Background: The coronavirus disease 2019 (COVID-19) has become a pandemic, placing significant burdens on the healthcare systems. In this study, we tested the hypothesis that a machine learning approach incorporating hidden nonlinear interactions can improve prediction for Intensive care unit (ICU) admission. Methods: Consecutive patients admitted to public hospitals between 1st January and 24th May 2020 in Hong Kong with COVID-19 diagnosed by RT-PCR were included. The primary endpoint was ICU admission. Results: This study included 1043 patients (median age 35 (IQR: 32-37; 54% male). Nineteen patients were admitted to ICU (median hospital length of stay (LOS): 30 days, median ICU LOS: 16 days). ICU patients were more likely to be prescribed angiotensin converting enzyme inhibitors/angiotensin receptor blockers, anti-retroviral drugs lopinavir/ritonavir and remdesivir, ribavirin, steroids, interferon-beta and hydroxychloroquine. Significant predictors of ICU admission were older age, male sex, prior coronary artery disease, respiratory diseases, diabetes, hypertension and chronic kidney disease, and activated partial thromboplastin time, red cell count, white cell count, albumin and serum sodium. A tree-based machine learning model identified most informative characteristics and hidden interactions that can predict ICU admission. These were: low red cells with 1) male, 2) older age, 3) low albumin, 4) low sodium or 5) prolonged APTT. A five-fold cross validation confirms superior performance of this model over baseline models including XGBoost, LightGBM, random forests, and multivariate logistic regression. Conclusions: A machine learning model including baseline risk factors and their hidden interactions can accurately predict ICU admission in COVID-19.