Machine Learning in Healthcare and the Value of Human Expertise

Chief innovation officer Adam GronskyWhen discussing artificial intelligence like NLP, the use of machine learning in healthcare inevitably comes up, so today we’re going to clearly articulate what machine learning is, and how it can be best used.

First, when people say machine learning (ML), they usually think of “unsupervised,” that is to say, build a machine learning algorithm, point it at a goal, and leave it to its devices.

We do not employ unsupervised machine learning for a couple of key reasons.

First, speaking very frankly, the current capabilities of unsupervised machine learning are limited, unpredictable, and often misrepresented. Unsupervised machine learning reduces AI technology to an actual black box, stopping anyone from being able to explain why a solution does what it does. This inhibits product QA, security, and providing a full picture for decision-making.

Second, the continued expansion of risk adjustment among providers means Lumanent® is at work in a healthcare setting. While we have faith and confidence in our solutions, based on the current state of the art of unsupervised machine learning, fully automating any AI for use in a live care setting is, in our opinion, not appropriate at this time.

Finally, our development strategy aims to empower human experts. Coders are specialized teams that can and should be supported with technology that helps them apply their nuanced expertise.

There is a counterpart to unsupervised machine learning, though.

Supervised machine learning is exactly what it sounds like, effectively using ML to perform complex, but cumbersome work, under the careful guidance of human expertise. In the past year alone, Google and Amazon have drastically simplified the process of enabling supervised machine learning frameworks. As a result, while we’re not augmenting the NLP itself with it, we’re very excited about what supervised machine learning is doing to further improve Lumanent Insights, specifically around confidence scoring.

Confidence scoring is Lumanent’s ever-evolving self-assessment of the accuracy of its outputs, and a critical part of any NLP solution. A challenge of clinical suspecting, (discerning undocumented diagnoses and presenting them to a care team) is that every condition presentation is, from a technology perspective, unique. Even if a dozen patients present with identical symptomology, there are innumerable ways physicians can formally document the diagnoses, not to mention the broader team. It’s why something as sophisticated as an NLP is necessary to do the job Pre-Encounter Prep does. That unique construction of each suspect also means there are varying degrees of certainty (confidence) with each output.

A higher confidence is nearly self-explanatory, it means Lumanent is very sure a suspect will be accepted by providers. Lower confidence is more complex because while there is lower certainty, there is still enormous value for patients and risk scores alike.

Through deploying supervised machine learning in healthcare, we can accelerate the refinement of our confidence scoring, meaning our provider partners can accomplish more for their patients with less time. They can also use the solution in more versatile, and customized workflows: For example, many organizations using Pre-Encounter Prep prefer to only have the highest confidence suspects sent to physicians, while clinical review specialists vet the lower confidence outputs before adding them to an encounter for confirmation.

Health Fidelity’s Natural Language Processor (NLP), Lumanent Insights, is one of a kind. Since its early development by Dr. Carol Friedman, professor, Department of Biomedical Informatics
at Columbia University, it has ingested 200+ million clinical records, each contributing to its refinement and maturation under the careful guidance of clinicians, linguists, and computer scientists. The opportunity to continue that development each year is one we approach each day with seriousness and enthusiasm. While we won’t be using unsupervised machine learning anytime soon, using supervised machine learning to refine its peripheral technology is an exciting next step, and we’re excited about how we can use it to offer more and more to our partners.