Many obesity indicators have been linked to adiposity and its distribution. Utilizing a combination of multidimensional obesity indicators may yield different values to assess the risk of moderate-to-severe obstructive sleep apnea (OSA). We aimed to develop and validate the performances of automated machine-learning models for moderate-to-severe OSA, employing multidimensional obesity indicators as compact representations. We trained, validated, and tested models with logistic regression and other 5 machine learning algorithms on the clinical dataset and a community dataset. Light gradient boosting machine (LGB) had better performance of calibration and clinical utility than other algorithms in both clinical and community datasets. The model with the LGB algorithm demonstrated the feasibility of predicting moderate-to-severe OSA with considerable accuracy using 19 obesity indicators in clinical and community settings. The useable interface with deployment of the best performing model could scale-up well into real-word practice and help effectively detection for undiagnosed moderate-to-severe OSA.