A Stacking Approach to Enhance K-Nearest Neighbors Performance for Autism Screening

Abstract

The increasing prevalence of autism spectrum disorders necessitates improved early screening methods for children to ensure timely intervention and support. While existing screening techniques play a vital role, they often face challenges regarding accuracy, accessibility, and scalability. This research addresses these gaps by enhancing the K-Nearest Neighbors (K-NN) algorithm by implementing a stacking model that integrates multiple distance metrics—Manhattan and Minkowski—to improve predictive performance. Utilizing a public dataset, the study employed K-Fold Cross-Validation with K=5 to ensure a robust evaluation of the models. The results demonstrated that the stacking model achieved an average accuracy of 86.67%, significantly surpassing the traditional K-NN approaches, which reported accuracies of 82.67% for Manhattan and 81.33% for Minkowski. A user-friendly web interface was also developed to facilitate real-world application, allowing users to input data and receive immediate predictive outcomes regarding autism risk. These findings confirm the effectiveness of the stacking method in enhancing K-NN performance and highlight its potential for practical use in autism screening. Future research may explore alternative machine learning algorithms and additional features to refine the predictive capabilities and user experience further.