An explainable deep learning framework for screening level autism severity using behavioural data

Vijayalaxmi N. Rathod

doi:10.1007/s42452-026-08352-3

An explainable deep learning framework for screening level autism severity using behavioural data

2026-03-04

0

OA

AI

V

Vijayalaxmi N. Rathod *

DOI：10.1007/s42452-026-08352-3

Original

Save

View PDF

Abstract

En 中文

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition characterized by heterogeneous behavioural manifestations that vary across age groups and severity strata. While questionnaire-based screening instruments enable large-scale ASD identification, reliable severity stratification remains challenging due to the absence of explicit clinician-annotated severity labels and the reliance of many existing approaches on fragmented or multi-stage pipelines. To address this gap, this study proposes a unified BiLSTM–CGBO framework for screening-level ASD identification and ordinal severity stratification under a controlled behavioural labeling paradigm. The proposed architecture integrates a Bidirectional Long Short-Term Memory (BiLSTM) network, employed as a high-capacity representation learner for modeling non-linear interactions among behavioural attributes, with a Chaotic Grey Ball Optimizer (CGBO) that adaptively modulates learning dynamics to enhance numerical stability and convergence. The framework is evaluated on harmonized ASD screening datasets from the Kaggle and UCI repositories, comprising 2154 participants across four developmental cohorts: toddlers, children, adolescents, and adults. Experimental results demonstrate consistently high classification accuracy and AUROC values across age groups, reflecting the model’s effectiveness in approximating a deterministic behavioural severity-scoring function rather than independent clinical diagnostic performance. Comparative analysis indicates that the proposed framework achieves stable improvements over classical machine learning and deep learning baselines, including SVM, CNN, and standard LSTM, under identical preprocessing and validation protocols. Model interpretability is examined using SHapley Additive Explanations (SHAP), which elucidate how behavioural attributes are utilized within the learned scoring structure without asserting causal or diagnostic attribution. Overall, the proposed BiLSTM–CGBO framework establishes a transparent, reproducible, and computationally efficient methodology for screening-level autism severity modeling, providing a robust foundation for future validation using independently clinician-annotated data and multimodal extensions.

Keywords:

Autism severity classification

Behavioral screening data analysis

BiLSTM-CGBO framework

Explainable AI

An explainable deep learning framework for screening level autism severity using behavioural data

Abstract

AI Summary

Journal

Organization

Cited Papers