By David Lareau
December 2017 – If you use Facebook, you’re likely aware of the option to designate your current relationship status. In addition to mainstream choices such as single, married, and in a relationship, Facebook has a choice called “It’s complicated” – which sounds like a pretty good label for a relationship that has some room for improvement.
When I hear about the growing use of artificial intelligence (AI) in healthcare, I’m left wondering what the best label would be for that relationship. If you were to believe the hype, you’d conclude that the AI and healthcare relationship was a match made in heaven. Yet AI depends heavily on discrete values to work its magic and healthcare does lend not lend itself well to discrete values – which is one reason why the relationship has yet to produce a wealth of deep, meaningful results.
Despite the hype about the power of AI for healthcare, I believe that currently the most appropriate label for the AI/healthcare relationship is, “it’s complicated.” Here’s why.
- A lack of high-quality data.
In order to produce high-quality insights, AI systems require high-quality, coded clinical data. Most of today’s health information systems were designed to track patients and facilitate billing and reimbursement – and not to capture critical clinical insights at the point of care. Unless an organization has ready-access to complete and accurate clinical insights, AI outcomes are of limited value.
- Healthcare is different than finance.
The greatest potential of AI is its ability to review hundreds of thousands of data points and quickly derive insights, compared to the hours or days required for a human to manually review and assess the same data. AI is particularly powerful in the finance world because most of its data is stored as discrete values.
Healthcare, however, is different than finance. Clinical systems contain a considerable amount of unstructured, yet critical, free-text information. Healthcare has thousands of CPT and ICD-10 codes, but sometimes those codes fail to capture certain nuances that are essential for understanding a patient’s health situation. For example, a physician can easily enter a patient’s complaint of chest pain into a structured note within the EHR. What’s less easy to capture in a structured format are the more obscure observations that may or may not be relevant, such as the fact the patient recently ran a marathon, was in a car accident, went to Mexico, and is getting a divorce.
EHRs also contain a fair amount of free text because many clinicians prefer to dictate their notes rather than document details directly into an EHR. Free-text information can be converted to a structured format using technologies like Natural Language Processing (NLP), but the downside is that some critical data will be missed. Though NLP continues to get better, even the most advanced NLPs are only 90-95% accurate – which may be acceptable if your goal is to calculate population risk, but it’s probably not adequate when customizing a patient’s cancer therapy.
- AI can’t replace the doctor.
Despite AI’s great potential to improve healthcare, the technology cannot replace a physician to diagnose every condition. For example, the patient complaining of chest pain could have a sore rib – or be in the midst of a massive heart attack. To make the correct assessment, we have to rely on the smartest computer in the room, which still happens to be the one between the doctor’s ears.
AI depends on quality clinical data to make its assessments. When the data is flawed or incomplete, we need physicians to identify the gaps and insistencies and make sense of any conflicting data. Simply put, it’s too risky to rely solely on a machine when a person’s life is on the line.
Healthcare won’t realize AI’s full potential until physicians have the ability to easily produce chart notes that are mostly structured. Additionally, providers need technology to transform the wealth of existing clinical data into formats that support AI’s advanced algorithms. This requires solutions to intelligently identify, interpret, and link medical concepts, and map them to standard nomenclature, such as ICD-10, SNOMED, RxNorm, and LOINC.
Yes, the AI/healthcare relationship is complicated – but as providers embrace new technologies to support AI’s need for quality data, healthcare will benefit from enhanced insights that improve the quality of patient care.
About the Author
Lareau’s work at Sinai led to the founding of a medical billing company that led, in turn, to his partnership with Medicomp. Realizing that the healthcare industry made less use of information technology than almost any other industry, particularly in the area of clinical care, Lareau immediately saw the potential for Medicomp’s powerful technologies and joined the company to help fulfill Peter Goltra’s vision.