AI Support for Patient Care Pathways in Oncology

Cancer treatments have improved significantly in recent years thanks to targeted therapies and immunotherapies, which have substantially increased survival rates but also made treatment more complex. Precision oncology—personalized medicine based on genetic tumor profiles—requires highly specialized treatment decisions. Oncology guidelines are intended to ensure comparable treatment for all cancer patients through evidence-based recommendations. In practice, however, there is often a discrepancy: studies show that, globally, only about 50% of patients are treated in accordance with guidelines.

In cooperation with the Kantonsspital Graubünden, the project group has set itself the goal of improving and standardizing the treatment of oncology patients. This is to be achieved through a data-driven analysis based on so-called patient or treatment pathways as the data foundation. Through structured modeling and visualization of oncology treatment pathways, deviations from existing guidelines—and thus bottlenecks in care—can be identified, and targeted solutions can be developed using artificial intelligence methods. In the long term, this will contribute to better quality of care and, at the same time, can lead to a reduction or optimization of treatment costs through standardization, as well as to a sustainable reduction in the burden on the healthcare system.

The steadily rising costs of health care—driven by an aging population, the increase in chronic diseases, and the use of increasingly complex medical treatments—pose major challenges for health care systems worldwide. This trend is also continuing in Switzerland: hospital expenditures are rising steadily, while at the same time the demand for high-quality and efficient care is growing.

This complexity is particularly evident in oncology. The introduction of targeted therapies, immunotherapies, and precision oncology has significantly advanced the treatment of cancer patients. While these advances improve survival rates, they also lead to complex treatment decisions that require precise and structured planning. Although clinical guidelines provide evidence-based recommendations for comparable care, they are often only partially implemented in practice. Reasons for this include institutional differences, varying treatment routines, and the predominantly unstructured documentation of medical procedures.

This lack of standardization complicates the systematic analysis of treatment pathways, the identification of deviations from guidelines, and the identification of opportunities for improvement. At the same time, there is a growing need for data-driven methods to improve the quality of care and use resources more efficiently. Technological approaches such as artificial intelligence (AI), natural language processing (NLP), and process mining offer new possibilities in this regard: They make it possible to convert unstructured medical texts into analyzable data formats, transparently map treatment steps, and thereby create a sound basis for decision-making aimed at the sustainable optimization of the healthcare system.

The goal of this project is to present and analyze patient processes in a structured manner based on clearly defined treatment steps. This should make it possible to map oncology treatment pathways in a transparent way and evaluate them systematically. Particular emphasis is placed on identifying deviations from medical guidelines as well as detecting potential gaps in care and inefficiencies.

As part of the project, the foundation was first laid by investigating how unstructured medical texts can be automatically converted into standardized coding systems such as ICD-10 or CHOP. This automated coding enables a uniform database on which further analyses can be based. Subsequently, initial treatment pathways were reconstructed and visualized using statistical and process-oriented methods (process mining) to illustrate the actual course of oncological treatments.

These results form the basis for further developments. In upcoming project phases, the developed analytical methods will be applied to routine clinical data and integrated into an interactive visualization. The goal is to provide healthcare professionals with an intuitive dashboard that facilitates the interpretation of treatment pathways and supports data-driven decision-making.

In the long term, this project will help standardize oncology care, improve the quality of treatment, and enable a more efficient use of available resources.

However, the path to data-driven patient pathway analysis is challenging. Medical treatments are primarily documented in unstructured text format. While this format is easily understandable to humans, it only allows computers to model treatment pathways in a structured way to a limited extent. Therefore, as a first step, the project team investigated how medical texts can be converted into structured data formats.

In medical practice in Switzerland, particularly with a focus on billing, treatments and diagnoses are documented using standardized codes, such as the Swiss Classification of Operations (CHOP) or the International Statistical Classification of Diseases and Related Health Problems, 10th Revision (ICD-10). There are several such coding systems in use internationally. Some of them cover only specific areas, such as outpatient, inpatient, or systemic therapy, while others take a more holistic approach, such as SNOMED CT (Systematized Nomenclature of Medicine – Clinical Terms).

The conversion of medical texts into such coding systems is referred to in research as “automatic coding systems.” In an international collaboration with the University of San Francisco, the project team trained small language models capable of performing this automatic coding. The focus was deliberately placed on small language models, as many hospitals do not have extensive IT resources and, due to data protection requirements, cannot readily access cloud services. Our goal was therefore to develop a high-performance model for the automated coding of medical text data that can be deployed locally with limited resources.

The resulting research was presented and published as the scientific study “Evaluating Fine-Tuned Small Language Models for ICD-10 Diagnosis Coding” by Victor Palacios, Curdin Marxer, B. Sunshine, and Prof. Dr. Yves Staudt at the Swiss Data Science Conference 2025 in Zurich.

After the automatic linking of text and medical codes was successfully implemented, treatment pathways were analyzed and visualized in a subsequent step using statistical and process-oriented methods, collectively referred to as “process mining.” An example of such a visualization was created for prostate cancer patients based on the 1998–2022 healthcare statistics from the Federal Statistical Office (see Figure 1).

We are currently working on implementing our approach using data from the Tumor Center at the Graubünden Cantonal Hospital. The next step will be to convert this into an interactive visualization via a dashboard to enable medical professionals to use it intuitively and in a practical manner. We would like to extend our special thanks to the Friends of the University of Applied Sciences of Graubünden and to the University of Applied Sciences of Graubünden for their financial support of this research project and for making it possible.

The project resulted in the following scientific article: “Evaluating Fine-Tuned Small Language Models for ICD-10 Diagnosis Coding” by Victor Palacios, Curdin Marxer, B. Sunshine, and Prof. Dr. Yves Staudt

The project was also supported by the following individuals:

• Prof. Dr. med. Roger von Moos – Director of the Tumor and Research Center, FMH Medical Oncology and Internal Medicine – Graubünden Cantonal Hospital
• Lucas Basler – Senior Radiation Oncologist & Data Scientist / Business Analyst & Project Manager – Artificial Intelligence (AI) – Graubünden Cantonal Hospital
• Rosaria Tino-Corrado – Assistant at the Tumor and Research Center – Graubünden Cantonal Hospital
• Prof. Dr. Joël Wagner – Full Professor at the University of Lausanne (HEC) | Director of the Department of Actuarial Science | Member of the Board of Directors of Retraites Populaires and LALUX Assurances
• Selina Steiner – Intern with Prof. Dr. Yves Staudt
• Raphael Brunold – Intern with Prof. Dr. Yves Staudt
• Victor Palcios – University of San Francisco, Director of the Data Science Partnership

Participants

The project was carried out by the Institute for Data Analysis, Artificial Intelligence, Visualization, and Simulation (DAViS) on behalf of the Kantonsspital Graubünden.