Decoding Neuronal Dynamics: Chaotic Deep Learning for Major Depressive Disorder Classification

1428 

Abstract Submission 

Yi-Ju Lee^1,2

¹Institute of Statistical Science, Academia Sinica, Taipei City, Taiwan, ²Smart Healthcare Project, Academia Sinica, Taipei City, Taiwan

Yi-Ju Lee, Dr.  
Institute of Statistical Science, Academia Sinica|Smart Healthcare Project, Academia Sinica
Taipei City, Taiwan|Taipei City, Taiwan

Navigating the intricacies of Major Depressive Disorder (MDD) proves challenging due to its roots in diverse neurotransmitters, neural circuits, and hormonal systems, making reliance on subjective self-reported symptoms for diagnosis difficult. In the realms of physical statistics, artificial intelligence (AI), and medical research, the amalgamation of complexity science emerges as a robust avenue for extracting vital health insights from intricate medical data, particularly within psychiatry. Innovations such as ChaosNet, an AI model rooted in chaos theory, show promise in faithfully simulating human neuronal firing patterns at a network scale. Concurrently, power law scaling serves to quantify the complexity of brain signal dynamics. In this research, leveraging fMRI data, we aim to scrutinize classification models for MDD diagnosis, integrating the concept of complex systems at different stages of the model analysis process. This comprehensive approach seeks to advance MDD diagnosis by embracing a holistic perspective grounded in the principles of complex systems.

Participants with structural and fMRI images, demographic and clinical data were selected from the Strategic Research Program for Brain Sciences (SRPBS) cohort. Brain imaging data of 400 age and sex-matched, right-handed MDD patients (age mean = 40.21 ± 7.35; male = 49.5%) and 400 health adults (age mean = 39.46 ± 8.01) were retrieved. The functional images were pre-processed. The power-law scaling of brain activity of each voxel was extracted and transformed into a heatmap for model training. The images were split randomly in the ratio of 8:1:1 for the training, testing, and validation data set. We use ChaosNet and DenseNet 121 as model backbone, and they were trained using Python 3.8 for 500 epochs with 1 Nvidia DGX A100 (40G) GPU.

In contrast to using typical fMRI BOLD signal as input to DenseNet 121, our result has suggested that the complexity-transformed image data with ChaosNet show significant decreased training time from 48.3 hours to 1.12 hours with similar classification results. In the best-performing model, the average testing accuracy is 92.3. We also identified the key brain regions that is related to MDD, such as prefrontal cortex, hippocampus and amygdala. under Bonferroni correction.

This study presents robust biological evidence surpassing previous methods in identifying MDD. We employ signal complexity within chaotic-based models to detect abnormal brain activities in MDD patients, necessitating expertise in statistical science, computer science, and neurosciences. The observed alterations in pathological hemodynamics in MDD suggest a significant loss of brain signal complexity, potentially contributing to precise clinical diagnosis. Our approach harnesses the unique properties of chaotic neurons, proving more efficient than alternative models. The future work will require exploring the integration of genetic data to subtype MDD patients, aiming to enhance our understanding of this complex disorder.

Psychiatric (eg. Depression, Anxiety, Schizophrenia)

Classification and Predictive Modeling ¹

Informatics Other ²

FUNCTIONAL MRI

Other - Signal Complexity; Chaotic-based Model; Major Depressive Disorder