Research Spotlights: May 2023

Using machine learning with electronic health records to help predict opioid use disorder among U.S. veterans

Although opioid prescription rates have decreased since 2012, many patients still receive opioids and are on long-term opioid therapy. It is estimated that one out of four patients receiving long-term opioid therapy in primary care have opioid use disorder (OUD) and 4.7% of all pain patients prescribed an opioid will develop prescription OUD. Because of challenges in identifying patients with OUD, these are likely underestimates of the true rates of OUD in the population. Being able to accurately predict the potential for developing OUD could enhance efforts to prioritize prevention of opioid overdoses, as well as other adverse health outcomes related to opioid misuse and dependence.

Recently published research supported by NIMH and others utilized electronic health records (EHRs) from the Veterans Health Administration (VHA) to explore factors predicting OUD during the period when opioid prescribing was increasing (2000-2012) compared to when it was decreasing (2013-2021). The team used VHA administrative EHR data to develop a predictive machine learning model of diagnosed OUD and determine if the set of predictors of OUD differed between patients receiving opioids from 2000-2012 (cohort 1) compared to those treated from 2013-2021 (cohort 2). Inclusion criteria consisted of regular VHA users (at least two visits annually) who had not filled an opioid prescription for two years prior to the start of each respective cohort. Exclusion criteria consisted of a diagnosis of HIV or experiencing cancer pain. Patients with missing demographic data were also excluded. The dataset for cohort 1 included 602,394 patients, 24,117 who were diagnosed with OUD, and cohort 2 was comprised of 141,734 patients, 4,221 who were diagnosed with OUD. The cohort 2 did not include new patients who were enrolled between 2013 and 2021. Each cohort was also stratified by age (18-34 years, 35-64 years, and > 64 years). Researchers examined a range of predictor variables including demographics, insurance coverage, psychiatric conditions, physical comorbidities, chronic pain, and prescription characteristics (e.g., total number of days early a patient refilled their prescription, duration of opioid use). The outcome variable of interest was an OUD diagnosis (using ICD-9 or ICD-10) and defined as diagnoses for opioid dependence only, or opioid dependence only, or both.

Among the first cohort (2000-2012), four of the five top features in predicting OUD were prescription-related variables including total number of days of early refills, duration of opioid use, maximum duration of early refills, and the total duration of late refills, with the fifth factor being age. The order of importance was nearly the same across age groups, and the predictive importance of these five variables was significantly higher than the next five variables included in the model. Age within each stratified group was also an important predictor of OUD, with prescription-related features being significantly more important for the group aged 65 and older, and prior substance use disorder diagnoses being significantly more important among younger (18-34) and mid-age (35-64) groups, compared to the older age (>64) group. Among cohort 2 (2013-2021), prescription-fill variables were also identified as important in predicting OUD.

The study findings have some limitations. Although the study used VHA records, it could not be determined if the opioids were taken as prescribed, if they were taken with other prescriptions obtained outside the VA or supplemented with any illicit opioids. It is also important the note that medical records may contain misclassifications. Additionally, the VA serves a predominantly male-identified population with high rates of comorbidities and may not be representative of the general population.

In summary, consistent with prior research, machine learning approaches that can leverage large, dense EHR data show great promise for identifying those at risk for OUD and other types of disorders. This work also highlights the need for research in diverse healthcare settings to ensure that algorithms can correctly identify risk among diverse populations.

Banks TJ, Nguyen TD, Uhlmann JK, Nair SS, Scherrer JF. Predicting opioid use disorder before and after the opioid prescribing peak in the United States: A machine learning tool using electronic healthcare records. Health Informatics Journal. 2023;29(2). doi:10.1177/14604582231168826

Brain maturation sequence sheds light on youth sensitivity to neighborhood impacts through adolescence

In a recent study, researchers supported by NIMH, NIDA, NINDS, NIBIB, NSF, and others, investigated how developmental processes from ages 8 to 23 occur across the human brain using magnetic resonance imaging (MRI). Over the course of development, it has been the general understanding that children have higher brain plasticity than adults. Brain plasticity refers to the capacity for neural connections and pathways in the brain to change or reorganize in response to internal biological signals or the external environment. Prior studies in animal models have shown that intrinsic brain activity, which occurs when the brain is at rest, or not being engaged by external stimuli or a mental task, is higher and more synchronized when a brain region is less developed and more plastic. As a result, measurements of brain activity waves show an increase in amplitude (or height). Using this information, allowed for the research team to study a functional marker of brain plasticity safely and non-invasively in youth and young adults.

To understand these developmental changes in brain structure and function, researchers used the demographic, fMRI, clinical, and environmental data from a cohort of 1,033 youth for this cross-sectional study. Participants were ages 8 to 23 years (self-reported, male n = 467, and female n = 566, 0.3% of the participants identified as American Indian or Alaskan Native, 0.7% Asian, 11% identified as multiracial, 41% Black or African American, and 47% as White. The sensorimotor–association cortical axis (S-A axis) is a prominent axis of cortical organization within the human brain that is rank ordered (lowest to highest), supporting key functions from primary sensations and responses to complex socioemotional functioning. Researchers modeled age-dependent changes by visualizing developmental trajectories for low and high environment factor scores per deciles of the S-A axis.

The researchers found that a higher neighborhood environment factor score (signifying positive, supportive socioeconomic circumstances) was associated with a more significant drop in sensorimotor region plasticity during childhood and adolescence and with a greater peak in association region plasticity during mid adolescence. Factor score-stratified developmental trajectories suggested that plasticity in key areas of social and emotional development does not decline till mid adolescence, providing key insight regarding healthy neurodevelopment in youth. Researchers investigated associations between spontaneous fluctuations in blood oxygenation level-dependent (BOLD) amplitude and neighborhood environment factor scores that index differences in neighborhood-level socioeconomic circumstances, using census-based American Community Survey geocoded data. When the researchers further divided the study sample into age groups aligned with childhood, adolescence, and young adulthood, they found that environmental effect estimates were the most strongly differentiated across the S-A axis during adolescence. A second age-resolved analysis was also conducted that sought to quantify the correlation between age-specific environment effects and S-A axis ranks at 1-month intervals between ages 8 and 23 years. This analysis revealed a maximal correlation between environmental effects and S-A axis ranks in adolescence, providing additional support for the researchers’ grouped developmental stage analysis.

Limitations to be considered include the cross-sectional stud design; the researchers themselves advocate for future longitudinal study designs that can more accurately characterize effects of the neighborhood environment on an individual’s neurodevelopment. Additionally, the researchers call for more direct measures of neural activity, and for a more comprehensive measure of the impact of puberty on regional fluctuation amplitude, given rodent study findings regarding the role of pubertal hormones on signaling. Researchers also call for a more nuanced approach to factor analysis regarding neighborhood environmental effects, suggesting that an enhanced understanding of the social and behavioral correlates of these environment effects can inform interventions that support healthy child and adolescent neurodevelopment.

In summary, this study showed that spatiotemporal variability in different regional development trajectories within the adolescent brain are organized along a sensorimotor–association cortical axis (S-A axis) during childhood and adolescence. This study provides neuroimaging data and factor analysis that supports the conceptualization of the S-A axis as an omnipresent map spatially coupled to the brain’s hierarchies, one that accounts for individual anatomical features, thus serving as a useful tool to better understand the nature of associations between neighborhood environment and fMRI activity amplitude from a developmental perspective.

Sydnor VJ, Larsen B, Seidlitz J, Adebimpe A, Alexander-Bloch AF, Bassett DS, Bertolero MA, Cieslak M, Covitz S, Fan Y, Gur RE, Gur RC, Mackey AP, Moore TM, Roalf DR, Shinohara RT, Satterthwaite TD. Intrinsic activity development unfolds along a sensorimotor-association cortical axis in youth. Nat Neurosci. 2023 Apr;26(4):638-649. doi: 10.1038/s41593-023-01282-y. Epub 2023 Mar 27. PMID: 36973514.

Study finds increase in smoking prevalence and age of initiation among US young adults from 2002 to 2019

Tobacco use is the leading cause of preventable death in the United States, which has driven several decades’ worth of public health efforts regarding prevention of smoking initiation, particularly among adolescents who are more likely to engage in risky behaviors. These prevention efforts have resulted in a significant decrease in the prevalence of smoking among adolescents in the United States. However, recent studies have also shown that the average age of smoking initiation has increased; in fact, the proportion of young adults beginning to smoke (among all US young adults who report tobacco use) has doubled from 20% to 40% over a 16-year period. It is less clear whether there are sociodemographic differences among young adults in the US with respect to smoking initiation.

A recent study supported by NCI, NIDA, and the FDA aimed to pursue this next level of inquiry by investigating whether prevalence of smoking and smoking initiation differs by race, ethnicity, and education among young adults (ages 21 to 25 years) in the US between 2002 and 2019. A team of researchers conducted a cross-sectional study using data from the National Survey on Drug Use and Health (NSDUH), an annual household survey administered by the Substance Abuse and Mental Health Services Administration (SAMHSA) that collects information on mental health, tobacco, alcohol, drug use, and related topics. All study variables were self-reported by survey respondents. The primary outcome variables were ever and daily smoking, smoking initiation during young adulthood, and daily smoking initiation during young adulthood. Other study variables included age, gender race/ethnicity, education, ever cigarette use, ever daily smoking, age of first smoking, and age of first daily smoking. Study analyses were weighted to ensure that results were generalizable to the US population of young adults.

A total of 187,821 young adults were included in the study sample, with noted race/ethnicity frequencies as follows: 14% non-Hispanic Black, 19% Hispanic, and 59% non-Hispanic White. In addition, 60% had at least some college education, 28% reported at least a high school education, and 12% reported having less than a high school education. Results showed that the prevalence of both ever smoking and daily smoking declined over the 17-year period among all racial/ethnic and education groups. However, this pattern belies the more troubling finding that the proportion of young adults who initiated ever and daily smoking increased among all education groups and racial/ethnic groups. The increase was highest among young adults with at least some college education, as well as among Asian and Black youth.

Study analyses were surprising for several reasons and suggest widening disparities in smoking trends. Young adults with a high school education or less as well as Black young adults showed a slower decline over time in smoking prevalence and age of smoking initiation. While tobacco control measures, including stricter regulatory measures on marketing and availability, have likely contributed to fewer adolescents initiating tobacco use, such measures may not be effectively reaching young adults. It is also possible that these public health interventions are not equitably conducted or received among some populations. Further work to better understand these nuanced differences in shifting smoking patterns over time could support more effective and targeted preventive efforts.

Harlow AF, McConnell R, Leventhal AM, Goodwin RD, Barrington-Trimis JL. Racial, Ethnic, and Education Differences in Age of Smoking Initiation Among Young Adults in the United States, 2002 to 2019. JAMA Netw Open. 2023 Mar 1;6(3):e235742. doi: 10.1001/jamanetworkopen.2023.5742. PMID: 36995718; PMCID: PMC10064249.