Machine Learning to Enhance Force Preservation

AI supporting leadership

>Capt Borinstein is an Intelligence Officer currently assigned to Company B, Marine Cryptologic Support Battalion at Fort Meade, MD. He holds a Master of Science in Data Analytics from the Georgia Institute of Technology.

Although generally not top of mind when considering the Marine Corps’ most pressing future warfighting challenges, mental fitness, and suicide prevention unquestionably remain a chief priority across the DOD. Today, suicide rates among service members are among the highest levels in their recorded history.1

In response to increasing suicide rates, the Marine Corps has resorted to requiring commanders to become more involved in Marines’ lives and applying the risk management process to those subjectively deemed at-risk through the Force Preservation Council (FPC) program. The FPC order directs commanders to “use engaged leadership and risk management guidance … to recognize and intervene early when stressors and potentially risky behaviors first develop in Service members in order to interrupt the chain of events that can lead to an adverse outcome.”2 Unfortunately, the Defense Suicide Prevention Office’s 2020 Annual Suicide Report shows that the Marine Corps’ suicide rate has increased on average since at least 2014, with suicide rates in 2020 being the highest ever recorded in the wake of the Coronavirus Disease 2019 Pandemic.3 This trend suggests that the Marine Corps will continue to battle with and for the mental health of its Marines well into the future, which poses significant challenges to the future force’s ability to remain ready to respond to our Nation’s calling.

Despite the Marine Corps’ good intentions, the FPC program in its initial form was riddled with flaws. One of its primary problems occurred when losing and gaining commands often failed to exchange information on Marines’ past and potential struggles. When they did exchange this information, it was often through informal, non-secure means. Although the Marine Corps FPC Order (MCO 1500.60) required losing commands to “ensure the gaining command is provided the necessary and relevant force preservation information,” there were no mechanisms by which to hold units accountable for failing to comply with policy.4 Such a lack of standardization and security meant that commanders rarely received all the information needed to contextualize Marines’ behaviors and issues and that Marines’ personal data was often put at risk through the unnecessary use of PowerPoints and other informal dissemination mechanisms.

In August 2020, the Marine Corps sought to resolve these issues by adopting the Command Individual Risk and Resiliency Assessment System (CIRRAS), which is essentially a standardized database for FPC data.5 Although certainly an improvement upon the legacy FPC process, CIRRAS will sell the Marine Corps short if it remains only a tool for data storage. Indeed, CIRRAS presents a unique opportunity for the Marine Corps to experiment with using artificial intelligence—and more specifically machine learning—to combat the threat of suicide within its ranks. The Marine Corps should examine the efficacy of using the CIRRAS database in conjunction with supervised classification machine-learning algorithms to help commanders better identify Marines who are most at risk for self-harm.

What is CIRRAS?
CIRRAS is a secure application developed by Marine Corps Systems Command that standardizes the FPC program across the Marine Corps, giving commanders the ability to monitor their Marines’ holistic health and combat readiness.6 It allows commanders and their representatives to input and track the various stressors that Marines regularly experience, including information regarding mental health, relationship disputes, alcohol- and drug-related offenses, and other significant issues that could impact operational readiness.7 Though it offers a new, more secure way of storing and transferring sensitive data about Marines, CIRRAS does not make any fundamental changes to the FPC program.

Although CIRRAS offers the means to standardize and secure Marines’ holistic health information, it does not seem to offer any additional analytical advantage to commanders. In other words, CIRRAS improves commanders’ abilities to securely communicate raw data, but it does not use that data to provide valuable insights to make better decisions.

The primary purpose of collecting standardized data in any capacity is to detect trends and patterns to better inform decision making. Human minds are very good at detecting simple, linear trends in two or three dimensions, but are very limited in their capacity to detect complex, non-linear trends, which can be common in multidimensional datasets such as those involving personal health information.

Machine Learning
Machine-learning algorithms happen to be especially adept at identifying complex, non-linear trends in vast amounts of data. They can take datasets on the scale of thousands of dimensions, identify their most important factors, and detect patterns that no human brain could hope to understand or recognize. These algorithms are regularly used in the private sector to determine which Netflix shows would best suit you, which songs you will most likely enjoy on Spotify, and which products you should next consider purchasing on Amazon.

At its most basic level, machine learning is using past data and consequent outcomes to identify complex patterns, generate models from those patterns, and then combine those models with future input data to quickly deliver predictions of future outputs. The machine-learning algorithms used by tech companies take the data you and others give them, such as browsing activity and personal information, to detect patterns and build statistical models that can quickly calculate high-probability outcomes.

By centralizing and standardizing FPC data in a single database, the Marine Corps has created a venue through which it could use machine-learning algorithms to identify under-the-surface trends common among Marines who have expressed suicidal or other life-threatening tendencies. If provided with the right types of data, these algorithms could prove useful in providing commanders indications of Marines who are more likely to engage in self-threatening behavior.

Among the many different types of machine-learning algorithms, the most useful for the purposes of predicting future behavior are classification prediction algorithms. These types of algorithms are trained to predict specific categorical outcomes (green/yellow/red), and not numerical ones (1, 2, 3). Among the most popular types of classification prediction algorithms are decision trees, random forests, k-nearest neighbor classifiers, logistic regression, and support vector machines. The Marine Corps should experiment with these types of algorithms to determine whether any of them can effectively predict Marine behavior.

Issues and Requirements
Using machine learning to make impactful decisions in Marines’ lives obviously presents several potential problems. The data science and tech worlds are alight with debate over the moral and ethical use of machine-learning algorithms with others’ personal information. Moreover, no model or algorithm is perfect and, if not properly understood, can result in unfounded dependence on “the numbers” and remove commanders’ responsibility to use their judgment.

First, one should note that no model is infallible. Models are abstract representations of reality and are optimized to represent historical data. They are susceptible to developing a narrow focus and will always produce some measure of error. No model or algorithm can perfectly describe previous forms of reality nor perfectly predict future ones.

Because of this, commanders using mathematical models to make decisions must remember that such models are tools designed to supplement decision making and should never replace well-informed human leadership and judgment. It seems too often that we settle for reducing complicated situations into PowerPoint slides with boxes colored green, yellow, or red. No Marine’s personal situation can be adequately captured by a simple color, and we should be wary of similar behavior when using other models to predict which Marines are most susceptible to suicidal behavior. Instead, commanders should use such tools to identify who they should be spending more time observing.

All prediction algorithms produce false positives and false negatives. The Marine Corps must avoid a zero-tolerance approach when it comes to using machine learning and artificial intelligence of all types. Tools that use such technologies are designed to inform better and faster decisions but are never intended to generate decisions in lieu of humans.

Garbage in, garbage out is a common saying among data scientists. Because machine-learning algorithms live on the data that they are given, poor data quality can easily result in models which fail to adequately reflect reality. Leaders responsible for inputting data into CIRRAS must do so properly. The notion of no data in is also worthy of consideration. Given that prediction of at-risk Marines is the ultimate goal, a lack of data on risk factors means some Marines could slip through the cracks.

Data used in machine learning must also be computable, meaning that it should be standard throughout the dataset (think multiple choice responses or numerical data with common formatting). Supervised classification learning algorithms work by identifying which characteristics were most prevalent among Marines who expressed self-harming inclinations, generating a model by appropriately weighing each of those characteristics based on their correlation with the outcome, and then applying that model to other Marines as needed. To make this work, however, these algorithms require standard data values, especially for the metric in question, which in this instance is whether a Marine has demonstrated a predisposition for self-harm. Machine-learning algorithms cannot easily interpret free-response data without additional processing, which often involves manual interaction. CIRRAS must provide standard datasets to generate effective models.

Not all models work well and there is no guarantee that these models will provide any value at all. It is very possible that none of the models listed would be able to accurately predict which Marines are most susceptible to self-harm, and in doing so could add unnecessary noise to an already-complicated FPC system. If, however, these models can generate correct predictions even as low as 50 percent of the time, they could prove very valuable to commanders.

Conclusion
In recent years, Marine Corps dialogue has become consumed with some of the Nation’s favorite tech buzzwords: artificial intelligence, machine learning, big data, and the like. Nevertheless, we have yet to find ways to implement these at scale in the same way multi-billion-dollar corporations have been doing for years. There is little question that we should be researching and experimenting with means to harness the power of these technological advancements. In reality, however, reluctance to adapt quickly and try new things at middle and lower echelons demonstrates that research in these fields may not truly be a top priority.

Exploring the use of machine learning in conjunction with CIRRAS’ database offers an easy opportunity for the Marine Corps to showcase its long-held reputation as the Nation’s most innovative force. Further research on this topic may prompt widespread use of this technology and could prove valuable to commanders by quickly providing automated actionable data in one of the Pentagon’s top challenges: service member mental health. If our people are truly our greatest strength, then we should leverage every advantage, technological or otherwise, to their benefit and that of the Naval Service.

Notes

1. U.S. Department of Veteran’s Affairs, 2020 Veteran Suicide Prevention Annual Report (Washington, DC: 2020).

2. Headquarters Marine Corps, MCO 1500.60 Force Preservation Council (Washington, DC: August 2016).

3. Under Secretary of Defense for Personnel and Readiness, U.S. Department of Defense, Calendar Year 2020 Annual Suicide Report (Washington, DC: 2020).

4. Stephen Losey, “Military Deaths by Suicide Jumped 25% at End of 2020,” Military.com, https://www.military.com/daily-news/2021/04/05/military-deaths-suicide-jumped-25-end-of-2020.html; and MCO 1500.60 Force Preservation Council.

5. Headquarters Marine Corps, “Announcement and Implementation of the Command Individual Risk and Resiliency Assessment System (CIRRAS),” Marines, August 12, 2020, https://www.marines.mil/News/Messages/Messages-Display/Article/2310545/announcement-and-implementation-of-the-command-individual-risk-and-resiliency-a.

6. Marine Corps Systems Command, “Marine Corps Develops Secure App to Monitor Holistic Health and Combat Readiness of Marines,” Marines, February 11, 2021, https://www.marines.mil/News/News-Display/Article/2500948/marine-corps-develops-secure-app-to-monitor-holistic-health-and-combat-readiness.

7. Ibid.

Share this: