| 摘要: |
| 目的 探讨基于循环神经网络（RNN）算法构建的动态预测模型用于创伤重症患者临床结局动态预测的价值，并研究动态策略和实时预测模型可行的搭建方案及路径。方法 本研究数据来源于美国重症监护医学信息数据库（MIMIC）-Ⅳ2.0。以创伤重症患者院内结局为预测目标，使用长短时记忆（LSTM）和门控循环单元（GRU）2种RNN算法分别在4、6和8 h时间窗下训练动态预测模型。使用灵敏度、特异度、F1值和AUC值对模型性能进行评价，并分析不同RNN算法和时间窗对模型性能的影响。在8 h时间窗下分别训练隐马尔科夫模型（HMM）、随机森林（RF）模型和logistic模型作为对照，横向比较2种RNN算法模型与对照模型的性能指标，并分析各模型的时间趋势变化。结果 在不同时间窗时，RNN动态模型在灵敏度、特异度、F1值和AUC值等4个性能指标上差异均有统计学意义（均P＜0.001），在8 h时间窗时模型的各性能指标均高于6 h和4 h时；不同RNN算法（LSTM和GRU）间仅特异度差异有统计学意义（P＝0.036）。横向比较结果显示，2种RNN算法模型和其他模型间各性能指标差异均有统计学意义（均P＜0.001），2种RNN算法模型各指标均高于HMM、RF和logistic模型；各算法模型灵敏度、特异度和F1值的ICC均小于0.400（95% CI未包含0），而AUC值的ICC在统计学上证据不足（95% CI包含0）。结论 基于RNN算法的动态模型对创伤重症患者临床结局的预测效果较其他常见模型具有一定优势，且时间窗对模型性能可能存在影响。 |
| 关键词: 循环神经网络 长短期记忆网络 门控循环单元 创伤 动态模型 临床结局 预测模型 |
| DOI：10.16781/j.CN31-2187/R.20240183 |
| 投稿时间：2024-03-22修订日期：2024-08-26 |
| 基金项目:上海市卫生健康委员会卫生行业临床研究专项（202340037），海军军医大学（第二军医大学）“三航”计划. |
|
| Dynamic prediction of clinical outcomes for critical trauma patients based on a recurrent neural network model |
| QI Geyao1△,XU Jin2△,JIN Zhichao1* |
(1. Department of Military Health Statistics, Faculty of Health Services, Naval Medical University (Second Military Medical University), Shanghai 200433, China;
2. No. 940 Hospital of Joint Logistics Support Force of PLA, Lanzhou 730030, Gansu, China
△Co-first authors.
* Corresponding author) |
| Abstract: |
| Objective To explore the value of dynamic prediction model based on recurrent neural network (RNN) algorithms for dynamic prediction of clinical outcomes in patients with critical trauma, and to study the feasible construction scheme and path of dynamic strategy and real-time prediction model. Methods The data of this study were derived from the US Medical Information Mart for Intensive Care (MIMIC) -Ⅳ 2.0. In order to predict the in-hospital outcomes of critical trauma patients, 2 RNN algorithms, long short-term memory (LSTM) and gated recurrent unit (GRU) were used to train dynamic prediction models under the time windows of 4, 6 and 8 h, respectively. The performance of the models was evaluated using the sensitivity, specificity, F1 value and area under curve (AUC) value; and the effects of different RNN algorithms and time windows on the performance of the models were analyzed. Hidden Markov model (HMM), random forest (RF) model and logistic model were trained under 8-h time window as the controls to compare the performances and the time trends horizontally with the 2 RNN algorithm models. Results There were significant differences in the 4 performance indexes of the RNN dynamic models including the sensitivity, specificity, F1 value and AUC value (all P＜0.001), and the performance indexes at 8-h time window were higher than those at 6 h and 4 h; there was only significant difference in specificity between different RNN algorithms (LSTM & GRU) (P＝0.036). The results of the horizontal comparison showed that there were significant differences in each performance index between the 2 RNN prediction models and other models (all P＜0.001), and each index of the 2 RNN algorithm models was higher than those of the HMM, RF model and logistic model. The intraclass correlation coefficients (ICCs) of each algorithmic model were less than 0.400 for the sensitivity, specificity and F1 value (0 was not included in 95% confidence interval ［CI］), while the ICCs for the AUC value were statistically under-evidenced (0 was included in 95% CI). Conclusion The dynamic models based on RNN algorithms have certain performance advantages over those based on other common algorithms, and the time window may have an impact on the model performance. |
| Key words: recurrent neural network long short-term memory gated recurrent unit trauma dynamic model clinical outcomes predicting model |
