Autism Spectrum Disorder (ASD) is a neurological condition that affects social and cognitive skills, and early detection is crucial for minimizing long-term effects. This study aims to develop a decentralized diagnostic framework for precise ASD detection, utilizing a federated learning simulation with raw data held at the client-side, but no specific privacy safety (DP, Secure Aggregation, Homomorphic Encryption). It integrates Ensemble Stacking with Deep Learning approaches, combining classifiers such as Decision Tree, Random Forest, XG Boost, and Logistic Regression (meta), and compares their effectiveness to deep learning models. The research used a Kaggle-based ASD dataset and included preprocessing steps such as missing value imputation, one-hot encoding, PCA for dimensionality reduction, oversampling for class balancing, and feature standardization. Two. models were implemented: a Federated Ensemble Model combining Decision Tree, Random Forest, and XG Boost in a stacked framework and a Federated Deep Learning Model for binary classification. Both models operated in a federated learning setup, with data distributed across two clients and aggregated via federated averaging. Metrics like accuracy, precision, recall, and F1-score were used for evaluation. The federated ensemble model achieved 96.0% accuracy with precision, recall, and F1-scores of 0.97 and 0.92. The federated deep learning model outperformed it with 97.0% accuracy, higher precision (0.98), and improved recall (0.94), highlighting better balance in sensitivity and precision for ASD detection in decentralized environments. Both federated models demonstrated high diagnostic accuracy while maintaining data privacy. The federated deep learning model was superior and is suited for clinical applications. These are initial results obtained and are only using Kaggle ASD screening dataset. Generalizability will be tested by using external validation of larger and clinically diverse datasets (e.g., ABIDE-I/II).
Das et al. (Mon,) studied this question.