Combatting the Opioid Epidemic with Next-Generation Risk Assessment Tools
Opioid drug overdoses are currently the leading cause of death among Americans under 50, outpacing guns and car accidents. With roughly 64,000 opioid-related deaths in 2016 and 91 Americans dying daily from overdoses (including prescription pain relievers and heroin), the opioid-related death rate has quadrupled since 1999.
Initially, prescription opioid pain killers (morphine, methadone, hydrocodone, oxycodone, etc.) were welcome innovations to treat acute pain, and were presented to clinicians, and to patients, as a low-risk way to effectively treat pain, a longtime challenge for clinicians. Aggressive marketing campaigns and an increased focus on relieving pain from regulatory boards and professional organizations changed both societal expectations about pain management and clinician prescribing patterns. Pain was introduced as a fifth vital sign in 1996; the current opioid problem has its roots around this time.
Risk prediction tools that leverage machine learning have the potential to change how clinicians prescribe opioids, with the goal of preventing overuse, misuse, and abuse.
Machine Learning Offers Deeper Pattern Recognition
Machine learning is relatively new to healthcare, but has already helped organizations improve outcomes and reduce costs:
- Machine learning applications have helped a large academic medical center reduce central line-associated blood stream infection (CLABSI) rates by 50 percent.
- Machine learning applications have produced models for readmissions, infection, and finance, helping clinicians and operations teams prioritize resources.
The Need for Better Opioid Risk Assessment Tools
Currently, clinicians have opioid risk assessment tools, but these tools aren’t sufficiently accurate, reliable, or available (they tend to be outside normal workflow):
Risk Assessment Before Prescribing
Clinicians can use safe prescribing tools before prescribing opioids to assess a patient’s risk for misuse. Many available prior-to-prescribing tools, however, don’t effectively identify patients at risk. For example, the Opioid Risk Tool and the Screening Instrument for Substance Abuse Potential use relatively few data points, are subjective, and are completed by the patient.
Risk Assessment After Prescribing
Instruments to assess misuse once opioid treatment has started use more data points than prior-to-prescribing tools, but don’t assess use of other substances (e.g., tobacco, alcohol, or marijuana), and most don’t include comorbid conditions that increase the likelihood of misuse (e.g. mental health conditions or history of substance abuse). One widely used tool, the Current Opioid Misuse Measure, is administered by the patient, and has a sensitivity of 0.76 (the portion of sick people who are correctly identified as having the condition) and a specificity of 0.66 (the portion of healthy people who are correctly identified as not having the condition). Relatively low specificity means that some patients will be incorrectly labeled as misusing (a false positive), when no problem actually exists.
Research Indicates Promise for Machine Learning in Opioid Misuse Risk Assessment
Machine learning can improve risk assessment tools’ ability to identify patients at risk for opioid misuse. While it’s too early to know exactly how well machine learning applications will help predict risk (the technology hasn’t been applied in practice yet), studies are showing significant potential:
Mining Twitter Data for Illegal Sales
Using machine learning to mine Twitter data, a team of researchers from the University of California San Diego successfully identified accounts used to illegally sell prescription opioids online. The study demonstrates that technology and machine learning could be used for active surveillance and detection of illegal online activities, and could be used to prohibit the online sale of controlled substances.
Predicting Potentially Problematic Use at the University of Pittsburgh
Researchers at the University of Pittsburg used machine learning to identify potentially problematic use of opioids in 194,148 fee-for-service Medicare beneficiaries between 2007 and 2012. The researchers identified five subgroups based on opioid use patterns and identified a small subgroup (less than one percent) who used higher dosages of opioids, received prescriptions from a higher number of prescribers, and obtained medication from numerous pharmacies. Most of these beneficiaries were disabled and had higher comorbid conditions and concurrent healthcare use. Having a better understanding of the factors influencing misuse could improve clinicians’ decision making and opioid prescribing patterns. Machine learning can also illuminate unknown patterns or confirm suspected patterns or pathways to abuse.
Identifying Patients at the Highest Risk of Overdose
Machine learning also has the potential to help clinicians identify which patients are at the highest risk of overdose. Using machine learning, researchers identified the variables most related to overdose and predicted which patients were more likely to overdose. This information could be valuable to clinicians as they make decisions regarding prescription opioids, and could help clinicians ensure that patients at a higher risk of overdose also receive a prescription for the overdose reversal medication, naloxone. A recent study demonstrated that natural language processing techniques could be used to extract unstructured data from the EHR, automating opioid risk assessments. In addition, machine learning could help identify which patients might benefit from non-pharmacologic, multi-modal therapies, or care management programs.
The Next Level: A Bigger Impact
These studies confirm machine learning’s potential to combat the opioid epidemic. More precise machine learning tools integrated into the workflow will take opioid misuse risk assessment to the next level and drive real change.
For example, one forward-thinking healthcare organization currently uses a client pain agreement. To reduce the patient’s risk of overuse and misuse, that patient signs a contract to only use one specific clinician for pain management and one specific pharmacy for prescription opioids and other prescribed medications.
The contract is a solid start, but adding a function that predicts when a patient will default on the pain agreement would take the agreement to the next level of proactive risk reduction. Integrated into the health system workflow, the ability to predict noncompliance will help clinicians plan appropriate interventions based on risk over time (e.g., scheduling follow-up appointments during the time of highest risk of default).
Machine Learning: A Multidimensional Solution for Reducing Opioid Misuse
Combatting the opioid epidemic is uniquely demanding because it’s a multidimensional problem with multiple factors, causes, and effects. Machine learning, however, meets the challenge of reducing opioid misuse risk by enabling a more comprehensive and complete view of the situation (comorbidities, other substance abuse, the amount of medication prescribed, the duration of opioid use, etc.).
Machine learning enables objective, multidimensional, and trainable risk reduction models (versus subjective, patient-administered tools) that will help clinicians make informed decisions about prescribing opioids, as well as finding other options in pain treatment, when appropriate. These insights will also help clinicians monitor patients and intervene to combat the escalating rate of opioid abuse in the United States.
Would you like to learn more about this topic? Here are some articles we suggest:
- Four Effective Opioid Interventions for Healthcare Leaders
- Combatting Opioid Abuse with Data-Driven Prescription Reduction
- How Allina Health Deployed Evidence-Based Decision Making and Reduced Variation
- The Benefits of Machine Learning in Healthcare
- 4 Essential Lessons for Adopting Predictive Analytics in Healthcare