|
| Challenges | Reason for challenges | Possible solutions |
|
| Confidentiality and safety | The fundamental issue with standard ML/DL models is that data from personal devices, sensors, and wearables from patients must be uploaded to a cloud server in order to train the data using the ML/DL models | In FL, instead of transferring data to the central servers, the ML model itself is deployed to each device to be trained on the data |
|
| Data heterogeneity | Healthcare data is heterogeneous for a number of reasons: | FL addresses the problem of heterogeneity by utilizing FedProx |
| (i) Differences among patient populations |
| (ii) Environments |
| (iii) Practices, and |
| (iv)Treatment protocols |
|
| Traceability and accountability | In FL, one of the biggest challenges is ensuring that the global ML model can be traced throughout the underlying ML process | (i) Traceability should be ensured during the training process to permit tracking of system events |
| (ii) Data access history and training configuration changes, such as hyperparameter tuning |
|
| System architecture | Using a client device that provides training and communication to the model can be difficult, which can lead to low-quality models | Healthcare institutions have usually better computing resources and high-speed networks compared to consumers, so they can run FLs at scale |
|
| Client management | Client management is an essential issue in FL, in contrast to the centralized ML architecture | Client management involves helping a patient or client develop a plan that coordinates and integrates essential support services for the most optimal results and outcomes |
|
| Health dataset issues | The majority of FL approaches are typically examined on a single dataset with a limited number of features | An inference strategy is presented to enable participants to use an ensemble of heterogeneous models without needing to explicitly join the data in a single place |
|
