Machine learning model for risk assessment of corona intensive care patients
2022
Charité - Universitätsmedizin Berlin, Berlin, Germany
Clinically established risk assessments, such as SOFA and APACHE II scores, show limited performance in predicting the survival of critically ill COVID-19 patients. In order to optimize the treatment and allocation of intensive care resources, plasma proteomes of two independent patient cohorts (Germany and Austria) were analyzed in order to predict the outcome (death vs. survival) in severe cases. In time series samples, 14 inflammatory proteins were identified that increased significantly in patients with fatal outcomes. In survivors, a decrease in these proteins in plasma was observed in parallel. The proteomic predictors showed high accuracy in the outcome prediction, while the established methods evaluated in parallel performed significantly worse. However, time series samples are time-consuming and labour-intensive and therefore unsuitable for clinical diagnostics and therapy decisions. Therefore, the research group developed a machine learning model whose predictions are based on single-time samples taken at the first point in time at the maximum treatment stage (WHO grade 7) of the patients. The AI analyzes individual protein pairs and calculates the probability of a lethal outcome by evaluating individual deviations from the overall population. The model identified 15 proteins of the coagulation system and 8 proteins of the complement cascade as highly relevant in a severe form with a fatal outcome. Based on proteomics data from the patient cohort from Germany, the AI correctly predicted the outcome for 18 out of 19 patients who survived and for 5 out of 5 patients who died in the Austrian cohort. The results show that the plasma proteome comprehensively reflects the host's response to COVID-19 and significantly improves the clinical diagnosis of corona-risk patients. The proteome analysis and development of further predictors can prove helpful in improving the diagnosis and therapy choice for other diseases.
A proteomic survival predictor for COVID-19 patients in intensive care
Florian Kurth
Added on: 08-11-2022
[1] https://journals.plos.org/digitalhealth/article?id=10.1371/journal.pdig.0000007[2] https://www.bionity.com/en/news/1174423/machine-learning-model-uses-blood-tests-to-predict-covid-19-survival.html
