ABSTRACT
Background: It has been difficult to distinguish between mild, moderate, and severe cases at an early stage for COVID-19 patients, making it challenging to decide an optimized treatment for each patient. This machine learning system could predict the clinical course and be used to develop a novel method to provide optimal treatment based on risk. Method: We applied machine learning techniques to international clinical data from a large cohort of patients with COVID-19 at 15 hospitals in Japan and three hospitals in New York City from January 1, 2020 to March 30, 2021. We analyzed clinical information of over 2000 COVID-19 patients comprising various races and ethnicities and built a severity and mortality prediction model. Furthermore, using a severity index with machine learning allowed early detection of patients most at-risk for developing severe illness to support the decision for the patient to receive optimized therapy. Findings: We developed an international COVID-19 early prediction model for use at the time of hospital admission that predicts disease severity and mortality with high accuracy, 0.88 (AUC). Using the novel method of severity-matched analysis to assess treatment effectiveness, in the high-risk group, the Kaplan–Meier estimates of mortality by Day 30 were 26% in the dexamethasone treatment group and 63% in the non-treatment group. The Kaplan–Meier estimates of mortality were low at 3% with remdesivir and dexamethasone in combination and 49% with no remdesivir and dexamethasone treatment by Day 30. There may be an add-on effect of remdesivir to conventional dexamethasone. Interpretation: The severity prediction index can be calculated, which can assist with an optimized treatment for COVID-19 patients in each risk group. The severity-matched treatment system could support the recommendation of optimized treatments, such as dexamethasone, remdesivir, or heparin, in high-risk groups by calculating the severity index predicted at the time of the first visit.Trial Registration Details: The trial registration number was 2020142NI. Funding Information: K.T., K.I., and Y.D. received funding from the Japan Agency for Medical Research and Development (AMED) (19fk0108153h0001). K.T. received funding from AMED (19jm0610015h0001). K.T. received funding from the Healthy Longevity Global Grand Challenge, Catalyst Award.Declaration of Interests: The authors declare no competing interests. Ethics Approval Statement: The study protocol was centrally reviewed by the Institutional Review Board of Tokyo University. The requirement for consent was waived given the retrospective and non-interventional nature of the study.