基于ATT-FC-LSTM模型的远程会诊服务时长预测  

Duration Prediction of Teleconsultation Services Based on the ATT-FC-LSTM Model

作　　者：翟运开 贾启硕 乔岩 赵杰 ZHAI Yunkai;JIA Qishuo;QIAO Yan;ZHAO Jie(Zhengzhou University,Zhengzhou,China;National Engineering Laboratory for Internet Medical Systems and Applications,Zhengzhou,China;Henan International Joint Laboratory of Intelligent Health Information System,Zhengzhou,China;The First Affiliated Hospital of Zhengzhou University,Zhengzhou,China)

机构地区：[1]郑州大学管理学院,郑州市450001 [2]互联网医疗系统与应用国家工程实验室 [3]河南省智能健康信息系统国际联合实验室,郑州市450001 [4]郑州大学第一附属医院

出　　处：《管理学报》2025年第3期568-576,共9页Chinese Journal of Management

基　　金：国家自然科学基金资助项目(72202217);河南省高等学校哲学社会科学基础研究资助重大项目(2022-JCZD-21);中国博士后科学基金资助项目(2023M743163)。

摘　　要：针对远程会诊服务时长异质性、动态性、随机性等特点,使用注意力机制与全连接层相结合的长短期记忆神经网络算法来预测远程会诊服务时长,利用回归预测和分类预测两种方法评估算法预测性能。该模型堆叠了一个注意力层、3个全连接层和两层长短期记忆神经网络,以实现对输入数据的选择性关注和更高级别的特征表示,从而提高模型的性能和表示能力。与4种流行的机器学习算法相比,该设计模型在分类预测中的多个预测性能评价指标上均表现更优,并在此基础上计算得出对远程会诊服务时长影响程度最大的4个变量,依次为会诊专家、会诊科室、专家是否迟到和专家职称。To address the characteristics of teleconsultation service duration,such as heterogeneity,dynamism,and randomness,a Long Short-Term Memory(LSTM)neural network algorithm combined with attention mechanisms with fully connected layers is used to predict the duration of teleconsultation services.The performance of the algorithm’s predictions is evaluated using both regression and classification methods.The model stacked an attention layer,three fully connected layers,and two LSTM layers to selectively focus on input data and achieve higher-level feature representation,thereby enhancing the model’s performance and representation capacity.Compared to four popular machine learning algorithms,the designed model performed better across multiple prediction performance evaluation metrics in classification predictions.Based on this,the four variables with the greatest impact on the duration of teleconsultation services were calculated,in order of significance:consulting expert,consulting department,whether the expert was late,and the expert’s title.

关 键 词：深度学习 远程会诊服务时长 长短期记忆神经网络 注意力机制 分类预测 

分 类 号：C93[经济管理—管理学]