Pediatric clinics face major challenges with rare diseases due to their complexity, limited research, and lack of treatment protocols. Diagnosing these conditions can take years, involving specialists, genetic testing, and inconclusive results — delaying care and distressing families. Clinics often rely on experimental or off-label treatments, adding further uncertainty. The ECHO (European Children’s Hospitals Organization) helps tackle these issues by sharing expertise, resources, and best practices across borders. This collaboration speeds up diagnoses, boosts research, saves costs, and improves clinical trial feasibility. Families benefit from faster, more accurate care, easing both emotional and financial burdens.
1 - Infrastructure setup
Here you can find an overview of the infrastructure setup comprised of Local Developer Environment, Feature Cloud, and on-premise server at the clinic.
The developer use their local environment to build their machine learning applications and provide them to the user at the clinic via the feature cloud app store, allowing the user at the clinic to securely retrieve the machine learning application and run them on their data locally.
This setup ensure that data resides on the clinic system and infrastructure at all times.
Developers obtain mock data from Neo4j in their local environment. The mock data mimics the data model on the clinic side.
Developers in the local environment follow the prepared template to build their machine learning applications and test them with the mock data. They can then upload their machine learning applications to the feature cloud app store.
The user at the clinic can select their machine learning application of interest from feature cloud app store and download them to their local environment via the pull process. They can then securely execute the machine learning application on their data in their local environment.
Depending on the performance of the application, the user at the clinic can choose to update the machine learning application residing on the feature cloud app store.
2 - What are graphs?
What exactly is a ‘graph’?
A graph is a modeling technique for complex environments. It thrives best in modeling relationships between entities.
Example: Two different objects, a patient and a disease, both are represented as a node in space. If the patient has that disease a connection between the two is made.
Why a graph?
For any kind of machine learning application the most important part is data. Being in the field of rare diseases this data unfortunately does not exist in suitable quantities. The idea is to leverage everything we have learned about human biology, disease, drugs, etc, to make up for this lack of data. Thus enabling the use of machine learning algorithms. All this background knowledge is best modeled via relationships between entities, hence the usage of the graph.
What kind of information is in our graph?
The graph used in this project is an evolution of the ‘clinicalknowledgegraph’ (https://ckg.readthedocs.io/en/latest/). It was originally designed for use in proteomics and was modified to fit the task at hand.
A detailed wiki on the graph, its structure, and the different node classes can be provided to you.
Please reach out to Henrik.Otterstedt@med.uni-muenchen.de.
What are the benefits for my clinic?
By incorporating the AMIGO approach into your clinic, we want to enable you to achieve the following goals:
Assist in the diagnosis of rare diseases. For immediate use in the clinic, the aim is to assist in the diagnosis and treatment of rare disease patients. The idea is to find similarities with existing patients and use the background to identify causal genes. This will be done not only in Munich, but also in partner hospitals throughout Europe.
Working with pharmaceutical companies to identify patients for clinical trials: Another goal is to identify patients who might be suitable for a clinical trial by a pharmaceutical company.
3 - About FeatureCloud
FeatureCloud is a lightweight pan-European AI development project aimed at mitigating cyber risks to healthcare infrastructure. Their focus is on creating innovative software toolkits employing federated learning with an all-in-one approach.
FeatureCloud addresses the main obstacles of AI and clinical data with two key characteristics: (1) no sensitive data is sent through any communication channels, and (2) data is not stored in one central point of attack.
The security and confidentiality of your data is our priority. For this reason, we use federated learning, a machine learning approach that does not require any data transfer to a centralized server or cloud. Instead, it allows the user to retrieve the model and run it on their data locally.
FeatureCloud provides an all-in-one federated learning platform and mediates the secure exchange and useability of the machine learning applications.
This ensures a complete separation of clinical data from the development environment and allows research partners from different clinics to collaborate without facing any data privacy restrictions or risking a central point of attack.
Inference, results and predictions of the machine learning models used for diagnosis is in full control of the clinics and they can choose to update the global machine learning model as they see fit. If clinics decide to share their results with other entities, they must do so explicitly.
