Ambulance response time can determine whether a patient requiring urgent care survives or not. Faster response times translate to higher survival rates.
According to a study conducted by Wilde, a one-minute increase in ambulance response time resulted in an 8-17% increase, on average, in mortality for patients in cardiac arrest or suffering a stroke.1 In another study by Peyravi, ambulances with a two-minute shorter response time had a mortality rate of 1.5%, 1.1-percentage points lower than the average ambulance response time.2
These studies highlight the importance of constricting ambulance response times. The challenge, however, is to do just that.
EMS organizations know that the key to reduce response time is to deploy ambulances near where the next emergencies are likely to happen. However, predicting where the next emergencies will occur is the main challenge. Some EMS organizations can generate a demand analytics report based on call volume associated with specific shifts, areas and time of day. They will simply look over the report and make a gut decision to predict when and where the next emergencies will occur. Other EMS organizations who are serious about reducing response time even employ statisticians to analyze historical demand, but even that is not accurate enough. That is why AI and machine learning are deployed to assist.
Demand prediction software using AI already exists. However, because they are commercial products, their AI methodology and predictive model is a business secret. This article sheds some light on various methodologies and predictive models that scientists from around the world have used.
In the first half of this article, you’ll read about four different approaches to shortening ambulance response time. In each approach, you’ll also find the results of various machine learning models.
Then, you’ll read about two commercially viable EMS software technology that are already successfully using AI and machine learning. Both help EMS improve planning and scheduling, freeing up resources so they can better respond to emergencies.
Ambulance Demand Prediction
In a study conducted by Kerakos3, they studied the possibility of using machine learning to build a “high-resolution” predictor, where results will eventually be used to predict ambulance demand so ambulances can be deployed closer to potential emergencies.
But what is “high-resolution?” In short, it is clusters, or sub-regions, based on geography. This method seeks to improve on an older method, where the city of Melbourne was divided into one square kilometers (2016). It was found that there was on average zero demand in more than 99% of grid cells over any given 1-hour period.4
For this study, Stockholm County, Sweden, was divided into clusters based on unsupervised machine learning algorithms that take into consideration the complex urban geography and other factors as various data points.
Four clustering algorithms were used. These algorithms takes into consideration the population density of a specific region using residential address, merging smaller clusters with neighboring clusters based on certain rules, building type (hospitals, schools, etc.). The goal is to know what types of activities takes place at a certain region and where people are at a given time-interval. Finally, dispatch data containing location coordinates were used to determine if dispatch were needed based on building type, time, and other factors.
Then, using those clusters, two models were used to predict demand over these clusters – logistic regression model and simple baseline model. The results support the argument that the logistic regression has better predictive power. However, that is not the case because evaluation results show that logistic regression is better at predicting positive samples (when a dispatch is actually needed) versus negative samples.
Although the logistic regression model outperformed the baseline in total (97.9% vs. 94.1), it was worse at predicting when dispatches actually happened (positive samples).3 The results were not as expected, and further investigation is needed.
Predicting Daily Ambulance Demand by Region
This study conducted by Lin5 used 10-year ambulance demand data recorded by Singapore Civil Defense Force (SCDF). Unlike the previous method of using clusters, this study used a massive dataset based on regional characteristics, historical ambulance records, and incident records. Their goal is to predict next-day demand for ALL regions, so staffing and other preparations could be easier.
For regional characteristics, spatial attributes were used such as demographic information (population, age within a region, etc.) and district classification (financial, residential, etc.). Temporal attributes were also used, such as day of the week and day of the month. Historical demands were also used, which take into consideration disease outbreaks and/or large sporting events.
Other datasets contain details of each ambulance call, which include time of incident, its classification, patient status, patient age, ambulance origin and destination hospital, gender, exact patient location, etc.
Using these massive datasets, this study used and compared several machine learning techniques.
Based on these results, Linear Regression and Light Gradient Boosting Machine (LightGBM) were the better performing models. LightGBM is one of the most efficient and high-performing gradient-boosting decision tree methods.5 It encompasses various individual regression trees and is more suitable in capturing non-linear dependencies, such as combining socioeconomic factors with other factors within a region (although socioeconomic features did not improve prediction performance).
However, LightGBM method is preferred because Linear Regression may not be suitable due to largely varying coefficients and not fitting for other datasets.5
Dispatching Using Decision Tree
This study was conducted using the knowledge, expertise and dispatch policy for the Dutch EMS region of Babant Zuid-Oost (BZO) in the Netherlands.6 The main goal is to help EMS deploy resources more efficiently so they can meet the national target of having response time within 15 minutes for highly urgent ambulance requests 95% of the time.
Unlike other approaches mentioned earlier that focused on predicting ambulance demand, this study focused on making better dispatch decisions. The goal is to reduce response time not by predicting demand, but by having the machine learn from the experiences and past decisions of the dispatchers. For instance, a dispatcher should immediately re-dispatch a transport that’s on its way to a “moderately urgent” to a “highly urgent” request. However, the decision to re-dispatch is not always clear-cut. This element of human decision and experience to re-dispatch is part of the “dispatch decision tree” for machine learning to make better dispatch decisions.
For the machine to learn how to make dispatch decisions, an initial decision tree is mapped out based on written dispatch policies that dispatchers had to abide to. Then, the decision tree is further improved with the historic dispatch decisions (experiences) of the dispatcher based on their unique situation. Data used in the dispatcher’s “experiences” include all information available to the dispatcher when making the dispatch decision, including whether ambulances are idle, driving, waiting, unloading patient at hospital, or dispatched to a less urgent request but did not arrive at the location. Then, simulations based on realistic spatial characteristics were used to enhance the model, factoring in the dynamic attributes of real-life events.
The results were significant. On-time response performance for highly urgent requests increased by 0.77 of a percentage point. The study claims that this performance gain is equivalent of adding more than seven weekly ambulance shifts.6
Ambulance Redeployment Using Scoring and Algorithms
This study is conducted in the city of Tianjin, China, which has a population of over 13 million. In Tiajin, “over one hundred thousand citizens are caught in emergent accidents or diseases every year, in need of ambulances to transport them to hospitals timely.”7 Emergent accidents include traffic accidents, whereas emergent diseases include cerebral hemorrhages and heart attacks. Their goal of this study is to shorten response time by redeploying ambulances at the right stations to shorten response times.7
To get to this goal, they needed to simultaneously consider each station’s spatial and temporal factors. These factors include (1) expected number of patients near a station, (2) number of ambulances a station currently has, (3) travel time and distance of the ambulance to a station from patient drop-off spot, and (4) the current status and destination of occupied ambulances.
What they used was a “deep neural network,” also called the deep score network, that incorporates all of the above dynamic factors of a station to generate a score, and an ambulance will be redeployed to a station with the highest score. Deep reinforcement learning framework was then deployed to learn this deep score network in an unsupervised manner.
For evaluation and simulation, real-world data is leveraged, which included (1) EMS request records, (2) ambulance stations, (3) hospitals, and (4) road networks. Below is a visual of the spatial (left) and temporal (right) distribution of EMS request in Tianjin.
When the deep score network was compared to traditional machine learning models and state-of-the-art baseline models, including ridge regression, lasso regression, and logistic regression, the deep score network achieved a clear advantage overall because of the complex dynamic factors. The deep score network can save ~100 seconds (~20%) of average pickup time of patients and improve the ratio of patients being picked up within 10 minutes from 0.786 to 0.838.7
IBM: Keeping Ambulance Transport On Schedule
When estimating ambulance transport time, it’s easy to estimate travel time with mapping apps. However, planners must also consider the required time to physically pickup and drop off patients, the patient’s condition, age, weight, and what floor the patient must be delivered to, whether wheelchair or stretcher is needed, and many other factors. IBM and its business partner IT’S…B2B developed an ambulance scheduling system called SoTras, a solution for transport logistics and patient booking.
The system uses IBM predictive analytics and machine learning to analyze 200,000 ambulance trips to estimate the total time of an ambulance transport, including the time needed to get a patient in and out of an ambulance. Datasets analyzed include patient demographics, medical conditions and the required medical services needed.
Their predictive analytics that IBM uses include many that we have mentioned in this article – nonlinear regression, neutral networks, support vector machines and decision trees. What’s more, they claim that their ambulance transport times prediction can be predicted with 98% accuracy.
Traumasoft: Maximizing Non-Emergent Medical Transports (NEMT) Capacity While Minimizing Disruptions
Even with all the demand prediction models mentioned earlier, unanticipated transport requests can still disrupt your entire dispatch schedule. Not only that, rerouting trips to adjust to sudden changes expends manpower, is error-prone, and can increase the mix of dead leg trips. Traumasoft, who provides an all-in-one EMS management system, developed real-time routes for NEMT scheduling that addresses this issue.
The system uses machine learning and artificial intelligence to enable the programmatic and dynamic routing and re-routing of trips, helping dispatchers create the best routes in seconds. Datasets from 100,000+ ambulance trips were used to establish baselines and artificial intelligence algorithms.
The results and benefits are evident. Organizations can avoid late, early and unneeded pickups. Staffing can be planned accurately, and route recalculation for additional trips can be accomplished dynamically. Optimized routes lower fuel costs, decrease vehicle idle-time, and free up resources for EMS to meet their 911 performance goals.
Conclusion
Improved ambulance dispatch and response time is essential to the well-being of the community. Further improvements in this area benefits all of us.
Although machine learning models that improve ambulance response time are promising, their results have yet to be perfected. Before implementing these models to real life-and-death situations, more development and data gathering are required. Further, the combined learnings of each of these approaches will pave the way for data scientists to design and create better models in the future.
Nevertheless, AI and machine learning have already began simplifying EMS management – evidenced by the studies and the two commercial products mentioned above. The momentum in this field will continue, and we can expect a future (notwithstanding current labor shortages) where EMS mortality rates will decline and servicing NEMT requests will become even more efficient.
References
1. Wilde ET. Do Emergency Medical System Response Times Matter for Health Outcomes? 2012. doi: 10.1002/hec. 2851.
2. Peyravi M, Khodakarim S, Ortenwall P, Manesh AK. Does temporary location of ambulances (“fluid deployment”) affect response times and patient outcome? 2015. doi: 10.1186/s12245-015-0084-1.
3. Kerakos E, Lindgren O, Tolstoy V. Machine learning for ambulance demand prediction in Stockholm County. Stockholm, Sverige: Skolan for Industriell Teknik Och Management; 2020.
4. Zhou Z, Matteson DS. Predicting Melbourne ambulance demand using kernel warping. In: The Annals of Applied Statistics 10.4 (2016), pp. 1977–1996. doi: 10.1214/16-aoas961.
5. Lin AX, Ho AFW, Cheong KH, Li Z, Cai W, Chee ML, Ng YY, Xiao X, Ong MEH. Leveraging machine learning techniques and engineering of multi-nature features for national daily regional ambulance demand prediction. Singapore: International Journal of Environmental Research and Public Health; 2020.
6. Theeuwes N, van Houtum GJ, Zhang YQ. Improving ambulance dispatching with machine learning and simulation. Netherlands: Eindhoven University of Technology.
7. Ji SG, Yu Z, Wang ZY, Li TR. A deep reinforcement learning-enabled dynamic redeployment system for mobile ambulances. Proc. ACM Interact. Mob. Wearable Ubiquitous Technol., Vol. 3, No. 1, Article 15; 2019.
