Search Papers | Poster Sessions | All Posters
Poster B113 in Poster Session B - Thursday, August 8, 2024, 1:30 – 3:30 pm, Johnson Ice Rink
DivfreqBERT : Encoding Distinct Frequency Ranges of Brain Dynamics Based on the Complexity of the Brain
Sangyoon Bae1 (), Junbeom Kwon1, Jiook Cha1, Shinjae Yoo2; 1Seoul National University, 2Brookhaven National Laboratory
The brain functions as a scale-free network, with fMRI BOLD signals demonstrating a power-law distribution. Traditional deep neural networks often overlook this aspect, leading to their underperformance due to a structure ill-suited for the complexity of brain signals. We introduce DivfreqBERT, an end-to-end model tailored for time series data, leveraging scale-free network properties for improved encoding of biological characteristics. Utilizing Lorentzian and multi-fractal functions, it segments whole-brain dynamics into three components, each consistent with the power-law function and displaying distinct small-world connectivity features. This method significantly enhances several downstream tasks, including predicting sex, age, intelligence, and depression in the Adolescent Brain Cognitive Development (ABCD) dataset, encompassing over 11,000 participants aged 9-10, and the UK Biobank (UKB) dataset, with data from over 500,000 participants aged 40-69. During pretraining, DivfreqBERT employs variations in small-worldness across frequencies to order nodes by communicability for masking, facilitating the learning of networks where highly communicable nodes play pivotal roles. Scaling up the model by pre-training on the extensive UKB dataset and fine-tuning on ABCD data markedly improved model performance. Additionally, DivfreqBERT provides interpretability by showing which connections between ROIs within each frequency range influenced the outcome and which ROIs were important. Overall, divfreqBERT demonstrates the potential of neural networks informed by complex system insights, emphasizing the benefits of integrating the brain's complexity into neural network models.
Keywords: Complex system fMRI BERT Knowledge-guided deep neural network