Artificial Intelligence-Based Tools Designed to Predict Future Mechanical Ventilation and/or Mortality in Hospitalised COVID-19 Patients: A Systematic Literature Review

OBJECTIVES: Artificial intelligence (AI) can be used to create personalised predictions for mortality and mechanical ventilation (MV) in patients hospitalised with COVID19. Such tools may offer valuable early warning alerts and spur prompt intervention to prevent poor outcomes. However, it is important to proactively test to ensure that these tools work fairly and equally well in all populations in which they are intended to be applied. This systematic literature review covers AI-based tools that are designed to predict mortality and/or MV in patients hospitalised with COVID19.

METHODS: This study searched articles from PubMed, MEDLINE and EMBASE. This study was performed using the AutoLit platform (Nested Knowledge). Papers had to be published between 01-Jan-2020 to 10-Jan-2024.

RESULTS: Originally 181 studies were identified, with 140 papers remaining after two human and one Nested Knowledge AI-powered screening. Most models focussed on mortality (87.1%), 6.4% on MV and 6.4% on both outcomes. Datasets were global, including from North America (30.7%) and Europe (21.4%). Where reported, average dataset size was 6589.7 individuals, 58.5% male, and average age of 60.2 years. Where reported, the average predictive accuracy was 87.3%. Most papers used tree-based methods (63.6%) and/or deep learning (29.3%). Most algorithms (91.4%) used demographic features and lab test data (81.4%), and some used vital signs (39.3%). Less than one-third of studies included external validation (27.3%). Most studies did not report ethnicity (78.6%) and only one explicitly reported an assessment of ethnicity bias.

CONCLUSIONS: These tools exhibit good predictive abilities and they may be able to meaningfully improve patient care and outcomes. However, accuracy is not the only dimension in which to understand AI tool performance. The presence of un-representative training data, lack of external validation data and reported bias assessments indicate significant room for improvement. These tools should work well for all patients whose care they are intended to inform/improve.