Philip Zachariah, MD, MSc, is Assistant Professor, Department of Pediatrics, Columbia University Vagelos College of Physicians and Surgeons, New York, New York. Elioth Sanabria, MS, is Graduate Research Assistant, Columbia University Fu Foundation School of Engineering and Applied Sciences, New York, New York. Jianfang Liu, PhD, MAS, is Assistant Professor, Quantitative Research (in Nursing), Columbia University School of Nursing, New York, New York. Bevin Cohen, PhD, MS, MPH, RN, is Associate Research Scientist, Columbia University School of Nursing, New York, New York. David Yao, PhD, is Piyasombatkul Family Professor, Columbia University Fu Foundation, New York, New York. Elaine Larson, PhD, RN, FAAN, CIC, is Professor, Columbia University School of Nursing, New York, New York.
Nurs Res. 2020 Sep/Oct;69(5):399-403. doi: 10.1097/NNR.0000000000000449.
Accurate, real-time models to predict hospital adverse events could facilitate timely and targeted interventions to improve patient outcomes. Advances in computing enable the use of supervised machine learning (SML) techniques to predict hospital-onset infections.
The purpose of this study was to trial SML methods to predict urinary tract infections (UTIs) during inpatient hospitalization at the time of admission.
In a large cohort of adult hospitalizations in three New York City acute care facilities (N = 897,344), we used two SML methods-neural networks and decision trees-to predict having a hospital-onset UTI using data available and accessible on the first day of admission at healthcare facilities in the United States.
Performance for both neural network and decision tree models were superior compared to logistic regression methods. The decision tree model had a higher sensitivity compared to neural network, but a lower specificity.
SML methods show potential for automated accurate UTI risk stratification using electronic data routinely available at admission; this could relieve nurses from the burden of having to complete and document additional risk assessment forms in the electronic medical record. Future studies should pilot and test interventions linked to the risk stratification results, such as short nursing educational modules or alerts triggered for high-risk patients.
准确、实时的模型来预测医院不良事件，可以促进及时和有针对性的干预，以改善患者的预后。计算技术的进步使得可以使用监督机器学习（SML）技术来预测医院获得性感染。
本研究旨在试用 SML 方法来预测住院期间入院时的尿路感染（UTI）。
在纽约市三个急症护理机构的一个大型成年住院患者队列中（N=897344），我们使用两种 SML 方法-神经网络和决策树-使用美国医疗机构入院第一天可用且可访问的数据来预测医院获得性 UTI。
神经网络和决策树模型的性能均优于逻辑回归方法。决策树模型的敏感性高于神经网络，但特异性较低。
SML 方法显示出使用入院时常规获得的电子数据进行自动准确的 UTI 风险分层的潜力；这可以减轻护士在电子病历中填写和记录额外风险评估表的负担。未来的研究应该试点和测试与风险分层结果相关的干预措施，例如针对高风险患者的短期护理教育模块或警报。
