ABSTRACT

A reliable and accurate estimate of the expected hospital length of stay (LOS) of a patient is important to patients, medical providers, and insurance companies. Predicting hospital Length of Stay (LOS) is a complex and ill-structured problem, driven by many factors such as a patient’s individual characteristics, treatment plans, and disease-interactions. In this paper, we develop a novel model to predict the expected LOS at the time of admission by combining network science and deep learning. We propose a two-dimensional construct of latent comorbidities comprising historical and probable comorbidities that a patient does not currently manifest but could likely develop during the course of hospital stay. The probable comorbidities are derived from a network comprising relationships among diseases observed in 3.2 million patient records in hundreds of US hospitals. We employ this construct of latent comorbidities in deep learning models to predict patients’ LOS using almost 10 million other patient visits belonging to various disease categories. Implementing these models and analyses required a high-performance computing (Big Data) facility. The average mean absolute percent error of our models across all categories of diseases was 29.8%, which is the best in the current state-of-the-art. Our primary contribution is in developing a generalizable method to create a predictor construct for recognizing underlying relationships through network analyses, which can then be used in a deep learning model to predict an exogenous dependent variable.

摘要

对患者的预期住院时间 (LOS) 进行可靠和准确的估计对于患者、医疗服务提供者和保险公司来说非常重要。预测住院时间 (LOS) 是一个复杂且结构不合理的问题，受患者个体特征、治疗计划和疾病相互作用等多种因素的驱动。在本文中，我们通过结合网络科学和深度学习，开发了一种新模型来预测录取时的预期 LOS。我们提出了一个潜在的二维结构合并症包括患者目前未表现出但可能在住院期间发生的历史和可能合并症。可能的合并症来自一个网络，该网络包含在数百家美国医院的 320 万个患者记录中观察到的疾病之间的关系。我们在深度学习模型中采用这种潜在合并症的结构，使用属于各种疾病类别的近 1000 万其他患者就诊来预测患者的 LOS。实施这些模型和分析需要高性能计算（大数据）设施。我们的模型在所有疾病类别中的平均平均绝对百分比误差为 29.8%，这是当前最先进技术中最好的。