P-060 how a learning model based on Artificial Intelligence can predict the pregnancy in relation to DNA spermatozoa damage

Abstract Study question The DNA spermatozoa integrity impacts on the male fertility, we aim to define if the damage status can predict the pregnancy. Summary answer The relation between the spermatozoa DNA damage and the pregnancy can support the doctor to recommend the couples to proceed in vivo or vitro conception What is known already The DNA spermatozoa is getting more and more important in the male fertility. Studies show that spermatozoa of infertile males have higher level of DNA damage than fertile males. DNA damages are often connected to poor seminal parameters such as sperm count, motility and morphology even if the 8% of males with normal seminal parameters show the same diagnosis. We are also skeptical about using the spermatozoa with damaged DNA in the ICSI, indeed there could be some consequences. The WHO suggests to evaluate the DNA damage in the routine exams. Study design, size, duration The aim of the study is to develop an Artificial Intelligence model that it can predict the pregnancies related on DNA quality of spermatozoa. During 2015 and 2017, we evaluated the DNA damage of 268 male after their partners had a natural pregnancy. The DNA spermatozoa damage is evaluated with p53 protein dosage so we developed a preliminary “learning model”. Participants/materials, setting, methods It has been enrolled 356 male partners (pm) of infertile couples, divided into: group A, 123 pm with 2-9 million spermatozoa/ejaculate (ICSI); group B, 108 pm with 10-38 million spermatozoa/ejaculate (IUI); group C, 125 pm with >39 million spermatozoa/ejaculate (natural conception). p53 dosage, with ELISA method, was performed for the groups A and B preliminary to the treatment ICSI and IUI; for group C, it was performed from 10 to 30 days afterwards pregnancy (betaHCG > 400 miU/mL) Main results and the role of chance Group A: 29 pregnancies (23,6 %) occurred out of 123 ICSI performed. According to our learning model: - group A1: 42 pm (p53 < 1.65), 26 pregnancies were assumed against the 25 obtained, forecast percentage of 96.2 %. - group A2: 81 pm (p53 > 1.66), 6 pregnancies were assumed against the 4 obtained with a forecast percentage of 66.6 %. Group B: 16 pregnancies (14,8 %) occurred out of 108 IUI performed. According to our learning model: - group B1: 38 pm (p53 < 1.65), 16 pregnancies were assumed against the 14 obtained, forecast percentage of 87.5 %. - group B2: 70 pm (p53 > 1.66), 2 pregnancies were assumed against the 2 obtained with a forecast percentage of 100.0 %. Group C: out of 125 couples, with natural conception, 28 pregnancies (22,4 %). According to our prediction model: - group C1: 71 pm (p53 < 1.65), 25 pregnancies were assumed against the 24 obtained, predicted rate of 96.0 %. - group C2: 54 pm (p53 > 1,66), 4 pregnancies were assumed against the 3 obtained with a 75,0 % forecast percentage. The forecast is accurate (AUC 0.70). Limitations, reasons for caution The comparison between the control “learning model” and the expectation is not totally comparable. It is necessary to add a detailed case history of the couple to minimize other factors that may interfere with conception. For this reason, it is still necessary to be cautious to adopt this model. Wider implications of the findings A preventive examination, such as evaluating sperm DNA damage, could minimize in vivo and in vitro conception failures. This “learning model” seems to have an excellent ability to predict a pregnancy. To achieve greater sensitivity, greater numbers are needed through its diffusion. Trial registration number Not Applicable