Journal Watch: AEDs by Drone

Reviewed This Month

Machine Learning-Based Dispatch of Drone-Delivered Defibrillators for Out-of-Hospital Cardiac Arrest

Authors: Chu J, Leung KHB, Snobelen P, Nevils G, Drennan IR, Cheskes S, Chan TCY

Published in: Resuscitation, 2021 Feb 22; 162: 120–7 

In this month’s Journal Watch we review a fascinating study that utilized machine learning to develop and evaluate dispatch rules for drone-delivered AEDs. Welcome to the future! 

As you might suspect, having drones deliver AEDs directly to scenes could be a great way to decrease time to first shock. In fact, recent studies have shown using drones in this way is feasible. However, dispatch policies for AED-carrying drones have not been well studied. Understanding which drones are available, where they are at any given time, and successfully delivering an AED to the scene of a cardiac arrest before the ambulance arrives requires a lot of coordination. 

Mathematical Optimization

The authors of this study used a mathematical optimization model to identify the best locations for the drone bases in the regional municipality of Peel, Ontario, to reduce the times to AED arrival on scene. The Peel region includes urban, suburban, and rural areas. The drone base locations were EMS, fire, and police stations. Drone response time was defined as the “interval between drone take-off and the AED hitting the ground at the scene of a suspected cardiac arrest location, using the straight-line distance between the two locations.” 

A lot goes into this calculation, including consideration of the drone ascent time (10 seconds), horizontal acceleration (19.6 m/s²), cruising speed (20 m/s), and descent velocity (2 m/s). Also, notice that the definition states the AED hit the ground. The drones did not land on scene but rather were dropped in protective cases from a height of almost 30 feet! Of course, the time it took for the AED to hit the ground after it was dropped (4.25 seconds) also had to be considered. 

The drone response time was compared to a predicted ambulance response time plus a response-time buffer to account for ambulance delays due to factors such as traffic congestion. Ambulance response times were calculated using data for all suspected cardiac arrest calls (defined as those dispatched as either cardiac arrest or unconscious) from January 1, 2015 to December 31, 2019. The arrival-on-scene time was used for these calculations, rather than the “at patient” time, because the authors assumed the time needed for a bystander to retrieve the AED and bring it to the patient was similar to the time it would take an EMS provider to arrive at the patient’s side from an ambulance. 

Statistical models were constructed to predict ambulance response times, adjusting for the distance of the ambulance from the scene, the day of week, and the time of day. Drone dispatch rules were developed dictating that the drone would only be dispatched if it were predicted to arrive no more than x seconds after the ambulance. The x was evaluated in 30-second increments from 0–150 seconds. 

Dispatch rules were also evaluated based on drone availability at the time of the cardiac arrest. To estimate drone availability, the authors assumed that, once dispatched, a drone would be occupied for the time it took to get from its base to the scene as well as the time it took to get from the scene back to its base. Thirty minutes was added to the round-trip travel time to allow for the drone to charge and be equipped with a new AED prior to its next mission. 

Each dispatch rule considered dispatching the drone closest to the scene. If the closest drone was unavailable, the next-closest drone was considered. If all drones were unavailable, then no drone was dispatched to the scene. Dispatch rules were also compared to a policy of simply dispatching the closest available drone to every suspected cardiac arrest call without consideration of arrival prior to EMS. 

Finally, because a decision to dispatch a drone can impact the availability of subsequent drones, the authors calculated a metric they called harmful false positives. A false positive was defined as dispatching of a drone to the scene that arrived after EMS. A harmful false positive was a false positive where a drone was occupied but, if available, that drone would have improved response time for a subsequent suspected out-of-hospital cardiac arrest. 

Results

There were 3,573 cardiac arrests in the analysis. Median patient age was 69. The population was 66% male. Approximately 14% of arrests occurred in public locations. There were 28% that were bystander-witnessed, and 25% had bystander-initiated CPR. Average, median, and 90th-percentile ambulance response times were 6.2, 5.8, and 9.5 minutes, respectively. All drone dispatch rules resulted in average, median, and 90th-percentile response times of 2–3 minutes less than the ambulance response times. 

The percentage of suspected out-of-hospital cardiac arrest calls where a dispatched drone would improve response times ranged from 62%–66% among all drone dispatch rules evaluated. The percentage of calls where the closest drone was busy was about 2% among all dispatch rules, and the number of harmful false positives ranged from 3–4 to 15–20. 

Conclusion

This was an interesting study that found that dispatch rules for AED-carrying drones that only dispatch drones predicted to reach the scene before EMS can achieve similar results to simply dispatching a drone to all suspected cardiac arrest calls. 

As with all studies, there are some limitations in this one. The study authors did not adjust for all factors, such as weather or obstacles in drone flight paths. They also did not evaluate response times for other first responders.

Nevertheless, they completed a very detailed and thorough study of drone dispatch policies that significantly adds to the literature. There will no doubt be more and more research that will help us imagine EMS in the future.

Antonio R. Fernandez, PhD, NRP, FAHA, is a research scientist at ESO and serves on the board of advisors of the Prehospital Care Research Forum at UCLA.