Classification of Fetal State through the application of Machine Learning techniques on Cardiotocography records: Towards Real World Application

Abstract

AbstractObjectiveAnticipating fetal risk is a major factor in reducing child and maternal mortality and suffering. In this context cardiotocography (CTG) is a low cost, well established procedure that has been around for decades, despite lacking consensus regarding its impact on outcomes.Machine learning emerged as an option for automatic classification of CTG records, as previous studies showed expert level results, but often came at the price of reduced generalization potential.With that in mind, the present study sought to improve statistical rigor of evaluation towards real world application.Materials and MethodsIn this study, a dataset of 2126 CTG recordings labeled as normal, suspect or pathological by the consensus of three expert obstetricians was used to create a baseline random forest model.This was followed by creating a lightgbm model tuned using gaussian process regression and post processed using cross validation ensembling.Performance was assessed using the area under the precision-recall curve (AUPRC) metric over 100 experiment executions, each using a testing set comprised of 30% of data stratified by the class label.ResultsThe best model was a cross validation ensemble of lightgbm models that yielded 95.82% AUPRC.ConclusionsThe model is shown to produce consistent expert level performance at a less than negligible cost. At an estimated 0.78 USD per million predictions the model can generate value in settings with CTG qualified personnel and all the more in their absence.