The information contained within a patient’s medical record is complex. Not only does it combine documentation from across the health care ecosystem, but it also contains various data types from each encounter. To further complicate things, the medical vocabulary is unique with a large lexicon of medical terms, concepts, synonyms, acronyms, and approved or unapproved abbreviations.
Reviewing and consolidating this data to form a comprehensive view of a member’s health is a significant challenge for most health plans, but it is essential to inform appropriate care management, analytics, research, and reporting.
Fortunately, advanced technology is available that can augment and accelerate the labor-intensive process of extracting relevant clinical insights for high-value use cases. Below are 3 common initiatives that health plans can apply natural language processing (NLP) technology to for increased efficiency and improved outcomes.
1. Deriving high-value insights across claims and unstructured clinical data
Health plans are receiving more and more clinical data. This will only increase as the health care ecosystem becomes more interoperable. The key to achieving maximum value is to be able to leverage all that rich clinical information found in unstructured text. Remember that claims data, while useful, are often incomplete and do not tell the entire story of a member. They are limiting in terms of understanding the complete clinical condition. But clinical data have their own set of challenges.
They come from many sources, in many formats, and are typically written in the voice of the provider. They are not structured or codified. All of this makes it very difficult to glean insights from them. To truly merge clinical and claims data for high-value use cases, health plans must extract and codify the clinical data first.
This is where advanced technology can play an important role. Augmented intelligence (AI) technology is capable of normalizing semistructured data like laboratory results to meaningful use standards like LOINC using lab domain-specific rules and machine learning to predict the correct LOINC. For completely unstructured text found in clinician notes, advanced clinical NLP, or cNLP, technology can help. cNLP can highlight and codify the information relevant to a condition or set of conditions, especially those that require ongoing care management such as diabetes or chronic obstructive pulmonary disease (COPD). It can highlight the encounter date, relevant patient information, diagnosis, procedures (like amputation), and even suspecting information like medications used to treat the conditions or symptoms that might indicate a complicating condition for diabetes.
2. Improving the clinical chart review process
Valuable information is scattered throughout a member’s chart, which can be hundreds of pages long, and that makes manually reviewing it for relevant clinical data time-consuming for health plans. Traditionally, plans hire expensive subject matter experts to manually browse multiple documents, search for keywords, and paste key findings into separate forms. In such an important exercise, ensuring that reviewers maintain a consistent level of accuracy and completeness is critical, adding to the administrative burden.
There's no technology that can completely replace this process today. As incredible as AI technology and machine learning are, a trained person still needs to look at the clinical data. But, by leveraging a foundation of terminology combined with NLP, the process can be significantly accelerated.
In reviewing NLP technology to help with the clinical chart review process, payers should look for partners that have a robust library of clinical terms and synonyms, acronyms, shorthand, and misspellings common in clinical notes to effectively detect cryptic clinical terms locked in unstructured text and properly codify them to standards. This will enable the identification of important context surrounding the information to gather a better clinical understanding.
The key is to only highlight the relevant information, however. Some NLP tools highlight everything which creates noise and can reduce the effectiveness of the solution. It is only slightly better than conducting manual chart reviews.
3. Enabling prospective risk adjustment
Currently, most risk adjustment coding is done retrospectively, months after a patient has been treated. It centers around validation of supporting documentation of ICD-10 codes found within claims data. Retrospective review requires health plans to send “chase lists” to providers to pull specific charts to send back to them, often by mail or fax. This is time consuming, expensive, and creates provider abrasion. Further complicating things, often a plan is asking for documentation from a single date of service, but because of the complexity of aggregating data surrounding 1 date of service, the provider sends an entire chart.
This adds to the overhead required to process that data and review it for the appropriate documentation. A robust cNLP solution incorporates the ability to organize the large amounts of data in a meaningful way. It should be able to quickly identify the date of service, pull the relevant sections from a chart, and point out the information relevant to the clinical domain in question.
As Medicare Advantage plans continue to gain momentum in the health care ecosystem, progressive organizations are wanting to move toward prospective risk adjustment to better care for their high-acuity patients. This requires near real-time access to clinical information allowing for suspect conditions to be identified through lab results, medications prescribed, or procedures performed. To reduce time, improve accuracy, assure completeness, and reduce provider abrasion, plans are beginning to leverage the electronic medical record (EMR) itself.
The benefits of an EMR-connected chart retrieval process are many, but 2 of the most critical are improved quality and reduced time to access information. Several vendors are providing this functionality today. Once this is in place and coupled with an accurate cNLP solution, it becomes easier to conduct proactive risk adjustment, providing suspecting lists geared toward capturing all appropriate diagnoses at the point of care and through patient outreach programs in the measurement year.
Additionally, as anyone involved in risk adjustment coding knows, there are complex rules at play. An important component of an efficient risk adjustment process includes accelerated chart review using AI-driven coding and suggestions based on industry and organizational coding rules and guidelines. This process benefits from a digital library of coding guidelines embedded in risk adjustment software so that risk adjustment can be prospective and timely to remove barriers between payers and providers, add efficiency into the process, and reduce overhead and burnout from reviewing endless pages of medical charts.
Many NLP technology solutions have already been developed, and they will continue to be optimized in the years to come. An important factor in the advancement of these solutions will be the integration of clinical informaticists with decades of experience using medical records and medical terminologies with data science teams. The strategic combination of these 2 teams makes it possible to rapidly enhance and configure the technology for maximum accuracy that can serve a multitude of important use cases.