Gaze-anchored cloud–edge facial behavioral learning for leakage-safe autism screening of students in educational environments

Early identification of autism spectrum disorder (ASD) in educational settings requires screening systems that are accurate, privacy-aware, interpretable, and computationally feasible under resource constraints. Here, we present a gaze-anchored cloud–edge framework for leakage-safe ASD screening support in educational settings. Two complementary but unpaired sources are used: the Autistic Children Face Dataset (ACFD), comprising 187 children, and an eye-tracking gaze-reference dataset comprising 59 participants. To prevent invalid sample-level multimodal fusion, dataset-specific encoders map facial/appearance and gaze–attention observations into modality-masked behavioral tokens and align them within a shared gaze–attention latent space. An edge-oriented pathway extracts compact local descriptors, while cloud learning performs label-blind cross-source alignment, uncertainty-aware multi-instance classification, probability calibration, and teacher–student optimization. Evaluation was conducted strictly at the child or participant level to prevent leakage across repeated frames, segments, temporal windows, or gaze rows. Under leave-one-child-out cross-validation on ACFD, the full framework achieved 95.19% accuracy, 95.65% sensitivity, 94.74% specificity, 95.14% F1-score, and an AUC of 0.9811. Participant-level calibration yielded an ECE of approximately 0.027 and a Brier score of approximately 0.035. Label-held-out leave-one-participant-out cross-source evaluation on the eye-tracking cohort achieved 86.44% accuracy, indicating useful discrimination under modality and protocol shift without constituting fully source-unseen external validation. The edge-oriented student model retained 94.64% accuracy while reducing parameters by 72.95%, peak memory by 64.08%, and inference latency by 63.59%. Interpretability analysis identified gaze concentration proxy, facial appearing time, and response latency as dominant behavioral cues. Overall, the framework provides a privacy-aware, interpretable, and computationally efficient screening-support pathway intended to assist professional referral rather than replace clinical diagnosis.