Explainable Artificial Intelligence to explore dynamic functional connectivity related to subjective
Poster No:
2038
Submission Type:
Abstract Submission
Authors:
Yi-Chia Wei1, Yi-Chia Kung2, Ching-Po Lin3, Kuan-Fu Chen1
Institutions:
1Chang Gung University and Chang Gung Memorial Hospital, Taoyuan & Keelung, Taiwan, 2Taiwan Society of Interventional Radiology, Taipei, Taiwan, 3Institute of Neuroscience, National Yang Ming Chiao Tung University, Taipei, Taiwan
First Author:
Co-Author(s):
Introduction:
Resting-state brain functional connectivity dynamics can reveal intrinsic brain activity in older adults. Dynamic features are more sensitive than static features in reflecting early changes, and both local and long-range dynamic connectivity relate to cognition. However, interpreting dynamic metrics of rs-fMRI connectivity features is challenging. Explainable artificial intelligence (XAI) (Meacham, Isaac et al. 2019) can clarify dynamic connectivity for a better understanding of underlying changes in subjective and objective cognition.
Methods:
This study enrolled 85 healthy individuals over 50 years old with normal cognitive function from the Northeastern Taiwan Community Medicine Research Cohort. The AD8 questionnaire (Galvin, Roe et al. 2005) and MoCA were used to evaluate subjective and structured global cognition, respectively. To be considered normal, the MoCA score needed to be above one standard deviation below age- and education-adjusted mean (Rossetti, Lacritz et al. 2011).
The DynamicBC toolbox (Liao, Wu et al. 2014) was used to generate the dynamic metrics of rs-fMRI by a sliding-window approach (Kung, Li et al. 2019). Greene's brain atlas was used to define 300 regions of interest (ROIs) that were classified into 14 predefined resting-state networks. Local dynamic connectivity of ROIs was measured by the mean dynamic amplitude of low-frequency fluctuation (mdALFF) (Liao, Li et al. 2019). Whole brain-based statistics were applied to test the dynamic functional connectivity (dFC) between ROIs.
The dynamic metrics (mean, variance, and coefficient of variation (CV)) of the 300 nodes and 45000 (i.e. 300x300/2) edges were ranked by the Boruta in R software (Kursa and Rudnicki 2010). The recommended features were used in the machine learning models to predict the score of AD8 and MoCA. Three machine learning models were used to test the model performance. They were Multilayer Perceptron (MLP) (Rosenblatt 1958), XGBoost (Chen and Guestrin 2016), and LightGBM (Ke, Meng et al. 2017). Grid search was used for fine-tuning model hyperparameters. During the model training and testing, a five-fold cross-validation technique was utilized.
The SHapley Additive exPlanations (SHAP) values of the selected dynamic metrics were yielded from the best-performed model (Lundberg and Lee 2017). The k-mean cluster analysis clustered the SHAP values of selected dynamic metrics features into three groups (Ikotun, Ezugwu et al. 2023). The contribution of dynamic metrics of nodes and edges to subjective and objective cognition is explained by the results of clustering.
The DynamicBC toolbox (Liao, Wu et al. 2014) was used to generate the dynamic metrics of rs-fMRI by a sliding-window approach (Kung, Li et al. 2019). Greene's brain atlas was used to define 300 regions of interest (ROIs) that were classified into 14 predefined resting-state networks. Local dynamic connectivity of ROIs was measured by the mean dynamic amplitude of low-frequency fluctuation (mdALFF) (Liao, Li et al. 2019). Whole brain-based statistics were applied to test the dynamic functional connectivity (dFC) between ROIs.
The dynamic metrics (mean, variance, and coefficient of variation (CV)) of the 300 nodes and 45000 (i.e. 300x300/2) edges were ranked by the Boruta in R software (Kursa and Rudnicki 2010). The recommended features were used in the machine learning models to predict the score of AD8 and MoCA. Three machine learning models were used to test the model performance. They were Multilayer Perceptron (MLP) (Rosenblatt 1958), XGBoost (Chen and Guestrin 2016), and LightGBM (Ke, Meng et al. 2017). Grid search was used for fine-tuning model hyperparameters. During the model training and testing, a five-fold cross-validation technique was utilized.
The SHapley Additive exPlanations (SHAP) values of the selected dynamic metrics were yielded from the best-performed model (Lundberg and Lee 2017). The k-mean cluster analysis clustered the SHAP values of selected dynamic metrics features into three groups (Ikotun, Ezugwu et al. 2023). The contribution of dynamic metrics of nodes and edges to subjective and objective cognition is explained by the results of clustering.
Results:
XGBoost performed superiorly, and the Boruta recommended features of 18 nodes and 16 edges for predicting the AD8 score and 10 node and 15 edge features for predicting the MoCA score. Further interpreting the k-mean cluster results of the SHAP values, subjective cognition was dependent on left SomatomotorDorsal mdALFF CV and right visual mdALFF variance (Figure 1A), as well as FrontoParietal-DefaultMode dFC CV, FrontoPairetal-DefaultMode variance, CinguloOpercular-SomatomotorLateral variance, and CinguloOpercular-DefaultMode variance (Figure 1B). The objective cognitive performance relied on mdALFF mean and variance of left CinguloOpercular node, mean of left DefaultMode node, and variance of left medial temporal lobe node (Figure 2A). Objective cognition was also influenced by dFC mean of DefaultMode-CinguloOpercular and CV of DefaultMode-Salience and FrontoParietal-SomatomotorDorsal edges (Figure 2B).
·Figure1
·Figure2
Conclusions:
By utilizing explainable AI techniques in machine learning models, this study has discovered the clinical significance of dynamic metrics of functional connectivity. The networks that are involved in subjective cognitive complaints include visual, motor-sensory, and triple (default-salience-executive) networks. Apart from the medial temporal memory network, the balance of the triple networks is also important for objective cognition.
Lifespan Development:
Aging
Modeling and Analysis Methods:
fMRI Connectivity and Network Modeling 2
Task-Independent and Resting-State Analysis 1
Keywords:
Cognition
Data analysis
Design and Analysis
FUNCTIONAL MRI
Machine Learning
1|2Indicates the priority used for review
Provide references using author date format
Galvin, J. E., C. M. Roe, K. K. Powlishta, M. A. Coats, S. J. Muich, E. Grant, J. P. Miller, M. Storandt and J. C. Morris (2005). "The AD8: a brief informant interview to detect dementia." Neurology 65(4): 559-564.
Ikotun, A. M., A. E. Ezugwu, L. Abualigah, B. Abuhaija and J. Heming (2023). "K-means clustering algorithms: A comprehensive review, variants analysis, and advances in the era of big data." Information Sciences 622: 178-210.
Ke, G., Q. Meng, T. Finley, T. Wang, W. Chen, W. Ma, Q. Ye and T.-Y. Liu (2017). "Lightgbm: A highly efficient gradient boosting decision tree." Advances in neural information processing systems 30.
Kung, Y. C., C. W. Li, S. Chen, S. C. Chen, C. Z. Lo, T. J. Lane, B. Biswal, C. W. Wu and C. P. Lin (2019). "Instability of brain connectivity during nonrapid eye movement sleep reflects altered properties of information integration." Hum Brain Mapp 40(11): 3192-3202.
Kursa, M. B. and W. R. Rudnicki (2010). "Feature Selection with the Boruta Package." Journal of Statistical Software 36(11): 1 - 13.
Liao, W., J. Li, G. J. Ji, G. R. Wu, Z. Long, Q. Xu, X. Duan, Q. Cui, B. B. Biswal and H. Chen (2019). "Endless Fluctuations: Temporal Dynamics of the Amplitude of Low Frequency Fluctuations." IEEE Transactions on Medical Imaging 38(11): 2523-2532.
Liao, W., G. R. Wu, Q. Xu, G. J. Ji, Z. Zhang, Y. F. Zang and G. Lu (2014). "DynamicBC: a MATLAB toolbox for dynamic brain connectome analysis." Brain Connect 4(10): 780-790.
Lundberg, S. M. and S.-I. Lee (2017). "A unified approach to interpreting model predictions." Advances in neural information processing systems 30.
Meacham, S., G. Isaac, D. Nauck and B. Virginas (2019). Towards Explainable AI: Design and Development for Explanation of Machine Learning Predictions for a Patient Readmittance Medical Application. Intelligent Computing: 939-955.
Rosenblatt, F. (1958). "The perceptron: a probabilistic model for information storage and organization in the brain." Psychol Rev 65(6): 386-408.
Rossetti, H. C., L. H. Lacritz, C. M. Cullum and M. F. Weiner (2011). "Normative data for the Montreal Cognitive Assessment (MoCA) in a population-based sample." Neurology 77(13): 1272-1275.