A Reliability-Oriented Hybrid Deep Learning Framework for Early Autism Screening

It proposes and formalizes a reliability-oriented parallel screening architecture in which two visual subsystems each output an independent posterior probability and a referral is triggered if either exceeds threshold, with system-level false-negative probability modelled as the product of the two subsystem false-negative probabilities. Unlike prior autism AI screening work focused on single-model accuracy or AUC, this work shifts the optimization objective from maximizing individual classifier accuracy toward minimizing the probability of simultaneous missed detection, and gives an explicit mathematical formulation of parallel fusion. The paper provides a full reliability derivation covering parallel sensitivity, parallel specificity, a correlated-error extension, and an ROC-space mapping, then substitutes measured subsystem sensitivities (0.9 and 0.871).

The facial-image pathway uses a weighted probabilistic average of ResNet50, EfficientNet-B0 and DenseNet121, with weights empirically set to 0.38, 0.31 and 0.31 based on individual AUC; the ensemble reached an AUC of 0.948, accuracy of 88.3% and F1 of 0.881, above the standalone AUCs of 0.919, 0.883 and 0.907. This indicates that architectural diversity yields complementary representations that improve discriminative performance and stability on the binary facial classification task, rather than relying on tuning a single model. Results come from a public Zenodo facial dataset (2,000 images per class in training, 500 per class in validation and test, 5,000 total), with ensemble results averaged over three independent training runs and reported with confusion matrix and ROC curve.

The static periocular pathway fine-tunes an ImageNet-pretrained ResNet18 into a four-class softmax (mild, low, medium, high), achieving 91.9% average accuracy, 90% sensitivity and a macro-average F1 of 0.90 under stratified 10-fold cross-validation, with Grad-CAM highlighting iris, sclera and eyelid contour regions. This pathway explicitly frames periocular images as static spatial phenotype data rather than fixation dynamics or scan-path time series, avoiding misreading static images as oculomotor behavioural evidence. The dataset contains roughly 1,000 224x224 colour images per class, 4,000 balanced samples in total, evaluated with image-level stratified 10-fold cross-validation and narrow dispersion of fold accuracies.

Parallel fusion produces the classic sensitivity-specificity trade-off: analytically estimated system-level sensitivity of 0.9871 while system-level specificity, computed as the product of the two subsystem specificities, drops to about 0.807, so the framework is positioned as a pre-diagnostic triage and referral-support tool rather than a diagnostic substitute. The work moves screening evaluation from isolated classification benchmarking toward system-level error-probability modelling, and discusses how high sensitivity raises negative predictive value at low prevalence while positive predictive value remains prevalence-dependent. Specificity is obtained by multiplying cross-validated estimates of 0.912 and 0.886; the paper also gives a correction formula for correlated errors, noting that positive correlation weakens the redundancy gain.

The framework targets pre-diagnostic risk indication and referral support, intended for settings such as primary care, school health, community health and telehealth where a static periocular image and a facial photograph can be captured; its output decides whether referral to a paediatric neurologist, developmental psychologist or other specialist is warranted, while final diagnosis remains with qualified clinicians. The paper states clearly that the two subsystems were trained and evaluated on two separate public Zenodo datasets rather than a unified cohort containing both periocular and facial samples from the same individuals, so the analytical fusion result should be read as a reliability model rather than a prospectively validated multimodal patient-level estimate.

Several open questions remain for a careful reader: the actual degree of conditional correlation between subsystem false-negative events is not quantified on data, and positive correlation would weaken the redundancy gain; reliable subject-level identifiers were absent from the public metadata, so image-level stratified folds cannot fully exclude the same child's images appearing across folds; the paper notes the datasets may not represent the full diversity of age, sex, ethnicity, camera conditions and clinical severity in real screening; the drop in system-level specificity implies higher referral volume, greater false-positive burden and more downstream clinical evaluation demand, whose operational impact is not yet evaluated; and this loaded text is a fast parse in which the confusion matrices and fold distributions of Figures 2 to 5 appear as textual descriptions, so checking exact value distributions still requires the original figures.