Λουτράδης Δημήτριος, Καθηγητής, Ιατρική, ΕΚΠΑ
Βλάχος Νικόλαος, Καθηγητής, Ιατρική, ΕΚΠΑ
Συριστατίδης Χαράλαμπος, Αναπληρωτής Καθηγητής, Ιατρική, ΕΚΠΑ
Χρέλιας Χαράλαμπος, Αναπληρωτής Καθηγητής, Ιατρική, ΕΚΠΑ
Δρακάκης Πέτρος, Αναπληρωτής Καθηγητής, Ιατρική, ΕΚΠΑ
Παναγόπουλος Περικλής, Επίκουρος Καθηγητής, Ιατρική, ΕΚΠΑ
Σιμοπούλου Μαρία, Επίκουρη Καθηγήτρια, Ιατρική, ΕΚΠΑ
Introduction: Medically Assisted Reproduction (MAR) has completed four decades of development and rapid evolution as a medical field, for the prevention and treatment of infertility, but with consistently low success rates worldwide, compared to the anticipated rate due to the level of medical intervention in the process of human reproduction. Following technological advances in other medical fields, in the MAR field, prediction models have been minimally utilized as a complementary tool for making medical decisions and for the assessment of possible factors that contribute to the positive or negative outcome of assisted reproduction, but with varying levels of precision and limitations that restrict their routine application. More sophisticated systems, such as Artificial Neural Networks (ANN) based on Artifical Intelligence, demonstrate flexibility with the ability to constantly learn from their surroundings and readapt, while demonstrating stability and efficiency, features that are superior in the context of improving the processes surrounding Assisted Reproduction Units, their respective success rates, and supporting medical decisions aimed at the individualized management of infertile patients.
Objective: The development of an innovative, efficient and reliable ANN. with clinical, laboratory and embryological parameters to predict the outcome of forthcoming MAR cycles and the statistical assessment of the included parameters to signify their role in the positive or negative clinical outcome of assisted reproduction, with the end-point of livebirth.
Participants and Methods: The present study was conducted through the retrospective and anonymous data recording on 118 parameters in a normalized population of 256 subfertile couples that underwent MAR 426 cycles in the Assisted Reproduction Unit of the Attikon University Hospital (July 2010 - February 2017), after relevant approvals by the Bioethics Committee and the Scientific Council of the Hospital and according to predefined eligibility and exclusion criteria. A dynamic database with demographic data, MAR cycle characteristics, clinical, laboratory and embryological parameters was created, including the clinical outcome for all included cycles. Data was re-adjusted and recorded according to its nature as numerical and qualitative data, followed by a statistical evaluation for the total of parameters included through SAS 9.4 software, by selecting the appropriate statistical test for the type of data, while logistic regression models were applied to highlight non-linear correlations with clinical outcome of MAR. On the basis of the statistically emerging factors, an innovative ANN was developed. Through programming in MATLAB software, which was validated for its stability by repeating the training and testing process in 10 groups of randomly divided data. The efficiency and accuracy of the developed ANN was assessed through the calculation of sensitivity and specificity values, Positive Predictive Value (PPV), Negative Predictive Value (NPV), False Positive Rate (FPR), False Negative Rate (FNR), Overall Accuracy(OA) and Odds Ratio (OR) for all 10 separate ANNs that descended for the originally developed system.
Results: Of the 118 parameters studied, the statistical analysis revealed that age of the woman, age group, age at menarche and particularly in a calculated threshold of 12 years, the total dose of gonadotropins in a single MAR cycle, endometrial thickness, top quality embryos on Day 3 of laboratory development and the corresponding ratio to the number of embryos available, the ratio of the number of embryos transferred to the total number of embryos of the cycle and “fresh” or “frozen” cycle, as evaluated parameters demonstrated statistical significance for their effect on the clinical outcome of MAR, with a further establishment of their relationship through the applied logistic regression models. In addition, an assessment of the correlation of sex ratio with age at menarche was performed in cycles with a positive clinical outcome, confirming a tendency for female offspring in women with earlier menarche. The statistically significant parameters were used as inputs to the developed ANN which was assigned a sensitivity of 77.68%, a specificity of 74.55% and total accuracy of 76.12% for the training set and a sensitivity of 71.05%, with a specificity 70.91% and a total accuracy of 70.95% for the test set. The system demonstrated a minimal standard deviation and difference in performance indicators between the training and test set. Following the training and testing of the system with ten different sets of random data, the sensitivity was 69.2% ± 2.36%, the specificity was 69.19% ± 2.8% (OR 5.21 ± 1.27) with a PPV of 36.96 ± 3.44 and a NPV of 89.61 ± 1.09, while exhibiting an OA of 69.19% ± 2.69%.
Conclusions: The findings of the study converge on repeatedly highlighting the role of certain parameters for which there is documented literature on their influence on the clinical outcome of MAR cycles, For others parameters the results presented here, came to add clinical value to previously presented data, sometimes with conflicting views and results. Furthermore and importantly, other parameters are genuinely documented and reported here for the first time in the scientific community, such as the role of age at menarche in the clinical outcome of MAR. In the final phase of the study and as defined by its initial design, an innovative ANN was developed, based on the concept of the use of artificial intelligence in assisted reproduction, demonstrating quality performance with high sensitivity and specificity values, even after testing and validating it with 10 random data sets and demonstrating increased stability in predicting livebirth following MAR. The presented system could act as a surrogate tool for infertility management by providing additional evidence on clinical decisions and through additional external validation and contribution with data from other IVF Units. with its simultaneous use, its performance could be enhanced by adapting to different MAR environments with differing practices.