Competence Unit for Federated AI & Bioinformatics Analysis (KIBCU)
AI and bioinformatics
Digitalization in healthcare and research is generating clinical data, image data, biosample information and molecular data at the BZKF that can be used for innovative bioinformatics processes and artificial intelligence methods. The aim is to improve oncological care and research processes in order to support clinical decision-making and optimize oncological patient care.
The IT strategy of the BZKF is based on concepts of the German Medical Informatics Initiative (MII) with the aim of keeping "Real World Data (RWD)" from patient care locally at the university hospitals and bringing the analysis to the data (federated analysis and federated machine learning). A basic infrastructure (based on DataSHIELD) was established for this purpose. The pipeline for data extraction, pseudonymization and mapping to the MII oncology core dataset module is continuously reviewed and further developed in order to transfer oncological data to the FHIR database of the BZKF sites. The AI & Bioinformatics Competence Unit (KIBCU) works closely with the clinical specialists, the Data Integration Centers (DIZ) at the BZKF sites and the BZKF IT working group.
- Heterogeneous IT systems at the sites (e.g. tumor documentation or biobank management system)
- Different data formats and data elements
- Central data consolidation of retrospective routine data (RWD) not permitted due to lack of consent
- No sustainable research infrastructure (secondary use)
- Lack of skills for distributed machine learning (AI)
- Lack of large, high-quality database (AI)
- Complexity of processing, storing and analyzing bioinformatic data
- As a result, no cross-site retrospective data use for scientific knowledge gain possible
- Data harmonization
- Development/revision of standardized pipelines
- Data quality and plausibility checks
- Data integration
- Federated evaluations
- Distributed machine learning (AI)
- BZKF-wide further education and training in the fundamentals and methods of AI development and concepts of federated machine learning
- Local IT infrastructures (data extraction and harmonization pipeline, DataSHIELD) for federated analysis and machine learning established at all sites IOS Press Ebooks - Towards a Bavarian Oncology Real World Data Research Platform
- Provision of the pipeline for data transfer, pseudonymization and mapping to bring basic oncological data in FHIR format into the DataSHIELD database
- Establishment of a BZKF ETL task force for task and knowledge sharing in the further development process of the pipeline
- Integration of data quality and plausibility checks into the pipeline
- Joint publications on distributed analyses using our own "onco-analytics-on-fhir" pipeline JMIR Preprints #65681: Bridging Data Silos in Oncology with Modular Software for Federated Analysis on FHIR: A Multisite Implementation Study
- Annual BZKF Summer Schools and regular AI webinars
- Establishment of a BZKF-wide research data repository with clinical data, image data, sample information and molecular data
- Development of innovative bioinformatics pipelines and new AI-based decision support
- Ensuring a governance structure for data use and data sharing
- Data quality analysis and continuous feedback for data improvement
- Support study groups and medical researchers in bioinformatics analysis and AI developments
- Tracking the development and analysis of federated AI models based on the established infrastructure
- Development of new algorithms for federated machine learning on different data types
- Continuous BZKF-wide education and training in federated machine learning concepts
Recording of the Machine Learning webinar series October 2025
To further strengthen machine learning skills at the BZKF locations, the AI and Bioinformatics Lighthouse held a two-week webinar series in October 2025. In the events, participants were taught the basics of machine learning both theoretically and through practical exercises - with the aim of creating an understanding of central concepts and techniques and enabling participants to carry out simple modeling independently.
To the YouTube recording: Day 1: Basics of machine learning and supervised learning
To the YouTube recording: Day 2: Machine Learning project steps and classification algorithms
To the YouTube recording: Day 3: Sources of error and noise
To the YouTube recording: Day 4: Unsupervised learning and application example
Day 5: Practical exercises: Repository with the practical exercise/slides is available at https://github.com/bzkf/ml-webinar-2025-uebung
Recording of the BZKF Lighthouse Lecture 2025
To the Youtube recording from 02.07.2025
To the presentation
Recording of the AI webinar series 2025
To the YouTube recording from 29.10.2025
AI webinar: "AI in colorectal cancer prevention - experiences from 6 years of clinical application"
Speaker: Prof. Alexander Hann (UKW); Professorship for Digital Transformation in Gastroenterology
Specialist in Internal Medicine and Gastroenterology
To the Youtube recording from 10.04.2025
Application of artificial intelligence (AI) in clinical diagnostics and the discovery of new therapeutic approaches in hematology, especially in hematologic neoplasms (blood cancer and related diseases). In detail, an AI-supported clinical framework will be presented that aims to facilitate the diagnostic process and accelerate the discovery of translational therapies.
Speaker: Anke Bergmann (University Hospital Würzburg)
Recording of the AI webinar series 2024
To the Youtube recording from 16.10.2024
Real World Federated Learning in Radiology: Hurdles to overcome and Benefits to gain
What obstacles need to be overcome in order to integrate Federated Learning into real radiology applications?
Speaker: Markus Bujotzek, DKFZ Heidelberg
To the Youtube recording from 13.11.2024
An Experimental Survey of Incremental Transfer Learning for Multicenter Collaboration - How can Incremental Transfer Learning solve data protection problems in multicenter collaboration?
Speaker: Dr. Florian Putz, University Hospital Erlangen
Publication: An Experimental Survey of Incremental Transfer Learning for Multicenter Collaboration
Publication: Multicenter privacy-preserving model training for deep learning BM autosegmentation
To the YouTube recording from 11.12.2024
Evaluation and mitigation of the limitations of large language models in clinical decision-making
Speaker: Paul Hager, TUM
