Federated Learning in Practice Networks
FLIP-IT develops a secure and trustworthy federated-learning infrastructure for primary care practices. It enables medical AI models to learn collaboratively from distributed clinical data while sensitive patient records remain stored locally.
1. Local Data Processing as a Controlling Tool
Individual Views for Different User Roles
Docport Analytics provides user-friendly data visualisation that presents both business and medically relevant indicators for a primary care practice in a clear and transparent form. Through a modern web interface, the dashboard accesses care-delivery and billing data from the practice management system in near real time. A range of interactive dashboards and charts gives each user role tailored insight into key performance indicators and statistics.
The dashboard provides detailed information on appointment scheduling, enrollment status in selective healthcare contracts, such as family-doctor-centered care and Disease Management Programmes, as well as other relevant key figures that help make everyday practice operations more efficient. Additional tools, such as the preventive care radar, specifically suggest patients who, based on age, gender, and pre-existing conditions, meet the enrollment criteria for complex preventive care modules and could benefit from extended diagnostics reimbursed outside the regular budget.
Physicians : Specialized disease-specific dashboards provide insights into quality indicators of service delivery, for example in the care of chronically ill patients, such as patients with type 2 diabetes mellitus, or in the area of medication prescriptions. In addition, treatment pathways are checked against current guidelines, and suggestions for additional diagnostic measures are displayed. This allows non-critical treatment pathways, such as DMP programmes, to be delegated within the practice to qualified staff and to be monitored in a standardized way using rule-based decision trees.
Comprehensive monitoring and controlling tools are available for the economic management of the practice. These tools provide a daily updated overview of case numbers, services delivered across different service groups, budget limits, review periods, and the achievement of business-related target values. A privacy-compliant comparison with other practices also enables continuous assessment of the practice’s own performance. The targeted transfer of tasks and responsibilities within the practice team can be transparently supported through visualized target indicators.
All data queries are carried out exclusively locally, meaning that data transfer to a cloud infrastructure can be avoided and structural protection against unauthorized access is ensured. Logins are performed using individual user accounts and are password-protected. Authorized users can access the statistics at any time and regardless of location, including from mobile devices, via secure VPN connections. Additional modules are also being implemented step by step to support guideline-compliant and preventive patient care by facilitating the identification of high-risk patients and enabling early interventions.
2. Federated Analytics
The dimensions of practice performance and quality of care described above can often only be meaningfully interpreted in relation to comparable practices. For this reason, practice statistics without individual patient-related features, such as case numbers, revenue statistics, as well as descriptive demographic statistics, including distribution by age cohorts, gender, insurance status, and similar indicators, should be shared in a privacy-compliant manner. Technical measures ensure that conclusions about individual patients are excluded.
In the future, the dashboard should also enable the management of and participation in university research projects and study collaborations. Practice owners who have agreed to participate will always receive a daily updated overview of study inclusion criteria and the number of patients in their practice who would meet these criteria. While practice owners can use this to identify individual patients for studies or new treatment options, with patient contact by the treating physician only if desired or consented to, federated analytics will return only aggregated numerical values, for example as total numbers or percentages, which can then be shared with the study center. This makes it possible to estimate the size of potential study cohorts and the necessary sample sizes. This extension not only supports the evidence-based advancement of medical care, but also promotes interdisciplinary exchange between medical practices and scientific research.
3. Federated Learning
AI tools could effectively support practice teams in identifying high-risk patients at the population level, selecting individually suitable interventions, and communicating with affected patients in a targeted way. While corresponding research activities already exist in the clinical context, for example “ACRIBIS – Personalized Risk Assessment for Cardiovascular Diseases” by the Medical Informatics Initiative, comparable projects for the outpatient sector are still lacking, according to the Institute for Medical Informatics and Biometry at TU Dresden.
General practices in particular currently play only a minor role in the training of contextualized AI models, despite their central role in healthcare delivery. Reasons for this include the technical challenges already mentioned, such as heterogeneous data silos, lack of interoperability, and insufficient concepts for data security, as well as the limited resources of many practices in terms of time, expertise, and financial means. The frequent structure of individual standalone practices limits the transferability of insights gained from their own data and makes it more difficult to systematically use local healthcare data for research purposes.
The consolidation, or centralisation, of health data by industry or research institutions is often presented as a solution, but it is subject to strict requirements for the protection of patient-related information and the associated consent obligations. Developing predictive models is further complicated because anonymising patient data generally entails at least a partial loss of information, while pseudonymisation requires the involvement of trusted third parties. Without new approaches to data use, dependence increases on multinational IT providers and on AI models trained outside the environments in which they will eventually be deployed.
The Federated Learning Approach
Federated Learning (FL) enables the efficient training of a wide range of AI models—including linear and logistic regression models, support vector machines, and neural networks—across decentralised practice databases without requiring personal medical data to be centralised or otherwise exchanged. In simple terms, the principle is: ‘The model moves to the data, rather than the data to the model.’ This represents a paradigm shift from conventional centralised data-processing approaches and is regarded by leading researchers as a key technology for AI research involving sensitive medical data (Rieke et al., Nature Medicine, 2020).
Provision of descriptive statistics for defined patient groups and benchmarking between participating practices (see ‘Federated Analytics’).
Training of logistic regression models for the early detection of disease complications, such as chronic kidney disease.
Local deployment and use of AI models trained across the practices’ decentralised datasets, for example through local early-detection dashboards.
Technically, the platform is based on the federated learning system Flower.ai. It supports data-protection-compliant work with decentralised and heterogeneous data sources, as well as the federated training of different AI models, including logistic and linear regression models. Data harmonisation and standardisation are achieved through the international interoperability standard HL7 FHIR, implemented through a local translation layer between the practice management system database and the Flower client.
Technical controls ensure that only aggregated numerical values-such as patient counts, quality indicators, and AI-model weights can be communicated between the clients in the practices and the server at the docport study centre. The accidental or intentional sharing of personal data points is excluded, preserving data confidentiality throughout the analytical process. Advanced privacy-enhancing technologies, including secure aggregation and differential privacy, further reduce the potential identifiability of participating practices and address this additional aspect of protecting sensitive information.