【作者】 张惠敏；

【导师】 张光华；

【作者基本信息】 河北科技大学 ， 计算机科学与技术， 2023， 硕士

【摘要】 通过临床电子健康记录,尽早识别患者的病情程度并有效降低死亡风险,是时下重症监护室（ICU）面临的重大挑战。急性呼吸窘迫综合征（ARDS）常伴随并发症发生,作为单病种患者群体,ICU常用的疾病评分系统以及传统的机器学习模型识别并预测患者死亡风险存在准确性低、效率低、可解释性不佳等问题,引起了国内外医护人员的广泛关注。本文基于重症监护医学信息数据库数据,针对严重程度识别的无创生理参数问题,死亡风险预测的医学小样本和可解释性问题,以及跨肺压指导下的影响ARDS患者的关键风险因素问题进行了研究,具体研究工作如下。（1）针对疾病评分系统和有创生理参数对ARDS病情发展评估存在主观性和非及时性的问题,结合无创参数提出了一种基于极度梯度提升树（XGBoost）的ARDS严重程度识别模型（XGB-SRM）。首先,将提取的患者生理参数进行统计分析,使用四分位距和合成少数过采样处理异常值和不平衡样本;然后,使用皮尔逊相关系数和随机森林作为混合特征选择方法对无创参数综合评分;最后,XGBoost结合网络搜索交叉验证确定模型最佳超参数,从而实现疾病程度的精确分类。实验结果表明,模型在区分疾病严重程度上曲线下面积（AUC）高达0.980,准确率达到0.901,从而保证模型精确度和识别效率。（2）针对医学小样本不均衡而特征空间较大和现有死亡风险预测模型缺乏可解释性的问题,提出了一种基于加权均衡分布适应（W-BDA）和多层感知机（MLP）的可解释性死亡风险预测方法（WM-IMRP）。首先,将提取的数据进行源域和目标域的划分并采用k近邻插补缺失值;其次,两域进行XGBoost特征选择,排除冗余特征以实现降维;然后,通过W-BDA将重新构成的源域和目标域映射到同一再生核希尔伯特空间中,引入平衡因子实现条件分布和边缘分布的加权均衡自适应;最后,由新源域训练MLP网络模型,在新目标域上实现具有可解释性的ARDS死亡风险预测。实验结果表明,该方法在预测死亡风险时AUC高达0.905,准确率为0.878,F1分数为0.921,从而能保证提供不错的预测精度和可靠的解释性。（3）针对跨肺压监测数据少而难以分析影响ARDS死亡风险因素的问题,提出了一种在跨肺压指导下探索影响ARDS的死亡风险关键因素的方法。首先,以跨肺压组为基准,将提取的ARDS患者进行倾向性评分1:1匹配对照组;然后,将通过秩和检验的指标进行全优子集回归筛选最佳特征组合集,经过线性和非线性特征分析获取重要特征;最后,建立机器学习模型表现最佳的死亡风险预测能力。实验结果表明,HCO₃和呼吸频率是影响ARDS的关键风险因素且在基于逻辑回归模型的28天死亡风险预测准确率高达0.921,可为医护人员提供诊疗建议。更多还原

【Abstract】 Identifying the severity of a patient’s condition as early as possible and effectively reducing the mortality risk through clinical electronic health records is a significant challenge currently facing intensive care units（ICU）.Acute respiratory distress syndrome（ARDS）often accompanies complications.As a group of single-disease patients,the commonly used disease scoring systems and traditional machine learning models in ICU have problems such as low accuracy,low efficiency,and poor interpretability in identifying and predicting the mortality risk of patients,which has attracted widespread attention from medical staff at home and abroad.Based on medical information mart for intensive careⅢdata,this thesis focuses on the non-invasive physiological parameter problem of ARDS severity recognition,the small medical sample and interpretability problem of mortality risk prediction,and the critical risk factors affecting ARDS patients under the guidance of transpulmonary pressure.The specific research work is as follows.（1）Given the subjectivity and non-timeliness of the disease scoring system and invasive parameters in evaluating the development of ARDS,combined with noninvasive parameters,this thesis proposed a model（XGB-SRM）of ARDS severity recognition based on extreme gradient boosting（XGBoost）.Firstly,the physiological parameters of patients were extracted from the database for statistical analysis,and the outliers and unbalanced samples were processed by the interquartile range and synthetic minority oversampling.Secondly,the Pearson correlation coefficient and random forest were used as hybrid feature selection to score the noninvasive parameters comprehensively.Finally,to realize the accurate classification of disease degree,XGBoost combined with grid search cross-validation to determine the best hyper-parameters of the model.The experimental results show that the area under the curve（AUC）of the model in distinguishing disease severity is as high as 0.980,with an accuracy of 0.901,thereby ensuring the accuracy and recognition efficiency of the model.（2）Given the characteristics of ARDS medical data with unbalance,small samples and large feature space,and the lack of interpretability of existing model,this thesis proposed an interpretable method（WM-IMRP）for mortality risk prediction based on weighted balanced distribution adaptation（W-BDA）and multilayer perceptron（MLP）.Firstly,k-nearest neighbor interpolation for missing values,the extracted data were preprocessed for the divided source and target domains.Secondly,feature selection based on XGBoost was performed in two domains to eliminate redundant features and achieve dimension reduction.Thirdly,the reconstructed domains were mapped to the same reproducing kernel Hilbert space（RKHS）through W-BDA,and the balance factor was introduced to achieve the weighted equilibrium adaptation of conditional and marginal distributions.Finally,the MLP network model was trained by the new source domain,and the mortality risk prediction of ARDS was achieved on the new target domain.The experimental results show that the method proposed in this article has an AUC of up to0.905,an accuracy of 0.878,and an F1 score of 0.921 in predicting the mortality risk,ensuring good prediction accuracy and reliable explanatory power.（3）Given the problem of limited data and difficulty in analyzing the ARDS mortality risk factors,this thesis proposed a method to explore the key factors affecting the ARDS mortality risk under the guidance of transpulmonary pressure monitoring.Firstly,based on the group detecting transpulmonary pressure,the extracted ARDS patients were scored with a propensity score of 1:1,matching with the control group.Secondly,the optimal feature combination set was selected through total optimal subset regression using the rank sum test indicators,and the critical risk factors affecting ARDS were obtained through linear and nonlinear feature analysis.Finally,the machine learning model demonstrated the best predictive ability for mortality risk.The experimental results show that HCO₃ and respiratory rate are the key risk factors affecting ARDS,and the accuracy of 28-day mortality risk prediction based on the logistic regression model is as high as0.921,which can provide medical staff with diagnosis and treatment suggestions.更多还原