Automatic Blind Sweep Laterality Detection

Blind sweep obstetric screening protocols enable data collection by novice users for deep learning-assisted fetal parameter estimation. Clinical reliability of the estimated parameters hinges on the protocol being followed strictly, such as acquiring the grid of blind sweeps in a consistent order and direction. This study tests whether a machine learning model can classify sweep laterality of cranio-caudal sweeps in a blind sweep protocol, using acoustic decoupling profiles, to provide assurance that sweep positioning is correct. Data from 22 patients were collected, with sweeps labeled as right, medial, or left. Decoupling masks were generated via a local peak detection algorithm. Candidate machine learning models were evaluated for classification of all sweep positions, and lateral sweep positions only. Sweep classification performance was greatest when predicting left/right laterality alone. These findings support this decoupling profile-based classification approach for detecting incorrect sweep placement by novice users. Future work will seek to improve performance on all sweep labels by increasing sample size, feature engineering, and exploring alternate model architectures.