Unlocking the Potential of Big Data to Reduce Opioid Overdose: Lessons from Maryland

Every day, nearly 200 Americans die of drug overdose—mostly from opioids. The opioid overdose crisis pervades all communities in the United States and affects people of all backgrounds. Maryland consistently has experienced one of the highest drug overdose rates in the U.S. This reflects many factors, including the longstanding heroin epidemic in Baltimore City, the rise of prescription opioid-related deaths that began in Maryland in the late 1990s, and the increasing presence of highly lethal fentanyl in the illicit drug supply.

A key challenge in the state is identifying the people who may be most at risk of overdose, and therefore likely to benefit from life-saving services. Motivated by a desire to advance the capacity of public health agencies to improve the targeting and delivery of evidence-based interventions, our study team at Johns Hopkins received a grant in 2015 to develop a data-driven model for predicting overdose risk. This project was a close partnership with the Maryland Department of Health and the Chesapeake Regional Information System for Our Patients (CRISP), the state health information exchange. 

The project led to the linkage of six disparate administrative databases in Maryland: the prescription drug monitoring program (PDMP), the Office of the Chief Medical Examiner, all-payer hospital discharges, specialty behavioral health, adult criminal justice records, and juvenile services.

Based on our analysis, we have developed a list of findings and recommendations, which we presented at a symposium at Johns Hopkins in November 2019.

Finding 1: Fatal opioid overdose is highly concentrated among a small group of Marylanders with identifiable risk factors

The top risk factors associated with a future fatal overdose included having high volume of opioid prescriptions in the year, a recent hospital encounter with a diagnosis of substance use disorders or a non-fatal overdose, being involved with the criminal justice system, and a recent history of addiction treatment. People in the top 5% of the risk score distribution account for 40% of all overdose deaths in our model. The concentration of risk suggests that there are likely to be opportunities to target outreach to these individuals. 

We recommend the creation of a statewide outreach program that sends caseworkers to contact vulnerable patients and assigns a responsible case team for each patient focused on risk reduction that may include Medicaid managed care, community providers, and criminal justice.

 

Finding 2: Patients at risk of fatal overdose can be accurately identified using only PDMP data

Although merged data from all sectors could be used to identify high-risk individuals accounting for 60% of all 2016 overdose deaths in the state, data from the PDMP alone could be used to help identify 40% of all overdose deaths in Maryland. We found that a prediction model created with PDMP data alone is very accurate, whether the outcome being predicted reflects overdose related to prescribed opioids or illicit opioids—which underscores that there is overlap between licit and illicit opioid use in this population. 

We recommend incorporating risk scores in the PDMP that are clear and transparent for prescribers who consult the PDMP, and pairing risk scores with outreach to prescribers of vulnerable patients.

Finding 3: Among the data sources considered, hospital records had the greatest overall utility in identifying people with overdose risk.

In a model that tested the predictive accuracy of the multiple study databases, the all-payer hospital records had the greatest ability to identify the largest cohort of those at high risk for future overdose death. This was the case because hospital records capture a large population with both prevalent indicators of moderate risk (e.g., visits to hospitals for injuries not explicitly coded as being overdose related), as well as other factors that are less common but more indicative of future risk, such as occurrences of nonfatal overdoses. 

We recommend that Medicaid managed care plans and other payers adopt a hospital-based model for care management interventions.

Finding 4: Patients with opioid addiction who receive medication treatment are substantially less likely to overdose than those who do not receive such treatment.

About 2/3 of the patients in state-sponsored specialty treatment within Maryland receive methadone or buprenorphine. Like other studies, we found that patients in addiction treatment receiving these medications had much lower risk of fatally overdosing while they received treatment. 

We recommend that the Maryland Department of Health ensure that access to medication treatment is available at all addiction treatment programs that are regulated by or receive grants from the Department of Health.

Finding 5: Persons involved with the justice system have much higher overdose risk yet lower rates of medication treatment in the community compared to other Marylanders.

We found that 200 per 100,000 people with an inmate record fatally overdose in a two-year period, compared to the statewide average of 49 per 100,000. Subgroups in the justice system with especially elevated risk include those with hospital encounters and opioid prescriptions. Among justice-referred individuals in treatment within state-sponsored community programs, about one-third receive medication, compared to two-thirds of those not justice-referred. 

Building on recent legislation, we recommend that all state prisons and jails provide medication treatment, and that programs are created to ensure linkage to care after release. 

Finding 6: It is feasible to develop an accurate model for opioid overdose using a multi-partnership model informed by informatics.

With appropriate resources, it is technically possible to extract, link, and curate these disparate databases in a secure and confidential fashion. A team with the appropriate mix of content knowledge and technical skills can apply the available databases to develop sophisticated public health analytic techniques to guide opioid risk-reduction interventions for individuals and communities, as well as add to the scientific knowledge base in this high-priority target area.  

We recommend that the types of cross-agency data sharing collaboration embodied within our project can and should be rapidly applied within the MDH and other cross-agency opioid death reduction activities within the state of Maryland.

Overall, our project provides a scientific “proof-of-concept” that it is operationally feasible to link disparate databases to predict opioid overdose. We are pleased to see that similar efforts have been launched in other jurisdictions, such as Massachusetts and Allegheny County, Pennsylvania. These partnerships can unravel some of the complicated system interactions that precede overdose and point toward potential interventions. Merging these efforts with sound interventions and overcoming the cultural, legal, and logistical barriers to more timely intervention looms as a major challenge. The hardest work is yet to come.

Note from the author: This blog post summarizes findings from a Johns Hopkins-led research study supported by the U.S. Department of Justice Bureau of Justice Assistance. The co-Principal Investigators at Johns Hopkins were Jonathan Weiner and Brendan Saloner. The Johns Hopkins team members included (in alphabetical order): Hsien-Yen Chang, Matthew Eisenberg, Lindsey Ferris, Molly Jarman, Noa Krawczyk, Klaus Lemke, Thomas Richards, and Kristen Schneider. This project involved multiple partnerships with Maryland state agencies, with special thanks to Kate Jackson and the team at the Maryland Department of Health. The findings and recommendations listed represent my own views and are not the official positions of any government agencies.