Posgrado en Ciencias de la Ingenieria, Tecnologico Nacional de Mexico/ Instituto Tecnologico de Tijuana, Tijuana, Baja California, Mexico.
Universidad Autonoma de Baja California, Tijuana, Baja Califronia, Mexico.
Comput Biol Med. 2021 Oct;137:104798. doi: 10.1016/j.compbiomed.2021.104798. Epub 2021 Aug 27.
This paper addresses the problem of estimating the response time to a medical emergency, specifically from the Red Cross of Tijuana (RCT), which provides most of the emergency medical services (EMS) in the city of Tijuana, Mexico. For institutions with low funding, such as the RCT, relying on free or open source mapping systems to estimate travel times is necessary but also error prone because these systems are not tuned for ambulance movements within a city. Therefore, this work formulates a supervised machine learning problem where the goal is to predict the difference in travel time between the ground truth travel time provided by a GPS and the approximation offered by two mapping systems, Google Maps (GM) and Open Source Routing Machine (OSRM). To this end, this work develops a new dataset based on the EMS logs of the RCT, considering calls from January 2017 to April 2017. The posed learning problem is solved under different scenarios, including using an off-the-shelf default configuration of a Random Forest classifier, applying a hyper-parameter optimization process and using an Auto Machine Learning (AutoML) system. Considering all of the dataset for GM, test accuracy was 69.6% for the first two learning approaches and 71.6% using AutoML. For OSRM, performance was 64.6%, 65.2% and 66.4% for each of the learning approaches, respectively. Results show that it is possible to predict the level by which a mapping system over or under estimates the true travel time of an ambulance. Finally, the impact of the model is demonstrated by using it to solve the ambulance location problem, with notable differences in ambulance deployments and percentage of double coverage achieved relative to using the standard mapping system. Results show that without correcting the travel time the percentage of double coverage is 83.90%; on the other hand, double coverage reaches 100% when applying travel time correction.
本文解决了估算医疗急救响应时间的问题，具体来说，是来自提华纳红十字会（RCT）的响应时间，该组织提供了墨西哥提华纳市大部分紧急医疗服务（EMS）。对于 RCT 等资金较少的机构来说，依靠免费或开源的地图系统来估算旅行时间是必要的，但也容易出错，因为这些系统没有针对城市内救护车的行驶进行调整。因此，这项工作提出了一个监督机器学习问题，目标是预测 GPS 提供的地面真实旅行时间与两个地图系统（Google Maps [GM] 和 Open Source Routing Machine [OSRM]）提供的近似值之间的旅行时间差异。为此，这项工作基于 RCT 的 EMS 日志开发了一个新数据集，考虑了 2017 年 1 月至 2017 年 4 月的呼叫。在不同场景下解决了提出的学习问题，包括使用随机森林分类器的现成默认配置、应用超参数优化过程和使用自动机器学习（AutoML）系统。考虑到 GM 的所有数据集，前两种学习方法的测试准确率为 69.6%，使用 AutoML 的准确率为 71.6%。对于 OSRM，每种学习方法的性能分别为 64.6%、65.2%和 66.4%。结果表明，可以预测地图系统对救护车真实行驶时间的高估或低估程度。最后，通过使用该模型来解决救护车位置问题，展示了模型的影响，与使用标准地图系统相比，救护车的部署和实现双重覆盖的百分比有明显差异。结果表明，在不修正行驶时间的情况下，双重覆盖的百分比为 83.90%；另一方面，当应用行驶时间修正时，双重覆盖达到 100%。
