Deep phenotyping to predict live birth outcomes in in vitro fertilization

Deep phenotyping to predict live birth outcomes in in vitro fertilization
Prajna Banerjeea,1, Bokyung Choib,1, Lora K. Shahinea,c, Sunny H. Juna,d, Kathleen O’Learya, Ruth B. Lathia, Lynn M. Westphala, Wing H. Wonge, and Mylene W. M. Yaoa,2
+ Author Affiliations

aDepartment of Obstetrics and Gynecology, Stanford University School of Medicine, Stanford, CA 94305;
bDepartment of Applied Physics, School of Humanities and Sciences, Stanford University, Stanford, CA 94305;
cPacific Northwest Fertility and In Vitro Fertilization Specialists, Seattle, WA 98104;
dDepartment of Obstetrics and Gynecology, Palo Alto Medical Foundation, Fremont, CA 94538; and
eDepartment of Statistics, School of Humanities and Sciences, Stanford University, Stanford, CA 94305
Edited* by Grace Wahba, University of Wisconsin, Madison, WI, and approved June 1, 2010 (received for review February 24, 2010)

↵1P.B. and B.C. contributed equally to this work.

Abstract
Nearly 75% of in vitro fertilization (IVF) treatments do not result in live births and patients are largely guided by a generalized age-based prognostic stratification. We sought to provide personalized and validated prognosis by using available clinical and embryo data from prior, failed treatments to predict live birth probabilities in the subsequent treatment. We generated a boosted tree model, IVFBT, by training it with IVF outcomes data from 1,676 first cycles (C1s) from 2003–2006, followed by external validation with 634 cycles from 2007–2008, respectively. We tested whether this model could predict the probability of having a live birth in the subsequent treatment (C2). By using nondeterministic methods to identify prognostic factors and their relative nonredundant contribution, we generated a prediction model, IVFBT, that was superior to the age-based control by providing over 1,000-fold improvement to fit new data (p < 0.05), and increased discrimination by receiver–operative characteristic analysis (area-under-the-curve, 0.80 vs. 0.68 for C1, 0.68 vs. 0.58 for C2). IVFBT provided predictions that were more accurate for ∼83% of C1 and ∼60% of C2 cycles that were out of the range predicted by age. Over half of those patients were reclassified to have higher live birth probabilities. We showed that data from a prior cycle could be used effectively to provide personalized and validated live birth probabilities in a subsequent cycle. Our approach may be replicated and further validated in other IVF clinics.

regression tree model in vitro fertilization prediction live birth prediction in vitro fertilization prognostics personalized medicine

Footnotes
2To whom correspondence should be addressed. E-mail: mylene.yao@stanford.edu. Author contributions: P.B., B.C., L.M.W., W.H.W., and M.W.M.Y. designed research; P.B., B.C., L.K.S., S.H.J., and M.W.M.Y. performed research; P.B., B.C., L.K.S., S.H.J., K.O., R.B.L., L.M.W., W.H.W., and M.W.M.Y. analyzed data; and P.B., B.C., L.K.S., S.H.J., K.O., R.B.L., L.M.W., W.H.W., and M.W.M.Y. wrote the paper.

Conflict of interest statement: M.W.M.Y and W.H.W. have cofounded a new company to make personalized prognostics accessible to IVF patients. The company is in the start-up, prefunding, precommercialization stage at the time of manuscript submission. L.K.S. is an employee at Pacific Northwest Fertility and IVF Specialists, and S.H.J. is an employee at Palo Alto Medical Foundation.

*This Direct Submission article had a prearranged editor.

See Commentary on page 13559.

This article contains supporting information online at
www.pnas.org/lookup/suppl/doi:10.1073/pnas.1002296107/-/DCSupplemental.

http://www.pnas.org/content/107/31/13570