The authors compiled ASD-FER, a pediatric facial image benchmark containing 2,779 images of children with Autism Spectrum Disorder. Three independent annotators labeled each image into six discrete emotion categories; labels were finalized using majority voting across annotators. The preprint states this dataset was created to fill an identified gap: most existing FER corpora are not ASD‑specific and lack emotion‑level annotations tailored to pediatric autistic populations. The authors provided a GitHub link for data and code, and they emphasize that the manuscript is a preprint not yet peer reviewed.
Nine models were evaluated against the ASD‑FER benchmark. Traditional and lightweight architectures included CNN, SVM, AlexNet, ResNet34, ResNet50, and MobileNetV2. The authors also tested three backbone configurations of a proposed task‑specific architecture termed EQFormer: ViT, Swin‑T, and a compact Hybrid CNN‑Transformer (HybridT). Evaluations used stratified 10‑fold cross‑validation and a held‑out 90/10 split; the paper reports cross‑validation means and held‑out metrics without further methodological details beyond these splits.
Across the stratified 10‑fold cross‑validation runs, ResNet50 led the conventional models with a mean accuracy of 84.89%. Transformer backbones outperformed the traditional baselines: ViT achieved 88.27% mean accuracy, Swin‑T reached 88.02%, and the HybridT configuration scored 86.58%. The authors reported these figures as mean cross‑validation accuracies; any additional metrics or statistical tests beyond those numbers were not described in the abstract.
Performance on the held‑out 90/10 split shifted leaderboards slightly. The ViT backbone produced the highest overall accuracy on the held‑out set at 91.01%. Swin‑T did not top overall accuracy but obtained the best balanced accuracy (74.47%) and the highest macro‑F1 score (0.7078). The HybridT remained competitive—89.57% accuracy on the held‑out set—while operating with a substantially reduced parameter count. The preprint specifies HybridT parameterization at 2.66M, presenting it as a parameter‑efficient alternative to larger transformer backbones.
These reported held‑out metrics indicate different tradeoffs: ViT maximized raw accuracy on the reserved sample, whereas Swin‑T produced more balanced class performance as reflected by the balanced accuracy and macro‑F1. The authors noted class imbalance in the dataset, which can disproportionately affect overall accuracy and make balanced metrics more informative for multi‑class FER in ASD.
The study team applied Grad‑CAM visualizations to inspect where models focused when predicting emotions. Grad‑CAM maps showed consistent attention on facial regions traditionally associated with emotion signaling: the eyes, eyebrows, and mouth. Swin‑T produced the most spatially concentrated attention maps among the transformer variants, according to the authors. The compact HybridT produced explainability maps as well but was emphasized for being parameter‑efficient rather than for superior localization. The preprint does not provide additional quantitative explainability metrics in the abstract; the description is limited to these qualitative observations.
The authors identified two remaining challenges. First, class imbalance within ASD‑FER persists and may bias both training and evaluation; the preprint signals this as an unresolved dataset limitation. Second, cross‑dataset generalization remains an open problem for FER in ASD: models trained on ASD‑specific images may not transfer robustly to other FER corpora or clinical settings. The paper lists these issues directly and frames them as directions for follow‑up work rather than problems fully addressed in the current report.
Limitations and unanswered questions reported in the preprint include the dataset composition imbalance and generalization constraints. The authors made the data and code available via GitHub (link in the preprint), enabling external validation and replication; however, the manuscript itself is an unreviewed preprint, and no peer‑review validations or external replication studies are reported in the source. The preprint does not report additional demographic breakdowns, detailed annotation agreement statistics, per‑class performance figures beyond the summary metrics cited, or the exact training hyperparameters used for each model in the abstract. These items remain unspecified in the source and represent concrete information gaps for readers and future reviewers.