Personalized prediction of live birth outcomes in IVF

Abstract

OBJECTIVE: The decision to pursue IVF after a failed attempt is challenging. Currently, IVF patients are primarily guided by age-based stratification. We aimed to develop a personalized prognostic tool that uses clinical data from prior, failed treatments to predict live birth (LB) probabilities in subsequent treatment cycles.DESIGN: We extracted IVF outcomes data from 5056 IVF cycles performed at the Stanford Hospital Clinic from 2003-08.MATERIALS AND METHODS: Using a non-deterministic approach, we generated a boosted tree model (IVFBT), by training the model with data from 1676 first cycles (C1s) from 2003-06. IVFBT was externally validated by 634 C1s from 2007-08. We further tested the ability of IVFBT to predict LB outcomes in the subsequent cycle (C2).RESULTS: We found that age alone was limited in predicting LB outcomes. Top non-redundant prognostic factors affecting LB outcome included percentage of embryos forming blastocysts, total amount of gonadotropins administered, the number of 8-cell embryos on Day 3, embryo cryopreservation. The IVFBT model was superior to the age-based control at a rate greater than 1000:1 in its ability to fit new data to predict the probability of live birth in a subsequent cycle (p<0.05). Receiver-operative characteristic analysis also showed that IVFBT increased discrimination by 17%. Finally, more than ∼60% of patients would receive a different yet more correct prediction of live birth outcomes in a subsequent cycle compared to an age-based prediction, and more than half of those patients were reclassified to have higher LB probabilities.CONCLUSION: We showed that data from a prior cycle could be used effectively to provide personalized LB probabilities in a subsequent cycle. Thus, the first IVF cycle is prognostic and therapeutic. Our approach may be replicated to support personalized prognostic counseling in other IVF clinics. OBJECTIVE: The decision to pursue IVF after a failed attempt is challenging. Currently, IVF patients are primarily guided by age-based stratification. We aimed to develop a personalized prognostic tool that uses clinical data from prior, failed treatments to predict live birth (LB) probabilities in subsequent treatment cycles. DESIGN: We extracted IVF outcomes data from 5056 IVF cycles performed at the Stanford Hospital Clinic from 2003-08. MATERIALS AND METHODS: Using a non-deterministic approach, we generated a boosted tree model (IVFBT), by training the model with data from 1676 first cycles (C1s) from 2003-06. IVFBT was externally validated by 634 C1s from 2007-08. We further tested the ability of IVFBT to predict LB outcomes in the subsequent cycle (C2). RESULTS: We found that age alone was limited in predicting LB outcomes. Top non-redundant prognostic factors affecting LB outcome included percentage of embryos forming blastocysts, total amount of gonadotropins administered, the number of 8-cell embryos on Day 3, embryo cryopreservation. The IVFBT model was superior to the age-based control at a rate greater than 1000:1 in its ability to fit new data to predict the probability of live birth in a subsequent cycle (p<0.05). Receiver-operative characteristic analysis also showed that IVFBT increased discrimination by 17%. Finally, more than ∼60% of patients would receive a different yet more correct prediction of live birth outcomes in a subsequent cycle compared to an age-based prediction, and more than half of those patients were reclassified to have higher LB probabilities. CONCLUSION: We showed that data from a prior cycle could be used effectively to provide personalized LB probabilities in a subsequent cycle. Thus, the first IVF cycle is prognostic and therapeutic. Our approach may be replicated to support personalized prognostic counseling in other IVF clinics.