First, artificial intelligence (AI) came for the code breakers, and broke more codes than code breakers ever could. Then it came for the Chess players, and new grandmasters emerged. The Credit Analysts and the Actuaries were the next to fall, as were the Meteorologists, Match Makers, and so many others. Now AI has come for the authors, as epitomized by OpenAI’s launch of ChatGPT. Writing text is soon to the purview of bots, from research papers, to novels, and everything in between. The growth of “Generative AI” like ChatGPT, whereby computers can create novel content, has forced healthcare operators to ask some very basic questions about text in general. Can computers help us process text we already create, while helping write new documents? The short answer is absolutely! A rapidly growing body of evidence is demonstrating that appropriate digestion of clinical text can streamline operations, improve care, lower costs, and improve provider satisfaction. This realization, shared by payers, providers, and healthcare investors alike, is shaping healthcare analytics landscape. Specifically, AI-enabled text analytics it will drive the following five trends:

“Control + F” (i.e., keyword search) is dead, AI can now “understand medical data.”

For better and for worse, the vast majority of healthcare data today exists as unstructured text in medical notes. While rich in detail and clinically useful, this data has historically been very difficult to leverage at scale. Analytics on textual data is not practical like it is on medical codes, lab values, and insurance claims. Hence, the text of the record does not inform clinical decision support tools, gaps-incare alerting, population health reporting, predictive models, and the myriad of analytic tools used by the healthcare industry today. Advances in natural language processing (NLP) for the first time make it possible to extract meaning from medical text, not simply find key terms without context. The sentence “patient A tested positive for diabetes, although her mother does not” means something very different than “patient B reported her mother did not have diabetes.” Innovations in AI now make it possible to confidently say that patient A has the disease, but patient B does not. While a seemingly trivial example, electronic medical record (EMR) textual data is rife with clinical concepts that do not actually describe the patient.

“Big Data” in 2023 will be in the form of paragraphs, not databases.

There is a lot of textual data on each patient in the U.S. Insurance payment requirements, obscure billing rules, older and sicker patients, and medical malpractice means each of the 1b clinical encounters that occur (not including hospitalizations) generates a lot of text. EMRs, with the ability to insert templated text, makes it even worse. Innovations in AI will mean this text can now be understood, and “structured” for analysis. That is wonderful, except of course for the people and budgets tasked with storing, managing, and organizing all this data. Analysis conducted at KAID Health, a Boston, MAbased AI-powered health care data analysis and provider engagement platform, suggest the text in an average medical chart can generate over 10,000 clinical data points. Compare that to each patient having ~20 medications on the medication list, or 10 problems on their problem list, and 100 or so medical claims. It becomes clear that text analytics will create a lot of data.

Pulling it all together, text analytics builds on current capabilities, rather than replace them.

AI will find the golden needles in the EMR haystacks.

Just as AI creates the new problem of finding so much data from the medical record, it also holds the solution. AI-based search technology will make these newly created massive databases analyzable, helping individual clinicians and analytics teams alike find the few needed clinical nuggets required to address a clinical or operational challenge.

Point AI solutions and general AI platforms will grow up together in the healthcare market.

Regulatory requirements and commercial pressures have up to now forced healthcare AI solutions in general, and text analytic solutions in particular, to be narrow in scope. A solution will be optimized to address a problem, be it detecting a missed diagnosis, predicting risk or a poor outcome, suggesting an over/under in billing, etc. While this will continue, we are starting to see more multi-use solutions of text analytics emerge as well. After all, if the AI is good at structuring the textual data, and the search makes finding what you are looking for easier, why not apply those insights across multiple groups in the healthcare enterprise?

Accepting all clinical data.

Clinical notes exist today in a variety of formats, from XML to PDF to scanned images. Further, each note can differ from another in structure and section headings. When providers are seeking solutions to analyze all data, it’s important to remember that nothing should be ignored because it is hard to read.

Pulling it all together, text analytics builds on current capabilities, rather than replace them.

Rather than just analyzing one note at a time, text analytics of the future will aggregate the insights captured over time in the medical record, including combing the already structured data, e.g., medications, labs, with NLP-generated insights. Such a bigger set is of course beneficial. One plus one equals two after all. However, if combined and leveraged together effectively, such a combination of structured data and NLP insights can be highly synergistic. The known data on the patient can make the NLP more accurate. Similarly, the NLP data can address errors and omissions in the structured data. Here, the combination of data sets is much more than the sum of their parts. This article was not written by ChatGPT, but it probably could have been.