Artificial intelligence (AI) has become a household word as its abilities and supposed applications continue to expand across numerous industries. The use of AI in medicine is no exception to this. Whether it is AI more broadly, or specific subgroups such as machine learning (ML) or deep learning (DL), researchers and clinicians are seeking new ways in which these technologies may improve our ability to synthesize data, reduce errors and enhance patient care. Within the field of spine surgery, for instance, AI is being applied to nearly every stage of the perioperative process, from image interpretation to intraoperative navigation to prediction of postoperative outcomes. In these domains, AI may, in some instances, approximate the abilities of a trained surgeon to various degrees. Overall, however, AI does not yet eclipse the role of the neurosurgeon in any meaningful capacity.
That said, regarding neurosurgical research, AI may come closer to eclipsing human abilities at the current time. Theoretically, AI can analyze massive amounts of data on a scale and timeline that is simply not possible using conventional means. AI algorithms may identify trends or predict outcomes that were not previously identifiable via non-automated means. Moreover, AI could augment our ability to curate and maintain large datasets, as information may be gleaned directly from electronic medical records, wearable technologies or myriad other methods by which patient data may be recorded. In general, it is understandable why using AI to conduct research using “big data,” genetic information or large outcomes datasets is attractive to researchers when the volume of data in these repositories may be on the scale of terabytes.
Nevertheless, while we acknowledge the promise of AI as a research tool, we also must highlight specific caveats. First and foremost is the reliability of current AI algorithms. AI can perform tasks more rapidly than humans alone, but the predictions enabled by ML or DL algorithms may be of questionable quality. The validity of the outputs of a particular algorithm are only as good as the data on which they were trained. As it stands, the research methodology and model training in neurosurgery research that utilizes AI is surely suboptimal. Many studies do not train their algorithms on sufficient numbers of datapoints or instances. Others may have simplistic models that do not produce results applicable to real-world scenarios where there are unaccounted-for variables.
Indeed, an overarching theme of AI models is their inherent “black box” nature whereby the inner workings of the model are obscure to all but perhaps those who are most familiar with its development. To that end, we must avoid becoming overly reliant on AI models to analyze data based primarily on its status as an emerging and exciting technological advancement. Just because an AI model predicts an outcome or recommends some treatment does not mean that it is inherently correct. Rigorous statistical analyses still apply, and iterations of algorithms must be continually improved – and the results of these improvements published – to add validity and predictive abilities to a model.
Furthermore, an under-considered aspect of AI in research is exactly which parties will ultimately be controlling the data used by AI algorithms, the algorithms themselves or a combination of the two. In neurosurgery, extensive resources are being put by industry partners towards research and data collection utilizing AI. These algorithms are being harnessed to both create datasets and to learn from datasets comprised of the very patients we treat. Just one example includes Medtronic’s UNiD ePRO data collection service and its Adaptive Spine Intelligence (ASI) AI platform, which are being used to glean and analyze data from electronic health records and the perioperative process.
In one scenario, the use of AI by industry or similar institutions to analyze patient outcomes data may lead to improved outcomes and more personalized medicine. In another scenario, it may introduce commercial bias into the literature and further cloud the reliability and biases that accompany the use of any AI in research. Little consideration thus far has been given to how companies use patient data, much of which is voluntarily ceded to industry by surgeons and patients themselves. Intentions may be good regarding this research, but we also must avoid a situation in which industry seeks to impede surgeon autonomy or decision-making or alter other important aspects of care delivery and clinical research.
This is not to say that all AI research is inherently flawed or that AI may easily be twisted to benefit its creators. But as more AI models are being applied to neurosurgical research, their methodology and results should be carefully scrutinized. AI models are human-generated, as with any statistical, animal or experimental model that are normally reviewed in our publications. The peer review process must not become blinded by AI’s novelty and relax its standards.
Finally (and, we know, somewhat fancifully), we would support the creation of open-source datasets and algorithms so that the scientific community or the broader public may scrutinize AI models, datasets and results themselves. This could help address both the internal obscurities of AI models as well as ensure that results or predictions created by these algorithms are valid. Important research in neurosurgery will no doubt be powered by AI. Much can be done at this stage to promote transparency in the effort of avoiding certain growing pains.
Alexander Yahanda, MD, is a PGY-4 Neurosurgery resident at Washington University in St. Louis. His research interests include spinal deformity and oncology, spinal cord injury, neurosurgical innovation and biomedical ethics. He plans to pursue a fellowship in spinal deformity after residency.
Camilo Molina, MD, FAANS, is a spine fellowship-trained neurosurgeon with specialized expertise in spinal tumors and deformities. He leads an active research laboratory focused on spinal cord injury, spinal oncology, spinal deformity and advanced surgical technologies, including robotics and augmented reality. As a leader in the field, he contributed to the FDA approval and implementation of the first augmented reality-mediated spine procedures in the United States
