Expert: AI can help in the fight against rare cancers

Thanks to machine learning and, in particular, deep learning using artificial neural networks and the increasing availability of digital health data, the diagnosis and treatment of cancer is making great progress. Biologist Marc Bovenschulte writes this in a short study recently published by the Office of Technology Assessment at the German Bundestag (TAB). According to him, people suffering from rare cancers could also benefit from such approaches based on artificial intelligence (AI). However, such further developments, which involve dealing with increasingly personalized and data-driven medicine, face numerous technical and regulatory challenges.

Rare tumors are defined as those that affect fewer than 6 in 100,000 people. In Germany, these include esophageal, laryngeal and thyroid cancer, Hodgkin's disease (malignant disease of the lymphatic system) and certain forms of leukemia. According to the study, considering the individual genetic, physical and morphological characteristics of patients increases the chance of "a precisely tailored treatment that is as effective as possible and has as few side effects as possible". AI approaches could be used here in the diagnosis, the selection of suitable therapeutic measures, the prognosis of the course of the disease and the therapeutic support of those affected. The technology also plays a role in the development of medication and new, personalized therapeutic approaches.

AI systems are particularly well suited to evaluating different data such as X-ray images, molecular biological information, sequence information from DNA analyses or literature databases, comparing them with each other, relating them to each other and drawing conclusions from them, explains Bovenschulte. CAD systems, for example, carry out an analysis of the image content in addition to humans and incorporate patterns from comparative or reference data to highlight conspicuous areas. However, this is less successful with small numbers of cases. In a study on the detection of heterogeneous tumors, however, a deep learning AI was almost twice as good at assessing the degree of aggressiveness of the disease based on computer tomography images compared to the classic method. The author also describes the creation of a digital twin to model treatment based on a virtual replica of patients as promising, as well as personalized mRNA vaccines in the fight against recurrent cancer.

Approval regime does not match individual therapy

Considering the boom in generative AI, experts are also exploring the extent to which ChatGPT & Co. can be used in precision medicine for cancer treatment, according to Bovenschulte. The background to this is that knowledge of the biology of a tumor also improves the possibilities for its treatment. In initial experiments at Charité, for example, the bots have provided "some useful suggestions and clues" and in two cases even "unique therapeutic approaches" that no one had come up with before. As a rule, however, the results have not yet come close to the quality of human experts. Sometimes fictitious information (hallucinations) is integrated. In addition, the consistency of the results across different versions of the large language models is low.

According to the author, however, the individualization of therapeutic approaches described is "difficult to fit into the framework of existing approval regimes", which have so far been based as far as possible on extensive clinical studies with numerous test subjects, standardized products and procedures. Experimental treatments are permitted within the framework of therapeutic freedom and as therapeutic trials within narrow limits to treat individual patients with novel and less tested approaches. Nevertheless, Bovenschulte considers "a reliable framework" to be necessary in order to prove the effectiveness of personalized diagnosis and therapy. This also applies to AI-based procedures. A "diffusion of responsibility" involving hospital management, IT specialists or manufacturers in addition to medical professionals should be avoided.

(vbr)