Autism detection based on multiple time scale model.

Area of Science:

• Neuroscience
• Medical Imaging
• Machine Learning

Background:

• Current autism detection methods are subjective and prone to misdiagnosis.
• Existing functional magnetic resonance imaging (fMRI) analyses for autism often oversimplify temporal information, limiting generalization.
• There is a need for objective, data-driven approaches to autism detection using neuroimaging.

Purpose of the Study:

• To develop an objective autism detection method by extracting multi-time scale brain features from fMRI data.
• To improve the accuracy and generalization ability of autism detection compared to existing models.
• To provide a novel framework for analyzing fMRI sequences in autism research.

Main Methods:

• Feature extraction using least absolute shrinkage and selection operator (LASSO) for a time scale of 1.
• Dynamic brain feature extraction using Hidden Markov Models (HMM) for a time scale of 2.
• High time-scale feature extraction via a Long Short-Term Memory (LSTM) auto-encoder, followed by recursive feature elimination and 1D Convolutional Neural Network (CNN) classification.

Main Results:

• The proposed multi-time scale fMRI feature analysis achieved an accuracy of 76.0%.
• The model demonstrated strong generalization ability, with an Area Under the ROC Curve (AUC) of 0.83 on independent test data.
• The method outperformed established models in autism detection accuracy and robustness.

Conclusions:

• Multi-time scale analysis of fMRI data offers a promising avenue for objective autism detection.
• The developed framework provides a novel approach for analyzing complex neuroimaging data in autism research.
• This study highlights the potential of advanced machine learning techniques for improving diagnostic accuracy in neurological disorders.