循环神经网络模型在腹膜透析临床预后预测中的初步应用

目的 应用深度学习模型循环神经网络(recurrent neural network,RNN)及其变体门控循环单元(gated recurrent unit,GRU),基于临床真实数据,构建腹膜透析临床预后预测模型,并比较其与医学研究中常用的逻辑回归(logistic regression, LR)模型的预测性能,探索预测结果中可能的医学意义。方法 使用北京大学第三医院腹膜透析门诊的常规诊疗数据,基于患者在开始透析时的基线数据、随访数据和预后数据构建RNN和GRU预测模型。使用受试者工作特征曲线下面积(area under the ROC curve,AUROC)、召回率(recall)、F1分数(F1-score)三个指标在测试集上评价比较模型对患者死亡风险的预测效果。结果 共纳入656例患者,其中死亡患者261例,共计13 091条诊断记录。经过十折交叉验证调整超参数并在单独的测试集测试结果显示,LR模型、RNN模型、GRU模型的AUROC分别为0.701 4、0.786 0、0.814 7,RNN和GRU模型的预测性能显著优于传统的LR模型。在召回率和F1分数方面,RNN和GRU模型的性能也均显著优于LR模型,且GRU模型表现最好。进一步分析显示GRU模型在不同预测窗口下对于不同死因或相同死因的召回率不尽相同。结论 RNN模型(尤其是GRU模型)相比于传统医学研究所使用的LR模型,对于腹膜透析临床预后预测具有更佳效果,可能有助于医生早期干预,提高医疗质量,具有很强的临床应用价值。

Application of recurrent neural network in prognosis of peritoneal dialysis

Objective: Deep learning models, including recurrent neural network (RNN) and gated recurrent unit (GRU), were used to construct the clinical prognostic prediction models for peritoneal dialysis (PD) patients based on routine clinical data. The performance of the RNN and GRU were compared with logistic regression (LR), which is commonly used in medical researches. The possible underlining clinical implications based on the result from the GRU model were also investigated.Methods: We used the clinical data from the PD center of Peking University Third Hospital as the data source. Both the baseline data at the beginning of dialysis, and the follow-up and prognostic data of the patients were used by the RNN and GRU prediction models. The hyper-parameters were tuned based on the 10-fold cross-validation. The risk prediction performance of each model was evaluated via area under the receiver ope-ration characteristic curve (AUROC), recall rate and F1-score on the testset. Results: A total of 656 patients with the 261 occurrences of death were included in the experiment. The total number of all diagnostic records were 13 091. The results on the testset showed that the AUROC of the LR model, RNN mo-del, and GRU model was 0.701 4, 0.786 0, and 0.814 7, respectively. The predictive performances of the GRU and RNN models were significantly better than that of the LR model. The performances of the GRU and RNN models assessed by recall rate and F1-score were also significantly better than that of the LR model, in which the GRU model reached the best performance. In addition, the recall rates were different among different causes of death or by different prediction time windows.Conclusion: The recurrent neural network model, especially the GRU model, is more effective in predicting PD patients’ prognosis as compared with the LR model. This new model may be helpful for clinicians to provide timely intervention, thus improving the quality of care of PD.