基于时空数据分析的深度学习慢性疼痛智能评估方法-《中国医学物理学杂志》

文章信息/Info

Title:: A deep learning method for intelligent chronic pain assessment based on spatiotemporal data analysis

Author(s):: ZHENG Rui1; WANG Cuiling2; 1. School of Medical Sciences, Shanxi Medical University, Taiyuan 030000, China 2. Department of Outpatient, Shanxi Cancer Hospital, Shanxi Medical University, Taiyuan 030009, China

Keywords:: Keywords: chronic pain spatiotemporal data deep learning physiological signal environmental information intelligent assessment

摘要:: 目的：利用时空数据（包括生理信号和环境信息）构建深度学习模型，实现对慢性疼痛评分的精准预测，为智能化疼痛评估提供支持。方法：提出一种基于改进通道注意力机制的残差模块与双向长短时记忆网络（BiLSTM）的疼痛评分预测模型。改进的通道注意力机制提取关键特征并减少冗余信息，BiLSTM建模时间序列的依赖关系，结合改进粒子群优化算法优化模型超参数。在实验中，模型性能通过均方误差（MSE）、平均绝对误差（MAE）、决定系数（R2）这3个指标进行评估。结果：本文模型的MSE为0.117，MAE为0.254，R2为0.932，优于卷积神经网络（CNN）-BiLSTM、CNN-LSTM等对比模型。消融实验验证了改进模块对模型性能的关键作用。多模态数据实验进一步表明环境信息的引入提升了模型预测能力。结论：本文模型能够高效捕捉疼痛评分的时空特征，具备较高的预测精度和稳定性，为慢性疼痛评估建立了新方法。

Abstract:: Objective To construct a deep learning model using spatiotemporal data (including physiological signals and environmental information) to achieve accurate prediction of chronic pain scores and provide support for intelligent pain assessment. Methods A pain score prediction model based on a residual module with an improved channel attention mechanism and BiLSTM was developed. The improved channel attention mechanism extracted key features and reduced redundant information, while BiLSTM models the temporal sequence dependencies. Improved particle swarm optimization algorithm was integrated to optimize models hyperparameters. In the experiment, the model performance was evaluated using 3 metrics: mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R2). Results The experimental results showed that the proposed model achieved an MSE of 0.117, an MAE of 0.254, and an R2 of 0.932, outperforming the comparison models such as CNN-BiLSTM and CNN-LSTM. Ablation experiment verified the critical role of the improved module in enhancing model performance, and the multimodal data experiment further demonstrated that the integration of environmental information improved the models predictive capability. Conclusion The proposed model can efficiently capture the spatiotemporal characteristics of pain scores, with high prediction accuracy and stability, and thus establishing a new method for chronic pain assessment.