Researchers have created an algorithm that correctly predicts drinking and sobriety among individuals experiencing homelessness four out of five times, using it to develop a smartphone app that delivers personalized intervention messages to people at risk of alcohol misuse.

The results on imminent drinking were described in a study published April 19 in the Journal of Substance Abuse Treatment. Researchers in Texas and Oklahoma surveyed 78 adults who were experiencing homelessness and living at a shelter. The team collected 4,557 completed assessments over the course of four weeks, asking simple prompts, written at a sixth-grade reading level, about craving, mood, alcohol availability and more. 

Adults experiencing homelessness are eight times more likely than the general population to abuse alcohol, but there is limited knowledge on what triggers their alcohol consumption. While treatment for substance use has been proved to lower rates of drinking, many adults drop out of such programs, citing a lack of motivation or transportation as two of the main reasons. 

"Alternative treatments that are more responsive to the needs of adults experiencing homelessness are sorely needed," the authors note in the paper, and this rings even more true during a global pandemic. Lead author and public health professor Scott T. Walters told The Academic Times, "I'm really worried about what's going to happen in three or four years' time. Drinking tends to go up during national disasters, but this is totally different; it's not like a flood, where people drink for a few weeks." Walters is concerned for the homeless population because substance abuse keeps them from reintegrating into society and increases their risk of experiencing violence or dying.

In the past, Walters did studies with college students, who, he said, have a very predictable pattern of drinking. Young adults in school are relatively stable, Walters said, and tend to drink most heavily on Thursday, Friday and Saturday nights while around their friends. Even if college drinkers aren't the most responsible under the influence, they are comfortable with technology and can easily complete the study from their own phones, unlike many adults in homeless shelters. 

Walters became interested in understanding adults experiencing homelessness because of how unpredictable their daily routines can be, compared with those of students. The individuals he now works with are most concerned with satisfying basic needs, such as food and shelter, on a daily basis, all while living outdoors. Though these complicated interactions can be impacted by alcohol use, the drinking habits of adults experiencing homelessness are not often studied. Walters saw this huge gap in the research as an opportunity to help people who are overlooked. "The approach is more difficult, [but] it's really rewarding to do this 'Wild West' side of research, where people don't normally go," Walters said.

The researchers created what they described in the paper as "phone-based prompts [to be] answered several times a day in a participant's natural environment." Their algorithm looked at data from these momentary assessments to assess a participant's risk of drinking. The authors were pleased to find that the algorithm "predicted 82% of imminent drinking episodes within 4 hours of the first drink of the day, and correctly identified 76% of non-drinking episodes." Three of the most relevant predictors were an urge to drink, easy access to alcohol and feeling depressed. 

One unexpected finding the authors noted was the pattern that emerged in adults' consumption of alcohol. The researchers originally thought that the drinking habits of participants would be random but soon noticed that most adults drank in the early to late afternoon. Walters attributes this pattern to the structure of the participants' day, which, somewhat surprisingly, "had a certain rhythm to it," as the shelter provides three meals a day and a counseling program.

Another surprising result, according to Walters, was how well the assessments were correlated to alcohol intake. He notes that academic reviewers were skeptical that adults would truthfully report their intoxication levels and even whether they'd had a drink. A transdermal alcohol sensor — similar to those people may wear while on parole for crimes involving alcohol — proved the skeptics wrong. This device takes a sample of a person's sweat at regular intervals and analyzes it for the presence of alcohol, which the researchers reviewed at the end of the study. The authors found self-reporting to be consistent with the alcohol monitors even when adults were intoxicated, further validating their data.

Future research on predictors of drinking is needed to assess a more diverse population, the authors note. Nearly 85% of adults in the current study were male, and 65% were Black, which is not an accurate reflection of homelessness in the U.S. or around the world.

The authors' next task is to test the smartphone application in the field. Developed using responses in the current study, the app takes the team's intervention techniques one step further. The new phase of the trial will skip the alcohol monitor, which Walters said is a costly and cumbersome device, and focus on personalized messages based on a perceived amount of risk for drinking. 

The authors are in the process of recruiting adults experiencing homelessness for a new study, which was temporarily paused due to the pandemic. This break gave the team time to change the visuals of the app for an extra boost of motivation. The new color scheme shows a green background when a user is identified as low risk, a yellow screen for medium risk, and a red screen for high-risk users or those who have already consumed alcohol. Walters thinks this pattern will help the intervention message "come alive with color."

The study, "Using machine learning to identify predictors of imminent drinking and create tailored messages for at-risk drinkers experiencing homelessness," published April 19 in the Journal of Substance Abuse Treatment, was authored by Scott T. Walters, Xiaoyin Li and Eun-Young Mun, University of North Texas; Michael S. Businelle and Emily T. Hebert, University of Oklahoma; and Robert Suchting, University of Texas.