Digital megavoltage imaging (DMI) detector is a standard imaging device installed on Halcyon treatment machines to measure the x-ray intensity transmitted through a patient during the treatment session and for daily quality assurance (QA) using the MPC system. In this study, we characterized the digital megavoltage imaging (DMI) response for the machine performance check (MPC) on Halcyon treatment machine and predicted the performance of DMI using the machine learning algorithms to improve the accuracy of DMI dosimetry. The measured flood-field image on the DMI segmented to 9 regular grids of 9 × 9 cm2 size with the central grid positioned at the isocenter to evaluate the uniform response over time across the different cross-section of the DMI surface area. The time-resolved MPC beam output changes evaluated through comparison with the ion chamber measurements over 30 days. Also, the image ghosting effect evaluated for the repeated and time-delay MPC beam output change, and the raw flood field data was analyzed. Four different machine learning (ML) algorithms (linear regression, support vector regression (SVR), k-nearest neighbor (KNN), and random forest) were trained and tested with 757 days measurement, aiming to predict our MPC output change and flood field. The accuracy of the predictive data calculated by assessing its deviation from the test data in terms of root mean square error (RMSE). The stability of DMI across the different cross-sections was a maximum of 0.8 % deviation over 30 days compared to the center grid. The time-resolved measurement of MPC output compare to the ion chamber showed a gradual increase in the output deviation up to 0.4% over 30 days. The test result of the ghosting effect for the repeat MPC measurements shows the maximum output difference - 0.95% compared to fresh MPC with 5 minutes time-delay. The recovering rate of MPC output was 0.02 for 5 - 40 minutes and with a gradual rise afterward. After an approximate of 10-hour delay, the MPC measurement will fully recover its full response. The predicted MPC output changes agree with test data to within RMSE of 0.281, 0.266, 0.270, and 0.302 % for linear, SVR, KNN, and random forest models, respectively. On the average, predictive raw flood field agrees with test data within RMSE of 2.030, 1.794, 2.943, and 3.979% for linear, SVR, KNN, and random forest models, respectively. Generally, the SVR ML algorithm has the highest degree of accuracy. From this study, the DMI panel will lose its response function by 1.799% (∼2%) in 365 days usage. The characteristic of the DMI for MPC measurement was presented, and ML algorithms will be useful for preventive maintenance. Further study warranted to develop DMI response correction based on the current study and also appropriate QA strategy of DMI needs to be in place.