Whether or not it is beneficial, ethical or risky, artificial intelligence (AI) and related data analysis methods are in wide use and quickly becoming a more and more prominent part of medicine and biomedical research. The National Institutes of Health is embracing funding studies focusing on a range of conditions using AI approaches . While there are clearly downsides and risks to be mitigated, AI is here to stay and will be a part of biomedical research in the foreseeable future.
As has been well described, the key advantage of AI in biomedical research relates to analysis of “big data.” For example, the development of reliable biomarkers, drug discovery, imaging review and other applications is already becoming a part of patient care and research studies. In such approaches, programs use AI related tools (machine learning and neural networks) to find associations buried within data to a point that humans may or may not detect the correct combination of variables with traditional hypothesis testing. Standard Large Language Models (LLMs) such as Chat GPT can use these large data sets to generate answers to questions that would probabilistically be the most likely answer. While a common critique of this approach is that such approaches are not necessarily aimed to find “truth” but rather what is the most likely answer based on the source data (which could be inaccurate), efforts within neurosurgery are underway to carefully curate these source data to strengthen their reliability. One example is the AtlasGPT, which uses only allows the LLM to access trusted peer reviewed sources when attempting to answer an input question.
Ultimately, however, the promise of AI in research and decision making has even higher potential as processing power continues to improve. Early in its rollout, Google AI famously suggested that a way to deal with feeling depressed might be jumping off the Golden Gate Bridge . A human reader understands the context that the user was searching for: A treatment to improve depression rather than a terminal one. So what is the distinguishing factor between us, who recognize that this is an absurd statement based on internet postings versus the LLM recommendation of jumping from the Golden Gate Bridge as a means with depression? One hopes that the answer is more information processing and the ability to have strengthening connections with time. In other words, as computing power increases, these models will have access to more resources, including books, newspapers and other writing that is not accessible on the internet. In addition, such models would speak and translate all known languages with high reliability and context, minimizing translation errors. This additional knowledge would lead a model to better understand the context of passionate but perhaps inaccurate online disinformation. One would hope that these developments would lead AI closer to “truth” as we understand it.
Finally, a more interesting question may still be coming in the medium term related to research questions and AI. While more and more authors (for better or worse) are using AI to assist in manuscript and grant preparation, this may in fact be missing the true value of generative AI approaches. Autonomous research laboratories with bench models are already developing approaches and using AI to design and robots to conduct experiments with higher throughput and objectivity than a human can achieve. Such processes could theoretically be expanded to clinical trial prioritization as well. Once a LLM has all available data on a given topic, an unbiased assessment of the priorities and needs for research can be better balanced. Is there a future where instead of simply helping with a literature review, AI is assessing the literature and determining which studies need to be performed? Such a frontier would have massive undeniable implications for how research is conceived. Investigators would still be needed to conduct AI prioritized research projects with designs developed by AI, but this would be a very different role. We also may be forced to grapple with the potential reality that an unbiased assessment of needs to improve health care (either individually or on a population basis) may not prioritize topics “near and dear” to investigators.
In summary, the “Pro” argument for AI in research is similar to the “Pro” argument of motorized transportation. It already exists, and it has enormous potential, even beyond ways that are currently being integrated and considered. It is also our responsibility to use such powerful technologies in a responsible, regulated fashion.
Dr. Andrew P. Carlson was born and raised in New Mexico. He is a neurosurgeon in Albuquerque, New Mexico and is affiliated with University of New Mexico Hospitals.
