Artificial Intelligence (AI) often is defined as machines or agents that display cognitive functions, that when applied within their environment, take actions designed to successfully achieve their goals. As these types of behaviors mimic the ways in which humans learn, think, and interact, there is great focus and excitement around the potential applications for healthcare. Indeed, AI already is being applied to a wide range of problems to help hospitals lower costs and improve outcomes, efficiencies, and the overall ability to manage care delivery.
At Kaufman Hall, we are continually building and applying AI and machine learning capabilities to our software, our management consulting offerings, and our Data and Analytics practice. At an enterprise level, AI shows promise to help optimize staffing, speed software implementations, provide greater accuracy of external comparative analytics, and much more.
In this month’s Spotlight, four Kaufman Hall data scientists each highlight examples of AI applications that they believe could have profound impacts for healthcare in the near future.
by Ian Crane, Data Scientist
Recent years have seen tremendous advances in natural language processing (NLP). This area of machine learning is focused on analyzing everyday human language so that a computer can gain an understanding of what was said. Smart assistants such as Amazon Alexa, Apple Siri, Microsoft Cortana, and Google Home all use NLP to determine what words were spoken, and then use those words to determine the intent and meaning of the spoken phrase. However, NLP is not just limited to audio. Once character recognition algorithms parse an image with text, the same types of NLP algorithms can derive meaning from written words as well. Perhaps the most obvious place for this technology in the medical setting is to assist with medical transcriptions.
At the most basic level, AI-powered medical transcription provides speech-to-text or handwriting-to-text translation of a doctor’s or other clinician’s notes. The text can be automatically inserted into an electronic health record (EHR), saving the time and expense of having a human manually provide the transcription. These types of services, at least for speech-to-text, have existed for years.
Recent offerings—such as VoiceBoxMD, Amazon Comprehend Medical, and others—go further in offering recognition of medical terms, pharmaceuticals, and linking to relevant ICD-10 codes to help minimize the time it takes to correctly populate health records. Manual corrections of the automated transcription still are necessary, but NLP algorithms are actively under development and continue to improve.
Long-term, the focus of firms ranging from small Silicon Valley startups to Microsoft1 is on having NLP be part of the patient-provider encounter. The premise is not only to have AI assist the provider with clinical decision making, but to automatically parse out demographic information, medical histories, symptoms, and potential diagnoses without manual interventions. Medical coding information could be determined by AI from the patient-provider conversation, minimizing incorrect data entry. While likely still some years away, NLP ultimately could handle most post-encounter record keeping and overhead.
Advances in Natural Language Processing
by John maxwell, Data Scientist
Unstructured data, such as text from EHRs, present challenges to healthcare organizations because it inherently requires more effort than structured data to make it actionable. Advances in NLP will benefit healthcare organizations in the near future in two primary ways: easing the conversion from unstructured to structured data, and adding capabilities from unstructured data that currently are out of reach for many organizations.
For example, medical coding requires converting unstructured medical records such as physician notes and lab results into structured codes in order to receive payment from payers.
According to Medical Billing Advocates of America, coding errors occur in 75% of the bills they review.
Incorrect coding causes problems for patients and can delay payments.
Machine learning advances with unstructured data also will expand organizations’ capabilities in the near future.2 Creating structured data requires distilling unstructured data, which can remove important details. Citizen Endo3 from Columbia University uses unstructured text to improve the understanding of endometriosis. Endometriosis has no known cure or biomarkers, even though it affects an estimated 6-10% of women of reproductive age. ICD codes alone do not provide sufficient information. The Citizen Endo project helps understand the patient experience and phenotyping of endometriosis in part through Phendo, a personal health tracking app. Similarly, other organizations can assess population health by data-mining text, or improve digital interactions with patients using NLP where human resources cannot be allocated.
In the last few years, the NLP community has made large leaps in interpreting free-text data. Harder challenges need to be continually developed as state-of-the-art methods exceed human baselines.4 Most NLP research to date has been on casual language from sources such as Wikipedia and Reddit, because they offer large amounts of freely available data, but language in healthcare is much different.
Luckily, there are a lot of academic and commercial efforts going into applying advances from the general domain into specialty areas such as healthcare. In November 2018, Amazon unveiled Amazon Comprehend Medical,5 tailored to healthcare-specific language. The barrier to entry for capitalizing on the latest techniques will continue to decrease over time as tools improve and advances spread.
Healthcare organizations can capitalize on these trends by understanding what the lower investment, higher impact projects are on unstructured data, preparing their IT infrastructure, and gathering the right expertise for unstructured data.
How AI/ML can Optimize Revenue Cycle
by Michael Voss, Data Scientist
Another example of how machine learning and AI are being applied in healthcare is within revenue cycle management. Hospitals and health systems often have to appeal payer denials, as payment guidelines constantly change.7
According to an analysis by Change Healthcare, an estimated $262 billion, or 9% of the estimated $3 trillion in claims submitted by hospitals in 2016 were initially denied. For a typical health system, that could be a potential loss of 3.3% of net patient revenue. With increasing pressure on Operating Margins, some financial leaders are considering a new strategy for revenue cycle management—one that incorporates machine learning and AI.
Traditionally, when a payer denies a claim, administrators are required to appeal the denial in hopes that the hospital can get paid. This often is a very time-consuming process, and requires administrators to work closely with payers and physicians. However, this process is not always rewarding, and the claims may still be denied. In addition, some hospitals lack the resources (staff or budget) to review and appeal every denial. In those cases, hospitals only focus on the high-dollar amounts, and the low-dollar amounts go unaccounted for and can start to add up.
To increase workflow efficiencies and significantly reduce the number of denials, many hospitals have begun incorporating machine learning and AI into the revenue cycle management process. By applying machine learning models to hospitals’ historical claims data, the models can identify patterns of denied claims. When a new claim exhibiting those same patterns comes in, the machine learning algorithm would flag it and give a reason as to what the issue might be (registration/eligibility, missing or invalid data, etc.).
For easy fixes, the machine learning algorithm can auto-correct the claim itself. For other cases, the administrators can review the claims and make the necessary changes prior to submitting them to the payer. Having error-free claims prior to submission for approval significantly reduces the number of denials that a hospital will experience. Fewer denials can then lead to faster payments, improved cash flow, and better use of administrators’ time, allowing them to spend that time on other important tasks.
AI’s Application of Predicting Disease Outcomes to Adjust Resourcing
by anika ghosh, Data Scientist
Hospital staff members need to regularly evaluate patients and their conditions. Care decisions often are based on medical records, but in some cases those records aren’t accurate, and could miss something that is vital for delivering the right patient care in a timely matter.
Medical records need to be updated when they are wrong, but how can a hospital prioritize which records to check? Machine learning algorithms could be used to identify errors. Such algorithms can analyze real-time data from patients and identify differences between real-time data, and what is in the medical records. Based on this information, an algorithm can create a prioritized list of patients for staff members to see. This already is being done successfully by ezDI.7
Machine learning algorithms can also help prioritize patients who need the most medical attention. In a perfect world, hospitals would have the resources to prevent all complications in every single instance. Unfortunately organizations don’t have access to unlimited and perfect resources.
An algorithm that can correctly predict negative outcomes can create a prioritization list of patients to monitor and check on based on anticipated outcomes that would require the most attention. Several hospitals, such as El Camino Hospital, are using machine learning models to predict heart attacks, sepsis, strokes, or other serious complications using patients’ medical records and real-time vital information to alert staff to take immediate preventive actions.9 Companies like Aidoc prioritize which radiology patients to see based on their given pathologies.10
Investment in these types of algorithms has resulted in the ability to somewhat accurately predict negative outcomes for certain diseases. In the future, such algorithms should be able to simultaneously predict unfavorable outcomes for a variety of different pathologies and diseases in the future, such as for stroke and heart disease patients.
It is impossible to predict all of the ways in which AI and machine learning ultimately will be used in healthcare. As discussed here, the possible applications for these new advancements are many, ranging from building new back-office, administrative efficiencies, to helping healthcare providers save lives by anticipating and preventing serious complications. One thing is certain—the pace of change is accelerating as AI and machine learning help to shape a new, uncharted future for healthcare.
Learn more about how Kaufman Hall is applying AI and Machine Learning to help clients solve strategic, financial, and clinical challenges.
1 Langston, J.: “Microsoft and Nuance Join Forces in Quest to Help Doctors Turn Their Focus Back to Patients.” Microsoft, The AI Blog, Oct. 17, 2019. (https://blogs.microsoft.com/ai/nuance-exam-room-of-the-future/)
2 Gooch, K.: “Medical Billing Errors Growing, says Medical Billing Advocates of America.” Becker’s Hospital CFO Report, April 12, 2016. https://www.beckershospitalreview.com/finance/medical-billing-errors-growing-says-medical-billing-advocates-of-america.html
3 Columbia University Irving Medical Center, Citizen Endo website: http://citizenendo.org
4 Hao, K.: “Baidu Has a New Trick for Teaching AI the Meaning of Language.” MIT Technology Review, Dec. 26, 2019. (https://www.technologyreview.com/s/614996/ai-baidu-ernie-google-bert-natural-language-glue/)
5 Amazon Comprehend Medical website: https://aws.amazon.com/comprehend/medical/
6 Change Healthcare: Change Healthcare Healthy Hospital Revenue Cycle Index. June 26, 2017. (https://www.changehealthcare.com/blog/wp-content/uploads/Change-Healthcare-Healthy-Hospital-Denials-Index.pdf)
7 Yadav, D.: “Worklist Prioritization: The Critical Element in Today’s CDI Operations.” ezDI Inc., October 2019. (https://www.ezdi.com/blog/worklist-prioritization-the-critical-element-in-todays-cdi-operations/)
8 Just, E., Thatcher, L., Lawry, T.: “Machine Learning in Healthcare: What C-Suite Executives Must Know to Use it Effectively in Their Organizations.” Health Catalyst, Aug. 7, 2018. (https://www.healthcatalyst.com/insights/machine-learning-in-healthcare-what-c-suites-must-know)
9 Parra-Novosad, N.: “Machine Learning in Healthcare: 5 Use Cases that Improve Patient Outcomes.” Anaconda, Sept. 5, 2019. (https://www.anaconda.com/machine-learning-in-healthcare-5-use-cases/)
10 Bloom, J., Dyrda, L.: “100+ Artificial Intelligence Companies to Know in Healthcare, 2019.” Becker’s Hospital Review, July 19, 2019. (https://www.beckershospitalreview.com/lists/100-artificial-intelligence-companies-to-know-in-healthcare-2019.html)