In a recent news article, a Health & Human Services Department’s audit revealed that one Medicare Advantage plan continually overstated just how sick patients actually were.
The overpayments were for medical conditions such as cancer or diabetes that have serious medical complications where Medicare pays an extra fee because these cases are more expensive to treat. The auditors examined a random sample of 200 medical charts to compare the payments to the actual care the patient received, and to determine if the conditions were as severe as claimed. Based on those 200 cases alone, auditors used a controversial technique called extrapolation to arrive at the $200 million claim.
This may be the tip of the iceberg of problems due to fraud or human error. A government report from 2020 estimated improper payments to the Medicare Advantage plans topped $16 billion the previous year.
Due to the inherent uncertainty with the concept of extrapolation, understandably AHIP, the health insurance plan industry trade group, has long opposed the use of extrapolation for payment errors, and in 2019, called a CMS proposal to start doing it “fatally flawed.”
One solution to the problem would be implementing a Big Data approach to increase the sample size of the audit by analyzing hundreds of thousands of medical charts. This would result in a far higher level of confidence in the outcome, both for the auditor as well as the company under audit.
The Problem with Medical Charts
But medical charts are very complex documents containing different medical records. This requires an inordinate amount of human review time to extract and compare the relevant data points needed to detect fraud and errors.
Much of the most useful qualitative data is stored within the complex structure of a medical chart where rarely is the required data uniformly located in the same position.
Additionally, to successfully analyze and extract the data in a medical chart requires treating each individual medical record as a separate entity. However, many charts exist as single monolithic PDF files containing many discrete records. Not only is there a challenge with accessing data from each record, but there is the challenge of separating each chart file into individual records.
Some attempts have been made to automate this process using computers and legacy capture software, but the unstructured nature of the medical chart has made it notoriously difficult for computers to read and understand the documents without continual human intervention.
Thus this process of uncovering errors has been resistant to both the Big Data approach and to legacy capture solutions.
The New Cognitive Capture Solution
Using advanced machine learning algorithms, it is now possible to train cognitive capture software to identify the key characteristics of each individual medical record within a medical chart. Different algorithms are often employed to evaluate various attributes, such as the presence of graphical information (such as a logo), textual data (such as facility names and addresses), and even spatial information (such as the distance between different dates on a page and use of specific language related to those dates).
All these attributes are then analyzed to identify the most reliable way to identify and separate one record from another, a capture process known as classification. Once the medical records within a chart are classified, then data extraction can begin with the highest degree of accuracy. The final step uses cognitive capture to further analyze and identify specific patterns in the text that might reveal various problems and conditions not only with a single patient’s chart, but that apply across the entire sample.
When finished, the data collected by the capture software is then ready for use in any Big Data application.
Simplifying Data Analysis
By greatly simplifying the analysis of medical chart data, cognitive capture software is an essential weapon in the government’s battle against fraud and error in medical payments. The software helps auditors to dramatically increase the sample size while at the same time reducing the time to completion. Larger sample sizes will go a long way towards the ultimate goal of recovering every overpaid dollar.