Evaluating the efficacy and site-specific performance of machine learning approaches: A comprehensive review of autism detection models

Abstract

As autism diagnoses rise globally, it is important to find a better approach for early and effective prediction. The primary objectives are to identify the models that provide the optimum balance of accuracy while taking age and data type considerations into account, as well as to identify shortcomings and recommend future directions. This study investigates the efficacy of various computational models in early autism detection, analyzing 22 distinct studies. From them, 18 studies are based on 14 popular machine learning (ML) models to identify the most effective prediction methods. And four of them are more progressive, sophisticated methods including the convolutional neural network (CNN) model, diagnostic autism spectrum disorder (DASD) strategy, Ensemble Diagnosis Methodology (EKNN), and Self-Organizing Maps (SOM). Some existing study find out that Gradient Boosting, Extreme Gradient Boosting (XGBoost), DecisionTree (DT), RandomForest (RF), and Light Gradient-Boosting Machine (LGB) demonstrated maximum accuracy scores of 100$\text{\%}$, while AdaBoost (AB), Logistic Regression (LR), Support Vector Machine (SVM), and Random Forest (RF) achieved accuracies of 100$\text{\%}$, 100$\text{\%}$, 96$\text{\%}$, and 96$\text{\%}$, respectively. In contrast to the most recent model, sophisticated CNN obtained 99.39$\text{\%}$ accuracy. For ML models, LR requires less processing time compared to others with high accuracy, making it a suitable choice for efficiency-driven applications, while CNN is optimal for neuroimaging-based autism detection. This study also suggests that the choice of model for autism prediction should be based on specific requirements of accuracy and processing time. This study contributes to the field by providing a comprehensive evaluation of current methodologies, guiding future researchers toward more precise and efficient early autism detection strategies.