Artificial Intelligence for Medical Evacuation in Great-Power Conflict
It is 4:45 a.m. in southern Afghanistan on a hot September day. A roadside improvised explosive device has just gone off and was followed by the call, “Medic!” Spc. Chazray Clark stepped right on the bomb, losing both of his feet and his left forearm. Clark’s fellow soldiers immediately provided medical care, hoping he might survive. After all, the unit’s forward operating base was only 1.5 miles away, and it had a trained medical evacuation (medevac) team waiting to respond to an event of this nature.
A 9-line medevac request was submitted just moments after the explosion occurred, and Clark’s commanding officer, Lt. Col. Mike Katona, had been assured that a medevac helicopter was en route to the secured pickup location. Unfortunately, that was not the case; the medevac team was still awaiting orders 34 minutes after the call for help was transmitted.
Although the casualty collection point was secure, the current policy in place required an armed gunship to escort the medevac helicopter, but none were available. It wasn’t until 5:24 a.m. that the medevac helicopter started to fly toward the pickup location, but it was too late. Clark arrived at Kandahar Air Field medical center at 5:49 a.m. and was pronounced dead just moments later.
No one knows if Clark would have survived his wounds if he had received advanced surgical care earlier, but most people would agree that his chances of survival would have been much higher. What went wrong? Why wasn’t an armed escort available during this dire time? Are the current medevac policies in place outdated? If so, can artificial intelligence improve upon current practices?
With limited resources available, the U.S. military ought to carefully plan how medevac assets will be utilized prior to and during large-scale combat operations. How should resources be positioned now to maximize medevac effectiveness and efficiency? How can ground and air ambulances be dynamically repositioned throughout the course of an operation based on evolving, anticipated locations and intensities for medevac demand (i.e., casualties)? Moreover, how should those decisions be informed by operational restrictions and (natural and enemy-induced) risks to the use of ground and aerial routes as well as evacuation procedures at the casualty collection points? Finally, whenever a medevac request is received, which of the available assets should be dispatched, considering the anticipated future demands of a given region?
The military medevac enterprise is complex. As a result, any automation of location and dispatching decision-making requires accurate data, valid analytical techniques, and the deliberate integration and ethical use of both. Artificial intelligence and, more specifically, machine-learning techniques combined with traditional analytic methods from the field of operations research provide valuable tools to automate and optimize medevac location and dispatching procedures.
The U.S. military utilizes both ground and aerial assets to perform medevac missions. Rotary-wing air ambulances (i.e., HH-60M helicopters) are typically reserved for the most critically sick and/or wounded, for whom speed of evacuation and flexibility for routing directly to highly capable medical treatment facilities are essential to maximizing survivability. Ground ambulances cannot travel as far or as fast as air ambulances, but this limitation is offset by their greater proliferation throughout the force.
Machine Learning to Predict Medevac Demand
More than 4,500 U.S. military medevac requests were transmitted between 2001 and 2014 for casualties occurring in Afghanistan. The location, threat level, and severity of casualty events resulting in requests for medevac influence the demand for medevac assets. Indeed, it is likely that some regions may have higher demand than others, requiring more medevac assets when combat operations commence. A machine-learning model (e.g., neural networks, support vector regression, and/or random forest) can accurately predict demand for each combat region by considering relevant information, such as current mission plans, projected enemy locations, and previous casualty event data.
Effective machine-learning models require historical data that is representative of future events. Historical data for recent medevac operations can be obtained from significant activity reports from previous conflicts and the Medical Evacuation Proponency Division. For example, one study utilizes Operation Iraqi Freedom flight logs obtained from the Medical Evacuation Proponency Division to approximate the number of casualties at a given location to help identify the best allocation(s) of medical assets during steady-state combat operations. Open-source, unclassified data also exist (e.g., International Council on Security and Development, Defense Casualty Analysis System, and Data on Armed Conflict). Although historical data may not exist for every potential future operating environment, it can still be utilized to generalize casualty event characteristics. For example, one study models the spatial distribution of casualty cluster centers based on their proximity to main supply routes and/or rivers, where large populations are present. It utilizes Monte Carlo simulation to synthetically generate realistic data, which, in turn, can be leveraged by machine-learning practitioners to predict future demand.
Demand prediction via a machine-learning model is essential, but it is not enough to optimize medevac procedures. For example, consider a scenario wherein the majority of demand is projected to occur in two combat regions located on opposite sides of the area of operations. If there are not enough medevac resources to provide a timely response for all anticipated medevac demands in both of those regions, where should medevac assets be positioned? Alternatively, consider a scenario wherein one region needs the majority of medevac support at the beginning of an operation, but the anticipated demand shifts to another region (or multiple regions) later. Should assets be positioned to respond to demand from the first region even if it makes it impossible to reposition assets to respond to future demand from the other regions in a timely manner? How do these decisions impact combat operations in the long run?
Optimization Methods to Locate, Dynamically Relocate, and Dispatch Medevac Assets
How do current decisions impact future decisions? The decisions implemented throughout a combat operation are interdependent and should be made in conjunction with each other. More specifically, to create a feasible, realistic plan, it is necessary to make the initial medevac asset positioning decisions while considering the likely decisions to dynamically reposition assets over the duration of an operation. Moreover, every decision should account for total anticipated demand over all combat regions to ensure the limited resources are managed appropriately.
How many possible asset location options are there for a decision-maker to consider? As an example, suppose there are 20 dedicated ground and aerial medevac assets that need to be positioned across six different forward operating bases. Moreover, suppose decisions regarding the repositioning of these assets occur every day for a 14-day combat operation. For any day of the two-week combat operation, any of the 20 assets can be repositioned to one of six operating bases. Without taking into consideration distances, availability, demand constraints, or multiple asset types, the approximate number of options to consider is over 10,000! It is practically impossible for an individual (or even a team of people) to identify the optimal positioning policy without the benefit of insight provided by quantitative analyses.
Whereas a machine-learning model can predict when and where demand is likely to occur, it does not inform decision-makers where to position limited resources. To overcome this, operations research techniques — more specifically, the development and analysis of optimization models — can efficiently identify an optimal policy for dynamic asset location strategies for the area of operations over the entire planning horizon. The objectives of an optimization model define the quantitatively measured goal that decision-makers seek to maximize and/or minimize. For example, decision-makers may seek to maximize demand coverage, minimize response time, minimize the cost of repositioning assets, and/or maximize safety and security of medevac personnel. The decisions correspond to when, where, and how many of each type of asset is to be positioned across the forward operating bases for the planned combat operation, as well as how assets are dispatched in response to medevac requests. It is necessary to have information about unit capabilities and dispositions to accurately inform an optimization model. This information includes the number, type, and initial positioning of medevac assets as well as the projected demand locations, threat levels, and injury severity levels. An optimization model also considers operational constraints to ensure a feasible solution is generated. These constraints include travel distances and time, fuel capacity, forward operating base capacity, and political considerations.
Medevac assets may need to be dynamically repositioned (i.e., relocated) across different staging facilities, especially as disposition and intensity of demand changes, despite the long-term and strategic nature of combat operations. For example, it may be necessary to reposition assets from forward operating bases near combat regions with lower projected demand to bases near regions with higher projected demand. Moreover, it is important to consider projected threat and severity levels when determining which type of assets to position. For example, it may be beneficial to position armed escorts closer to combat regions with higher projected threat levels. Similarly, air ambulances should be positioned closer to combat regions with higher projected severity levels (i.e., life-threatening events). Inappropriate positioning of assets may result in delayed response times, increased risks, and decreased casualty survivability rates. One way to determine the location of medevac assets is to develop an optimization model that simultaneously considers the following objectives: maximize demand coverage, minimize response time, and minimize the number of relocations subject to force projection, logistical, and resource constraints. Trade-off analysis can be performed by assigning different weights (i.e., importance levels) to each objective considered. Given an optimal layout of medevac assets, another important decision that should be considered is how air ambulances will be dispatched in response to requests for service.
The U.S. military currently utilizes a closest-available dispatching policy to respond to incoming requests for service, which, as the name suggests, tasks the closest-available medevac unit to rapidly evacuate battlefield casualties from point of injury to a nearby trauma facility. In small-scale and/or low-intensity conflicts, this policy may be optimal. Unfortunately, this is not always the case, especially in large-scale, high-intensity conflicts. For example, suppose a non-life-threatening medevac request is submitted and only one air ambulance is available. Moreover, assume high-intensity operations are ongoing and life-threatening medevac requests are expected to occur in the near future. Is it better to task the air ambulance to service the current, non-life-threatening request, or should the air ambulance be reserved for a life-threatening request that is both expected and likely to occur in the near future?
Many researchers have explored scenarios in which the closest-available dispatching policy can be greatly improved upon by leveraging operations research techniques such as Markov decision processes and approximate dynamic programming. Dispatching decision-makers (i.e., dispatching authorities) should take into account a large number of uncertainties when deciding which medevac assets to utilize in response to requests for service. Utilizing approximate dynamic programming, military analysts can model large-scale, realistic scenarios and develop high-quality dispatching policies that take into account inherent uncertainties and important system characteristics. For example, one study shows that dispatching policies based on approximate dynamic programming can improve upon the closest-available dispatching policy by over 30 percent in regards to a lifesaving performance metric based on response time for a notional scenario in Syria.
Ethical Application Requires a Decision-Maker in the Loop
Optimization models may offer valuable insights and actionable policies, but what should decision-makers do when unexpected events occur (e.g., air ambulances become non-mission capable) or new information is obtained (e.g., an unmanned aerial vehicle captures enemy activity in a new location)? It is not enough to create and implement optimization models. Rather, it is necessary to create and deliver a readily understood dashboard that presents information and recommended decisions, the latter of which are informed by both machine learning and operations research techniques. To yield greater value, such a dashboard should allow its users (i.e., decision-makers) to conduct what-if analysis to test, visualize, and understand the results and consequences of different policies for different scenarios. Such a dashboard is not a be-all and end-all tool. Rather, it is a means for humans to effectively leverage information and analyses to make better decisions.
The future of decision-making involves both artificial intelligence and human judgment. Whereas humans lack the power and speed that artificial intelligence can provide for data processing tasks, artificial intelligence lacks the emotional intelligence needed when making tough and ethical decisions. For example, a machine-learning model may be able to diagnose complex combat operations and recommend decisions to improve medevac system performance, but the judgment of a human being is necessary to address intangible criteria that may elude quantification and input as data.
Whereas the effectiveness and efficiency of the U.S. military medevac system has been very successful for recent operations in Afghanistan, Iraq, and Syria, future operating environments may be vastly different from where the United States has been fighting over the past 20 years. Artificial intelligence and operations research techniques can combine to create effective decision-making tools that, in conjunction with human judgment, improve the medevac enterprise for large-scale combat operations, ultimately saving more lives.
The Way Forward
The Air Force Institute of Technology is currently examining a variety of medevac scenarios with different problem features to determine both the viability and benefit of incorporating the aforementioned artificial intelligence and operations research techniques within active medevac operations. Once a viable approach is developed, the next step is to obtain buy-in from senior military leaders. With a parallel, macroscopic-level focus, the Joint Artificial Intelligence Center, the Department of Defense’s Artificial Intelligence Center of Excellence, is currently seeking new artificial intelligence initiatives to demonstrate value and spur momentum to accelerate the adoption of artificial intelligence and create a force fit for this era.
Capt. Phillip R. Jenkins, PhD, is an assistant professor of operations research at the Air Force Institute of Technology. His academic research involves problems relating to military defense, such as the location, allocation, and dispatch of medical evacuation assets in a deployed environment. He is an active-duty Air Force officer with nearly eight years of experience as an operations research analyst.
Brian J. Lunday, PhD, is a professor of operations research at the Air Force Institute of Technology who researches optimal resource location and allocation modeling. He served for 24 years as an active-duty Army officer, both as an operations research analyst and a combat engineer.
Matthew J. Robbins, PhD, is an associate professor of operations research at the Air Force Institute of Technology. His academic research involves the development and application of computational stochastic optimization methods for defense-oriented problems. Robbins served for 20 years as an active-duty Air Force officer, holding a variety of intelligence and operations research analyst positions.
The views expressed in this article are those of the authors and do not reflect the official policy or position of the U.S. Air Force, the Department of Defense, or the U.S. government.