Anatomical brain MRI markers of Suicidality in Bipolar Disorder Using Deep Learning
Poster No:
680
Submission Type:
Abstract Submission
Authors:
Melanie Garcia1, Joan Camprodon1, Benjamin Wade2
Institutions:
1Harvard Medical School, Massachusetts General Hospital, Charlestown, MA, 2Massachusetts General Hospital, SOMERVILLE, MA
First Author:
Co-Author(s):
Introduction:
Bipolar Disorder (BD) is characterized by extreme mood swings, including episodes of depression and mania, and is associated with an increased risk of suicidal ideation (SI) and psychological anxiety. Characterizing neural biomarkers of such symptoms in BD is crucial for prevention and early intervention. Artificial Intelligence (AI) methods may significantly improve the detection of biomarkers of depression symptoms in BD and offer new insights into complex mental health issues[1]. Here, we developed Deep Learning models to detect depression symptoms in BD patients through structural Magnetic Resonance Imaging (sMRI) scans. To this end, we utilized an open-source dataset from the "UCLA Consortium for Neuropsychiatric Phenomics LA5c Study,"[2,3]. The code of this project will be openly available on GitHub.
Methods:
We used an open-source dataset available as ds000030 on OpenNeuro. This study employed anatomical data preprocessed with fmriprep[4] in native space, preserving maximum information as compared to registration to template space.
Our approach involved training a 3D Convolutional Neural Network (CNN) with a DenseNet architecture[5]. The target variables for the model were derived from the Hamilton Depression Rating Scale (HAM-D), including for instance the presence or absence of depression and SI. Each model was trained with Adam optimizer and Cross-entropy loss.
To interpret the model's predictions, we implemented a pipeline that included guided grad-CAM[6] and the HighRes3DNet[7] algorithms. This approach identified specific brain anatomical regions on sMRI scans that were important in predicting the presence of SI.
Our approach involved training a 3D Convolutional Neural Network (CNN) with a DenseNet architecture[5]. The target variables for the model were derived from the Hamilton Depression Rating Scale (HAM-D), including for instance the presence or absence of depression and SI. Each model was trained with Adam optimizer and Cross-entropy loss.
To interpret the model's predictions, we implemented a pipeline that included guided grad-CAM[6] and the HighRes3DNet[7] algorithms. This approach identified specific brain anatomical regions on sMRI scans that were important in predicting the presence of SI.
Results:
From the dataset, we selected 41 subjects with BD, ensuring the inclusion of good quality scans and complete HAM-D scores. The sample comprised 22 males and 19 females, with a mean age of 34.7 years (standard deviation = 8.9). The dataset was divided into training (25 subjects), validation (5 subjects), and testing sets (15 subjects).
Certain HAM-D subscales were underrepresented (small sample size) and could not be used. Others led to overfitting as target variables as the number of modalities was too high compared to sample size. Here, we focus on results on presence/absence of depression, of SI, and levels of psychological anxiety. Model accuracy is detailed on Figure 1.
The most effective model in the prediction exhibited an accuracy of 52% in training, and 100% in both validation and testing sets. Despite low training accuracy - possibly due to cross-entropy loss penalization of low confidence predictions or to BD representation imbalance across datasets - the high performance on both validation and testing sets is promising, suggesting that the model may have successfully identified relevant patterns of neuroanatomy predictive of SI.
Notably, the model identified biologically plausible brain regions, all in the right hemisphere, that were predictive of suicidal thoughts. In more than 80% of cases, these areas included the parietal operculum, planum temporale, transverse temporal gyrus, superior temporal gyrus, supramarginal gyrus, central operculum, middle temporal gyrus, posterior insula, and temporal white matter. Figure 2 represents these regions.
The regions identified as important are well-aligned with pevious reports[8]. Future work will directly investigate this set of regions and their association with depressive symptoms in BD.
These results could improve the diagnosis and early detection of SI in BD, leading to timely interventions that save lives and enable healthcare providers to tailor treatment plans more effectively.
Certain HAM-D subscales were underrepresented (small sample size) and could not be used. Others led to overfitting as target variables as the number of modalities was too high compared to sample size. Here, we focus on results on presence/absence of depression, of SI, and levels of psychological anxiety. Model accuracy is detailed on Figure 1.
The most effective model in the prediction exhibited an accuracy of 52% in training, and 100% in both validation and testing sets. Despite low training accuracy - possibly due to cross-entropy loss penalization of low confidence predictions or to BD representation imbalance across datasets - the high performance on both validation and testing sets is promising, suggesting that the model may have successfully identified relevant patterns of neuroanatomy predictive of SI.
Notably, the model identified biologically plausible brain regions, all in the right hemisphere, that were predictive of suicidal thoughts. In more than 80% of cases, these areas included the parietal operculum, planum temporale, transverse temporal gyrus, superior temporal gyrus, supramarginal gyrus, central operculum, middle temporal gyrus, posterior insula, and temporal white matter. Figure 2 represents these regions.
The regions identified as important are well-aligned with pevious reports[8]. Future work will directly investigate this set of regions and their association with depressive symptoms in BD.
These results could improve the diagnosis and early detection of SI in BD, leading to timely interventions that save lives and enable healthcare providers to tailor treatment plans more effectively.
·Figure 1: Performance of the models.
·Figure 2: Regions important to predict suicidality.
Conclusions:
This study highlights how psychiatric research can leverage AI, particularly in identifying brain regions associated with suicidal thoughts in BD. These findings align with existing research and open avenues for more focused treatment approaches. Future work will include performing experiments on larger datasets and coupling sMRI with fMRI information.
Disorders of the Nervous System:
Psychiatric (eg. Depression, Anxiety, Schizophrenia) 1
Modeling and Analysis Methods:
Classification and Predictive Modeling 2
Other Methods
Novel Imaging Acquisition Methods:
Anatomical MRI
Keywords:
Computational Neuroscience
Data analysis
Machine Learning
Modeling
Multivariate
Open-Source Code
Psychiatric
Statistical Methods
STRUCTURAL MRI
Other - Deep Learning, Bipolar Disorder, Suicide, Depression
1|2Indicates the priority used for review
Provide references using author date format
[2] Preprocessed Consortium for Neuropsychiatric Phenomics dataset—PMC. (n.d.). Retrieved 13 November 2023, from https://www-ncbi-nlm-nih-gov.elib.tcd.ie/pmc/articles/PMC5664981/
[3] Poldrack, R. A., Congdon, E., Triplett, W., Gorgolewski, K. J., Karlsgodt, K. H., Mumford, J. A., Sabb, F. W., Freimer, N. B., London, E. D., Cannon, T. D., & Bilder, R. M. (2016). A phenome-wide examination of neural and cognitive function. Scientific Data, 3(1), 160110. https://doi.org/10.1038/sdata.2016.110
[4] Esteban, O., Markiewicz, C. J., Blair, R. W., Moodie, C. A., Isik, A. I., Erramuzpe, A., Kent, J. D., Goncalves, M., DuPre, E., Snyder, M., Oya, H., Ghosh, S. S., Wright, J., Durnez, J., Poldrack, R. A., & Gorgolewski, K. J. (2019). fMRIPrep: A robust preprocessing pipeline for functional MRI. Nature Methods, 16(1), Article 1. https://doi.org/10.1038/s41592-018-0235-4
[5] Huang, G., Liu, Z., van der Maaten, L., & Weinberger, K. Q. (2017). Densely Connected Convolutional Networks. 4700–4708. https://openaccess.thecvf.com/content_cvpr_2017/html/Huang_Densely_Connected_Convolutional_CVPR_2017_paper.html
[6] Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., & Batra, D. (2019). Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization. arXiv:1610.02391 [Cs]. https://doi.org/10.1007/s11263-019-01228-7
[7] Li, W., Wang, G., Fidon, L., Ourselin, S., Cardoso, M. J., & Vercauteren, T. (2017). On the Compactness, Efficiency, and Representation of 3D Convolutional Networks: Brain Parcellation as a Pretext Task. In M. Niethammer, M. Styner, S. Aylward, H. Zhu, I. Oguz, P.-T. Yap, & D. Shen (Eds.), Information Processing in Medical Imaging (pp. 348–360). Springer International Publishing. https://doi.org/10.1007/978-3-319-59050-9_28
[8] Ellard, K. K., Zimmerman, J. P., Kaur, N., Van Dijk, K. R. A., Roffman, J. L., Nierenberg, A. A., Dougherty, D. D., Deckersbach, T., & Camprodon, J. A. (2018). Functional Connectivity Between Anterior Insula and Key Nodes of Frontoparietal Executive Control and Salience Networks Distinguish Bipolar Depression From Unipolar Depression and Healthy Control Subjects. Biological Psychiatry: Cognitive Neuroscience and Neuroimaging, 3(5), 473–484. https://doi.org/10.1016/j.bpsc.2018.01.013