A novel approach based on machine learning analysis of flow velocity waveforms to identify unseen abnormalities of the umbilical cord.

Introduction: A Doppler ultrasound (DUS) is essential for detecting blood flow abnormalities in the umbilical cord (UC). Any morphological abnormalities of the UC may lead to morbidity and stillbirth. Some abnormalities such as torsion, strictures and true-knot, however, may only be discovered at birth. This study proposes a novel approach of using machine learning analysis of flow velocity waveforms to improve the diagnosis of UC abnormalities.

Methods: A dynamic in-vitro simulator for DUS and three UC models, each representing a different morphology: true-knot, straight and coiled, were designed. DUS flow field images were captured from four cases of flow through the models: straight, coiled, at mid- and exit of the UC true-knot. The images were transformed into vector profiles of average flow signals that were segmented into 250 flow waves, each comprising 120 samples, for each of the four cases. Three sets of features were extracted from each flow wave and different machine learning algorithms were used for dimensional reduction and binary and multiclass classification.

Results: Significant differences were obtained between flow signals measured at the mid-knot compared to all other cases, which were also reflected in the average high accuracy rates of 97.5%-99.2%. Good accuracy rates of ∼80% and up were also generated, allowing the differentiation between the straight, coiled and exit true-knot.

Discussion: Our dynamic simulator can produce an unlimited database, and combined with the proposed machine learning analysis, may be used as decision support system and increase the ability to diagnose unseen pathologies of the UC.