Deep Learning-based Stress Determinator for Mouse Psychiatric Analysis using Hippocampus Activity

Benjamin C. M. FungarXiv 2020 · arXiv

TL;DR

This paper introduces a deep learning model, specifically a Bidirectional Long Short-Term Memory (BiLSTM) network, to decode mouse hippocampus activity and predict their stress levels. By analyzing neural signals, the model can accurately determine if a mouse is interacting with a bullying mouse, a proxy for stress, outperforming traditional machine learning methods.

What Problem Does This Paper Address?

Understanding the neural basis of behavior, particularly stress, is a fundamental goal in neuroscience. While traditional methods for analyzing neural activity exist, they often lack the sophistication to capture the complex temporal dynamics of neural signals. Furthermore, the relationship between specific neural activity patterns in the hippocampus and an animal's stress level, especially in social contexts, remains an area ripe for deeper investigation. This paper aims to bridge this gap by applying state-of-the-art deep learning techniques to neural decoding for psychiatric analysis, specifically focusing on stress in mice.

Methods

The core of this research lies in applying a Bidirectional Long Short-Term Memory (BiLSTM) network for neural decoding. BiLSTMs are chosen for their ability to process sequential data by considering information from both past and future time steps, making them well-suited for analyzing time-series neural activity. The model takes raw neural activity data from the mouse hippocampus as input.

For comparison and benchmarking, the authors also implemented and evaluated other machine learning models, including:

  • Bidirectional Gated Recurrent Unit (BiGRU): Similar to BiLSTM but with a slightly different architecture.
  • Support Vector Machine (SVM): A classical supervised learning model for classification.
  • Random Forest: An ensemble learning method that builds multiple decision trees.
  • Naive Bayes: A probabilistic classifier based on Bayes' theorem.
  • LSTM-Softmax and LSTM-SVM: Variations combining LSTM with other classifiers.

The model architecture includes a Dropout layer to prevent overfitting during training. The training process utilizes the Adam optimizer with a cross-entropy cost function and a learning rate of 1e-4.

Datasets

The study utilizes a dataset derived from recorded mouse neural activity and behavioral data. The neural activity data is extracted from the hippocampus using the CNMF-E method. Behavioral data, specifically the distance between two mice, is processed using DeepLabCut. The "interaction" status is defined by this distance: less than 15 centimeters indicates interaction (and thus potential stress), while over 15 centimeters indicates no interaction.

The dataset is split into training, testing, and validation sets with proportions of 0.56, 0.3, and 0.14, respectively. The paper refers to specific experimental scenarios labeled as "SI A" and "SI B," where "A" signifies no bullying mouse present and "B" signifies the presence of a bullying mouse. Different mouse annotated numbers (e.g., 1058, 3470, 1056) are used to denote specific experimental groups.

Key Insights and Ideas

  • Deep Learning for Neural Decoding: The paper demonstrates that deep learning, particularly BiLSTMs, can effectively decode complex neural activity patterns from the hippocampus for behavioral analysis.
  • Hippocampal Activity and Stress: The research strongly suggests that hippocampus activity is a reliable indicator of stress in mice, especially in social contexts involving potential threats (bullying mice).
  • Social Stress Detection: The model can accurately predict whether a mouse is stressed based on its neural activity, with the presence of a bullying mouse being a significant factor.
  • Environmental Equivalence: The experiments provide evidence against the idea that mouse stress levels are equivalent across diverse environments, suggesting that specific contextual cues (like the presence of a bully) are crucial.

Key Results

  • High Accuracy: The proposed BiLSTM model achieved an overall prediction accuracy of approximately 0.985, significantly outperforming traditional machine learning models like SVM, Random Forest, and Naive Bayes.
  • Statistical Significance: Student's t-tests confirmed statistically significant differences in neural activity between groups of mice under different conditions (e.g., presence vs. absence of a bullying mouse, interacted vs. non-interacted states).
  • BiGRU Efficiency: While BiLSTM and BiGRU achieved similar accuracy, BiGRU demonstrated faster convergence (300-400 epochs vs. 500-600 epochs for BiLSTM) with the same hyperparameters, suggesting it might be a more computationally efficient choice.
  • Stress Indicators: The presence of a bullying mouse significantly influenced defeated mouse neuron activity, leading to higher neuron firing rates when the bully was visible, indicating increased stress.

Limitations and Future Work

  • Dataset Size: The authors acknowledge that the limited dataset size might prevent the BiLSTM from reaching its absolute best performance.
  • Feature Engineering: The study did not extensively focus on feature engineering, and the authors suggest that extracting the importance of individual neurons could further improve model performance.
  • Interpretability: While the model shows high accuracy, improving interpretability is a future goal, with suggestions to incorporate attention layers and feedforward neural networks.
  • Oversampling Techniques: To address the limited dataset size and potential class imbalance, the authors propose merging data from different sessions and exploring oversampling techniques like SMOTE.
  • Distance Threshold: The current distance threshold of 15 cm for defining interaction is a parameter that could be further investigated, with a suggestion to explore a 10 cm threshold in future work.
  • Other Scenarios: The authors plan to apply the deep learning model to other experimental scenarios (e.g., Def, Hab) to explore its feasibility in broader neural decoding applications.
  • Place Cells: The potential influence of place cells on neuron activity, which could confound stress detection, is noted as an area for future investigation.

Summary updated

BibTeX
@misc{fung2020deep,
  author = {Benjamin C. M. Fung},
  title = {{Deep Learning-based Stress Determinator for Mouse Psychiatric Analysis using Hippocampus Activity}},
  year = {2020},
  eprint = {2006.06862},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2006.06862},
}
Figure