Smartphone Detection of Fetal Movements Using Artificial Intelligence

OBJECTIVE: To evaluate whether machine learning could be used with audio recordings from a smartphone to detect fetal movements that create disruptions of the amniotic fluid environment. METHODS: We conducted a prospective study to simultaneously record fetal movements seen on ultrasound and audio recordings using a smartphone placed on the maternal abdomen and to compare these with maternal perception of fetal movements. Smartphone audio segments were preprocessed to reduce noise, window the signal, and divide them into typed 1-second audio snippets. These were subsequently converted into visual representations of their acoustic features known as Mel-frequency cepstral coefficients (MFCCs). Selected MFCCs were examined to evaluate how feature characteristics vary with gestational age and body mass index (BMI). Fetal movement detected on ultrasonography was considered the gold standard to estimate the accuracy of model prediction applied to tagged audio segments and maternal perception of movement. The area under the receiver operating characteristic curve (AUROC) was used to evaluate the accuracy of the binary classifier to detect the presence or absence of any fetal movement. Macro F1 scores were used to evaluate the accuracy of more refined movements (gross movement, breathing, and hiccups). Isolated trunk and limb movements were marked with a single timestamp, and continuous or repetitive gross fetal movements were annotated with a continuous timestamp spanning the duration of the activity. RESULTS: Overall, 136 participants were included; 30 patients were followed longitudinally, and 106 received only one study visit. Generalized additive models were applied to selected MFCCs and analyzed separately for cohort recordings and fetal movement types. Results revealed nonlinear associations with gestational age (adjusted P