Vidivelli S, Padmakumari P, Shanthi P
Department of Computer Science and Engineering, School of Computing, SASTRA Deemed to be University, Thanjavur, Tamilnadu, 613402, India.
Department of Computer Science and Engineering, School of Computing, SASTRA Deemed to be University, Thanjavur, Tamilnadu, 613402, India.
Comput Methods Programs Biomed. 2025 Mar;260:108492. doi: 10.1016/j.cmpb.2024.108492. Epub 2024 Dec 18.
Social communication difficulties are a characteristic of autism spectrum disorder (ASD), a neurodevelopmental condition. The earlier method of diagnosing autism largely relied on error-prone behavioral observation of symptoms. More intelligence approaches are in progress to diagnose the disorder, which still demands improvement in prediction accuracy. Furthermore, computer-aided design systems based on machine learning algorithms are extremely time-consuming and difficult to design. This study used deep learning techniques to develop a novel autism detection model in order to overcome these problems.
Preprocessing, Features extraction, Improved Feature level Fusion, and Detection are the phases of the suggested autism detection methodology. First, both input modalities will be preprocessed so they are ready for the next stages to be processed. In this case, the facial picture is preprocessed utilizing the Gabor filtering technique, while the input EEG data is preprocessed through Wiener filtering. Subsequently, features are extracted from the modalities, from the EEG signal data, features like Common Spatial Pattern (CSP), Improved Singular Spectrum Entropy, and correlation dimension, are extracted. From the face image, features like the Improved Active Appearance model, Gray-Level Co-occurrence matrix (GLCM) features and Proposed Shape Local Binary Texture (SLBT), as well are retrieved. Following extraction, enhanced feature-level fusion is performed to fuse the features. Ultimately, the combined features are fed into the hybrid model to complete the diagnosis. Models such as Convolutional Neural Networks (CNN) and Bidirectional Gated Recurrent Units (Bi-GRU) are part of the hybrid model.
The suggested MADDHM model achieved an accuracy of about 91.03 % regarding EEG and 91.67 % regarding face analysis meanwhile, SVM=87.49 %, DNN=88.59 %, Bi-GRU=90.02 %, LSTM=87.49 % and CNN=82.02 %.
As a result, the suggested methodology provides encouraging outcomes and opens up possibilities for early autism detection. The development of such models is not only a technical achievement but also a step forward in providing timely interventions for individuals with ASD.
社交沟通困难是神经发育障碍孤独症谱系障碍（ASD）的一个特征。早期诊断孤独症的方法很大程度上依赖于容易出错的症状行为观察。目前正在研发更智能的方法来诊断该疾病，但其预测准确性仍有待提高。此外，基于机器学习算法的计算机辅助设计系统极其耗时且设计困难。本研究使用深度学习技术开发了一种新型孤独症检测模型，以克服这些问题。
所提出的孤独症检测方法包括预处理、特征提取、改进的特征级融合和检测阶段。首先，对两种输入模态进行预处理，使其为后续阶段的处理做好准备。在这种情况下，面部图片利用伽柏滤波技术进行预处理，而输入的脑电图（EEG）数据通过维纳滤波进行预处理。随后，从这些模态中提取特征，从EEG信号数据中提取诸如共同空间模式（CSP）、改进的奇异谱熵和关联维数等特征。从面部图像中也提取诸如改进的主动外观模型、灰度共生矩阵（GLCM）特征和提议的形状局部二值纹理（SLBT）等特征。提取后，进行增强的特征级融合以融合这些特征。最终，将组合后的特征输入到混合模型中完成诊断。卷积神经网络（CNN）和双向门控循环单元（Bi - GRU）等模型是混合模型的一部分。
所提出的MADDHM模型在EEG方面的准确率约为91.03%，在面部分析方面的准确率为91.67%，而支持向量机（SVM）为87.49%，深度神经网络（DNN）为88.59%，Bi - GRU为90.02%，长短期记忆网络（LSTM）为87.49%，CNN为82.02%。
因此，所提出的方法提供了令人鼓舞的结果，并为孤独症的早期检测开辟了可能性。此类模型的开发不仅是一项技术成就，也是朝着为ASD患者提供及时干预迈出的一步。
