Clinical neurosurgery, research and education will all see significant benefits from the incorporation of artificial intelligence technologies. These benefits will not be immediate, will not all come at once and will not be without potential risks and drawbacks. In this article, I will outline the spectrum of benefits that may result. The major categories of benefits will include data management/documentation, neurosurgery large language models (LLMs) and clinical discovery.
Data Management
The immediate impact for neurosurgeons will be in clinical documentation and data management. The implementation of electronic health records allowed for massive amount of data storage over the last 15 years. This data storage has brought great benefits to patients, but has created difficulties for clinicians. Patel, et al showed that an on-call neurosurgery resident spends an average of nine hours of their 24-hour shift interacting with EHR. These clerical duties are dreaded by all active clinicians. AI is creating breakthroughs to help clinicians lessen this burden. Ambient listening technologies are being trialed now in many major medical centers and are demonstrating significant value in shortening the time needed for documentation of clinical interactions. AI technologies are now being introduced that can accurately summarize patient charts, both active and past. AI enabled documentation of operative reports, discharge summaries and coding are being trialed. In fact, this article was dictated using an AI enabled listening technology that has impressive accuracy and speed.
The next breakthrough for neurosurgeons may be a more accurate reading of digitizable data such as radiology and pathology images. These technologies are already being trialed in many health care systems throughout the United States. Clinical care algorithms are being implemented in both pre and postoperative neurosurgical care. Mayo Clinic already has more than 200 AI algorithms in use. These algorithms are being used with caution. All AI algorithms should be carefully trialed and assessed for accuracy and safety as they are rolled out.
Neurosurgical LLMs
AI technologies for surgical consultation are now commercially available. The Atlas GPT development created by the Neurosurgical Atlas is a remarkably impressive LLM based on RAG (Retrieval Augmented Generative AI) programming. This RAG structure allows the LLM to benefit from broad training LLM’s such as ChatGPT but draws its answers from a limited data set of trusted neurosurgical literature. This can potentially improve the accuracy of answers given by the algorithm. This program has performed with remarkable accuracy on standardized neurosurgery exams. Atlas GPT is easily accessible, can be utilized by neurosurgical experts, trainees and even patients, with the depth of explanation matching the user. I have tested this technology, entering specific cases and scenarios. The resulting analysis of the case and suggestions for surgical approaches, description of approach concerns as well as potential morbidity, was very impressive. The next AI application will be the volumetric analysis of preoperative films and clinical data in individual cases. There is no doubt that this application is imminent.
Clinical Discovery
The application of AI to clinical data finds its greatest utility in analyzing vast data sets where subtle associations may be overlooked using standard statistics. The application of AI in neurosurgery was recently summarized by Levy et al. Examples of recent AI applications in both vascular and spinal neurosurgery are available. Of course, the rules of standard statistics and the need for clean, complete, large datasets still apply to AI. Publications by Emani et al and Warman et al give excellent guidance on how to interpret papers where AI was used in the analysis. At this point, the major limitation of AI literature is the “black box” question. Can we trust the outcome of a program when we don’t fully understand how it reached a conclusion? In Warman et al, the major deficiency in AI publications over the past several years is the lack of external validation. This deficiency was recently addressed by Xianyu, et al with the introduction of the concept of hypothesis driven AI. This new adaptation limits the inputs into the algorithm to known factors, and thus limits the problems with interpretability.
Pace of AI development
Since the inception of the concept of artificial intelligence the development and pace of AI has come in fits and starts. Slow development and implementation of AI typically been due to development of computer power, computer energy consumption and cost; AI requires massive amounts of data entry for training and testing sets as well as extraordinary computer power, time and money. As major breakthroughs in computer power arise, AI enthusiasts and investors create breakthroughs. These breakthroughs are limited by further needs for data, money and energy consumption. These limitations lead to what has been historically known as AI winters.
Although we are in a time of major AI breakthrough, the last six months have shown a slowdown in AI development and implementation. This is characteristic of the ebb and flow of AI development over the last 70 years. Given the incredible hype around AI in the media since the release of ChatGPT, the public and some in health care have expected too much of AI, too soon. The need for safety and accuracy and health care, especially in fields like neurosurgery, have appropriately slowed the pace of AI development in our field.
In many ways, our patients will benefit from AI, and neurosurgeons need to embrace it. AI is coming into our practice, whether we are ready or not!
Richard Byrne, MD, FAANS, is the Monica Jacoby Chair of Neurosurgery at Mayo Clinic Florida. He is a member of the AANS Board of Directors, treasurer of the ABNS and president-elect of the SNS.
