Reliable Detection of Autism Spectrum Disorder in Children Using Conformal Prediction

ICLR 2026 Conference Submission20407 Authors

Keywords: Conformal Prediction, Machine learning, Autism spectrum disorder, Logistic regression, SVM, Random forest

TL;DR: Conformal prediction, a machine learning framework, enables Autism Spectrum Disorder detection with zero false predictions

Abstract: Autism Spectrum Disorder (ASD) is a neurological condition affecting communication and daily functioning. Early intervention can reduce challenges in learning and behavior, motivating the use of machine learning techniques for ASD detection. Although models with high accuracy and F1 scores may appear promising, they can be misleading in low-prevalence settings. By Bayes’ theorem, low prevalence substantially reduces the positive predictive value (PPV), meaning that even models with strong traditional metrics may yield unreliable predictions in practice. False positive ASD detections can lead to unnecessary psychological stress, including anxiety and depression, while false negatives may delay intervention and make treatment more difficult later. In this paper, we integrated conformal prediction into the classification pipeline. Unlike standard classifiers, conformal methods provide prediction sets that include the true label with a specified confidence level ($1-\alpha$), thereby reducing the risk of false predictions. Results show that Conformal predictors occasionally left cases unpredicted, thereby abstaining in situations where reliability could not be guaranteed. Among the evaluated models, SVM achieved the best performance with 86% correct predictions and 14% abstentions, followed by Logistic Regression (84% correct, 16% abstentions). These results demonstrate that conformal prediction offers a more trustworthy approach for ASD screening.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 20407