Faculty Bibliography

Prior applications of machine learning to population health have relied on conventional model assessment criteria, limiting the utility of models as decision supports for public health practitioners. To facilitate practitioner use of machine learning as decision support for area-level intervention, this study developed and applied four practice-based predictive model evaluation criteria (implementation capacity, preventive potential, health equity, and jurisdictional practicalities). We used a case study of overdose prevention in Rhode Island to illustrate how these criteria could inform public health practice and health equity promotion. We used Rhode Island overdose mortality records from January 2016 to June 2020 (N=1,408) and neighborhood-level Census data. We learned two disparate machine learning models, Gaussian process and random forest, to illustrate the comparative utility of our criteria to guide interventions. Our models predicted 7.5-36.4% of overdose deaths during the test period, illustrating the preventive potential of overdose interventions assuming 5-20% statewide implementation capacities for neighborhood-level resource deployment. We described the health equity implications of predictive modeling to guide interventions along urbanicity, racial/ethnic composition, and poverty. In sum, our study discussed considerations to complement predictive model evaluation criteria and inform the prevention and mitigation of spatially dynamic public health problems across the breadth of practice.

In the United States, cannabis is increasingly used to manage chronic pain. Veterans Health Administration (VHA) patients are disproportionately affected by pain and may use cannabis for symptom management. Because cannabis use increases the risk of cannabis use disorders (CUDs), we examined time trends in CUD among VHA patients with and without chronic pain, and whether these trends differed by age. From VHA electronic health records from 2005 to 2019 (∼4.3-5.6 million patients yearly), we extracted diagnoses of CUD and chronic pain conditions (International Classification of Diseases [ICD]-9-CM, 2005-2014; ICD-10-CM, 2016-2019). Differential trends in CUD prevalence overall and age-stratified (<35, 35-64, or ≥65) were assessed by any chronic pain and number of pain conditions (0, 1, or ≥2). From 2005 to 2014, the prevalence of CUD among patients with any chronic pain increased significantly more (1.11%-2.56%) than those without pain (0.70%-1.26%). Cannabis use disorder prevalence increased significantly more among patients with chronic pain across all age groups and was highest among those with ≥2 pain conditions. From 2016 to 2019, CUD prevalence among patients age ≥65 with chronic pain increased significantly more (0.63%-1.01%) than those without chronic pain (0.28%-0.47%) and was highest among those with ≥2 pain conditions. Over time, CUD prevalence has increased more among VHA patients with chronic pain than other VHA patients, with the highest increase among those age ≥65. Clinicians should monitor symptoms of CUD among VHA patients and others with chronic pain who use cannabis, and consider noncannabis therapies, particularly because the effectiveness of cannabis for chronic pain management remains inconclusive.

BACKGROUND:Despite evidence that the U.S. "War on Drugs" is associated with increases in drug-related harm and other negative outcomes, all U.S. states have long criminalized most drug possession. In early 2021, both Oregon and Washington became exceptions to this rule when they fully (Oregon) or partially (Washington) decriminalized possession of small amounts of all drugs. METHODS:We obtained arrest data for 2019 to 2021 for intervention states (Oregon and Washington) and control states (Colorado, Idaho, Montana, and Nevada). We calculated monthly rates for arrests overall and for violent crimes, drug possession, equipment possession, non-drug crimes, and a set of low-level crimes termed displaced arrests. Using an interrupted time series analysis, we examined changes in monthly arrest rates after the implementation of policy change in Oregon and Washington compared to control states. RESULTS:In Oregon, there were 3 fewer drug possession arrests per 100,000 in the month after the policy change; the rate decreased throughout the post-implementation period. In Washington, there were almost 5 fewer drug possession arrests per 100,000 in the month following policy change, and the rate remained stable thereafter. Both declines were significantly greater than in comparison states. There were also statistically significant reductions in arrests for possession of drug equipment in Washington and a significant increase in displaced arrests in Oregon. There were no significant changes in overall arrests, non-drug arrests or arrests for violent crime in either state, relative to controls. CONCLUSION:This analysis demonstrates that it is possible for state drug decriminalization policies to dramatically reduce arrests for drug possession without increasing arrests for violent crimes, potentially reducing harm to people who use drugs and their communities. Additional research is needed to determine whether these legal reforms were associated with changes in overdose rates and other drug-related harms.

BACKGROUND:In the United States, methadone treatment for opioid use disorder is only available at opioid treatment programs (OTPs). In addition to federal regulations, states can enact laws which shape access to OTPs. We aimed to define classes of states according to restrictiveness of state OTP laws and examine population characteristics associated with class membership. METHODS:A set of laws was extracted from a database of statutes and regulations governing OTPs in 49 states and the District of Columbia as of June 2021. Latent class analysis of laws was used to estimate the probability of class membership for each state. Class-weighted multinomial logistic regression analysis assessed state-level correlates of class membership and adjusted Relative Risk Ratio (aRRR) and 95% confidence intervals (95%CI) were generated. RESULTS:States (n = 50) were assigned to three classes; Class 1) High restrictiveness on patient experience, low restrictiveness on access to service (n = 13); Class 2) Medium restrictiveness on patient experience, high restrictiveness on access to service (n = 14); Class 3) Low restrictiveness on patient experience, low restrictiveness on access to service (n = 23). States with a higher probability of membership in Classes with higher restrictiveness had higher rates of unemployment (Class 1 vs Class 3, aRRR:1.24; 95%CI:1.06-1.45), and Black residents (Class 2 vs Class 3, aRRR:1.10; 95%CI:1.04-1.15), and lower likelihood of Medicaid coverage of methadone (Class 1 vs Class 3, aRRR:0.25; 95%CI:0.07-0.88). States with a higher probability of membership in Classes with higher restrictiveness also had higher rates of potential indicators for opioid use disorder treatment need, including rates of opioid dispensing (Class 1 vs Class 3, aRRR:1.06; 95%CI:1.02-1.10, Class 2 vs Class 3, aRRR:1.07; 95%CI:1.03-1.11) and HIV diagnoses attributed to injection (Class 1 vs Class 3, aRRR:3.92; 95%CI:1.25-12.22). CONCLUSIONS:States with indicators of greater potential need for opioid use disorder treatment have the most restrictions, raising concerns about unmet treatment need.

Trauma, violence, and their consequences for population health are shaped by complex, intersecting forces across the life span. We aimed to illustrate the strengths of agent-based modeling (ABM), a computational approach in which population-level patterns emerge from the behaviors and interactions of simulated individuals, for advancing trauma research; Method: We provide an overview of agent-based modeling for trauma research, including a discussion of the model development process, ABM as a complement to other causal inference and complex systems approaches in trauma research, and past ABM applications in the trauma literature; Results: We use existing ABM applications to illustrate the strengths of ABM for trauma research, including incorporating interactions between individuals, simulating processes across multiple scales, examining life-course effects, testing alternate theories, comparing intervention strategies in a virtual laboratory, and guiding decision making. We also discuss the challenges of applying ABM to trauma research and offer specific suggestions for incorporating ABM into future studies of trauma and violence; Conclusion: Agent-based modeling is a useful complement to other methodological advances in trauma research. We recommend a more widespread adoption of ABM, particularly for research into patterns and consequences of individual traumatic experiences across the life course and understanding the effects of interventions that may be influenced by social norms and social network structures. (PsycInfo Database Record (c) 2022 APA, all rights reserved).

BACKGROUND/UNASSIGNED:The risk for cannabis use disorder (CUD) is elevated among U.S. adults with chronic pain, and CUD rates are disproportionately increasing in this group. Little is known about the role of medical cannabis laws (MCL) and recreational cannabis laws (RCL) in these increases. Among U.S. Veterans Health Administration (VHA) patients, we examined whether MCL and RCL effects on CUD prevalence differed between patients with and without chronic pain. METHODS/UNASSIGNED:Patients with ≥1 primary care, emergency, or mental health visit to the VHA and no hospice/palliative care within a given calendar year, 2005-2019 (yearly n=3,234,382 to 4,579,994) were analyzed using VHA electronic health record (EHR) data. To estimate the role of MCL and RCL enactment in the increases in prevalence of diagnosed CUD and whether this differed between patients with and without chronic pain, staggered-adoption difference-in-difference analyses were used, fitting a linear binomial regression model with fixed effects for state, categorical year, time-varying cannabis law status, state-level sociodemographic covariates, a chronic pain indicator, and patient covariates (age group [18-34, 35-64; 65-75], sex, and race and ethnicity). Pain was categorized using an American Pain Society taxonomy of painful medical conditions. OUTCOMES/UNASSIGNED:In patients with chronic pain, enacting MCL led to a 0·14% (95% CI=0·12%-0·15%) absolute increase in CUD prevalence, with 8·4% of the total increase in CUD prevalence in MCL-enacting states attributable to MCL. Enacting RCL led to a 0·19% (95%CI: 0·16%, 0·22%) absolute increase in CUD prevalence, with 11·5% of the total increase in CUD prevalence in RCL-enacting states attributable to RCL. In patients without chronic pain, enacting MCL and RCL led to smaller absolute increases in CUD prevalence (MCL: 0·037% [95%CI: 0·03, 0·05]; RCL: 0·042% [95%CI: 0·02, 0·06]), with 5·7% and 6·0% of the increases in CUD prevalence attributable to MCL and RCL. Overall, MCL and RCL effects were significantly greater in patients with than without chronic pain. By age, MCL and RCL effects were negligible in patients age 18-34 with and without pain. In patients age 35-64 with and without pain, MCL and RCL effects were significant (p<0.001) but small. In patients age 65-75 with pain, absolute increases were 0·10% in MCL-only states and 0·22% in MCL/RCL states, with 9·3% of the increase in CUD prevalence in MCL-only states attributable to MCL, and 19.4% of the increase in RCL states attributable to RCL. In patients age 35-64 and 65-75, MCL and RCL effects were significantly greater in patients with pain. INTERPRETATION/UNASSIGNED:In patients age 35-75, the role of MCL and RCL in the increasing prevalence of CUD was greater in patients with chronic pain than in those without chronic pain, with particularly pronounced effects in patients with chronic pain age 65-75. Although the VHA offers extensive behavioral and non-opioid pharmaceutical treatments for pain, cannabis may seem a more appealing option given media enthusiasm about cannabis, cannabis commercialization activities, and widespread public beliefs about cannabis efficacy. Cannabis does not have the risk/mortality profile of opioids, but CUD is a clinical condition with considerable impairment and comorbidity. Because cannabis legalization in the U.S. is likely to further increase, increasing CUD prevalence among patients with chronic pain following state legalization is a public health concern. The risk of chronic pain increases as individuals age, and the average age of VHA patients and the U.S. general population is increasing. Therefore, clinical monitoring of cannabis use and discussion of the risk of CUD among patients with chronic pain is warranted, especially among older patients. RESEARCH IN CONTEXT/UNASSIGNED:

OBJECTIVES/OBJECTIVE:This study aimed to examine trends in cannabis-positive urine drug screens (UDSs) among emergency department (ED) patients from 2008 to 2019 using data from the Veterans Health Administration (VHA) health care system, and whether these trends differed by age group (18-34, 35-64, and 65-75 years), sex, and race, and ethnicity. METHOD/METHODS:VHA electronic health records from 2008 to 2019 were used to identify the percentage of unique VHA patients seen each year at an ED, received a UDS, and screened positive for cannabis. Trends in cannabis-positive UDS were examined by age, race and ethnicity, and sex within age groups. RESULTS:Of the VHA ED patients with a UDS, the annual prevalence positive for cannabis increased from 16.42% in 2008 to 27.2% in 2019. The largest increases in cannabis-positive UDS were observed in the younger age groups. Male and female ED patients tested positive for cannabis at similar levels. Although the prevalence of cannabis-positive UDS was consistently highest among non-Hispanic Black patients, cannabis-positive UDS increased in all race and ethnicity groups. DISCUSSION/CONCLUSIONS:The increasing prevalence of cannabis-positive UDS supports the validity of previously observed population-level increases in cannabis use and cannabis use disorder from survey and administrative records. Time trends via UDS results provide additional support that previously documented increases in self-reported cannabis use and disorder from surveys and claims data are not spuriously due to changes in patient willingness to report use as it becomes more legalized, or due to greater clinical attention over time.

BACKGROUND:Cannabis legalization for medical and recreational purposes has been suggested as an effective strategy to reduce opioid and benzodiazepine use and deaths. We examined the county-level association between medical and recreational cannabis laws and poisoning deaths involving opioids and benzodiazepines in the US from 2002 to 2020. METHODS:Our ecologic county-level, spatiotemporal study comprised 49 states. Exposures were state-level implementation of medical and recreational cannabis laws and state-level initiation of cannabis dispensary sales. Our main outcomes were poisoning deaths involving any opioid, any benzodiazepine, and opioids with benzodiazepines. Secondary analyses included overdoses involving natural and semi-synthetic opioids, synthetic opioids, and heroin. RESULTS:Implementation of medical cannabis laws was associated with increased deaths involving opioids (rate ratio [RR] = 1.14; 95% credible interval [CrI] = 1.11, 1.18), benzodiazepines (RR = 1.19; 95% CrI = 1.12, 1.26), and opioids+benzodiazepines (RR = 1.22; 95% CrI = 1.15, 1.30). Medical cannabis legalizations allowing dispensaries was associated with fewer deaths involving opioids (RR = 0.88; 95% CrI = 0.85, 0.91) but not benzodiazepine deaths; results for recreational cannabis implementation and opioid deaths were similar (RR = 0.81; 95% CrI = 0.75, 0.88). Recreational cannabis laws allowing dispensary sales was associated with consistent reductions in opioid- (RR = 0.83; 95% CrI = 0.76, 0.91), benzodiazepine- (RR = 0.79; 95% CrI = 0.68, 0.92), and opioid+benzodiazepine-related poisonings (RR = 0.83; 95% CrI = 0.70, 0.98). CONCLUSIONS:Implementation of medical cannabis laws was associated with higher rates of opioid- and benzodiazepine-related deaths, whereas laws permitting broader cannabis access, including implementation of recreational cannabis laws and medical and recreational dispensaries, were associated with lower rates. The estimated effects of the expanded availability of cannabis seem dependent on the type of law implemented and its provisions.

BACKGROUND AND AIMS:Low outcome prevalence, often observed with opioid-related outcomes, poses an underappreciated challenge to accurate predictive modeling. Outcome class imbalance, where non-events (i.e. negative class observations) outnumber events (i.e. positive class observations) by a moderate to extreme degree, can distort measures of predictive accuracy in misleading ways, and make the overall predictive accuracy and the discriminatory ability of a predictive model appear spuriously high. We conducted a simulation study to measure the impact of outcome class imbalance on predictive performance of a simple SuperLearner ensemble model and suggest strategies for reducing that impact. DESIGN, SETTING, PARTICIPANTS:Using a Monte Carlo design with 250 repetitions, we trained and evaluated these models on four simulated data sets with 100 000 observations each: one with perfect balance between events and non-events, and three where non-events outnumbered events by an approximate factor of 10:1, 100:1, and 1000:1, respectively. MEASUREMENTS:We evaluated the performance of these models using a comprehensive suite of measures, including measures that are more appropriate for imbalanced data. FINDINGS:Increasing imbalance tended to spuriously improve overall accuracy (using a high threshold to classify events vs non-events, overall accuracy improved from 0.45 with perfect balance to 0.99 with the most severe outcome class imbalance), but diminished predictive performance was evident using other metrics (corresponding positive predictive value decreased from 0.99 to 0.14). CONCLUSION:Increasing reliance on algorithmic risk scores in consequential decision-making processes raises critical fairness and ethical concerns. This paper provides broad guidance for analytic strategies that clinical investigators can use to remedy the impacts of outcome class imbalance on risk prediction tools.