Federated Learning Analysis: Revolutionising global research data
- Riddhi Laijawala
- 2 days ago
- 5 min read
I am a mental health researcher working on the HappyMums project, a European consortium that focuses on understanding depression in pregnancy. At King’s College London, we are leading a clinical study involving the use of a smartphone application called the HappyMums App. Research is often a collaborative approach across different countries, and we, in this project are collaborating with institutions across Europe. Since the start of our project, much of our discussions have been about privacy and data sharing. The idea of having a large-scale dataset encompassing a thousand participants, across seven different sites, and finding a way to analyse this data, always seemed like a gargantuan task. Today, I'm writing this piece for Inspire the Mind, to highlight how clincial research data can be analysed to advance screening and monitoring of mental health symptoms across the world in a way which does not violate data sharing rules.
Our collaborators from the Artificial Intelligence in Medicine Lab at the University of Barcelona proposed the idea of developing a federated learning (FL) platform as a solution, which I had never heard of before.
So, what exactly is it, and how does it differ from a centralised approach?
Traditionally, centralised approaches to data analysis have been used in mental health research. Data is stored at one site, and each site runs its own analyses, or has to go through hours or days of paperwork to even start thinking about sharing data.
The solution? Federated analysis.
The federated analysis method is a way to train Ai models without data ever leaving the device it is collected on. It gives collaborators the opportunity to “remotely share their data to collaboratively train a single deep learning model”. Speaking in practical terms, an environment downloaded on a device from each data collecting site will enable an Ai model to be developed without this data ever leaving their server.
Let me explain this in Stranger Things terms (I’m loving season 4 and can't stop thinking about this). In federated analysis terms, it would be like all the children looking for ways to defeat Vecna in their own time, in their own homes. Instead of bringing actual clues and materials to their key hideout spot, they do the detective work in their own spaces and bring their conclusions to their hideout and brainstorm ways to defeat Vecna. So, above all, this method allows researchers to utilise large amounts of data while respecting participant privacy and legal obligations.
The federated analysis approach was first introduced in 2016, with published research in mental health appearing in 2019. This does show that this approach to mental health research is still in its very early stages, with the number of publications using this approach rising from 7 in 2021 to more than 10 in 2023. The approach has been explored mostly in relation to depression, but other work has looked at its implementation with Bipolar Disorder and Obsessive Compulsive Disorder. To advance the usage of this approach in mental health, researchers have suggested finding practical implementation solutions, such as an understanding of the technical software and hardware, computational resources, and organisational limitations. Additionally, due to participant heterogeneity, it is recommended to ensure that models adapt to individual characteristics, to ensure that patients can benefit from Ai models equally.
The novelty brought by the HappyMums project:
In the field of perinatal mental health, published research is even more limited. A 2023 study has used population birth data in Europe and developed a Common Data Model to understand perinatal indicators from routine medical information.
To the best of my knowledge, based on the current state of the literature, no research project has used this approach for a clinical study; i.e., newly generated real-world participant data, not available in existing healthcare databases. What is unique about our project is that clinical and digital data from numerous sites across Europe will be analysed through the FL platform.
Applications within the broader sphere of women’s health
Within the wider space of women’s health, this approach can have transformative potential. The field of FemTech (tech-driven products like apps and wearable devices to address female health, like pregnancy and menopause) has seen a boom in recent years, and federated analysis could have numerous benefits in this regard. In fact, in early 2025, an ovulation tracking app called FLORA used the FL approach to provide personalised health insights, while also addressing the privacy concerns of users as a case study in this field. Ai models have also been applied in the context of menstrual care and breast cancer screening. Federated analysis takes this progress further by integrating diverse datasets from across the globe. This allows for the creation of more representative and inclusive models that truly serve all women, everywhere.
Benefits and drawbacks:
Coming back to my personal experience with FL platforms, I must say that I am in no way an expert in Ai or computer science, with no prior experience with coding. I was initially sceptical about this approach as it had never been used in our lab. With the immense help from our collaborators in Barcelona, I was able to download the environment on my laptop, and we jokingly said that by the end of this, I would have a joint PhD in maternal mental health as well as computer science.
I am excited to implement this approach for the HappyMums clinical study, because it will allow for improved data diversity and representation from our Europe-wide sample. This model will allow us to collaborate and share data without relying on legal delays and data transfer agreements. With news that sensitive patient data is breached in cyberattacks, I do hope that federated analysis will be a solution to data privacy, seeing as the data never leaves secure computer servers.
One of the drawbacks of this approach is that it involves a large amount of technical setup, which can be difficult for those outside of the field of technology and computer science. For this, researchers need to make sure their organisations have the correct infrastructure and permissions, which, if not in place, can add a great deal of complexity. Additionally, if a data collection plan has not been decided beforehand, it can be quite challenging to harmonise the data, and inconsistencies in data collection can lead to delays in this process.
With that being said, from my emerging involvement with this approach, I do believe that the benefits outweigh the drawbacks. If used correctly, FL can lead to great advancements in women’s health and overall mental health research. I no longer have a feeling of dread when I hear Ai, and machine learning, but rather, a sense of excitement.
I will leave you with a quote from Ursula von der Leyen, President of the European Commission from Davos 2024.
"I am a tech optimist and, as a medical doctor by training, I know that AI is already revolutionizing healthcare. That's good. AI can boost productivity at unprecedented speed. First movers will be rewarded, and the global race is already on without any question.”
So…Where to, next?
