Featured Learning together without exchanging data

The more an artificial intelligence learns, the better it can support doctors. To do this, it requires large amounts of data. But how can this be reconciled with the protection of sensitive patient data? A study led by the Schulthess Clinic shows how several hospitals can train AI models together without exchanging sensitive data.

Artificial intelligence is increasingly assisting doctors in the interpretation of X-ray images. In spinal surgery, for example, it helps to automatically measure the spine. Such measurements provide important information for diagnosis, treatment planning and the assessment of surgical outcomes, and help to improve treatments. 

For such AI systems to work reliably, they must be trained using large amounts of image data. However, this poses particular challenges in the healthcare sector: for data protection reasons, patient data often cannot be shared between clinics. This presented our researchers and medical staff with the question: how can several hospitals benefit jointly from large amounts of data without sharing sensitive information?

A research team led by the Schulthess Klinik investigated an approach known as ‘federated learning’. Under this approach, patient data remains within the respective clinic. Instead of consolidating the data at a central location, the AI model is trained directly on-site. Subsequently, only the learning outcomes are pooled, not the data itself. 

The key question of the study was: Can an AI system designed to automatically measure the spine still learn reliably even if the data never leaves the individual clinics? The result: Yes. This privacy-friendly method achieved almost the same level of accuracy as a system in which all data was collected centrally.

The study highlights the opportunities offered by this form of collaboration in medicine:  

• No compromises on data protection: Sensitive health data does not leave the respective clinic. This enables research and innovation without compromising data protection.
• Better quality of care: The AI can learn from a wider variety of data. This makes automated spinal measurements more accurate and reliable, which supports diagnosis and treatment planning.
• Benefits for smaller clinics too: It is not only large centres that benefit from such systems. Smaller clinics can also share in the collective knowledge and offer their patients modern AI-supported care without needing to have vast amounts of data at their disposal.

For the study, researchers from the Schulthess Klinik collaborated with four partner clinics from the European Spine Study Group consortium. University hospitals in Barcelona, Madrid, Bordeaux and Istanbul were involved. A total of 3,064 X-rays from 904 patients were included in the study. The data remained within the respective clinics at all times. The project was supported by the research fund of the Schulthess Foundation.

The study demonstrates that international research projects involving artificial intelligence are possible even under strict data protection regulations. Hospitals can pool their knowledge and learn from one another without exchanging sensitive patient data. 

For orthopaedics and spinal medicine, this opens up new opportunities to better connect research institutions and develop AI applications based on a broader data set. In the long term, this will benefit research and, in turn, patients.