Researchers from the Berlin Institute for the Foundations of Learning and Data (BIFOLD) at TU Berlin, the Faculty of Medicine at the University of Duisburg-Essen (UDE), and LMU Munich have created a novel artificial intelligence (AI) tool that can decipher complex interrelationships. The study was published in Nature Cancer.
Professor Frederick Klauschen is Director of the Institute of Pathology at LMU. Image Credit: Ludwig Maximilian University of Munich
The goal of personalized medicine is to customize care for each patient. The progression of a disease has been predicted using a limited set of parameters. These few criteria, however, are frequently insufficient to comprehend the complexity of illnesses like cancer.
Leveraging the smart hospital infrastructure at University Hospital Essen, the researchers have combined data from various modalities—including medical history, laboratory values, imaging, and genetic analyses—to enhance clinical decision-making.
Although large amounts of clinical data are available in modern medicine, the promise of truly personalized medicine often remains unfulfilled.
Jens Kleesiek, Professor, Institute for Artificial Intelligence in Medicine (IKIM), University Hospital Essen
Jens Kleesiek is also associated with the Cancer Research Center Cologne Essen (CCCE).
Interaction of 350 Parameters Examined
Individual differences like sex, nutritional status, or comorbidities are not taken into consideration by the rather rigid assessment systems used in oncological clinical practice today, such as the classification of cancer stages.
Modern AI technologies, in particular explainable artificial intelligence (xAI), can be used to decipher these complex interrelationships and personalize cancer medicine to a much greater extent.
Frederick Klauschen, Professor, Director and Research Group Leader, Institute of Pathology, Ludwig Maximilian University of Munich (LMU)
This approach was developed together with Prof. Klaus-Robert Müller.
The AI was trained using data from over 15,000 patients with a total of 38 distinct solid tumors. A total of 350 parameters were analyzed, including genetic tumor profiles, imaging procedure data, laboratory values, and clinical data.
We identified key factors that account for the majority of the decision-making processes in the neural network, as well as a large number of prognostically relevant interactions between the parameters.
Dr. Julius Keyl, Clinician Scientist, Institute for Artificial Intelligence in Medicine (IKIM)
Transparent Decisions
The detected interactions were then confirmed by successfully testing the AI model on data from more than 3,000 patients with lung cancer. The AI compiles the information and determines each patient's overall prognosis. Being an explainable AI, the model explains to clinicians how each parameter affects the prognosis, making its decisions transparent.
“Our results show the potential of artificial intelligence to look at clinical data not in isolation but in context, to re-evaluate them, and thus to enable personalized, data-driven cancer therapy,” said Dr. Philipp Keyl, LMU.
This kind of AI approach could also be applied in emergencies where it is critical to evaluate all diagnostic parameters as soon as possible.
Using traditional statistical techniques, the researchers hope to find intricate cross-cancer relationships that have not yet been identified.
“At the National Center for Tumor Diseases (NCT), together with other oncological networks such as the Bavarian Center for Cancer Research (BZKF), we have the ideal conditions to take the next step: proving the real patient benefit of our technology in clinical trials,” said Martin Schuler, Managing Director, NCT West site and Head of the Department of Medical Oncology, University Hospital Essen.
Journal Reference:
Keyl, J., et al. (2025) Decoding pan-cancer treatment outcomes using multimodal real-world data and explainable artificial intelligence. Nature Cancer. doi.org/10.1038/s43018-024-00891-1