A machine learning model for assessing fetal health during pregnancy.

There is a global imperative to end stillbirths, particularly in low middle income countries (LMICs), which suffer from disproportionate incidence. Sudden changes in fetal movement (FM) patterns often precede a crisis, which, if flagged, can trigger life-saving intervention. Existing means of FM tracking, however, are based on outdated understanding which have remained unchanged for decades. The current standard for monitoring FM out-of-clinic remains maternal perception, which suffers from subjectivity and has little impact in reducing stillbirth or poor perinatal outcomes. Ultrasound can trace FM and trigger intervention but is administered sporadically over the course of pregnancy and demands clinician expertise and resources; frequent use is not feasible, particularly in LMICs. Wearable FM monitors have been proposed for FM empirical movement monitoring, however, clinical impact remains negligible due to homogeneous sensing modalities and lack of clinical validation. Herein, a multimodal vibrational-acoustic sensor array consisting of piezoelectric and acoustic sensing modalities is validated for use in a wearable FM. 25 pregnant participants were recruited to record vibrational-acoustic data from the array in parallel with ultrasound scanning. Categorised fetal movements were recorded by a clinician, and several machine learning models were investigated to validate the sensors to track FM. An ensemble RUSBoost model combined with concatenated sensor data inputs was implemented, yielding FM prediction with precision and recall of 0.44 and 0.61, demonstrating the feasibility of the vibroacoustic sensor array to monitor FM. Inexpensive off-the-shelf sensors comprising the array provide a basis for the development of a fully wearable FM monitor that can be used in LMICs.