Federated learning applied to medical document classification

Within the context of clinical text mining a real issue rests on data-sharing due to condentialityregulations. WhileDeep Neural Networksare strong, big data are required, however, gettinghealth records from dierent hospitals is barely feasible. To cope with this issue, we resort tofederatedlearning. This enables to infer a model from each hospital and combine the models’parameters without the need of sharing data. In this project the model purpose is classifymedical documents within a standard classication.The aim of this projectis to assess the performance of individual models learned fromdierent cohorts or hospitals and compare to a federated model.Learning outcomesThe student will acquire background in information extraction reinforc-ing the following areas:•deep learning applied to information extraction•multi-label classication•federated approaches•submit the results to a research paperGoalsthe student will cope with extreme multi-label classication and will evaluate andcompare the base-models with respect to a federated approach.Languages:this work can be carried out in English, Spanish or Basque.
Arantza Casillas & Alicia Pérez