Various factors are behind the forces that drive hospitals toward more sustainable operations. Hospitals contracting with Medicare, for instance, are reimbursed for the procedures performed, regardless of the number of days that patients stay in the hospital. This reimbursement structure has incentivized hospitals to use their resources (such as their beds) more efficiently to maximize revenues. One way hospitals can improve bed utilization is by predicting patients’ length of stay (LOS) at the time of admission, the benefits of which extend to employees, communities, and the patients themselves. In this paper, we employ a data analytics approach to develop and test a deep learning neural network to predict LOS for patients with chronic obstructive pulmonary disease (COPD) and pneumonia. The theoretical contribution of our effort is that it identifies variables related to patients’ prior admissions as important factors in the prediction of LOS in hospitals, thereby revising the current paradigm in which patients’ medical histories are rarely considered for the prediction of LOS. The methodological contributions of our work include the development of a data engineering methodology to augment the data sets, prediction of LOS as a numerical (rather than a binary) variable, temporal evaluation of the training and validation data sets, and a significant improvement in the accuracy of predicting LOS for COPD and pneumonia inpatients. Our evaluations show that variables related to patients’ previous admissions are the main driver of the deep network’s superior performance in predicting the LOS as a numerical variable. Using the assessment criteria introduced in prior studies (i.e., ±2 days and ±3 days tolerance), our models are able to predict the length of hospital stay with 86 % and 91 % accuracy for the COPD data set, and with 74 % and 85 % accuracy for the pneumonia data set. Hence, our effort could help hospitals serve a larger number of patients with a fixed amount of resources, thereby reducing their environmental footprint while increasing their revenue, as well as their patients’ satisfaction.

促使医院朝着更可持续发展的方向发展的力量背后有多种因素。例如，与Medicare签约的医院将为执行的手术程序报销费用，无论患者在医院停留的天数如何。这种报销结构激励医院更有效地利用其资源（例如病床）来最大化收入。医院可以提高床位利用率的一种方法是通过预测入院时的住院时间（LOS），其好处扩展到员工，社区和患者本身。在本文中，我们采用数据分析方法来开发和测试深度学习神经网络，以预测患有慢性阻塞性肺疾病（COPD）和肺炎的LOS。我们所做努力的理论贡献是，它将与患者先前入院有关的变量确定为医院LOS预测的重要因素，从而修改了当前的范例，在该范例中很少考虑患者的病史来预测LOS。我们工作的方法论贡献包括开发数据工程方法以扩充数据集，将LOS预测为数字（而不是二进制）变量，对训练和验证数据集进行时间评估以及对COPD和肺炎住院患者预测LOS的准确性。我们的评估表明，与患者先前入院相关的变量是深层网络在将LOS预测为数值变量方面优越性能的主要驱动力。使用先前研究中引入的评估标准（即±2天和±3天耐受性），我们的模型能够以COPD数据集的86％和91％的准确度以及74％和90％的准确度来预测住院时间肺炎数据集的准确度为85％。因此，我们的努力可以帮助医院以固定数量的资源为更多的患者提供服务，从而减少他们的环境足迹，同时增加他们的收入和患者的满意度。