Autism Spectrum Disorder (ASD) is a neurodevelopmental condition that impacts social interaction and behavior, where early detection is crucial for optimizing intervention outcomes. This study proposes a machine learning-based screening tool for Autism Spectrum Disorder (ASD) using psychological electronic medical records (EMR) for children aged 1–8 years. Two datasets were utilized: (1) a public dataset with approximately 300 records and (2) a clinical EMR dataset with approximately 600 records. Both were labeled by psychology experts based on 18 behavioral evaluation criteria, including eye contact, pointing, and imitation. Ten significant features were selected as inputs for training four machine learning models: Decision Tree, XGBoost, CatBoost, and Overall Local Accuracy (OLA). OLA demonstrated superior adaptability but faced challenges with class imbalance (90% ASD) and feature bias due to high-accuracy EMR data. To address these, SMOTE-IPF was applied to balance the data, and OLA was enhanced with dynamic weighting. The proposed system shows high feasibility for ASD screening.