By Dr. Neil Kudler, Chief Medical Officer at Pixel Health
AI is not a new concept in the halls of the hospital. Before the advent of large language models (LLMs), AI had its roots in data collection and rules-driven algorithms. Consider electronic health records (EHRs), which sort and organize enormous amounts of data for end users. Beyond serving as attractive substrates for data mining and predictive analytics, these data stores also represent the foundation for clinical decision support, alarming and alerting, and medical devices, such as ECG machines.
Examining the early possibilities—and shortcomings—of AI in healthcare
My interest in AI started over a decade ago, when I was a Chief Medical Information Officer (CMIO) responsible for the oversight of Health Information Management. We had a robust clinical documentation improvement (CDI) team of nurses and physicians supported by the best computer-assisted coding (CAC) available.
At the time, natural language processing (NLP) and machine learning (ML) platforms were being marketed to healthcare systems. I was excited about the possibility to help improve the identification of complex cases and optimize the capture of diagnostic specificity and complexity for coding and risk stratification.
From the perspective of a fee-for-service environment, I also recognized how the technology could deliver significant dollars to healthcare’s bottom line. The goal of the CDI reviewer is to recognize gaps in documentation to optimize, not maximize, their coding. Our goal was to deploy an NLP/ML platform to augment the work of human reviewers to enable the hospital to recoup their total costs more effectively, and, from my perspective as an internist, good documentation is good care.
Sadly, the state-of-the-art technology for clinical use was not ready for delivery. The semantics, symbols, and context were poorly interpreted by NLP; it lacked the ability to process unstructured data, especially the abbreviations or colloquialisms in physician documentation.
Considering the hurdles of implementing operational changes
Sometime after exploring NLP for CDI, I sponsored the implementation of a predictive analytics platform to help identify high and rising risk patients and worked to design interventions to improve patient outcomes and experience. I observed that once the data were structured and normalized, which was not an easy feat, the technology performed well, and the patient insights gained were useful.
However, the health system was not ready for change, primarily because of competing reimbursement paradigms. While the organization was ahead of the curve in value-based contracting, the bottom line was ultimately driven by fee-for-service compensation. Added to this was the reluctance to change of the physician community. Proper documentation can be considered the most grueling part of a physician’s job; the patient is competing with the computer for the provider’s attention.
Now, as we fast forward to today, I believe generative AI can offer new solutions to coding, documentation and more complex tasks in the clinic—and that healthcare providers will adopt these changes. The AI landscape has accelerated dramatically, paving the way for a myriad of applications that will revolutionize clinical care.
Continue reading in part two of this article >