[论文解读] Depression Diagnosis and Drug Response Prediction via Recurrent Neural Networks and Transformers Utilizing EEG Signals

Abdolkarim Saeedi, Arash Maghsoudi+1 more|arXiv (Cornell University)|Mar 9, 2023

EEG and Brain-Computer InterfacesNeuroscience被引用 3

一句话总结

本研究提出了一种基于Transformer的深度学习模型，利用脑电图（EEG）信号诊断重度抑郁障碍（MDD）并预测个体患者对药物治疗的反应。该方法在MDD分类任务中实现了97.14%的准确率和99.41%的召回率，在区分药物反应者与非反应者方面实现了97.01%的准确率和97.76%的召回率，优于CNN、LSTM和CNN-LSTM基线模型，且参数量更少。

ABSTRACT

The Early diagnosis and treatment of depression is essential for effective treatment. Depression, while being one of the most common mental illnesses, is still poorly understood in both research and clinical practice. Among different treatments, drug prescription is widely used, however the drug treatment is not effective for many patients. In this work, we propose a method for major depressive disorder (MDD) diagnosis as well as a method for predicting the drug response in patient with MDD using EEG signals. Method: We employ transformers, which are modified recursive neural networks with novel architecture to evaluate the time dependency of time series effectively. We also compare the model to the well-known deep learning schemes such as CNN, LSTM and CNN-LSTM. Results: The transformer achieves an average recall of 99.41% and accuracy of 97.14% for classifying normal and MDD subjects. Furthermore, the transformer also performed well in classifying responders and non-responders to the drug, resulting in 97.01% accuracy and 97.76% Recall. Conclusion: Outperforming other methods on a similar number of parameters, the suggested technique, as a screening tool, seems to have the potential to assist health care professionals in assessing MDD patients for early diagnosis and treatment. Significance: Analyzing EEG signal analysis using transformers, which have replaced the recursive models as a new structure to examine the time dependence of time series, is the main novelty of this research.

研究动机与目标

• 开发一种可靠、非侵入性的方法，利用EEG信号实现对重度抑郁障碍（MDD）的早期诊断。
• 利用基于EEG的机器学习方法预测个体患者对抗抑郁药物的治疗反应。
• 评估Transformer架构在EEG基MDD及药物反应分类任务中相对于传统深度学习模型（CNN、LSTM、CNN-LSTM）的性能表现。
• 建立一种筛查工具，通过提升诊断准确率和治疗个性化水平，支持精神卫生领域的临床决策制定。

提出的方法

• 本研究采用一种经过改进的Transformer架构，旨在有效建模EEG时间序列数据中的时序依赖关系。
• 该模型通过多头自注意力机制处理原始或预处理后的EEG信号，以捕捉长距离和复杂的时序模式。
• 该架构在健康对照组和MDD患者EEG数据上进行端到端训练，其中诊断和药物反应分类任务分别使用独立的输出头。
• 性能通过与成熟模型（CNN、LSTM、CNN-LSTM）进行对比评估，采用相同的训练/测试数据划分和评估指标。
• 该方法利用注意力权重解释模型决策过程，增强临床可解释性。
• 模型通过标准深度学习损失函数（如交叉熵）和训练协议进行优化。

实验结果

研究问题

• RQ1 与传统RNN和CNN相比，基于Transformer的模型在利用EEG信号区分MDD患者与健康对照组方面是否能实现更优性能？
• RQ2 同一EEG基模型是否能准确预测MDD患者对药物治疗的阳性反应？
• RQ3 Transformer中的注意力机制如何有助于识别与MDD及药物反应预测相关的EEG模式？
• RQ4 与传统深度学习架构相比，Transformer模型是否能在参数量更少的情况下仍保持高性能？

主要发现

• 该Transformer模型在利用EEG信号区分MDD患者与健康对照组时，实现了97.14%的准确率和99.41%的召回率。
• 在药物反应预测方面，该模型在分类药物反应者与非反应者时达到了97.01%的准确率和97.76%的召回率。
• 尽管参数量相近，该Transformer模型在诊断和反应预测任务中均优于CNN、LSTM和CNN-LSTM模型。
• 注意力机制使模型能够聚焦于临床相关的EEG模式，从而提升可解释性与性能表现。
• 结果表明，Transformer在早期MDD诊断和个性化治疗规划方面具有作为筛查工具的潜力。