ADAPT: Provenance-Aware Cross-Dataset Facial Autism Classification with Domain Alignment and Prototype Reasoning

Machine-learning models can perform exceptionally well within a dataset yet fail to generalize when the data source changes, a critical concern in biomedical applications where datasets may differ in acquisition conditions, population composition, and provenance. Facial-image-based autism spectrum disorder (ASD) classification provides a useful test case for such cross-dataset shifts. We introduce ADAPT, an Adversarial Domain-Aligned Prototype Transformer for cross-dataset facial ASD classification. ADAPT combines EfficientNet-B4 and ViT-B/16 representations with domain-adversarial alignment, facial action unit (AU)-guided feature fusion, and a prototype-mediated pathway that maps the fused representation to eight activations before classification. These activations are the sole input to the ASD/typically developing (TD) classifier, providing an eight-dimensional decision bottleneck that can be inspected per image. We evaluate transfer from FADC to a Zenodo target dataset using either labelled FADC data alone (source-only) or additional unlabelled non-test Zenodo images for domain alignment (transductive adaptation), while keeping the test set held out. In the pre-audit transductive setting, ADAPT achieves an apparent target accuracy of 96.09%. A subsequent provenance audit identifies 848 exact cross-dataset image overlaps, including 79 images in the originally defined Zenodo TD test subset. After controlling exact source-target overlap, ADAPT achieves 51.87±1.19% mean Zenodo accuracy, while the best-evaluated variant reaches 55.84±0.69% across three seeds. Source-domain accuracy remains above 99%, revealing a substantial source-target generalization gap. Component analysis shows that AU-guided fusion and prototype-mediated reasoning do not resolve this residual shift. These findings show that unrecognized dataset overlap can substantially inflate estimates of cross-dataset generalization and that provenance auditing is necessary for credible evaluation.