Combating Infertility: Machine Learning for Clinical Decision Support

Abstract

One in eight couples is affected by infertility in the United States. In vitro fertilization (IVF) is the most popular treatment option. Unfortunately, the process is time-consuming (1-2 months), costly ($12,000-15,000), emotionally taxing, and physically demanding without guaranteed success. Hence, often the first question patients ask is their probability of live birth after treatment. Unfortunately, clinicians fail to effectively predict and communicate this likelihood, resulting in high IVF dropout rates. Clinical decision support systems (CDSSs) powered by machine learning (ML) can address this issue through their abilities to prognosticate clinical outcomes based on patient features and enhance clinician decision making with these predictions. However, to date, the vast majority of CDSSs have failed to achieve adoption and impact. Research suggests these failures stem from problems in the ML life cycle of data preparation, model creation, and deployment. Thus, we present three contributions to tackle shortcomings in each phase of the ML life cycle. For data preparation, we developed IVF Explorer, an interactive visualization tool, to onboard new researchers and clinical partners. It surfaces clinical terminology and the IVF process while enabling dynamic exploration and pattern identification in the data. We verify usability through a think-aloud study with three new IVF researchers. For model creation, we preprocess the new data from the Boston IVF Fertility Clinic and conduct EDA to confirm the viability of prognostication. Then, we put forth a novel phase-by-phase modeling architecture that incorporates the prediction of intermediate clinical outcomes into the holistic prognostic model. We demonstrate its improvement in performance at predicting IVF clinical outcomes compared to standard ML prognostic models for pregnancy and live birth. For deployment, we present methods to improve patient risk communication through visualizations of summary statistics and patient cohort comparisons. We then evaluate the impact of visual explanations for patient risk on clinical decision making under uncertainty and emotional affect. We propose a novel experimental framework of a think-aloud protocol and MTurk user study to evaluate our hypothesis that the risk communication method postively affects patient decision making.