Harvard Catalyst Profiles

The goal of this project is to develop and assess AI algorithms that predict the need for interventions during birth using electronic fetal monitoring (EFM) data. Electronic fetal monitoring (EFM) is currently used in 99% of all hospital births in the United States to monitor the fetal heart rate. Despite its ubiquity, substantial limitations persist in the efficacy, reliability, and accuracy of EFM in accomplishing its primary intended goal of preventing intrapartum fetal injury. One of the greatest challenges that obstetricians face is interpreting EFM; how to distinguish fetal distress that warrants an emergency cesarean delivery (CD) from a false alarm that can safely be ignored. After over 70 years of attempts to refine obstetricians interpretation of EFM, the results remain dismal. Up to 30% of CDs performed in the United States are thought to be done unnecessarily due to false-positive interpretations of fetal distress from EFM. Currently, EFM interpretation is dependent on clinicians visual inspection of the monitor strips as they print out at the patient bedside or scroll across a computer monitor. Doctors and nurses are taught to examine 4 particular features of the EFM and to categorize the tracing into one of three categories: normal (Category I), indeterminate (Category II), or abnormal (Category III). This method is deeply fraught. A large majority of intrapartum strips are categorized as indeterminate, which then leads to often unnecessary and potentially morbid interventions. Furthermore, interpretation of EFM is extremely subjective and highly vulnerable to biases; both inter-observer and intra-observer reliability of EFM interpretation are poor. Studies have shown that different doctors agree on interpretation of the same strip only 20-30% of the time, and doctors only agree with themselves approximately 70-80% of the time when asked to re-interpret a strip on a different day. The lack of consistency and accuracy in EFM interpretation has been proven to lead to unnecessary interventions without improved fetal outcomes, and this challenge plagues obstetrics. In recent years, machine learning approaches have shown great promise in facilitating data processing and pattern recognition in medicine, and we believe that the application of machine learning techniques to the problem of EFM interpretation has the potential to improve the predictive accuracy and reliability of intrapartum fetal monitoring. Student role: students will work closely with Dr. La Cava and other lab members to understand EFM data and develop deep learning models, trained on the time series data. Students will assess the efficacy of these methods in comparison to traditional approaches and other machine learning approaches. Prior skills: experience with data science tools and software development, especially Python, is preferred. Students will gain additional experience in code development using open source libraries and perform data visualization and statistical analysis.

Our lab develops machine learning methods that optimize clinical risk prediction models to be fair to patients they are deployed on. In this project, the student will gain experience in this development process and assist in applying these methods to clinical applications. Application areas include: - emergency room admission risk prediction - fair resource allocation for hypertension management An ideal candidate will have some background or experience programming in Python, a solid grasp of statistics and familiarity machine learning. Tasks include data analysis and visualization, contributing to open source software, and running computational experiments on large sets of health records. Students will develop technical skills and a deeper understanding of clinical applications of AI.