From Ars Technica, by Fintan Burke:
- The concept of “federated learning”, inspired by the privacy-focused structure of new social media platforms, is being adopted by medical researchers to train AI in spotting disease trends. In this approach, user data is hosted on independent servers instead of a single corporate entity, which promotes data privacy and enables selective sharing of information.
- Instead of pooling patient data from various hospitals into one database, which raises privacy concerns and legal complications, researchers send their AI models to individual hospitals. These models can then analyze the data within the hospital’s firewall, maintaining the privacy of sensitive patient information.
- The training process involves doctors identifying eligible patients, selecting necessary clinical data, and organizing it on a local database. The AI software then uses this data to identify disease trends. The trained model is periodically sent to a central server, where it is combined with models from other hospitals to update the original model.
- The updated “consensus model” is sent back to each hospital to be trained further, and this cycle continues until the final model is deemed accurate enough. This process ensures data privacy, as the information sent back to the central server is anonymized and remains within the hospital’s firewall.
- Federated learning has seen significant growth in medical research. For instance, in 2021, a study successfully used this method to predict diabetes from CT scans of abdomens, potentially identifying at-risk patients up to seven years prior to their diagnosis.