Fifty years into the “Computational Age” and digitization has advanced nearly every aspect of our lives. Furthermore, society’s need to negate the impact of the first global pandemic of the era has converted what might previously have been considered a quiet background revolution into something much more conspicuous. Acceptance of digital developments that were anticipated to take a generation or more has happened in mere weeks. And, of course, no industry nor profession has experienced this change more acutely than healthcare.
In pathology, digitization has generated significant advancements in terms of how efficiently and effectively we allocate, collaborate, and execute our casework. Artificial intelligence, machine learning, deep learning, and countless new algorithms also provide fabulous new near-term potential. But have we tapped the real value yet? No. Not even close.
Data Flows In…
In most laboratories, casework is still syndicated to the pathologist for review by hand-delivering, with all the risk of delay, loss, and breakages this involves, cardboard trays of glass slides. With digital pathology eliminating this inefficiency through the instant availability of whole slide images, many early-adopters are prone to boast about how they now ‘never wait for glass’, can access their work from anywhere, and enjoy shorter case turn-around-times as a result. And to be fair, given the leap of faith taken by some very early first-movers, perhaps a degree of self-congratulation or gentle told-you-so ribbing of those peers previously less convinced is to be excused! But I would argue that what we have achieved thus far, however commendable, is only half of the real digital opportunity in our field.
Information Flows Out…
In most practices, the outbound deliverable from a clinical sign-out is the pathology report. Final diagnostic reports. Molecular pathology reports. Amendments. Addendums. Information in many forms. But are we using this data as well as we should?
Some of this data is structured. Synoptic checklists are now readily used and reporting templatization is increasing in currency. This obviously provides immediate and exciting opportunities for computer-mining at which we should not sneeze. However, the overwhelming majority of data coming out of pathology are still largely unstructured narrative text. The implications? Every day, in every health system, in almost every part of the world, we are building, but neglecting to put to good use, a veritable treasure trove of voluminous past knowledge and insights capable of catapulting the pathology practice to the center of the precision medicine ecosystem and introducing a step-change in how we diagnose and treat our patients. What is more, there are both zero barriers and zero good reasons why we couldn’t be utilizing that data right now, today, through natural language processing (NLP) AI.
Creating Knowledge is Key…
NLP is neither a new nor necessarily complex technology. What it is, however, is criminally underused in a pathology context given its status as a proven, scalable, and relatively inexpensive way of unlocking and bringing immediate structure to the amorphous data embedded within our reports, notes, impressions, and other documents.
Earlier this spring I was delighted to be joined by some of the individuals I admire most within our profession as I hosted a virtual fireside chat on pathology innovation. As part of our conversation, the panelists and I discussed some of the common obstacles inhibiting new breakthroughs in computational pathology. On review, there were not many challenges identified which couldn’t in some way be remedied via the effective deployment of NLP.
How can we more efficiently extract key data from clinician notes to accelerate the training of new image analysis toolsets? NLP. How can often over-stretched clinicians contribute to the development of new algorithms in a way which doesn’t conflict with the day job? Yep, NLP. How can image analysis teams get access to critical clinical data while respecting patient privacy? You’ve guessed it, NLP.
The data and metadata in the countless reports we generate in pathology are, by any measure, valuable. Still, when combined with the whole slide image, and images and values extracted from other patient encounters, their value becomes incalculable. Why? Because they finally afford us an opportunity to formulate longitudinal records. These information timelines can then be mined to establish trends, elucidate similarities, and help physicians better understand an individual’s disease. And that there, should be our digital zenith. Let’s not be satisfied with just going digital. Let’s create new knowledge and elevate clinical care. Consider the critical coupling of digital pathology with natural language processing.
Digital in. Knowledge out.