Assessment of machine learning strategies for simplified detection of autism spectrum disorder based on the gut microbiome composition

Abstract

Many studies relating the gut microbiota composition and autism spectrum disorder focus on finding the statistical differences in microbiome composition between neurotypical and autistic subjects. Since microbiota composition involves high-dimensional variables, establishing inferential or causal relationships using only statistical information is complex, hindering advances toward early functional treatment. Complementary machine learning strategies related to the study of autism spectrum disorder are focused on early diagnosis, substituting the expensive screening tests without providing a possible guide to future alternatives to reduce autism spectrum disorder symptoms. Such techniques may offer better outcomes as a direct approach complemented with statistical analysis to optimize patient healthcare based on an early and simplified detection process. This work evaluates several classic machine learning models, including random forests, support vector machines, k-nearest neighbors, Naïve Bayes, and artificial neural network models. The developed models were assessed to identify less-known patterns and their underlying structures to prior published research on the relationship between gut microbiome composition and an autism spectrum disorder. The differences and similarities between the discovered patterns and existing research are discussed to detect a minimal set of strains that may define the presence of autism spectrum disorder. The best-evaluated models were an artificial neural network and a k-nearest neighbor model, reaching an accuracy of 94.7% in the testing partition with only two missed classifications from 38 previously unseen testing samples. These outcomes support the potential of machine learning strategies to construct a useful pre-diagnostic tool for autism spectrum disorder based on relative gut microbiome distribution.