Introduction: Machine learning algorithms may be beneficial in the setting of prostate cancer screening. Our group has developed one such algorithm based on PSA, age, free:PSA and other variables based on the data sets from the Prostate, Lung, Colorectal and Ovarian Cancer and PCPT trials. Limited data exists examining the effect of machine learning algorithms on prostate cancer diagnosis prospectively. We aimed to report our early experience using this algorithm in prostate cancer screening on an Australian cohort.
Methods: 1223 men without previous history of prostate cancer were tested using an AI clinical decision support tool in a digital health clinical setting. Clinicians assessed demographic factors as well as PSA, Free-PSA, Prostate Health Index and MRI results alongside an AI based clinical decision support tool. Men determined to be high risk were referred for urologic assessment and potential biopsy. When results were discordant with guidelines but flagged by AI the case was reviewed by a second clinician. The results of those men who underwent biopsy were collected and compared to statistics in a recent publication from the Australian Prostate Cancer Outcomes Registry.
Results: Of the 1223 men tested prospectively, 49 were referred to a urologist on the basis of their clinical parameters and AI risk score. These men underwent MRI which was then assessed. Of those that underwent MRI 26 continued to a prostate biopsy. Of the 26 biopsies completed 19 were found to have prostate cancer, 17 of the 19 (89%) cancers were found to be Gleason 7 or above on surgical histopathology. This is above a recently published rate of 74%. The median PSA at diagnosis was 2.3 µg/L, lower than the published median of 7.3 µg/L. No biopsied men were found to have metastatic disease compared to a published rate of 7%. Overall initial results show early detection of clinically significant disease with a reduced rate of biopsies in low risk men.
Conclusions: Our machine learning model appears to reduce the need for biopsy and reduces the diagnosis of low-risk prostate cancer. Further, we highlight that PSA at the time of diagnosis was considerably lower than comparable screening cohorts. We provide a proof-of-principal highlighting the utility of machine-learning algorithms for prostate cancer screening in a real-world population.
Source of Funding: MP is funded by a Fulbright scholarship through the Australian-America Fulbright Association