Introduction The purpose of the present study was to develop artificial intelligent (AI) algorithms for analyzing genome-wide cell-free DNA (cfDNA) for early detection and prediction of prognosis of epithelial ovarian cancer.

Methods Whole blood samples from epithelial ovarian cancer patients (n=120) stored in a biobank were used to develop AI algorithms for genome-wide analysis of cfDNA. Convolutional neural network (CNN) and multilayer perception (MLP) deep learning methods were used for algorithm development. Another batch of whole blood samples from the patients who were newly-diagnosed with ovarian tumor (both benign and malignant) were prospectively collected and run through the developed algorithms. Sensitivity and specificity of the developed algorithms in differentiating malignant tumors from benign tumors were explored.

Results A total of 219 whole blood samples from the patients who were newly-diagnosed with ovarian tumor were run through the algorithms and the probability scores of malignancy were calculated. The probability scores calculated by the analysis of DNA fragmentation size, patterns of sequence of end motif, regional mutation types and their density were found to be significantly higher in cancer patients than those with benign tumors. Furthermore, these scores became increasingly higher as the extent of disease assessed by the FIGO staging system increased. This machine-learning model incorporating genome-wide cfDNA analysis had sensitivities of detection at 92% at 98% specificity, with an overall area under the curve value of 0.99.

Conclusion/Implications The use of AI algorithms for analyzing cfDNA yielded high diagnostic accuracy for epithelial ovarian cancer demonstrating the potential value of precision oncology based on whole-genome analysis.