创新背景

传统工具只对当天的整体数字进行预测，提供的数据很笼统，无法高效率地进行急诊床位的分配。

创新过程

在这项研究中，研究小组表明，该工具比规划者使用的传统基准更准确，基于前六周一周同一天所需的平均床位数量。

该工具还考虑了尚未到达医院的患者，也提供了比传统方法更详细的信息。该工具不是对当天的整体数字预测，而是包括四小时和八小时内需要多少张床位的概率分布，并每天提供四次预测，通过电子邮件发送给医院规划人员。

研究团队现在正在与UCLH合作完善模型，以便他们可以估计医院不同区域需要多少张床位（例如，医疗病房或外科病房的床位）。

主要作者Zella King博士（伦敦大学学院临床运营研究部门和伦敦大学学院健康信息学研究所）表示，该人工智能模型提供了关于全天床位可能需求的更丰富的图片。他们在记录患者数据的那一刻就利用了这些数据。UCLH患者流动和应急准备。

研究人员使用UCLH在2019年5月至2021年7月期间记录的患者数据训练了12个机器学习模型。这些模型根据从年龄和患者如何到达医院，到测试结果和咨询次数的数据，评估了每位患者从急诊科入院的可能性，并将这些概率结合起来，对所需的床位数量进行总体估计。

然后，他们将模型的预测与2019年5月至2020年3月之间的实际录取进行了比较，发现它们优于传统方法，中心预测平均比实际数字高出四次，而传统方法平均为6.5次入院。在Covid袭击后，研究人员能够调整模型，以考虑到达的人数和他们在急诊室花费的时间的显着变化。

自患者到达以来，12个模型中的每一个都专注于不同时间间隔的数据：第一个模型仅关注到达时记录的数据，第二个模型关注前15分钟内记录的数据，而模型12专注于12小时内记录的数据。这是因为各种因素的重要性各不相同，这取决于已过去的时间和可用的其他数据量。例如，在模型1中，到达医院的方法是一个重要因素，但在后来的模型中却变得不那么重要。研究人员发现，将12个模型放在一起使用比使用更少的模型更准确。

创新关键点

人工智能在医疗保健中的大多数应用都集中在临床问题上，而该智能工具旨在帮助医疗保健的运营方面——即如何运行和管理。

该工具考虑了尚未到达医院的患者，也提供了比传统方法更详细的信息。该工具不是对当天的整体数字预测，而是包括四小时和八小时内需要多少张床位的概率分布，并每天提供四次预测，通过电子邮件发送给医院规划人员。

创新价值

这可以帮助规划人员管理患者流量——这是一项复杂的任务，涉及平衡计划安排紧急入院的患者。这对于减少取消手术的数量和确保高质量的护理非常重要。这个人工智能工具在帮助管理入院人员及患者流量方面将非常有价值。
它可以响应患者需求和特征的实时变化。这些变化可能是短期的，也可能是长期的。例如，该工具可以适应疾病流行期间的使用。

AI tool was used to optimize the allocation plan of emergency beds

In the study, the team showed that the tool was more accurate than traditional benchmarks used by planners, based on the average number of beds needed on the same day in each of the previous six weeks.

The tool, which also takes into account patients who have not yet reached the hospital, also provides more detailed information than traditional methods. Instead of an overall numerical forecast for the day, the tool includes probability distributions of how many beds will be needed over four and eight hours, and provides forecasts four times a day, emailed to hospital planners.

The research team is now working with UCLH to refine the model so that they can estimate how many beds are needed in different areas of the hospital (for example, in a medical ward or a surgical ward).

Lead author Dr Zella King (UCL Clinical Operations Research Unit and UCL Institute of Health Informatics) said the AI model provided a richer picture of the likely demand for beds throughout the day. They make use of patient data the moment they record it. UCLH patient mobility and emergency preparedness.

The researchers trained 12 machine learning models using patient data recorded by UCLH between May 2019 and July 2021. These models assess the likelihood of each patient being admitted from the Ed, based on data ranging from age and how the patient arrived at the hospital, to test results and number of consultations, and combine these probabilities to produce an overall estimate of the number of beds required.

They then compared the model's predictions with actual admissions between May 2019 and March 2020 and found that they outperformed the traditional method, with central forecasts beating the actual number by an average of four times, compared with 6.5 admissions by the traditional method. After the Covid attack, the researchers were able to adjust the model to account for significant changes in the number of arrivals and the amount of time they spent in the emergency room.

Each of the 12 models focused on data at different time intervals since the patient arrived: the first model focused only on data recorded at arrival, the second model focused on data recorded within the previous 15 minutes, and Model 12 focused on data recorded within 12 hours. This is because factors vary in importance, depending on the amount of time that has elapsed and the amount of other data available. For example, the method of reaching the hospital was an important factor in model 1, but became less important in later models. The researchers found that using 12 models together was more accurate than using fewer models.