“The Machine Will See You Now”: A Clinician's Perspective on Artificial “Intelligence” In Clinical Care

Abstract

Intelligence is the ability to think logically, to conceptualize and abstract from reality.1 Its companion, wisdom is the capacity to grasp human nature, which is paradoxical, contradictory, and subject to continual change.1 Both of these constructs are key to the practice of medicine and often improve with age and clinical experience. Artificial “intelligence” (AI) refers to the ability of machines to recognize and “learn” patterns from complex data, predict outcomes and help in decision-making.2 AI has been heralded as a new age in medicine, in which machines will take over medical diagnosis and management. With this background, we draw attention to the shortcomings and potential dangers of AI in our own specialty of movement disorders. The term Artificial Intelligence was coined in 1956 but in the last few years has made considerable progress and is no longer science fiction.3 A quick google search of even reputable and generally trustworthy sources reveals frequent promotional slogans such as “AI is revolutionizing healthcare as we know it” and “2023: a year of groundbreaking advances in AI and computing”. Meta launched “Galactica”, a large language model (LLM) based on a training dataset of 48 million examples of scientific articles, textbooks, websites, lecture notes, and encyclopedias. The purpose behind Galactica was to have a single AI-based tool to summarize all academic articles and write scientific code and annotate molecules. It lasted a total of 3 days online after it was found to be unable to distinguish truth from fiction and that it could “hallucinate” data.4 The adoption of AI has led to increased concerns about the absence of published negative results and some top AI researchers are concerned about companies adopting “shiny products” over safety.5, 6 Leading researchers have expressed alarm about the low regulatory bar for AI adoption in medicine and the total absence of transparency.7 The skill of code writers was questioned when the emphasis on hiring personnel with an AI aptitude led to a 142X increase in AI as a skill on LinkedIn.8 The diagnostic accuracy of AI carries great importance when the consequence of error is harm to patients. A quick search reveals that research on “Artificial intelligence” was funded by the NIH to the tune of ~$1.1 billion in 2023. Correspondingly, the number of articles on Medline on the topic has greatly increased. A recent study showed that only 20% of studies using neuroimaging in Parkinson's disease passed minimal quality criteria, with only 8% using external test sets where accuracy was even lower.9 A systematic review of fifty-five relevant studies on the use of AI for diagnosis of PD found that only three studies were validated with external data sets and only five studies had a low risk of bias.10 The field of movement disorders relies on a detailed clinical history and focused neurological examination to arrive at the diagnosis. The gold standard diagnostic criteria for the most common movement disorders such as Parkinson's disease and essential tremor depend on clinical acumen A large part of what we observe in the clinic is based on intuition and tacit knowledge. Key documented data includes case reports, case series, videos and studies with a small sample size, given the rarity of the diagnoses. As such, these may not serve as appropriately comprehensive training datasets for an AI model. Different forms of chance have played a major role in advances in the field including the story of the introduction of L-DOPA into medicine. Dr. Langston encountered MPTP-induced parkinsonism while he was enjoying his coffee and was annoyed at being interrupted by his residents.11 Amantadine was originally introduced and utilized as an antiviral medication.12 While entertaining the thought of an AI algorithm replacing a neurologist, we must remember that diagnosis is the first and easiest aspect of clinical care. The telling of the diagnosis however requires nuance, grace and empathy. Perhaps the only question that matters with regard to machine learning is: would we trust it with diagnosing our family members? The answer to this question was highlighted in a recent study that compared AI, AI + human physician and human physician in diagnosis. While comprehensibility was similar across the three groups, empathy, reliability and willingness to follow advice was significantly better with the human physician.13 A survey of 1400 US adults revealed that 69% of them were uncomfortable with a diagnosis made only by AI.14 As such, we need to continue investing in training and supporting meticulous, caring physicians and not just state-of-the-art diagnostic technology. The demise of a need of human expertise in fields like pathology and radiology at the hands of AI has long been predicted. However, certain pitfalls with AI make this unlikely. Missing data may lead to bias. Rare diagnosis may be missed or be overcompensated for in the model, leading to overdiagnosis and misdiagnosis. LLMs based on published data, may reinforce outdated practices.15, 16 Unlike physicians, AI may be insensitive to impact of their decisions and therefore, may not demonstrate such safety behavior and/ or know its limitation. There is also the question of accountability should an AI-led calculation lead to an error. Perhaps, the most voiced concern is that of a mismatch between the training dataset and real world, leading to known and unknown errors.15, 16 Such discrepancy was noted in a study conducted at Stanford in Gastroenterology.17 Preliminary studies had indicated that an AI-based tool could help detect polyps during colonoscopy. The subsequent trial conducted in the clinic was negative. The authors acknowledged the difference and added that more real world data can lead to more noise and alter the efficacy of AI-based tools.17 Concerns and disappointment with opacity of AI-based research have been expressed with computational scientists at the University of Toronto noting a trend of excitement around new research which feels like an “advertisement for cool technology” instead of having a basis in science.16 There is substantial excitement towards the incorporation of LLMs into electronic health records to alleviate documentation burden.18 However, this is not without risks. In addition to hallucination of data, AI may misinterpret recorded text. A noted example is that issues with hands, feet and mouth scribed as a diagnosis of hand, foot and mouth disease. The use of LLMs may also lead to chart bloat requiring additional time for screening errors. Overall, while LLMs may be used to summarize the discussion, there is greater risk of summarizing the information in the chart and synthesizing data.19 As we look for options to expand movement disorders care across the world, we must ensure that the tools we use are not biased and inadequately tested, thereby introducing a new inequity in care between the resource-rich and resource-poor nations. Recognizing these issues, the World Health Organization has called for careful examination of AI tools, especially LLMs and screening for potential biases before they are adopted in low-resource settings with the intent of reducing inequity. The FDA has called for “nimble regulation” to avoid being “swept up in something we hardly understand”. The FDA commissioner also recognizes that these models will likely “evolve” after implantation requiring “continuous adjustment to remain accurate”.20, 21 The “Intelligence” in AI is a poor surrogate of human intelligence, wisdom and interaction. It performs poorly even on standard screening cognitive tests.22 While it is clear that AI will never replace physicians, it can paradoxically help in improving the patient-doctor relationship. In recent years, an environment of high regulatory burden has stifled the practice of medicine and innovation, leading to an increase in physician burnout and job dissatisfaction.23-25 Several studies have noted regulatory burden as a primary contributor to burnout and physician exodus.26 “Boring AI” as it has been termed offers some hope in lowering such burdens thereby making the practice of medicine more sustainable and pleasurable.27 Quality measures cost hospitals over 5 million dollars annually, in addition to precious manpower. It took researchers less than an hour to draft abstractions with >90% accuracy using AI.28 The use of AI along with video filming does offer an interesting future for data collection and synthesis which can then be utilized by the neurologist.29 Such an approach should seek not to supplant history taking and physical examination, but to compliment them for good patient care. AI can help streamline clinical workflows while reducing costs up to 17-fold and preserving reliability.30 Once ready and with caution, smart systems can potentially use AI to help with scheduling, patient risk stratification and decisions on resource allotment. As new medications are approved by FDA and require prior authorization for coverage, AI can help with letters of medical necessity and create relevant reading material for patients.31, 32 AI can therefore help physicians be physicians and spend more time with the patient. The patient must always be at the center of any decision related to healthcare and the practice of medicine. It is important to acknowledge the narrative, subjective and human aspect of medicine. Despite its high potential, AI is currently not consistent or accountable for independent clinical diagnostics. It is important that we understand that generative AI is at best, a predictive tool which may make errors and can never replace expert human judgment.33 The accuracy of AI models further drops substantially when these diagnoses are made on a conversation with a simulated patient instead of documented data, thereby highlighting its limitations as an independent entity.34 The demise of neurologists was once predicted with the advent of brain imaging.35 Instead, physicians embraced its use, based on good evidence, to improve patient outcomes across multiple neurological disorders. As we incorporate AI into improving patient care and the practice of medicine, it is imperative that researchers partner with clinicians in the development and exhaustive testing of AI before rushing into their premature and dangerous deployment in direct patient care. (1) Research Project: A. Conception, B. Organization, C. Execution; (2) Statistical Analysis: A. Design, B. Execution, C. Review and Critique; (3) Manuscript: A. Writing of the First Draft, B. Review and Critique. A.M.: 1A, 1B, 1C, 3A A.J.L.: 1A, 1B, 1C, 3B Ethical Compliance Statement: This document was written following ethical guidelines, in the absence of an IRB approval. Informed patient consent was not necessary for this work. All authors have read and complied with the Journal's Ethical Publication Guidelines. We confirm that we have read the Journal's position on issues involved in ethical publication and affirm that this work is consistent with those guidelines. Funding Sources and Conflicts of Interest: The authors declare no funding for this effort. The authors declare no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Financial Disclosures From the Last 12 Months: AM served on the scientific advisory board of adaptive biosciences. He serves as an associate editor for MDCP. AJL reports consultancies from Britannia Pharmaceuticals and BIAL Portela. He also reports honoraria from Britannia Pharmaceuticals, BIAL, Convatec and FMQ Brazil. Data sharing not applicable to this article as no datasets were generated or analysed during the current study.

Ähnliche Arbeiten