Poster
Executive Summary:
In the United States, opioids are the mainstay of pain management and recent years have observed a spike in overdose-related deaths. Our study proposes to utilize machine learning algorithms on medical claims data to identify risk factors of opioid use disorder (OUD) and accurately predict an overdose event. We will implement five different algorithm models on over 200 possible risk factors to find a consistent optimal solution. It can then be applied in the clinical setting to identify high-risk patients and design targeted intervention strategies, thereby contributing to efforts in mitigating opioid misuse.
Research Motivation:
Each year, over 10 million Americans report misusing prescription opioids, with approximately 20% diagnosed with OUD. From 1999 to 2017, the number of opioid overdose deaths increased fivefold, and it is currently estimated that 115 people die from opioid overdoses every day. The misuse or abuse of opioids results in an annual economic loss exceeding $78 billion, including healthcare costs, productivity losses, substance abuse treatment, and criminal justice expenses. Despite ongoing improvements to opioid prescription guidelines and repeated warnings from scholars about the severity of the issue, the problem remains unresolved.
Research Objective:
This study aims to develop and validate one (or multiple) machine learning algorithms to predict the risk of OUD in patients with chronic opioid prescriptions. Through these algorithms, we hope to predict and assess the risk of OUD based on comprehensive patient information, supporting clinical decision-making and enhancing the effectiveness and targeting of interventions.
Research Methodology:
To achieve the above objectives, the project will be divided into two main parts: data extraction and model implementation. We will use the Centers for Medicare & Medicaid Services (CMS) dataset as the primary source, extracting relevant patient information and their historical therapy records. Unlike conventional machine learning projects, determining the study population in health analytics is particularly challenging and requires strict criteria. First, we will review a large body of literature to learn how previous scientific studies have defined their study populations. Then, we will implement this core logic into our SQL database, using SQL queries to identify the target population accurately.
Regarding machine learning models, we will implement and fine-tune four to five different algorithms, such as regression, boosting, bagging, deep learning, and reinforcement learning-based optimization algorithms. This approach will allow us to identify the optimal solution across various scenarios.
Results:
We will analyze over 200 possible features for how accurately they predict OUD, many of which are consistent with previous published studies. These include demographic factors, inpatient and outpatient quantities such as length of stay or number of visits, RX features for prescription data, and other features such as mental health history. We can implement this comprehensive list of risk factors into about five different machine learning algorithms to identify the optimal solution.
Conclusions and Future Work:
While traditional statistical techniques have been applied in assessing opioid-related risks, the complexity and interactions within the data present significant challenges. Machine learning offers unique advantages in addressing this issue. By leveraging machine learning, we can identify hidden patterns in large datasets, uncover new risk factors, and generate more precise and personalized predictions. This will provide clinicians with a powerful tool to more effectively identify patients and design more targeted interventions for these patients, thereby mitigating the problem of opioid misuse to some extent.