Request for Clarification and Correction of Internal Inconsistencies in “Neuromimicnet: A Multimodal Neuromorphic Framework for Early Screening and Cognitive Pattern Recognition in Children with Autism” [Letter]

Dear editor

I write to raise concerns regarding several internal inconsistencies in the above-referenced article, published in the Journal of Multidisciplinary Healthcare (Volume 19, Article 605039, published 14 August 2026).¹ These inconsistencies affect the reproducibility and interpretability of the reported results and, in my view, warrant clarification or correction by the authors. I outline the principal concerns below.

Contradiction Regarding Use of EEG Data

The Abstract states that “Performance is evaluated using accuracy, F1-score, precision, recall, and computational efficiency across benchmark EEG-autism datasets.” However, the Dataset Description section identifies the sole dataset used as the “Multimodal Autism Dataset,” comprising only audio recordings and facial expression images, with no EEG component described anywhere in the Methods, Preprocessing, Feature Extraction, or Results sections. Furthermore, the Conclusion explicitly lists EEG as a direction for future work: “Future work will focus. on the integration of more modalities, including EEG or eye-tracking signals, to enhance the diagnostic accuracy further.” This directly contradicts the Abstract’s claim that EEG datasets were already used for evaluation. I would ask the authors to clarify which claim is accurate and to correct the Abstract accordingly if EEG data were not, in fact, used.

Discrepancy in Reported Sample Sizes Across Confusion Matrices

The Dataset Description states that the dataset contains “approximately 3000+ samples, including nearly 1500 facial images and 1500 audio instances.” However, the row totals in the confusion matrices reported in Figure 8 imply substantially different total dataset sizes depending on the split reported:

• ●70:30 split (Figure 8A): row totals sum to 987 test samples. If this represents 30% of the dataset, the implied total is approximately 3290 samples.
• ●80:20 split (Figure 8B): row totals sum to 878 test samples. If this represents 20% of the dataset, the implied total is approximately 4390 samples.

These two implied totals differ by roughly 1100 samples (approximately 33%), despite both splits purportedly being drawn from the same underlying dataset. Neither figure reconciles cleanly with the “3000+” total stated in the Dataset Description, and the 80:20 figure in particular substantially exceeds it. I would ask the authors to clarify the actual number of samples used in each experimental split and to reconcile this with the stated dataset size.

Unexplained Conversion from Binary to Four-Class Labels

The Dataset Description states that the dataset “is organized into two diagnostic categories, ASD and Non-ASD.” However, the entire classification pipeline, including all reported confusion matrices, performance metrics, and the ablation study, is structured around four severity classes: Typical, Mild, Moderate, and Severe. The manuscript does not describe how the original binary ASD/Non-ASD ground-truth labels were converted or re-annotated into four graded severity categories. Given that this labeling scheme underlies every reported result, I would ask the authors to describe the annotation or labeling procedure used to derive the four severity classes, including whether this was performed by clinical raters, and if so, using what criteria and with what inter-rater reliability.

Inconsistent Naming of the Facial Feature Extraction Backbone

The Abstract, Introduction, and Proposed Methodology sections consistently refer to “ResNet50” as the pretrained CNN used for facial feature extraction. However, the subsection “Facial Feature Extraction Using RestNet50” and the subsequent Feature Fusion and Classification section refer at two points to “restNet15”/“ResNet15.” As ResNet-15 and ResNet-50 are architecturally distinct networks with different depths and parameter counts, this inconsistency should be corrected to confirm which architecture was actually implemented.

Characterization of ResNet50 as “Lightweight”

The Facial Preprocessing and Facial Feature Extraction sections describe the pretrained CNN as “a powerful face recognition and alignment algorithm” (MTCNN) and, separately, as “a pretrained lightweight CNN” (in reference to ResNet50/ResNet15). ResNet50 (approximately 25 million parameters, 50 layers) is not conventionally described as lightweight relative to architectures purpose-built for efficiency (eg, MobileNet, SqueezeNet). The authors may wish to revise this characterization or clarify the basis for it.

Scope Mismatch in Table 1

Table 1 is titled “Summary of Existing Neuromorphic and SNN Approaches in Healthcare.” However, two of the six entries in this table—Oh et al and Ahmed et al—are described in the table itself as using “Conventional Deep Learning” and “Deep Learning Models,” respectively, with no neuromorphic or spiking-neural-network component indicated. Their inclusion in a table specifically scoped to neuromorphic and SNN approaches appears inconsistent with the table’s stated purpose and may confuse readers assessing the novelty of the proposed approach relative to prior neuromorphic work.

Conclusion

Given that NeuroMimicNet is proposed as a clinical screening framework for a pediatric population, precision in reporting dataset composition, sample sizes, labeling methodology, and architectural details is particularly important for reproducibility and for any subsequent clinical validation efforts. I respectfully request that the editorial office invite the authors to respond to these points, and that a correction or clarification be published as appropriate following review of their response.

Disclosure

The author declares that there are no competing interests in this communication.