Machine Learning Architecture to Improve Predictability of Borderline Cases in Health Applications

Abstract

The practice of medicine has evolved significantly during the past decade, with the emergence of new diagnostic and prognostic tools allowing for a better implementation of “precision medicine”. Among these tools is Machine Learning (ML) that offers the opportunity of personalized patient-tailored care. However, just as humans, ML models still face some difficulty when classifying patients in certain applications where clear-cut boundaries between classes are not easy to identify (e.g., when diagnosing patients with intermediate pretest probability or estimating level of stress of healthcare workers during a pandemic). In this work, we propose an ML architecture to improve the sensitivity of the model to detect patients in intermediate “hard-to-classify” classes and boost the overall performance. This architecture replaces a single classifier by a group of cascaded specialized classifiers that we refer to as: the Human-like Classifier, the Segregating Classifier, and the Deep Classifiers. By doing so, it flags the points that are hard to classify and then develops more specialized models to segregate them. To test its effectiveness, 8 machine learning algorithms were used to predict the feeling of protection among healthcare workers during the COVID-19 pandemic, based on a global online survey, using the traditional and the proposed architectures. The results show, for most algorithms, an enhanced sensitivity for points belonging to intermediate classes, as well as an overall improvement in the models’ accuracies. To validate the results and check for generalizability, the new architecture is tested on a different output of the public health dataset (to predict respondents’ perception of being valued by their community), and on another public dataset (Wine Quality Dataset), and yielded similar results with improved accuracies for most algorithms when compared to the old architecture. This architecture is proving to be a very promising tool to assist physicians in their decision making especially that it is fully automated and does not depend on the algorithm or dataset used.