Construction accidents can easily cause massive casualties and property losses. This research uses machine learning technique to analyze 16 critical factors and assess the impact of diverse combinations of factors on the performance of predicting the severity of construction accidents. The prediction is carried out with eight algorithms: Logistic regression, Decision tree, Support vector machine, Naive Bayes, K-nearest neighbor, Random forest, Multi-Layer Perceptron and AutoML. The results show that (1) Based on 16 accident factors, Naive Bayes and Logistics regression achieve the best F1-Score of 78.3 % on raw data set. (2) With AutoML method, severity classification can achieve an average F1-Score of 84 %. (3) The analysis of the confusion matrix shows that the subjective classification of the original data and specific unusual accidents are the sources of misprediction. (4) The "Type of accident" and "Accident reporting and handling" are the most critical factors and "Emergency management" and "Safety training" are important subsystems, both of which greatly affect the severity of the accident. (5) Based on the Decision tree, a set of assessment rules for the severity of construction accidents can be extracted. The prediction models and conclusions obtained from this study can be used to enhance the experience of safety professionals in urban construction and to make the safety intervention measures more efficient. (C) 2020 Institution of Chemical Engineers. Published by Elsevier B.V. All rights reserved.