In 2019, U.S. healthcare spending totaled $3.8 trillion[1] — that’s $11,582 per person and 17.7 percent of our gross domestic product. And yet, the U.S. continually had the lowest life expectancy and highest suicide rates among high-income countries. So, what’s the issue? A closer look at the data gets to the heart of what is problematic, from both an ethical health and a cost perspective: without the tools to capture a holistic patient view, providers struggle to pick up the warning signs necessary for preventative interventions to achieve patient-centered, comprehensive care.
The disappointing truth about EMRs
The problem isn’t that providers are ignoring the signs; it’s that they are buried within the mountains of electronic data that exists about each patient – more often than not, in the unstructured documents. The race to digitize health records has meant that often, clinical findings are spread across multiple disparate sources – with different clinics and specialties using multiple EMR systems. What EMRs certainly do well – is provide a mechanism for efficient claims data capture. For many people – “EMR data” and “claims data” are synonymous, such is the black hole that the 80% of healthcare data which is free text or unstructured is more often than not confined to. The 80% of the data that tells us who is at risk is there – it’s just hard to get at.
A recent study from Kaiser Permanente[2] highlighted this gap between structured and unstructured data. In research which aimed to identify the disease burden of valvular heart disease, researchers showed that billing codes only existed for two thirds of the patients who had information in their medical records diagnostic of Aortic Stenosis. As Aortic Stenosis, the most common valvular heart disease, can be complex and costly to manage, using structured data alone to inform population health policy would result in the omission of significant volumes of information rich data
IDNs, NLP and the future of interoperability
The good news is that times are changing rapidly. Providers today are increasingly entering value-based contracts and forming Integrated Delivery Networks (IDNs) as a means of improving care quality and lowering costs through better care coordination. Under an IDN, one or more acute-care hospitals and several associated outpatient clinics and ambulatory facilities consolidate to offer a continuum of care to patients that none could deliver alone.
This is important progress, but the IDN concept is only effective if the network can ensure interoperability between each facility’s multiple data sources to effectively capture SDoH. Given the average U.S. health system operates 16 different EMR systems[3], with each housing rich insights that must be united for a complete patient view, flexible systems that can rapidly surface insights across systems are essential for the IDN concept to succeed. That’s why many IDNs are now leveraging artificial intelligence (AI) technologies such as natural language processing (NLP). This enables providers to quickly access the data buried in unstructured sources to improve treatment plans, make better use of clinical informatics and advance organizational goals under value-based arrangements.
NLP, in particular, enables provider groups such as IDNs to automate the capture of specific insights and terms within clinical documentation at scale, and then normalize that captured data to make it more digestible for clinicians and data analysts, boosting provider efficiency when contrasted with time-consuming manual searches for the same information.
By better understanding the disease burden of their population – more accurate risk adjustment can be made, and more appropriate preventative care initiated to ensure the most cost efficient care can be delivered.
Natural Language Processing to unlock the missing puzzle pieces
The best way to convey the breadth of what NLP can do is to illustrate its outcomes in real world settings – from an example in the acute emergency setting, through to mining nursing notes to extract social determinants of health in order to identify at risk patients.
Going beyond billing codes – understanding the true disease burden
Referring back to the research from Kaiser Permanente above – this particular IDN is already using NLP to make up for the gaps that billing codes leave behind. In research published earlier this year, they demonstrated that NLP could identify 50% more patients with Aortic Stenosis than they were aware of through billing codes. This ability to rapidly surface insights from mountains for free text medical records (nearly 1 million records from 54,000 patients with aortic stenosis were identified in minutes) can lead to far more data drive population health initiatives being implemented.
Creating a safety net in busy Emergency Departments
A large U.S. health system, wanted to ensure follow-up on potential early signs of lung cancer found incidentally in radiology reports, many of which can be missed in busy emergency room settings due to urgent care demands. Using Linguamatics’ NLP, every radiology report in the system was screened nightly, with algorithms in place to flag questionable results for review by care coordinators so no worrisome signs were missed. In 14 months, the software reviewed 1,212 radiology reports with 64 biopsies ordered and 37 malignant cases diagnosed. When you consider the treatment costs for early stage lung cancer ($27,000 per year on average) versus complex late state ($178,000 per year on average), it’s easy to see that NLP can not only deliver better patient care, but also substantially improve costs.
Identifying Social Isolation in prostate cancer patients
In this example, an academic medical center knew that prostate cancer patients with certain social factors, like social isolation, were likely to miss appointments. But identifying this risk manually by scouring physician notes for mentions of social isolation would take months. Using Linguamatics’ NLP software, the medical center was able to comb through more than 55,000 clinical notes from 3,138 prostate cancer patients in just eight seconds[4]. Once trained, NLP was able to analyze a new set of documents and identify those who were socially isolated with 90 percent accuracy. This approach to using rich data buried in clinical notes to identify at risk patients is an excellent example of the potential of the unstructured data to augment structured data in risk identification.
These are just two of multiple potential applications for NLP to support providers. Whether applications are clinical, operational or research focused – unlocking the unstructured data provides each of these areas significantly more high quality data to improve processes and outcomes.
As IDNs partner to define interoperability standards that can achieve improved data accuracy and completeness, NLP can be a powerful resource. If you are interested in what NLP can do for your practice, reach out to Linguamatics today for a demo or watch our webinar to learn how you can power the delivery of comprehensive care with clinical NLP.
References:
[1] https://www.healthdatamanagement.com/news/nlp-to-help-mercy-health-better-treat-heart-failure-cases
[2] https://www.healthcare-informatics.com/article/ehr/its-mission-capture-unstructured-ehr-data-mercy-leaders-realize-value-nlp
[3] https://www.healthcareitnews.com/news/how-mercy-using-nlp-its-epic-ehr-improve-analytics-cardiac-care
[4] https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6416852/
