Physician-centered informatics

In the past year since ChatGPT first opened up to new accounts, it seems that all business are looking for ways to incorporate it into their practice, not only to look for ways to increase efficiency and save money, but perhaps also to appear forward-looking and to join the trend and not miss out.  I've mentioned that medicine and healthcare delivery have been one of the ways that those in the executive suite have been looking to incorporate into their business – or practice.  It's helpful to review that artificial intelligence is not just chatbots or Natural Language Processing.  It is also computer vision and deep learning, which itself comprises classification, modeling and prediction.  We've seen how computer vision advances have help the radiologists, cardiologist and neurologists interpret their visual clinical data. Time sequence models, such as LSTM, have helped predict usage trends, which is of value to accountants and planners. The New England Journal of Medicine, in their article on Artificial Intelligence in U.S. Health Care Delivery summarized the state of the technology in healthcare in 2023.  As I have discussed before, the areas where it has been implemented have been in the insurance reimbursement area, to help with claims tracking and audits.  The article discusses how modeling has helped improve efficiences in operating room utlization.  In the clinical world, the use of A.I. discussed in the article has been in using deep learning models to predict sepsis, or predict clinical outcomes in the ICU or emergency department, looking for factors that predict readmission or death.  These are the low-hanging fruit scenarios, and mainly represent the application of data science techniques to various deep learning architectures, rendering operations more efficient.

The article mentioned the slow adaptation of A.I. in healthcare delivery.  The reasons for this?  One is the "variability and heterogeneity" of the data, which is understandable. This includes data generated by all the numerous sensors and imaging modalities (sometimes with audio as well as visual components), and the massive corpus of text information (both handwritten and printed). Before any data can be used by an A.I. system, there is need for preprocessing. Everything must be translated into a language that the learning architecture and a database, can process. This usually mean conversion to vectors (or tensors), but the question then becomes, whose format shall be used? There is now a trend to vector databases, which would be especially helpful in the medical world, since the old system of classifying things by human-created categories is laborious and slow. Furthermore, it would be difficult to correlate someone whose disease was given one of the R category of ICD-10 codes with more specific and precise codes. Vector or tensor databases promise to hold multimedia data, and enable trainable queries that will find hidden associations between conditions. At present, I suspect that most stored information is human-entered, and represents only a small subset of the information generated in patient encounters. All other data is still likely represented in formats standard to their media, such as PDF, JPG, WAV or MP4 files. EHR vendors, such as Epic, still store patient data in SQL-based servers, however Oracle is one of their database providers, and Oracle is investing in vector databases, so the technology may change.

But apart from this is lack of trust in artificial intelligence, both from patients as well as doctors.  In this article from the New York Times, there is brief mention of Google's Med-PALM 2 chat model, which is geared to help healthcare workers, but concerns were raised over privacy and informed consent. I worry about the data from which it was trained.  Content gets out of data quickly, and these models will need continual updating to stay current enough to be trusted by clinicians. Although connections to the Internet are now being incorporated with Retrieval Augmented Generation with langchain, so that problem may be tractable.  But there are still ways to go before generative A.I. is ready to replace the physician.  In an emergency department in Boston, GPT4 (so far, the best in class chat model) was so-so in diagnosing a woman's painful knee.  The correct diagnosis was considered by the chatbot, but so was the human physician's differential diagnosis.  At this time, it is probably true that a physician with A.I. is better than a physician without A.I., but this is mainly restricting to formulating a diagnosis. I don't think it's quite where it needs to be in terms of accuracy in predicting outcomes or prescribing treatment.  It can provide suggestions, but skilled and experienced physicians may not find the occasion to consult the chatbot very often. And chatbots still struggle to explain why they came up with a conclusion, although the hallucination problem is improving with new fine-tuning algorithms.

At this time, I think it's still difficult to know how far chatbots and Natural Language Processing will be used in the clinic to perform duties of clinicians. A promising effort was made previously with the app Babylon, but it failed spectacularly.  Sadly, it was going down in flames just as the new chat technology was in ascendance.  I suspect that the developers unable to convince investors to give them more time to utilize the new transformer model in Natural Language Processing, which was showing the world just how awesome it was compared to previous chatbot technology.