Addressing Opioid Epidemic Crisis Leveraging AI/OR
The opioid crisis in the United States has reached alarming proportions, as evidenced by staggering statistics. In 2023, opioid overdoses claimed the lives of more than 230 people in the U.S. each day. Although this number is beginning to decline, the epidemic persists, with over 70,000 fatal overdoses recorded last year. Beyond these tragic deaths, over 6 million people in the U.S. live with opioid addiction. While medications are available to help treat this condition, many who begin treatment face challenges in staying engaged with their care. This crisis has become a critical issue demanding urgent attention and comprehensive solutions.
Muhammad Noor E Alam, an associate professor of Industrial Engineering from Northeastern University have been leading multiple research projects pertinent to leveraging Artificial Intelligence (AI) and Operations Research (OR) to address opioid epidemic crisis. This work has led him to invest efforts in the following key areas (e.g., blue dashed box in the Figure) to address this multifaceted critical challenge and how to address by integrating AI and OR:
· Decision Analytics for Pro-actively Identifying Vulnerable Patients
· Design Policies for Improving Treatment Retention
· Design Optimal Allocation Policies for Overdose Prevention Toolkit
· Prescription Opioids Diversion Reduction Policy
Alam’s research team is actively publishing in Industrial Engineering/Operations Research/Medical Informatics/Public Health journals pertinent to this topic. Here we share a story from one of his recent publication:
Evaluate Health Policy to Improve Opioid Used Disorder Treatment Outcome:
While Medication for Opioid Use Disorder (MOUD)—including buprenorphine, methadone, and naltrexone—remains the gold standard for OUD treatment, its effectiveness is hampered by premature discontinuation rates, typically ranging from 30% to 50%. Increased treatment retention directly correlates with better health outcomes, including significant reductions in mortality.
One promising strategy to improve adherence and treatment completion is integrating self-help groups, such as Narcotics Anonymous (NA) and Alcoholics Anonymous (AA), into MOUD treatment protocols. These groups provide critical psychosocial support, fostering social connections and addressing the psychological and spiritual aspects of recovery. This study examines the impact of combining MOUD with self-help group participation, leveraging a nationally representative dataset and causal inference methodologies to uncover actionable insights and inform policy and practice.
The analysis utilized a large-scale dataset from the Substance Abuse and Mental Health Services Administration (SAMHSA), spanning multiple years and incorporating diverse patient demographics, treatment modalities, and MOUD outcomes. Initially, classical statistical methods, such as logistic regression, were employed to explore associations between various features and MOUD treatment completion rates. However, the high dimensionality and complexity of the data caused convergence issues, prompting the adoption of machine learning (ML) techniques.
ML methods, including Lasso Regression, Random Forest, Decision Trees, and XGBoost, facilitated the removal of noise variables and the identification of key predictors significantly associated with treatment outcomes. These algorithms demonstrated robust predictive performance, consistently indicating a strong link between self-help group participation and treatment completion.
To ensure the causal validity of these findings, the study introduced a machine learning-assisted causal inference framework. A key component of this framework was the outcome-adaptive elastic net (OAENet), a two-stage ML method developed in Alam’s lab, to identify confounding variables and predictors critical for estimating treatment effects. OAENet minimized selection bias and variance, enabling a more reliable estimation of causal relationships.
To calculate the Average Treatment Effect on Treated (ATT), OAENet-selected variables were used to match treatment and control distributions. Propensity score matching with the nearest neighbor method yielded high-quality matches, and the ATT was calculated as 0.260, indicating involvement in self-help groups led to a 26.0% improvement in treatment completion rates among patients with opioid use disorder (OUD).
Finally, the study implemented a robust matching framework, the robust McNemar’s test, grounded in discrete optimization methods, to address uncertainty arising from the matching process and ensure robustness in causal effect estimation. This test produced a highly significant result (p < 0.0001), confirming that participation in self-help groups significantly increased the likelihood of treatment completion among individuals undergoing MOUD. These findings underscore the importance of incorporating self-help group interventions into broader treatment plans to enhance retention and adherence.
By combining machine learning and discrete optimization-based causal inference techniques, this study provides a robust framework for analyzing high-dimensional observational data. This methodology has applications beyond OUD treatment, offering valuable insights for evaluating the effects of policies and interventions across various domains in healthcare and social sciences.
More information about the work can be found on Alam’s webpage: mnalam.com