AI tools include anomaly detection, predictive analysis, and social network analysis.
Medicare fraud is a serious issue and an expensive one. “Improper payments” amount to more than $52 billion per year.
Our government uses subcontractors to do its work. They have corrected around $500 million (half a billion dollars) in claims each year. About half of that is returned to the Medicare Trust Fund.
The U.S. healthcare system is incredibly vast, easily the largest and most complex in world history. Medicare spending is 3 percent of our GDP, and this figure should rise to more than 4 percent in a few years.
There are more than 6,000 U.S. hospitals, which include short-stay psychiatric units, rehabilitation units, swing-bed hospitals, children’s hospitals, long-term care, critical access facilities, and religious non-medical institutions. There are numerous other types of providers: more than 15,000 skilled nursing facilities, 10,000 home health agencies, more than 210,000 independent and clinical labs, more than 5,000 ambulatory surgical centers, more than 5,000 end-stage renal disease treatment facilities, 3,000 or so outpatient physical therapy and outpatient speech pathology providers, and around 600 portable X-Ray providers. There are others, too: rural health clinics, organ procurement organizations, transplant centers, hospices, and more. Medicare services more than a million beds for patients.
Around 60 million citizens are enrolled in Medicare, around 71 million are using Medicaid. Each year approximately 7 million people are hospitalized, 33 million use physician/DME (durable medical equipment) services, 26 million are outpatients, and 1-2 million receive hospice services.
Most of the budget is paid out to physicians, outpatient hospitals, clinics, and for prescription drugs. There are around 1.3 million providers, including primary care, surgery, anesthesiology, OBGYN, pathology, psychiatry, radiology, and so on. There are 85,000 DME providers.
The Centers for Medicare & Medicaid Services (CMS) has 16 or so third-party administrators that process around 5 million claims each day, so you can imagine all of the computer power being used. Claims are filed using the Healthcare Common Procedure Coding System (HCPCS). There are approximately 6,500 codes.
With so many providers, so many different institutions and parties involved, and so many codes, there is ample opportunity for fraud. Yet the sheer volume of the transactions is so large, and the complexity of the system so great, any manual-based solution to accomplish fraud detection quickly would spiral out of control. It would take a giant army to manually monitor this amount of electronic paperwork, and the cost quickly would outstrip any potential benefits.
Underneath these activities is a basic principle of economics. There are tradeoffs. The government must expend resources in tracking down fraud, yet it must strike a balance between the amount spent on investigations and the amount recovered. Each year, CMS reports to Congress how many dollars were recovered for each dollar spent in fraud detection. Like all investments, there are diminishing returns. Eventually, the additional money spent on investigating fraud results in proportionally less being detected.
Artificial intelligence (AI) is about efficiency. Any technology that would tend to either raise the amount recovered from fraud or decrease the cost of detecting fraud would increase efficiency.
We read about AI every day. It is one of those strange new technologies that seem to be growing everywhere, but no one can quite understand what it really is. AI is controversial. Many technology leaders have expressed concerns about the social dangers of AI. Part of the concern comes from the history of technology. The factory took away the livelihood of the artisanal shop; the machine took away the soul of the industrial worker; automation stripped away the function of factory-floor labor; the computer took away the jobs of untold numbers of clerical workers. In the same way, AI threatens to take away the livelihood of large segments of the professional class.
There also are concerns regarding the use of AI for destructive purposes. In the future, AI will be in charge of operating deadly weapons during wartime: killer robots. AI will literally be able to kill you. Many have called AI the greatest threat to humanity.
Medicare Fraud Detection
CMS is working with data scientists and entrepreneurs to harness AI in the detection of healthcare fraud. Fraud detection is done through a panoply of agencies. The operation includes a Medicare Integrity Group, Medicaid Integrity Group, Data Analytics Group, provider enrollment operations, and Program Integrity Group. Approximately 500 persons are involved (not including the army of sub-contractors). CMS looks for the following:
- Errors –– including mistakes in coding;
- Abuse –– such as up-coding;
- Waste –– such as the ordering of excessive tests or services; and
- Fraud –– the billing for services or supplies that are not provided.
How is this work being done? Artificial intelligence is used through the application of rules-based systems, anomaly analysis, predictive analysis, and social network analysis.
In the rules-based approach, a set of conditions are defined that, if satisfied, tag a claim as fraudulent. Examples would include “impossible day of admission” services or a claim for a previously stolen Medicare number.
Anomaly detection is a data-mining technique that identifies rare items, events, or observations that raise suspicions by differing significantly from the majority of the data. In Medicare, anomaly detection involves scrutiny of the deviation of a claim from the average or from its peer group. For example, a provider may file significantly more claims in a single day than 99 percent of its peer group.
Predictive analysis is based on regression models that establish characteristics of a fraudulent provider. This is a type of electronic “profiling.”
Social network analysis is based on the identification of links between health providers. For example, if a health provider is linked to an address of a different provider who in the past committed fraud, the system will be alerted.
In addition to using these methods to analyze all incoming claims, other information is used, such as data from complaints.
All of these approaches are employed simultaneously. They are implemented completely in silico because the volume of claims is so great. The AI system is able to generate a series of alerts. There are so many alerts that it is impossible to follow up on all leads. In order to prioritize what actions to take, the system assigns to each alert a level of seriousness or risk.
These, in turn, are studied by experts, who on a case-by-case basis may dispatch “boots on the ground” activities, including site visits and medical chart review.
For further reference, you may wish to read the World Bank discussion paper “Preventing, Detecting and Deterring Fraud in Social Health Insurance Programs: Lessons from Selected Countries” (November 2018).
Compared to other uses of AI, the search for Medicare fraud is relatively unsophisticated. At this point, there is little evidence of the type of “machine learning” that would take place if the AI was able to make up its own approaches to finding bad claims. But give it time.