VA Uses Machine Learning for Suicide Prevention

The United States Department of Veterans Affairs developed a machine learning model to predict veterans’ risks of suicide.

Reading Time: 3 minutes

Suicide is the 10th leading cause of death among all ages in the United States. An average of 132 Americans commit suicide every day, with nearly one-fifth of them being veterans. According to the United States Department of Veterans Affairs (VA), veteran suicides are on the rise and totaled over 6,000 in 2019. In an attempt to lower veteran suicide numbers, the VA is working with researchers in Berkeley, California to develop Reach Vet, a computer program that uses machine learning to identify veterans at risk of taking their own lives.

Machine learning is a subject within the field of artificial intelligence that teaches computers how to make predictions from data. Put simply, computer programs are fed massive sets of data that are classified into different categories. The program analyzes the characteristics of the data and their classifications, and after studying enough data, it is capable of classifying data on its own. The Reach Vet machine learning model uses data from around 700,000 veterans to learn demographic, medical, and biological trends among veterans who committed suicide and previously suicidal veterans. After it is trained with this massive dataset, it can analyze a veteran’s characteristics and medical records to predict whether they may be suicidal or at risk of becoming suicidal.

Like many other machine learning models, Reach Vet uses a neural network to process inputted data. Neural networks are composed of several layers of artificial neurons, which are coded to resemble the functions of a biological neuron. Similar to a biological neuron, which receives, processes, and sends electrical impulses throughout the body, artificial neurons receive data, perform functions on it, and pass on an output. They are organized in layers and communicate with each other in order to learn about the data they are given, which mimics the way neurons communicate with each other in the human body. To effectively analyze a veteran’s data, Reach Vet uses its neural networks to analyze 61 specific factors, including previous suicide attempts, arthritis, and drug prescriptions. Each layer has a different weight in the model’s predictions, with the heaviest factor being previous suicide attempts. After analyzing each factor, the program assigns a score to each veteran in the database indicating their risk of suicide. Every month, the model flags veterans with the highest 0.1 percent of scores as high-risk.

Once a veteran is flagged as high-risk, his or her local VA clinic is notified, and a doctor contacts the veteran. The doctor explains the meaning of the flag and encourages the veteran to schedule an appointment to discuss safety plans and potential treatment options. Doctors emphasize that Reach Vet’s flag does not mean they are in danger; rather, it is more of a warning sign and an opportunity for veterans to check in with the VA about their health.

Though Reach Vet is changing the face of suicide prevention efforts, it is not perfect. Some veterans, for example, feel wary about having an inanimate machine learning model determine their risk level. In a New York Times article, one veteran who was flagged noted that he doesn’t like the idea of being put on a list by a computer. However, he appreciated that the VA was making an effort to give him the support he needs, and he now regularly goes to therapy at a VA hospital. Another complication with the machine learning model is that it does not make perfect predictions all the time. This is potentially due to the model overfitting the data: in other words, making extreme predictions or finding patterns in the data where there are none. In addition, not all data is perfect: Reach Vet has to combine electronic health records with physical records such as prescriptions and doctor’s notes, which can be difficult to locate and organize. Reach Vet’s miscalculations of high-risk veterans could scare some veterans, who may be flagged inaccurately, and cost the lives of veterans who are overlooked and miss the treatment they need.

Currently, Reach Vet cannot show why a veteran was flagged and what type of care is best for each veteran. Without this extra information, doctors can find it difficult to interpret the model’s findings, but they hope, with future improvements to Reach Vet, to gain a deeper understanding of each patient. Nevertheless, Reach Vet is an important development in the intersection of healthcare and technology, providing suicide prevention care to veterans who might not know they need it.

Stuyvesant: If you are in a crisis or are struggling, call the National Suicide Prevention Hotline at 1-800-273-TALK (8255), for help 24 hours a day, seven days a week. Your call is confidential.