AI-Based Multi-Modal COVID-19 Risk Score Improves Severity Prediction of Hospitalized COVID-19 Patients

Risk scores for identifying predictors of disease severity combine several factors including age, sex, and comorbidities. Some risk scores also include additional markers of severity such as the dyspnea symptom, clinical examination variables such as low oxygen saturation and elevated respiratory rate, as well as biological factors reflecting multi-organ failures. Beyond clinical and biological variables, CT scans also contain prognostic information, as the degree of pulmonary inflammation is associated with clinical symptoms, and the amount of lung abnormality is associated with severe evolution. However, the extent to which CT scans at patient admission add prognostic information beyond what can be inferred from clinical and biological data is unresolved.

Owkin scientists conducted a study to integrate clinical, biological, and radiological data to predict the outcome of hospitalized patients. By processing CT scan images with a deep learning model and by using a radiologist report that contains a semi-quantitative description of CT scans, the team evaluated the additional amount of information brought by CT scans. Owkin scientists collected 58 clinical and biological variables, and chest CT scan data, from 1003 coronavirus-infected patients. The team trained a deep learning model based on CT scans to predict severity and then constructed the multimodal AI-severity score that includes five clinical and biological variables (age, sex, oxygenation, urea, platelet) in addition to the deep learning model. Their study showed that neural network analysis of CT scans brings unique prognosis information, although it is correlated with other markers of severity (oxygenation, LDH, and CRP) explaining the measurable but limited 0.03 increase of AUC obtained when adding CT-scan information to clinical variables. The scientists found that when comparing AI severity with 11 existing severity scores, the prognosis performance improved significantly, suggesting that AI severity can rapidly become a reference scoring approach.

Related Links:
Owkin