Abstract Study question How machine learning assisted in generating patient-friendly corifollitropin alfa protocol in normal responders? Summary answer In retrospective experiments, our machine learning model integrated physiological measurements of patients and clinical experience to generate a patient-friendly corifollitropin alfa protocol. What is known already Long-acting corifollitropin alfa can simplify the regimen, minimizing injections during the whole cycle. The previous study has described the patient-friendly protocol using corifollitropin alfa without routine pituitary suppression in normal responder can result in non-compromised clinical outcomes. Some studies showed machine learning can help with making clinical decisions and have the ability to learn from physiological measurements. Those methods effectuate certain points throughout short-acting menotropin protocols, however, there are still no robust AI tools for long-acting corifollitropin alfa protocols. Study design, size, duration 1,309 cycles were collected at Stork Fertility Center from November 2016 to October 2019, and 1,221 cycles were available after data cleaning and applying exclusion criteria, which Anti-Mullerian Hormone (AMH) is lower than 2. The data from electronic medical records (EMRs) consisted of age, AMH, body weight, luteinizing hormone (LH), and estradiol (E2) concentrations measured on revisit. Evaluation is performed by one physician who has more than 20 years of experience in IVF. Participants/materials, setting, methods: The protocol generator consisted of 5 parts: doses of Elonva, trigger type, doses of recombinant follicle-stimulating hormone (rFSH), doses of recombinant luteinizing hormone (rLH), and day of oocyte retrieval. The protocol was predicted by age, AMH, and weight firstly, then fine-tuned by LH and E2 after the first revisit. We used the gradient boosting decision tree algorithm to learn the protocol. The dataset was randomly split into 80% for training and 20% for testing. Main results and the role of chance In classification, the model predicted the dose of Elonva achieved an accuracy of 0.913 and an AUC of 0.946, and trigger type got an accuracy of 0.901 and AUC of 0.852 only using features on stimulation day (SD) 1 and gained 0.012 and 0.056 in accuracy and AUC correspondingly after adding features on the first revisit day. In regression, the mean absolute error (MAE) of rFSH dose, rLH dose, and oocytes retrieved day was 156.30 IU, 232.75 IU, and 0.80 days respectively, and after refining, the MAE dropped to 92.37 IU, 100.07 IU, and 0.46 days. The error of predictions in rFSH and rLH was almost equal to half increments of rFSH (150 IU) and one increment rLH (75 IU). This indicated that our model could provide a better prediction of these clinical decisions with one revisit only. Limitations, reasons for caution The present study was a single-center retrospective, and only analyzed the data from normal responders, whose AMH was equal or greater than 2. Though, the recommendations of our system act as references, the physician will make the final decision. Wider implications of the findings: Our result showed the potential of machine learning in generating protocols is promising. Recommendations generated by our model can provide the junior clinical teams to optimize the clinical plans and learn from the experience of experts. We look forward to applying our machine learning model to different protocols. Trial registration number Not applicable
Read full abstract