Robust DNN-based Decoder Model with an Embedded State-Space Model Layer

Area of Science:

• Computational Neuroscience
• Machine Learning in Biology
• Biobehavioral Data Analysis

Background:

• Neuroscience data analysis relies on characterizing complex biobehavioral time series.
• Traditional Deep Neural Networks (DNNs) face limitations with noisy, small datasets common in neuroscience.
• Existing DNNs are sensitive to data noise and require large sample sizes, hindering their application.

Purpose of the Study:

• To introduce a novel framework, State-Space Model Deep Neural Network (SSM-DNN), to address DNN limitations in neuroscience.
• To demonstrate SSM-DNN's capability to overcome sample size and noise sensitivity issues.
• To apply SSM-DNN for decoding participant phenotypes from biobehavioral data during a Death Implicit Association Test (D-IAT).

Main Methods:

• Integration of a State-Space Model (SSM) within a classic Deep Neural Network (DNN) architecture.
• Development of the SSM-DNN framework for training and inference on biobehavioral time-series data.
• Application to a dataset from a Death Implicit Association Test (D-IAT) designed for phenotype decoding.

Main Results:

• SSM-DNN achieved a decoding accuracy of 78%, outperforming state-of-the-art DNN models by 20%.
• The model demonstrated a high Area Under the Curve (AUC) of 0.8, indicating excellent specificity and sensitivity.
• The framework proved scalable to high-dimensional time-series data.

Conclusions:

• The novel SSM-DNN framework offers a robust solution for analyzing complex, noisy neuroscience time-series data.
• SSM-DNN significantly enhances decoding accuracy compared to traditional DNNs, especially with limited or noisy datasets.
• This approach provides a broadly applicable and accurate method for biobehavioral data analysis in neuroscience research.