| 引用本文: | 单武林,彭文举,许鑫鑫,阚劲松,张家云,陈继明.基于机器学习构建妇科恶性肿瘤患者院内大肠埃希菌感染预测模型[J].中国临床新医学,,():-. |
| Shan Wulin,Peng Wenju,Xu Xinxin,Kan Jinsong,Zhang Jiayun.基于机器学习构建妇科恶性肿瘤患者院内大肠埃希菌感染预测模型[J].中国临床新医学,,():-. |
|
| 摘要: |
| 目的 分析妇科恶性肿瘤患者院内感染的病原学分布特征，并构建主要致病菌大肠埃希菌的感染预测模型，为临床抗感染治疗和风险评估提供依据。方法 回顾性收集146例妇科恶性肿瘤患者的临床资料，分析感染病原菌分布特征；分析主要致病菌大肠埃希菌耐药情况及感染的影响因素并构建预测模型。使用R4.3.2和i-Research软件进行统计分析。结果 共检出180株病原菌，中段尿液样本分离率最高（65.56%，118/180）。革兰阴性菌占比80.56%（145/180），以大肠埃希菌为主，占革兰阴性菌70.34%（102/145）及总病原菌56.67%（102/180）。药敏结果显示，大肠埃希菌对头孢曲松、环丙沙星、复方新诺明及左氧氟沙星的耐药率均超过55%，对碳青霉烯类、哌拉西林/他唑巴坦、呋喃妥因及阿米卡星耐药率均低于10%。纳入全部样本类型时，Logistic回归显示大肠埃希菌感染仅与中段尿样本类型呈显著正相关（P=0.004）；基于上述结果，后续选择中段尿液样本进行因素筛选和感染模型构建。单因素和多因素回归显示大肠埃希菌感染与肿瘤类型、血液WBC总数、Lpa、及尿亚硝酸盐定性结果呈显著相关（P < 0.05）；与尿液细菌数量呈一定相关性。接下选取上述指标组合基于多机器学习方法构建大肠埃希菌感染评估模型。结果显示，decision_tree和logistic_regression模型表现稳定，decision_tree模型训练集与测试集AUC分别为0.84和0.82；logistic_regression模型训练集与测试集AUC分别为0.78和0.74。此外，模型校准曲线良好，预测风险与实际风险一致；决策曲线显示在10%-80%阈值范围内，模型净收益显著优于极端策略，具备良好临床实用性。结论 妇科恶性肿瘤患者院内感染以泌尿系统革兰阴性菌为主，大肠埃希菌为显著优势菌株。基于临床分期和常规检测指标等易获参数构建的机器学习预测模型，在区分中段尿样本大肠埃希菌感染时表现出优异性能，能够为临床早期识别高危患者、指导抗生素精准使用提供实用工具，对减少耐药发生和改善抗感染治疗结局具有重要价值。 |
| 关键词: 妇科肿瘤 机器学习 病原菌 感染模型 |
| DOI： |
| 分类号: |
| 基金项目:国家自然科学(编号:82404092)；常州市卫健委重大科技项目(编号:ZD202314) |
|
| A Machine Learning Framework for Predicting Nosocomial?Escherichia coli?Infections in Cervical Cancer |
|
Shan Wulin,Peng Wenju,Xu Xinxin,Kan Jinsong,Zhang Jiayun
|
|
The First Affiliated Hospital of USTC, Division of Life Sciences and Medicine, University of Science and Technology of China
|
| Abstract: |
| Objective This study sought to characterize the pathogen distribution of hospital-acquired infections in patients with gynecological malignancies and to develop a predictive model for infections caused by the predominant agent,?Escherichia coli, thereby guiding antimicrobial therapy and improving risk stratification. Methods We performed a retrospective analysis of clinical and microbiological data from 146 patients with gynecological malignancies. The study characterized the distribution of pathogenic isolates and the antibiotic resistance profile of the primary pathogen,?Escherichia coli. Subsequently, we identified risk factors for?Escherichia coli?infection and developed a predictive model, which was constructed and evaluated using both multivariate logistic regression and machine learning algorithms. All statistical analyses were conducted using R (version 4.3.2) and i-Research software. Results A total of 180 pathogenic isolates were identified, predominantly from midstream urine specimens (118/180, 65.56%). Gram-negative bacteria constituted 80.56% (145/180) of the isolates and were overwhelmingly dominated by?Escherichia coli, which represented 70.34% (102/145) of Gram-negative isolates and 56.67% (102/180) of all pathogens. Antimicrobial susceptibility testing of?E. coli?revealed high resistance rates (> 55%) to ceftriaxone, ciprofloxacin, trimethoprim-sulfamethoxazole, and levofloxacin, in contrast to minimal resistance (< 10%) to carbapenems, piperacillin-tazobactam, nitrofurantoin, and amikacin. As an initial logistic regression identified a significant association between?Escherichia coli?infection and midstream urine sample type (P = 0.004), we focused subsequent model development on this cohort. Univariate and multivariate regression analyses indicated that?Escherichia coli?infection was significantly associated with tumour type, total white blood cell count, Lpa and qualitative urinary nitrite test results (P?< 0.05), and was also associated with urinary bacterial count. A predictive model for?infection was subsequently developed using these indicators through multiple machine learning approaches. The decision tree and logistic regression models demonstrated stable performance, with area under the curve values of 0.84 and 0.82 for the training and test sets of the decision tree model, respectively, and 0.78 and 0.74 for the logistic regression model, respectively. Furthermore, the model's calibration curve demonstrated excellent agreement, indicating that the predicted risk aligned closely with the observed risk. Its clinical utility was further validated by decision curve analysis, which showed that the model yielded a superior net benefit across a clinically relevant threshold probability range (10-80%) compared to the treat-all and treat-none strategies, affirming its potential for practical implementation. Conclusions In summary, our analysis establishes that hospital-acquired infections in gynecological oncology patients are predominantly urinary tract infections caused by Gram-negative bacteria, with?Escherichia coli?as the predominant etiological agent. To address this, we constructed a machine learning prediction model based on readily available clinical parameters, which demonstrated high discriminatory power for identifying?Escherichia coli?in midstream urine samples. This model provides a clinically applicable tool for the early identification of high-risk patients and to guide targeted antibiotic therapy, with significant potential to curb antimicrobial resistance and improve patient outcomes. |
| Key words: Gynecological Cancer Machine Learning Pathogen Infection Model |
