Detecting fraud, waste, and abuse in substance use disorder treatment

1. INTRODUCTION

As the opioid crisis re‐intensifies, ¹ the potential for fraudulent, wasteful, or abusive substance use disorder (SUD) treatment practices increases. Unethical SUD treatment practices include conducting or billing for unnecessary tests, upcoding for more complex or time‐consuming visits than were provided, and unbundling one lab test into several individual tests. ² , ³ , ⁴ They also include patient brokering, in which an individual receives a kickback in return for directing patients to a specific provider, who may not provide treatment. ⁵ Although high‐profile cases have been highlighted in US Department of Justice reports, ⁶ , ⁷ , ⁸ , ⁹ the true scope of unethical SUD treatment practices is unknown.

Improved detection of unethical SUD treatment is critical to enhancing care quality and limiting unnecessary spending. At best, a patient exposed to unethical SUD treatment receives an unnecessary, but harmless, test (increasing costs). At worst, a patient may fail to receive appropriate treatment and suffer bad outcomes, including death. ² Improved detection of potentially unethical behavior would help payors, law enforcement, and regulators target fraud mitigation resources.

In this commentary, we outline the role of claims data‐based algorithms in unethical SUD treatment detection, potential algorithm inputs, and limitations in algorithm‐based fraud detection, alternate detection strategies, and recommendations for improving fraud detection.

2. ROLE OF CLAIMS DATA‐BASED ALGORITHMS IN UNETHICAL SUD TREATMENT DETECTION

Prior investigations ⁶ , ¹⁰ , ¹¹ suggest that private insurance finances a substantial portion of unethical SUD treatment. Therefore, an examination of private health insurance claims may identify unethical actors. Claims data‐based algorithms aid detection in other instances of public and private health care fraud, waste, and abuse, including unethical practices related to opioid prescriptions, dental care, and implantable cardiac defibrillators. ¹² , ¹³ , ¹⁴ , ¹⁵ In many cases, these algorithms seek to identify unusual patterns in claims submitted by a provider, including outliers in frequency or dollar amounts of claims. ¹⁶

For claims data‐based fraud detection to be successful, a large comprehensive database is required. Markers of fraud that rely on a small sample of claims may be biased if the sample is not representative or if indicators of overpayment are imprecisely measured. ¹⁷ Patterns of multiple suspicious behaviors are likely to be better markers of potential fraud, ¹⁶ , ¹⁸ , ¹⁹ particularly if behaviors can be measured across multiple payors. ²⁰

3. POTENTIAL ALGORITHM INPUTS

To our knowledge, no publicly available algorithm exists for detecting unethical SUD treatment. The first step in developing one is to identify salient features in claims data that could serve as fraud signals. The literature suggests some general indicators of fraud that may be useful for detecting unethical SUD treatment practices. These include outliers in types and frequency of claims for services or visits, and in numbers, demographics, and diagnoses of patients seen. ¹⁶ , ¹⁸ , ²⁰ Common indicators of fraud include duplicate or near‐duplicate billing for procedures, unbundling of lab tests, a greater than expected number of patient encounters per provider, and patterns of treatment that do not meet the needs of a patient's diagnosis. ¹³ , ¹⁴ , ¹⁶ , ²¹ Billing for an impossible number of hours is an additional potential fraud indicator used elsewhere, including in Medicare. ²² , ²³

For numerical claims data fields that can be falsified, Benford's Law may be useful. Benford's Law states that in a series of data, the probability distribution of leading digits of numbers is not uniform. ²⁴ Leading digits are most likely to be one or two. If distributions of the leading digits of numbers deviate from what is predicted by Benford's law, that may signify falsified claims. ²⁵

3.1. Fraud indicators identified by SUD treatment stakeholders

To identify fraud indicators that may be most salient to SUD, as well as to identify potential indicators not previously reported in the literature, we conducted interviews with three regulators, five insurers, and one SUD treatment facility leader. Details of recruitment and interviews are available in the Appendix S1.

Stakeholders mentioned many of the indicators described in the literature on broader fraud detection, including outliers in service frequencies, unbundling of lab tests, multiple same‐day drug tests, and large numbers of ancillary services. They also mentioned billing for an impossible number of hours (e.g., 36 h in a day). If the same service is billed by multiple entities (e.g., by both the ordering physician and a lab) or if a provider exhibits a rapid increase in services rendered, that also may be an indicator of unethical behavior.

In addition to what has been previously described in the literature, stakeholders identified indicators that may reflect suspicious patterns of patient acquisition, including patient brokering. Stakeholders listed an unusual number of newly insured members (especially if the unusual number is from one specific geographic area or from a different age demographic than that which normally enrolls in the plan) and a large number of high‐cost members with SUD diagnoses as potential indicators of unethical behavior. Stakeholders explained that spikes in the number of newly insured patients may reflect patient brokers recruiting individuals to seek care in locations where claims for the care can be charged to a marketplace plan.

Stakeholders also suggested screening for a high proportion of out‐of‐state utilization in the marketplace or general commercial plans, which could signal the involvement of a patient broker. They noted that when a provider is billing out‐of‐network, the health plan has less oversight over the provider, and there may be fewer limits on allowable charges—a combination that unscrupulous providers can exploit. The extent to which patient acquisition indicators correspond to instances of fraud remains to be examined.

4. LIMITATIONS IN ALGORITHM‐BASED FRAUD DETECTION

An examination of claims data alone is unlikely to identify all cases of fraud or unethical SUD treatment practices. Stakeholders listed several potential indicators of fraud that require knowledge of intent or clinical context and the ability to distinguish between falsification and exaggeration: billing for services not rendered, patients referred to services from which they cannot benefit, low follow‐up rates, and requiring patients to step through every possible level of care. However, from claims data, one cannot ascribe intent—claims may reflect data entry errors, be outright false (no services provided), be an exaggeration of the number or intensity of services provided, or may reflect provided, but unnecessary, services. ²¹ , ²² , ²³ Moreover, statistical outliers in billing frequencies may reflect providers' real concerns about undertreating a patient. ²² , ²⁶ , ²⁷

Other potential indicators of unethical SUD treatment practices are not recorded in an insurance claim and thus would be missed in a claims data‐based algorithm. In our interviews, stakeholders mentioned several such potential indicators: incongruence between the listed and actual owner of a treatment facility; lack of appropriate certification or licensure; the presence of different limited liability companies for physician groups, facilities, and laboratories; laboratory name changes; unwillingness to provide medical records to insurers; same accounts being used to pay multiple insurance premiums; incentives offered to attend specific facilities; misleading media claims or online advertisements; and the absence of staff names on facility websites.

Moreover, fraud detection algorithms are reactive by nature, and their development relies on past experiences with fraud—there may be other ongoing fraudulent behavior related to SUD treatment that has not yet been detected by the field. As insurers adapt their coverage policies to react to known types of fraudulent activities, unethical actors will likely adapt as well.

5. ALTERNATE DETECTION STRATEGIES

Audit or mystery shopping studies are one alternative to claims data‐based algorithms for fraud detection. ² By using a standardized script, mystery shoppers (simulated patients or family members) can contact treatment facilities by telephone or in person. Scripts can be flexibly tailored to detect emerging fraudulent schemes and can address the patient experience more directly than a claims data‐based algorithm. For instance, a mystery shopper can ask questions about incentives to attend a treatment facility or willingness to provide services to an out‐of‐state resident. Such an approach has been used to monitor admission practices at SUD treatment facilities, ²⁸ and it could be expanded to assess a wider list of activities. Mystery shopping can identify potentially suspicious responses, but a determination of whether a suspicious response indicates fraud would require a targeted investigation by payors or by regulators of SUD treatment services.

6. RECOMMENDATIONS TO THE FIELD

Methods for improved fraud detection would be enhanced by greater availability of timely data of two types. One is recent data on the results of investigations into potential fraud, waste, and abuse by payors or by law enforcement. Without these, one cannot validate claims through data‐based algorithms. Groups like the Healthcare Fraud Prevention Partnership, which is comprised of members from federal agencies, state and federal law enforcement, private insurers, and state Medicaid plans, share best practices for fraud detection and provide a venue for pooling public and private data. ²⁹ However, these data are not generally available to the research community. Second, there is a need for more widely available cross‐payor data with which researchers with expertise in predictive modeling and substance use disorder treatment could contribute to the development of enhanced detection methods. If these data are recent—no more than one or two years old—researchers may be able to aid in the detection of ongoing fraudulent schemes, rather than reacting to previous ones.

If de‐identified claims data could be linked to gold standard data on confirmed cases of fraud, waste, or abuse, one could use these datasets to train and test a fraud detection algorithm. To test the model in the real world, this algorithm could then be shared with the Healthcare Fraud Prevention Partnership or other groups with access to identifiable claims data from multiple payors. They could run the model on their own data, flag potentially suspicious providers, and then use mystery shopping or targeted investigation on flagged cases. If the algorithm is successful at flagging unethical providers, the costs of investigations may be outweighed by savings associated with recovered payments as well as savings associated with prevented fraud, waste, and abuse. ¹⁵

7. CONCLUSIONS

Algorithms based on claims data are one potential tool to detect unethical SUD treatment practices, but they are limited by an inability to ascribe intent and by their reactive nature. However, even an imperfect claims‐driven algorithm could help target regulatory and enforcement resources. To the extent that algorithms can be validated against known cases of unethical SUD treatment practices, they will be more useful. A combination of quantitative surveillance methods and more patient‐focused audit studies may be needed to accurately identify providers suspected of fraud, waste, or abuse.

Ensuring access to reputable, high‐quality SUD treatment is necessary to improve outcomes. Improving methods to detect, mitigate, and prevent fraud among SUD treatment providers is one key step in this process. ³⁰

FUNDING INFORMATION

This work was supported by Arnold Ventures. Arnold Ventures was not involved in the study design; collection, analysis, and interpretation of data; writing of the report; or submission of the article for publication.

CONFLICT OF INTEREST

The authors declare no potential conflict of interest. The views expressed in this article are those of the authors and do not necessarily reflect the position or policy of the Department of Veterans Affairs or the United States government.

Supporting information

Appendix S1 Supporting information.

Garrido MM, Jones DK, Woodruff A, et al. Detecting fraud, waste, and abuse in substance use disorder treatment. Health Serv Res. 2022;57(5):997‐1000. doi: 10.1111/1475-6773.14046