Implementing Machine Learning and NLP for Review Processes

May 25, 2022

Machine learning (ML) isn’t just an algorithm that queries information. Machine Learning learns and adapts over time through a data feedback loop. Normal queries require human intervention to capture behavior changes, whereas machine learning recognizes this and makes modifications itself. Natural language processing (NLP) looks for variants in language. For example, myocardial infarction and heart attack are interchangeable terms for the same event, and Natural Language Processing understands this. It can group these terms to appear together in a search. Natural Language Processing analyzes the context of a sentence; this allows the machine to recognize if the patient currently has cancer, has a history of cancer, or the doctor has ruled out cancer.

Structured versus Unstructured Data

Data constraints can be a challenge when utilizing Machine Learning and Natural Language Processing. In healthcare, there is a shared desire to increase efficiency while reducing cost without sacrificing the quality of care. So much data resides in unstructured documents, like medical records, which continues to be the primary source for reviewing claims for medical necessity and billing accuracy.

How do you get structured data from unstructured data? Named entity recognition (NER) is a list of words and phrases that you want to extract from unstructured data. Ontology is a set of concepts and the relationship between them. For example, insulin could be NER, and within that, you would have keywords such as brand-name drugs like Humalog, etc. The ontology would be the relationship between the use of insulin to a diagnosis such as diabetes.

Targeted Machine Learning

Targeted machine learning is a tactic that is becoming more widely used. For example, itemized bill (IB) reviews are considered semi-structured data because IBs have similar components across the board: description, amount, units, quantity, etc., but they aren’t all the same. There may be fewer or additional components than what’s listed above. Some may have grouped by revenue codes or include unit cost. This is why there is variance within IB documents.

Machine learning can be taught the rules of logic to IBs, creating a definition of IB to guide the data structure. It will understand the IB pattern over a period of time, allowing it to understand layouts it hasn’t seen before. Targeted machine learning takes a target with no significant variance. It sets rules based on industry best practices to develop the structured data vital to streamlining complex review types like cost outlier.

The more information machine learning reads, the more it learns and can adapt to new data. For medical records, structured data looks slightly different from IB reviews since the medical record is a rich source of data. A DRG reviewer may want to know if a diagnosis was documented.

Today’s coding review includes evaluating the clinical documentation that supports the diagnoses and procedures billed. For a machine learning model to help predict codes billed in a medical record, a model would need to include labs, medications, procedures; the list goes on. Pair that with the fact that no two patients will ever look exactly the same due to demographics, illness severity, or other comorbidities.

How TREND Can Help

Is your organization interested in utilizing machine learning and NLP to augment your review process?  TREND Health Partners’ CAVO® technology is an enterprise technology platform that streamlines the medical record review process by providing solutions to your payment integrity challenges in one place. The integration of ML into IB and DRG reviews enhances the review process even further by supporting your review team with document-specific machine learning algorithms. Schedule a 20-minute demo to learn how CAVO and machine learning can improve your team’s productivity.