Artificial Intelligence in Cancer Clinical Research: I. Introduction

“insights that are available to anyone who can think logically with understanding and imagination lie beyond anything that can be formalized in a set of rules. Rules can sometimes be a partial substitute for understanding but they can never replace it entirely.”

- Roger Penrose

Human intelligence

In this first of the series on Artificial Intelligence in Cancer Clinical Research, it is important to attempt to define the concept of intelligence and the important features we associate with human intelligence. Intelligence is a construct that we associate most notably with human beings and our understanding of human intelligence including perception, cognition, thought, conceptualization, pattern recognition, symbolic processing, creativity, and problem solving. While we acknowledge apparent elements of intelligence in other living species, we most commonly think of intelligence as a manifestation of our conscious awareness and ability to form concepts along with our capacity for symbolic processing, the development of language and ultimately our ability to reason, make decisions and use judgement as we navigate our world (1–4). At the same time, while consciousness and self-awareness seem fundamental and unique to the human experience, these concepts remain profound and elusive mysteries. How can the complex array of atoms and molecules that ultimately behave the physical laws of motion, along with the intricate network of neurons in the central nervous system, give rise to our internal conscious experience and ability to think and reason along with experiences such as joy, sadness, love and beauty? While conceptual processes may be amenable to machine learning (ML), the human intellect is capable of generating new, abstract concepts that further organize his or her environment and, importantly, discover abstract ideas. Such processes provide new insights beyond the actual world and, in a sense, create his or her own world of meaningful symbolic systems with an active curiosity striving for meaning and true understanding of the world around us (2). While a sign represents a characteristic of the external world amenable to algorithmic analysis and mimicking, a symbol evokes an abstract internal response separate from a particular concrete event (2). Human symbolic systems are not only used for communication of cognitive information enabling language but the communication of emotional information. In fact, many of humanity’s most important and highest qualities relate to the development and creative use of symbolic systems resulting in the development of culture, myth, idealism and empathy. While a sign can be thought of as a context free characteristic of an object or situation, a symbol refers to experiences within ourselves. These experiences gain meaning within a context of emotions and interests resulting from the intersection of our prior experience and memory and the present situation to generate feelings or emotions within ourselves as we participate as active contributors to the creation of our world (1).

Artificial intelligence

Advances in computing technology have emerged with astounding and accelerating speed. This has given rise to the question of our ability to replicate or even exceed the capacity of our living neural circuitry to think and achieve features we associate with intelligence and the concepts of ‘artificial intelligence’, ‘machine learning’ and ‘large language models’ capable of responses we would normally associate with thought, learning and perhaps even consciousness. Features of artificial intelligence are emerging and are now blossoming across virtually every aspect of human activity including medicine and medical research. Artificial intelligence (AI) represents a broad array of considerations of automated processing of tasks previously relegated to humans. Machine learning (ML) is that aspect of AI related to the learning of patterns in data usually in the form of supervised learning based on mathematical algorithms that can learn from previous results over time performing future tasks using mathematical optimization methods. These can range from more routine repetitive tasks to generative AI neural networks. ML approaches are generally differentiated as either ‘supervised’ or ‘unsupervised’ techniques. In supervised approaches, labeled data are used to train a model and often used for classification and in prognostic or prediction models. In unsupervised methods, the modeling attempts to discover patterns of structure within the data without any prior labelling. ML has been applied in oncology to image analysis, natural language processing (NLP), clinical decision support and large real world data analysis among others. While ML broadly relates to methods of automated learning with minimal human involvement, deep AI or deep learning utilizes advanced neural networks in an attempt to simulate the function and responses of the human brain. Deep learning (DL) based on deep neural networks often based on billions of parameters has gained considerable attention for diagnostic and predictive modeling of treatment effects and toxicities and gene sequencing among many others.

Large language models (LLM) represent deep machine learning models previously trained on extremely large datasets transformed by neural networks working in parallel. These attempt to extract meaningful relationships and establish a grammar from sequences of text through unsupervised learning representing words as multidimensional vectors with statistical relationships. In essence, such models attempt to ‘understand’ and ultimately generate output similar and largely indistinguishable from human language. Although the output clearly does not represent actual thought as we experience it, it may appear to do so to an outside observer.

The term strong AI is often used for the aspirational goal of computer responses based on algorithms that could not be distinguished from those of a human being. A debate continues as to whether such a computer program, if achieved, would actually be conscious and self-aware in the same sense that the human mind is. Those supporting the affirmative argue that our sense of consciousness and self-awareness arise naturally from the immense complexity of interconnected neurons and resulting impulses that exist within the human brain. Therefore, they argue, all human thought and awareness is ultimately computational and results from the countless computations occurring simultaneously in the brain. Based on this presumption, strong AI seems destined, at some point in the future to have the ability to replicate human thought and consciousness. At that point, the behavior of strong AI software would become indistinguishable from human cognition and responsiveness.

However, the likelihood of this scenario emerging is challenged by the Incompleteness Theorem of Kurt Godel who demonstrated that there are computations that no algorithm can yield and truths that no formal system can provide. Likewise, Turing has demonstrated in the ‘Halting Problem’ that no computer algorithm can reveal when it will go into an infinite loop or prevent an infinite loop. The mathematician Roger Penrose has concluded that insights available to anyone thinking logically lie beyond what can be formalized as a set of rules or calculations and, therefore, aspects of higher intelligence and insight represent a non-algorithmic function of the brain (4). Aspects of higher intelligence associated with the human mind appear to be beyond anything reducible to a true formal or algorithmic system. This would lead one to conclude that human consciousness is beyond that which is simulated by a computer algorithm and, therefore, no algorithmic computer can ever completely demonstrate conscious thought or self-awareness.” (5)

General Limitations of AI: Admittedly, the rate of progress and output associated with advances in AI technology have exceeded nearly all expectations. However, like all rapidly advancing technologies, important limitations have been identified that are not always fully appreciated and may impact on the validity and accuracy of the applications and their output. LLMs have been successful in accessing huge amounts of clinical data and often perform quite well in relatively structured scenarios based on increasingly sophisticated algorithms. Nevertheless, challenges remain concerning their accuracy with both false positive and false negative results providing convincing but factually erroneous information reinforcing the need for continued clinical oversight (6,7). Despite the remarkable rate of progress and often impressive results with current LLMs, careful evaluation of LLMs continues to reveal frequent factual errors and even fabricated information suggesting that much additional development and validation is needed before clinicians, patients, or policy makers can, if ever, rely entirely on the output of such models (8). A joint Working Group of the German Society of Hematology and Oncology, the German Association for Medical Informatics, Biometry and Epidemiology and the German Informatics Society have put forward a roadmap for the development and implementation of AI in Hematology and Oncology (9). They highlight several of the more apparent general limitations of modern AI including data security and privacy, The potential for access to personal data has necessitated enhanced criteria for ethical review, informed consent and prior approval for the use and application of patient-related data (9). Also highlighted, is the impact of potential biases of data-driven supervised ML approaches limiting the robustness and generalizability of results built on patterns learned from training data sets and their specific inherent limitations. It has been demonstrated by several authors that the performance of AI-generated predictive models declines in other data sets that differ from those used for development or early validation studies (10,11). If the population represented in the training data are not representative of the broader population to which the model is to be applied or certain variables are defined or collected in a biased fashion due to technical or logical errors, the conclusions drawn for other populations may not be valid. As Ghassemi and colleagues have shown, AI algorithms applied to chest x-rays display underdiagnosis bias in underserved patient populations different than the populations used in the training of the algorithm thus exacerbating already existing care biases (12).

Another important caveat for complex modeling based on AI may lie in an inability to adequately explain the results and confirm model validity, as the needed statistical modeling methodology, which ensures the scientific rigor of results, tend not to be considered by these models. Importantly, the suggestion that AI may largely replace clinicians in providing the needed emotional support and empathy expected along with the important clinical knowledge patients expect appears inherently flawed. In a recent study by Ayers and colleagues, they report that chatbot generated quality and empathetic responses to patient questions posed in an online forum actually surpassed those from physicians (13). However, these conclusions have been challenged by other authors who point out that the assessment of the quality and empathy reflected in those encounters were provided by physicians and not by patients. In the clinician-patient interaction, it is essential that the clinician truly feels empathy for a patient as opposed to receiving an optimized empathetic-appearing response from an algorithm (14). Cadiente and colleagues argue that the proper evaluation of the patient experience lies with the patient and not with the clinician. Clearly, patients should be involved in the evaluation process and judge whether their experience was understood, shared and satisfactorily acted upon.

A remaining challenge in assessing AI models lies with the complex ‘black box’ nature of AI data and algorithms. At the same time, there appears to be continued pervasiveness of digital and AI illiteracy among the population including clinicians and patients. Moving forward, it is essential that the integration of AI methods into clinical practice and research be primarily driven by close collaboration between knowledgeable data clinician scientists and modeling methodologists in order for clinical modeling efforts to identify and attempt to mitigate the major challenges involved. The benefits and limitations of utilizing AI-driven approaches must be evaluated in well designed, comparative, large-scale population-base clinical trials before such models can be safely used for guiding patient care or forging healthcare policies in the future (9).

Conclusions

For clinicians, the most critical challenge will not be that of designing or analyzing clinical research studies based on ML but will rather be truly understanding, critically appraising and applying the results of such studies. Liu and colleagues from Google Health have presented a ‘user’s guide’ on how to read articles that use ML (15). The authors lay out the new language of AI related to ML methods and the principles of evaluating and applying the results of modeling studies while contrasting traditional decision methods with those emerging from ML approaches. The authors provide a checklist for reading manuscripts based on ML studies to make clinically relevant predictions. Nevertheless, a fundamental challenge for all developers and users of AI including ML studies is that the results nearly always lie somewhere between always being correct or often being incorrect. In reality, such studies appear to provide increasingly accurate albeit imperfect and sometimes completely erroneous predictions. Experienced clinicians and scientists will continue to be called upon to critically appraise and hopefully refine such models in order to improve their own imperfect predictions based on training and experience. While being cautiously optimistic and hopeful, we must remain vigilant of the limitations of AI methods for clinical research. It is absolutely essential that AI methods, like conventional research methods, be carefully scrutinized and utilized appropriately so as to complement rather than replace the essential role of human intelligence with its important intuitive, ethical, and empathetic abilities. It is imperative that clinicians and investigators limit the potential inappropriate or mistaken use of these methods no matter how enticing (3). In the end, the goal must be that of carefully, responsibly and appropriately utilizing the best tools available including clinical acumen and experience to make the very best decisions for our patients and society.

Declaration of interest

The authors report no conflict of interest. The authors alone are responsible for the content and writing of the article.

Additional information

Funding

The author(s) reported there is no funding associated with the work featured in this article.