Recently, non-invasive cancer diagnosis using ctDNA has emerged as a new cancer diagnosis method. However, the low fraction of ctDNA limits the use of ctDNA in patients in which mutations cannot be detected. Methylation profile of cancer patients differs from that of healthy people in a broad region of genome. Thus, several studies reported that the potential utility of cancer diagnosis using methylation profiles. but the low sequencing depth of the methylation profile and the selection of effective markers are still obstacles in cancer diagnosis using the methylation profile of ctDNA. In this study, we obtained methylation profiles from a total of 223 cancer patients and non-cancer patients using the enzymatic conversion-based method. To increase the accuracy of cancer detection model,
an effective methylation marker was selected by generating a probabilistic model considering characteristics of ctDNA. Interestingly, we observed that the accuracy of cancer diagnosis was significantly higher in the marker set selected by applying the model compared to when the probability model was not used. The performance of cancer detection model from two cancer types showed an AUC of 0.945 in the ROC curve in the analysis using cfDNA of colorectal cancer and non-cancer patients, and 0.950 in the analysis using cfDNA of lung cancer and non-cancer patients. These results indicate that cancer diagnosis using methylation profiles is effective, and noise removal for marker selection will play an important role in improving the accuracy of cancer diagnosis using ctDNA.