	(6,601, 2,101)	0.9128	An algorithm based on patients’ noninvasive physiological parameters to estimated P/F ratio was developed