Abstract

Deep learning models hold the potential to be integrated into clinical workflows, providing clinicians with a useful tool to help diagnose and plan the most appropriate action for their patients. However, for such models to be effective, generalisable and ethically promising, they must train on extensive datasets. These datasets are difficult to acquire in the healthcare setting, as a result of privacy concerns. In this article, I will explore a framework to get around this challenge and the important caveats to consider when implementing this, which arise due to the nuances of healthcare data. Finally, I will conclude with how, despite these caveats, this framework offers much promise for the future of AI in healthcare.

Deep Learning’s Clinical Promise

Deep learning models have been important to healthcare by acting as tools that doctors and clinicians can utilise to improve patient care. For example, deep learning models can provide assistance in identifying disease presence, or in extracting meaningful information from healthcare data, to enhance decision-making. 

While the real-world utility of some deep learning models remains a focus of future work, many models have displayed promising results. An example of this is the Enhanced Transformer for Health Outcome Simulation (ETHOS) model that predicts future patient health timelines. The potential integration of ETHOS in clinical workflows has exciting implications as clinicians could use a predicted patient health timeline to flag any concerning diagnostics or events (e.g. hospital mortality), and take action appropriately. 

The Privacy Barrier to Better Models

The effectiveness of deep learning models relies on extensive training from rich datasets to achieve optimal results. Training on only one dataset from a single healthcare institution risks potentially biased results due to socioeconomic, geographic and other factors inherent to that institution. Therefore, it is important that data can be combined cross-institutionally to improve its diversity and quantity, allowing for more generalisable and ethically acceptable models. 

Although there is a vast amount of digital health data, leveraging this for model training has proven difficult due to privacy and security concerns. For example, gathering extensive datasets within healthcare is often challenging due to the strict rules of the Health Insurance Portability and Accountability Act (HIPAA), which makes it difficult to combine data samples from multiple clients (e.g. hospitals). Given these limitations, models are often trained through secondary data sources, limiting their effectiveness despite rapid learning from healthcare data holding the potential to transform healthcare practices and policy. 

The Federated Approach – Learning Together, Separately

This is why deep learning models for healthcare require a way to train on extensive datasets without the need to centralise this data. Federated learning is an example of how deep learning models can be trained in this way. 

Before considering how federated learning operates, it is important to first understand how a deep learning model works. An example of deep learning model architecture can be seen in the figure below. This shows how the architecture of a deep learning model can be represented as a series of nodes and weighted edges, where the weights influence how the inputs of the model interact to predict the output. During the training phase of a deep learning model, the weights are adjusted as the model learns how the inputs influence the output from its training data. 

The federated learning (FL) setup involves a global model held by a central aggregator. Step 1 of the setup, as shown in the figure, involves the initial weights of the global model being sent to each client (e.g. a hospital). Each client has their own local model, which shares the same model architecture as the global model (i.e. the number of nodes and edges of each local model is exactly the same as in the global model, which can also be seen in the figure below, as each client has the same graph as the global model). The local model of each client is initialised with the sent weights of the global model (i.e. for each local model, the edges are weighted to match the weighted edges of the global model). 

Then, in step 2 of the figure, each client trains their initialised local model on the respective local data. During this training, the model weights of each client are updated, and once completed, each client then sends their updated model weights to the central aggregator. 

Upon receiving the updated model weights of all the clients, the aggregator updates the global model weights as an aggregation of these local model weights, as shown in step 3. This can occur in many ways through a number of different aggregation methods with varying complexity. A simple example is FedAvg, which simply updates the global model weights as a weighted sample average of the received local model weights, with weighting for this average usually determined by the size of their local data. 

The updated global model weights are then sent to the clients again for the next round of local training (i.e. step 1 of the figure below is reinstated), and this process repeats until a stopping condition is reached (e.g. the global model converges or a maximum number of training rounds have been completed).

In the framework described above, the local data of each client never leaves their local environment. Therefore, this is a promising technique to address the need of training deep learning models with extensive healthcare datasets as such a technique removes the necessity to collate data across clients in one central location.

The Nuances of Healthcare Data

However, implementing FL in healthcare is not straightforward as the nature of patient data requires special consideration and adjustments to this methodology.

Considering each client as a hospital, if each has a different patient population with different disease prevalence, this inconsistency can lead to downstream modelling issues. For example, if one hospital has a substantial elderly population and another hospital consists mostly of infant patients, then if the former hospital has significantly more data than the latter, the shared global model may be more heavily influenced by the first hospital’s data. Consequently, the global model produced may not be generalisable to patients across a broader age range, and is only suitable for an elderly population bracket. 

Similarly, different hospitals may follow different clinical practices, which drives differences in data collection techniques and handling of missing data. For example, if one hospital stores measurements of blood pressure for all their patients, the second hospital does not measure blood pressure and a third hospital holds measurements of blood pressure for only some of their patients, then for a global model, which requires blood pressure as an input, each hospital’s local model may not conform to this requirement. This would result in a significant mismatch between the architecture of the global and some of the local models. 

Another challenge with FL is what determines the best aggregation method to synthesise local model weights. Although it has been mentioned that clients with more data usually have a higher weighting when aggregating locally learned weights, this can be problematic. For example, if one hospital has more data than another hospital, but the other hospital has higher quality, more recently collected and reliable data, then this approach would not be optimal. 

Furthermore, despite FL ensuring that the local data of each client is not exchanged with other clients, it is still critical that data privacy is upheld when local model weights are transferred between clients or to the aggregator. Such transfers can be vulnerable to adversarial attacks, which risk uncovering sensitive information of the local data despite their nature. 

The Road Ahead

The transformative power of deep learning models in healthcare is undisputed in providing tools that clinicians can utilise to aid decision-making. However, for such models to be reliable, robust and ethically promising, we require them to train on extensive datasets, which ideally span across different hospitals. FL could allow for the training of a deep learning model across different locations with preserved data privacy at each location. Despite this, healthcare data has its own inherent challenges, which such FL methodologies must consider and adjust to, such as missing data across hospitals, varying data populations across hospitals and variations in data quality across hospitals. Therefore, while one can implement FL to preserve the privacy of datasets when training deep learning models in healthcare, these adjustments must be accounted for before considering integrating these models into clinical workflows, to ensure the models are generalisable, robust, and reliable.

Definitions

1. Deep Learning (DL): A branch of artificial intelligence which learns patterns from large datasets and makes predictions or decisions.
2. Generalisability: The ability of a model to perform accurately on new data from different populations, institutions or data unseen during training. 
3. Federated Learning (FL): A machine learning framework in which a model is collaboratively trained across institutions, whilst the data at each institution used to train the model is stored locally and therefore preserves its privacy. 
4. Global model: The central model in an FL system that aggregates information learned from multiple local models and sends updated parameters back to participating clients. 
5. Local model: A copy of the global model that is trained using data stored by a single client/institution before contributing updates back to the global model. 
6. Client: A participating organisation (e.g. hospital) in an FL setup, which trains a local model using its own data.
7. Model weights: Numerical values within a neural network determining how strongly different inputs influence the model’s predictions.

References

1. Sharma S, Guleria K. A comprehensive review on federated learning based models for healthcare applications. National Library of Medicine. 2023;146:102691.
2. Renc P, Grzeszczyk MK, Oufattole N, Goode D, Jia Y, Bieganski S, et al. Foundation model of electronic medical records for adaptive risk estimation. GigaScience. 2025;14:giaf107.
3. Sheller MJ, Reina GA, Edwards B, Martin J, Bakas, S. Multi-institutional deep learning modeling without sharing patient data: a feasibility study on brain tumor segmentation. In: Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries. Vol. 11383; 2019. p. 92-104.
4. Cremonesi F, Planat V, Kalokyri V, Kondylakis H, Sanavia T, Miguel Mateos Resinas V, et al. The need for multimodal health data modeling: a practical approach for a federated-learning healthcare platform. Journal of Biomedical Informatics. 2023;141:104338.
5. Zhang F, Kreuter D, Chen Y, Dittmer S, Tull S, Shadbahr T. Recent methodological advances in federated learning for healthcare. Patterns. 2024;5(6):101006.
6. Roland T, Böck C, Tschoellitsch T, Maletzky A, Hochreiter S, Meier J, et al. Domain Shifts in Machine Learning Based Covid-19 Diagnosis From Blood Tests. Journal of Medical Systems. 2022;46:23.
7. Alam MU, FedSepsis RR. A Federated Multi-Modal Deep Learning-Based Internet of Medical Things Application for Early Detection of Sepsis from Electronic Health Records Using Raspberry Pi and Jetson Nano Devices. Sensors. 2023;23:970.
8. Sun W, Chen Y, Yang X, Cao J, Song Y. FedIO: Bridge Inner and Outer-hospital Information for Perioperative Complications Prognostic Prediction via Federated Learning. In: Proceedings of the 2021 IEEE International Conference on Bioinformatics and Biomedicine (BIBM); 2021. p. 3215-21.
9. GeeksforGeeks. What is Federated Learning? GeeksforGeeks; 2025. Available from: https://www.geeksforgeeks.org/machine-learning/collaborative-learning-federated-learning/ [Accessed 6 Jun 2026]