Does artificial intelligence have a place in precision oncology?

The strength of artificial intelligence (AI) in precision oncology is that it can unlock the value of routinely available real-world data of millions of cancer patients. In recent years, AI has enabled oncologists to leverage large datasets from multiple sources of information, including next-generation sequencing and medical imaging, thus helping improve tumour profiling and providing a more comprehensive understanding of cancer (NPJ Precis Oncol. 2023;7:43).

Despite having the potential to transform precision oncology, the AI field is still in its early infancy and barriers to the implementation of AI tools in clinical practice still exist.

Large language models (LLMs) are a type of AI that can process large and diverse volumes of text at human-level competency to deal with the rapidly expanding amount of patient-specific data in oncology, and then generate outcome text from the processed information (NPJ Prec Oncol. 2024;8:72). Routine data, which are available for all patients, such as medical reports, pathology results and images, are the best data that AI can use to extract clinically meaningful evidence, while experimental data would have very limited benefits as it would be too costly to obtain and only available for select individuals.

LLMs can capture unstructured data and use them to provide useful information for clinical practice, such as to give guidance during surgery, aid diagnosis, predict response to treatment, recommend treatment strategies, predict treatment-related complications and provide physicians with medical knowledge (J Cancer Res Clin Oncol. 2023;149:7997–8006). The quality of an AI tool, however, depends on the quality of its training data, and racial, ethnic and gender biases can be a problem. If there is bias in the training data – the information used for training is not representative, contains errors or is incomplete, for instance – AI systems will learn and replicate the biased patterns resulting in unfair or discriminatory outcomes (Sci. 2024;6:3).

Accuracy, reliability and clinical relevance of AI models are key features that may drive their adoption in oncology practice, but these qualities are sometimes difficult to prove.

Most AI tools in academic publications, in fact, are not yet approved medical devices and must undergo approval according to the Medical Device Regulation (MDR) and also be supported with clinical evidence to ensure that they are effective and that their output is correct. Recently, a number of AI tools for oncology, mostly for radiological and pathological image analysis, were approved in the USA and the EU (NPJ Prec Oncol. 2024;8:72), and many others are currently under evaluation. Regulatory approval and the capacity of regulatory frameworks, however, can be a limiting step to adoption of AI-based personalised therapies. For instance, while AI-based healthcare tools are regulated under medical device laws, these or the laws that govern combinations of medicines and devices are inadequate for the greater complexities of AI-based personalised therapies. There is a clear need to adapt regulatory processes and the laws followed by regulatory bodies in order to assess AI-enabled personalised therapies appropriately and at a pace that parallels the development of novel tools (NPJ Prec Oncol. 2024;8:23).

Current AI systems in healthcare are designed for a single purpose, such as for analysing X-ray images or laboratory data, owing to the technical and regulatory constraints of medical devices. AI tools are evolving so that we can now integrate complementary data as a simultaneous input, such as images with text, or images combined with genomic data, which better reflects the decision-making processes employed by oncologists, who consider multiple pieces of information and corresponding data prior to prescribing treatments (NPJ Prec Oncol. 2023;7:43). Therefore, we must start thinking about how to implement multipurpose AI tools that process information from more than one source into the design of clinical trials for their validation.

Once patient data are in the appropriate format for the clinical validation of AI tools, AI has the potential to be used routinely and improve the treatment of our patients. However, the perception and trust of patients and healthcare professionals in AI tools are essential to translate research progress into clinical practice. The transition to AI support is generally viewed positively in the oncology community, but there are barriers that still need to be overcome in some individuals, including scepticism of AI among healthcare professionals with limited understanding of the technology, and concerns about cyber security, accuracy and decisions made without human compassion among patients (Technol Cancer Res Treat. 2022;21:15330338221141793).

Oncology societies, such as ESMO, can act as major driver to encourage further progress in the field, thus boosting oncology professionals’ confidence in investing in AI research and its applications. In November 2024, the first edition of the ESMO Artificial Intelligence and Digital Oncology Congress will be held in Berlin, representing a unique opportunity for oncology stakeholders to discuss relevant aspects including how to validate AI tools, the design of clinical trials, what the standards should be and how the outputs should be used. After the launch of the ESMO Real World Data and Digital Oncology journal in 2023, ESMO continues to deliver on its promise to be by the side of oncologists in a field that, in the digital era of oncology, can be a true game-changer.