Predicted-gender information in ASD facial-image classification: multitask performance, shuffled-label controls, and representation analysis

Autism spectrum disorder (ASD) facial-image classifiers may benefit from an auxiliary gender-prediction task, but an improvement alone does not establish that meaningful gender information is responsible. This study tested that distinction using a public ASD-labeled facial-image dataset and ImageNet-pretrained EfficientNet-B0 networks. An exploratory comparison found no significant advantage from training separate predicted-male and predicted-female classifiers rather than one general classifier (accuracy: 79.07% versus 82.56%; p = 0.3075). A subsequent comparison across 40 matched random seeds on a cleaned 168-image test set found that adding a predicted-gender auxiliary task increased mean accuracy from 77.3% to 79.4% and specificity from 59.0% to 63.5%, while sensitivity remained nearly unchanged. Crucially, an otherwise matched model trained with shuffled gender labels also improved performance over baseline and did not differ significantly from the genuine-label model on any evaluated metric. Thus, meaningful predicted-gender supervision was not shown to explain the observed gain. A separate exploratory analysis removed five-dimensional attribute-associated subspaces from the learned image representations. Predicted-gender subspace removal reduced ASD classification AUC more than removal of comparison-attribute subspaces on average in a pooled analysis, but this association did not establish the cause of the multitask improvement. Together, these findings distinguish the benefit of adding a training task from evidence that its intended information produces that benefit, and distinguish both from information encoded in the learned representation. Results are limited to this dataset and do not establish clinical diagnostic validity.