AI in Healthcare: Cognitive Payment Integrity
Payment integrity is a main focus area for a number of compelling AI use cases. From the algorithms and models that can help detect fraud, waste and abuse (FWA) to data aggregation and analytics tools, AI is positioned to improve detection, investigation and recovery processes - thereby reducing the amount of claims paid in error and the costs of investigations. At CognitiveScale, we are building out multiple AI-enabled solutions in support of more robust payment integrity initiatives.
In the United States, payment integrity programs are targeting billions of dollars in FWA (estimates ranging from $60-250B). BCG reports that there is less than 1% chance of recovery of funds once a claim is paid - and McKinsey projects that there is $20-30 Billion in potential savings from implementing Machine Learning (ML) as part of payment integrity improvement programs. In Gartner’s “Healthcare Payer CIOs, Look in the Mirror to Improve Payment Integrity,” (here) the key recommendation is to focus on root cause issues like poor data quality and earlier transaction surveillance. A key trend in the industry is “multi-dimensional analysis,” a fairly broad term meaning bring as much data from as many disparate sources as possible to the determination of FWA (an AI-use case in itself).
CognitiveScale is focused on specific AI-powered solutions that form some of the building blocks of a comprehensive FWA strategy, especially as it relates to root cause and multi-dimensional analysis and remediation:
- Data & Document Intake, Interpretation & Insights. Simply getting the data and documents in one place can take ML and AI to “read” and interpret unstructured data or documents, convert them into formatted data, classify and sub-classify documents, codify and standardize specific data elements, etc. In payment integrity, this can include adding data and documents such as patient records and claims attachments, for example, to a Member or Provider’s profile or claim history.
- Insights: “Predict Prioritize Prescribe.” Once data is “prepped” for insights, a basic hierarchy of AI capabilities would include things like predicting if a claim will pend to an investigation unit work queue based on known or learned attributes, prioritizing why a claim pends (based on scoring each prediction), and then prescribing solutions (e.g. “we see X in the data as root cause, so look at solution Y or confirming data set Z”). With prescriptive solutions, AI could even be used to then automate next best actions. For example, CognitiveScale has developed match exception models that can determine a confidence level for a transaction that can lead to auto payment (e.g. “Model is 99.9% sure this is an accurate invoice so the recommendation is to auto-pay it”).
The following diagram displays some of the foundational elements of 1) Data Intake, Interpretation, and Insights, and 2) Predict, Prioritize, Prescribe - and how these could be leveraged in a payment integrity program_
- Time & Learning Elements of Cognitive AI. In prior posts, I have talked about CognitiveScale’s Dial Loop (Data, Insights, Action, Learning) as a key hierarchy that extends beyond predictive analytics and base-level algorithmic science, adding a time and learning dimension. For example, in the prior confidence score that said “auto-pay the invoice,” the model would have matured on both invoice data and subsequent feedback on accuracy, learning over time and adjusting the confidence score. Additionally, new data sources (like unstructured patient records data) might be introduced to improve the training data set and eventually model accuracy.
- Trusted, Accurate, Robust AI Solutions. CognitiveScale’s clients have many algorithms and models working across the payment integrity spectrum_ fraud, waste and abuse - and even slivers of those. We are now working to help clients understand if they can trust these models if they are ‘black box’ solutions (e.g. from external vendors), and also to work to help understand model accuracy and robustness (among other scored model attributes). We suspect this will be particularly impactful in areas like payment integrity where there are many models and algorithms in use (build internally and consumed from external, 3rd-party solutions).
Member, Provider & Claims Profiles. Where this comes together for CognitiveScale is in hyper-personalized solutions based on what we call Profile of One. A “profile” is basically a construct for building out everything that it takes (intake, interpretation) to derive insights (predictions, prescriptions) and manage temporal analysis (the time element) and learning needs - all in a trusted manner - for “personalized solutions.” As it relates to payment integrity, this may not sound like a critical need, but there is a lot that can be derived from these profiles in support of FWA solutions. For example:
- Member Profiles_ All of a Member’s claims, benefits, utilization, costs, patient record, etc. data in one place - curated - as an enabling element of all subsequent analysis.
- Provider Profiles_ Similarly, and based on Member Profiles and all of a Member’s claim/remit/payment data, a provider should be “known” at a deep level by all of their claim, patient, contract, prescription, etc. data. A provider can be a person, a practice, an organization, or other types of entities that engage in billing and collections.
- Claims Profiles_ Profile of One can be leveraged to deliver numerous AI solutions to transactions as well as personas and business entities. In one example, a claim profile could include linkages to all other related claims for a specific patient or a provider.
A trend in payment integrity is to engage in multi-dimensional analysis - meaning, more data, earlier in the process, with more robust models and algorithms (e.g. “super models” that are combinations of multiple models and algorithms focused on different data sets). Member, Provider and Claim profiles are an example of where CognitiveScale is helping to address several key components of multi-dimensional analysis_ some root causes of fraud, waste and abuse that rely on more data (perhaps AI and ML to get user-friendly data); enabling predictions and prescriptions off of disparate data sources; and, codifying and normalizing training data over time (in support of learning models). Then, the data and algorithmic scientists (ours, yours, and 3rd parties) can develop models that drive FWA performance improvements off of multi-dimensional profiles.
In a future post we will dig further into what we are doing to help with trusted, robust AI since payment integrity is looking to leverage so many different models and AI-powered solutions.