IBM AI Introduces ‘VanDEEPol’: A Hybrid Model That Combines VDP with Recurrent Neural Networks (RNNs) To Predict Brain Activity

Source: https://research.ibm.com/blog/predicting-brain-activity?social_post=5769167617&linkId=136505565

VanDEEPol’, a hybrid AI/mechanistic model to predict brain activity and structure from imaging data, is IBM’s most recent development in the field of brain activity predictions. Compared to earlier methods, the model greatly improves predicted accuracy and promises to one day aid in detecting medical diseases or the construction of brain-computer interfaces by predicting brain activity from sparse imaging data.

Our brains are constantly undergoing complex interactions among billions of neurons, which determine our functions and behaviors. We can map these intricate connections with unprecedented detail using advanced techniques like functional magnetic resonance imaging (fMRI) and calcium imaging (CaI). Models like these could potentially help with the development of neurotechnological devices like brain-computer interfaces. However, they still confront significant difficulties in simulating extremely complicated brain functions.

Shortcomings of existing models

Multiple variables are included in imaging data, and brain activity exhibits nonlinear dynamics that may be missed by autoregressive models that assume a linear link between past and future states. Generic nonlinear models, such as recurrent neural networks (RNNs), which identify the sequential nature of input and use patterns to forecast future states, necessitate vast amounts of training data, which aren’t always available for brain imaging, and their results can be difficult to interpret. As a result, developing accurate, predictive models of brain activity from imaging data continues to be a challenge.

The approach behind VanDEEPol

VanDEEPol is a hybrid approach for reliably predicting brain activity based on imaging data that combines nonlinear van der Pol (VDP) oscillators with an RNN. VDP fits imaging data of many sorts and species accurately and can generalize to previously unseen data. It can provide a limitless amount of simulated data to enhance real imaging data for training an RNN, and it can identify anatomically meaningful relationships between brain areas, providing insights into their functional connectivity.

According to studies, the VanDEEPol model performs substantially better than each component alone in terms of prediction. The model has the potential to help with accurate brain activity predictions, which could be useful in medical diagnosis and neurotechnology.

Functioning of the model

CaI data from zebrafish was used to create the dataset. Singular value decomposition (SVD) analysis was used to identify the top six spatial and temporal components in order to extract functionally meaningful information from the data.

A more advanced parameter estimation approach was devised, which alternates between stochastic search and deterministic optimization to find the hidden variable (excitability) from the observable one (activity). The stochastic search strategy is simply a random walk through both the parameter and hidden state space, considerably improving the model’s fit to the training data.

Predictive Accuracy

VanDEEPol frequently outperforms the baseline autoregressive (VAR) model and an RNN on the zebrafish data set; in the worst scenarios, it performs similarly well. In the short-term prediction window for the zebrafish data set, the RNN outperforms the VDP in terms of correlation. In terms of correlation, all approaches perform similarly for short-term predictions on the human data set. However, VanDEEPol outperforms VAR by a substantial margin for long-term forecasts, implying that VDP captures additional aspects of the underlying dynamics that improve even long-term predictions when paired with the RNN.

Contribution to neuroscience

This research contributes to both the basic and applied aspects of neuroscience, providing fundamental insights into how our brains function and allowing us to use that knowledge to improve our quality of life. Its capacity to effectively forecast future brain activity based on imaging data should aid physicians in diagnosing medical conditions and predicting disease progression. Because it can deduce people’s intents from their brain signals, convert them into commands, and convey the instructions to output devices that achieve those intentions, such brain-computer interfaces could be useful in restoring brain function damaged by injury, disability, or aging.

Source: https://research.ibm.com/blog/predicting-brain-activity

Reference Paper: https://direct.mit.edu/neco/article-abstract/33/8/2087/101867/Learning-Brain-Dynamics-With-Coupled-Low?

https://www.youtube.com/watch?v=lppAwkek6DI&t=3s