TL;DR
This paper introduces a novel graph-based framework, an extension of Non-Parametric Heterogeneous Graph Scan (NPHGS), to detect high-risk and potentially illicit networks of opioid prescribers and dispensers. By modeling prescription data as a heterogeneous graph and integrating advanced anomaly detection techniques, the method effectively identifies suspicious patterns that individual-level analysis would miss, offering a powerful tool for public health agencies.
What Problem Does This Paper Address?
The opioid epidemic remains a critical public health crisis, with prescription opioids contributing significantly to overdose deaths and addiction. While Prescription Drug Monitoring Programs (PDMPs) collect vast amounts of data on dispensed prescriptions, their potential for proactively identifying high-risk prescribing and dispensing networks is largely untapped. Current methods often focus on individual anomalies, failing to capture the collective, network-based nature of illicit activities. This is a significant gap because criminal behavior often thrives in groups, and individuals may only appear suspicious when linked to others. State governments and law enforcement agencies need more effective tools to identify and disrupt these networks, which can divert prescription opioids to the black market and exacerbate the crisis.
Methods
The core of this paper is an extension of the Non-Parametric Heterogeneous Graph Scan (NPHGS) algorithm. The authors model prescription data as a heterogeneous graph where nodes represent prescribers and dispensers, and edges connect them based on prescription volume. Each node is characterized by ten "red flag" attributes derived from public health literature on opioid misuse, abuse, and diversion (e.g., cash payments, distance between patient and prescriber, high MME dosages).
The novel contributions lie in two key extensions to the original NPHGS:
- Flexible Anomaly Detection Integration: Instead of relying on a single method for the initial stage of anomaly scoring, the authors integrate a flexible mechanism that allows for the use of various point-based anomaly detection algorithms (like Local Outlier Factor, One-Class SVM, and Isolation Forests) to generate empirical p-values for each node. This allows for better adaptation to different datasets and problem domains.
- Graph Sparsification for Meaningful Patterns: The authors introduce a method to reduce graph density by setting a minimum threshold for creating directed edges between nodes. This threshold is tied to the scan statistic's significance level ($\alpha_{max}$), ensuring that only stronger connections are considered. This sparsification helps NPHGS focus on more meaningful relationships and avoid detecting spurious patterns in dense graphs.
Datasets
The study utilizes approximately 15 million de-identified prescription records for controlled substances from the state of Kansas, spanning from 2014 to 2015. This data was provided by the Kansas Board of Pharmacy, which oversees the state's PDMP, K-TRACS. The dataset includes detailed information about each prescription, such as dates, drug names, quantities, dosages (in MME), payment methods, and anonymized identifiers for patients, prescribers, and dispensers.
Key Insights and Ideas
- Network-centric approach is crucial: Individual-level anomaly detection misses the collective nature of illicit opioid diversion and abuse. A network perspective is essential for uncovering these patterns.
- Heterogeneous graphs capture complex relationships: Modeling prescribers and dispensers as distinct node types in a graph allows for a more nuanced representation of the prescription ecosystem.
- Integrating diverse anomaly detection methods enhances robustness: The ability to plug in different anomaly detection algorithms makes the framework more adaptable and powerful for various data characteristics.
- Graph sparsification is key to finding meaningful signals: By filtering out weak connections, the method can better identify significant clusters of anomalous activity, preventing dilution of signal by noise.
- Anomalous networks can be hidden in plain sight: Individual providers within a detected cluster might not appear highly anomalous when examined in isolation, but their collective behavior and connections reveal a significant risk.
Key Results
The proposed NPHGS extension, particularly when using Isolation Forests for initial anomaly scoring, significantly outperformed competing graph-based anomaly detection methods (SUBDUE, spectral clustering, and ranked pairs) across multiple metrics. The NPHGS method identified clusters with significantly more anomalous values for patient populations, including higher average and maximum patient MME, and a greater percentage of prescriptions paid in cash and concurrent opioid/benzodiazepine prescriptions.
Case studies highlighted the practical utility:
- Case Study A: A cluster of four prescribers and two dispensers exhibited extremely high-risk prescribing and dispensing behaviors, with dosages and cash payments far exceeding norms. Notably, three of the prescribers and one dispenser were located outside of Kansas, suggesting interstate diversion. The analysis revealed that while individual providers might not flag as top offenders, their combined activity formed a highly suspicious network.
- Case Study B: Another cluster of three prescribers and one dispenser showed a high percentage of cash payments and concurrent opioid/benzodiazepine prescriptions, with MME dosages significantly above average. This cluster also demonstrated strong interconnection, with the dispenser receiving 80% of its prescriptions from the cluster's prescribers. The study also noted that even when providers attempted to create new identities, the network approach could still identify them by their patient overlap.
Limitations and Future Work
The authors acknowledge that their study is based on data from a single state (Kansas) and that future work should involve applying the approach to data from multiple states to ensure generalizability. They also plan to develop an open-source tool to simplify the implementation of their NPHGS framework, aiming to accelerate its adoption by public health and regulatory agencies. The current work focuses on detecting anomalous networks; further research could explore predicting future anomalous behavior or understanding the causal mechanisms behind these detected patterns.
Summary updated