应用人工神经网络评价老年重症监护室患者的预后

Artificial neural network in the mortality prediction for patients in geriatric intensive care unit

【摘要】目的比较人工神经网络(ANN)和急性生理学及慢性健康状况评价Ⅱ(APACHEⅡ)评价老年重症监护室(GICU)患者死亡风险的预测能力。方法回顾分析北京大学第一医院GICU2005年1月至2006年12月收治的177例年龄≥65周岁患者的病例资料。记录患者入GICU当日的资料,计算APACHEⅡ评分和预期死亡风险。随机抽取117例患者的资料作为ANN的训练集,其余60例患者的资料作为ANN的检验集。根据APACHEⅡ的22个变量建立ANN22,根据单个变量对于预后预测的影响大小,去除影响较小的变量,分别利用10个和8个变量建立ANN10和ANN8。应用APACHEⅡ和ANN22、ANN10、ANN8预测检验集患者的死亡风险。应用受试者工作特性曲线下面积(aROC)判断ANN和APACHEⅡ的预测能力。结果ANN22和APACHEⅡ的aROC分别为0.943和0.949,差异没有统计学意义(P=0.829)。ANN10和ANN8的aROC分别为0.968和0.926。结论应用ANN预测GICU患者的死亡率与应用以Logistic回归为基础的APACHEⅡ相比,结果相当。应用较少的变量所建立的ANN预测准确性仍很高。

基金项目:

【Abstract】Objective To compare the ability of artificial neural networks(ANN) and the acute physiology and chronic health evaluation Ⅱ (APACHE Ⅱ) to predict mortality for patients in geriatric intensive care unit (GICU). Methods The purpose of this retrospective case series was to compare ANN and APACHE Ⅱ in the mortality pridiction for a cohort of patients admitted to a seven-bed GICU in a Beijing teaching general hospital. All 177 patients were older than 65 years and consecutively admitted to our GICU from Jan 2005 to Dec 2006.The 22 variables used to obtain APACHE Ⅱ score and risk of death were collected from each patient on admission. All data were randomly allocated to either the training (n=117) or validation set (n=60). Three ANN models were developed using the data from the training set, namely ANN22 (trained with all the 22 variables), ANN10 (trained with the 10 highest information gain variables) and ANN8 (trained with the 8 highest information gain variables). Three ANN models and APACHE Ⅱwere used to predict mortality in the validation set. The accuracy of ANN and APACHE Ⅱ was assessed by area under the receiver operator characteristics curve (aROC). Results The aROC was 0.943 for ANN22 and 0.949 for APACHE Ⅱin predicting GICU mortality(P=0.829). For ANN10 and ANN8, the aROC was 0.968 and 0.926, respectively. Conclusion Both ANN and APACHE Ⅱhave similar performance in predicting GICU outcome. ANN uses fewer variables and yet is comparable to APACHE Ⅱ.