Identifying Diagnostic Biomarkers for Autism Spectrum Disorder using the PED Algorithm

In the field of neuroinformatics, research on Autism Spectrum Disorder (ASD) predominantly focuses on the bidirectional connectivity between brain regions, with fewer studies addressing higher-order interaction anomalies among brain regions. To explore the complex relationships among brain regions, the authors adopted the Partial Entropy Decomposition (PED) algorithm to capture higher-order interactions by computing the high-order dependence of triads (three brain regions). This paper proposes a method based on PED and surrogate test methods to examine the influence of individual brain regions on triads, identifying critical triads. Further, the hypergraph modularity optimization algorithm revealed higher-order brain structures, showing that connectivity between the right thalamus and left thalamus was looser in ASD compared to typical controls (TC). The interaction of key redundant triads (left cerebellum, left precuneus, and right inferior occipital gyrus) showed significant attenuation, while the interaction of key synergistic triads (right cerebellum, left postcentral gyrus, and left lingual gyrus) significantly declined. The results of the classification model further confirmed the potential of critical triads as diagnostic biomarkers.

Research Background and Issues

Autism Spectrum Disorder (ASD) is a neurodevelopmental, non-focal brain disorder characterized mainly by social communication impairments, restricted interests, and repetitive behaviors. ASD diagnosis typically relies on behavioral observations, clinical interviews, and questionnaires, which can sometimes lead to misdiagnosis. Finding objective biomarkers for diagnosing ASD has become crucial. In elucidating the neural changes in ASD, it is considered a connectome dysfunction syndrome, manifesting as abnormal intrinsic functional connectivity in the brain. Functional connectivity, described as the correlation between two brain regions, is widely used in ASD-related research. Additionally, another type of brain connectivity called effective connectivity, with common methods such as Granger causality and conditional entropy, helps researchers understand the information transmission paths between different brain regions. However, both effective and functional connectivity only describe the influence between pairs of brain regions. Given the complex relationships within brain regions, analyzing the higher-order interactions among multiple brain regions is indispensable.

Paper Source

This paper was authored by Hao Wang, Yanting Liu, and Yanrui Ding from Jiangnan University and was accepted by Springer Science+Business Media, LLC, part of Springer Nature on March 23, 2024, for publication in the journal “Neuroinformatics”.

Research Methodology and Workflow

The research mainly comprises the following steps: 1. Dataset: The study used resting-state functional magnetic resonance imaging (rs-fMRI) data of ASD and TC from the NYU Langone Medical Center, from the first phase of the Autism Brain Imaging Data Exchange (ABIDE) project. 2. Data Preprocessing: BOLD signals of the brain were preprocessed with slice timing correction, intensity normalization, image realignment, etc. 3. Partial Entropy Decomposition (PED) and High-order Dependence Measurement: Using the PED algorithm, the redundancy and synergy of all triads were calculated for ASD and TC participants, and critical triads differentiating ASD from TC were identified. 4. Classification Model Construction: The redundancy and synergy information of critical triads were used as features to build a Support Vector Machine (SVM) classifier, optimizing parameters using grid search and evaluating the classification robustness with ten-fold cross-validation.

Dataset and Preprocessing

The study used rs-fMRI data from the ABIDE I project, anonymized and not containing protected health information as per HIPAA guidelines. The dataset was screened by three anthropologists, excluding subjects with incomplete brain coverage, peak motion, ghosting, and scanner artifacts. A total of 172 participants were finally selected, including 74 ASD and 98 TCs. Brain regions were defined using the Craddock 200 (CC200) atlas, which divides the brain into 200 distinct areas, and preprocessing was done with the Configurable Pipeline for the Analysis of Connectomes (C-PAC).

Partial Entropy Decomposition (PED) and High-order Dependence Measurement

The PED algorithm was used to compute the redundancy and synergy information of all triads for ASD and TC subjects. Surrogate data tests evaluated the importance of each brain region in the triads. When the time series of a brain region was replaced with random data, the high-order dependence measurements were recalculated to determine if they significantly differed from the original values, identifying critical triads.

Classification Model based on PED

Constructing the Classification Model: - Extracting redundancy and synergy information of key triads as feature vectors and inputting them into an SVM classifier for training. - Five different methods were used to divide the data into training and testing sets, and ten-fold cross-validation was utilized to assess the robustness of classification results. Classification quality was measured by accuracy, sensitivity, specificity, F1 score, and area under the curve.

Research Results

Intergroup Differences in Redundancy and Synergy Structures

Using the hypergraph modularity maximization algorithm, redundancy and synergy structures were generated for each participant, and association analyses were conducted with the classical seven-network Yeo system. Results showed significant differences in redundancy and synergy information for specific brain region pairs between ASD and TC participants, particularly in areas such as the left cerebellum and right inferior occipital gyrus.

Intergroup Differences in High-order Dependence Measurements

Two-sample t-tests revealed increased redundancy interactions in the 25-183-190 triad for ASD, and decreased redundancy and synergy interactions in multiple triads like 18-31-42 and 20-62-102.

Intergroup Differences in Key Triad Patterns

Key triad pattern changes between ASD and TC were notably observed, especially in critical triads like 18-108-150 and 25-183-190, where the pattern (0,0,0) was more common in ASD, indicating these brain regions exhibit common states in ASD.

Performance of the ASD-TC Classification Model

When using the redundancy information of key triads as features, the accuracy of the classification model was 85%, and using synergy information as features, the accuracy was 80%. The overall accuracy reached 83% when both redundancy and synergy information were used as features, with the highest accuracy of 97% under different training/testing set splits.

Conclusion and Research Significance

This paper proposes a method combining PED with surrogate tests, addressing the limitations of PED in studying single brain regions. The findings indicate changes in triad interactions or information transmission paths in ASD, stemming from abnormal states of individual brain regions. The validated classification model shows that key triads have potential diagnostic biomarker value, providing critical evidence for ASD identification. These results not only deepen the understanding of the brain functional organization and cognitive-behavioral changes in ASD but also provide important directions for subsequent research.