The future of federated learning in healthcare

M.S. Dinesh, Senior Scientist, Philips Innovation Campus, Bengaluru talks about the role of federated learning in healthcare

In the last few years, we have all come to acquaint the meaning and impact of emerging technologies on various aspects of our lives. Whether its applications on our mobile phones learning our preferences or using Artificial Intelligence to assist doctors in diagnosis, the significance of technology in our lives is high.

In healthcare, we have leapfrogged innovation in artificial intelligence, particularly in Machine Learning (ML) and Deep Learning (DL) leading to disruptive innovation to gain insights from medical data derived from radiology, pathology, genomics, oncology, etc.

However, in healthcare, data is not just highly sensitive but also not well organised. Furthermore, data privacy, data security, data access rights need to be followed with strict adherence to country specific regulatory aspects like GDPR (Europe), FDA(USA) and CDSCO(India), NMPA (China) etc. Apart from the massive volume, data also has a large proportion of unstructured content, patient privacy sensitivity, wide geographical distribution, and lack of interoperability between different devices and vendors.

For instance, AI has demonstrated the potential to assist radiologists in performing computer-aided analysis and diagnosis. However, it remains challenging to build good models without bias from small datasets and given the challenges in collecting, curating, and maintaining a high-quality data with diverse group takes considerable time, effort, and expense. To overcome this unique problem, experts in healthcare and life sciences can leverage the benefits of Federated Learning (FL) to address the problem of data governance and privacy by training algorithms collaboratively where data stays within the firewalls of the hospital and only models are shared[1].

Federated learning captures larger data variability and analyses patients across different demographics. For example, with the access to electronic health records, FL can help to find clinically similar patients and predict hospitalizations due to cardiac events, mortality and ICU stay time. Originally FL was developed for different domains such as mobile and edge device use cases [3], it recently gained traction for healthcare applications. Recent research has shown that models trained by FL can achieve performance levels comparable to ones trained on centrally hosted data sets and superior to models that only see isolated single-institutional data.

In the field of medical imaging, FL can help to develop models for organ segmentation in X-ray, Ultrasound, Computed Tomography, Magnetic Resonance Images and Positron Emission Tomography as well as disease-specific tumor characterization. By providing an opportunity to capture larger data variability and to analyze patients across different demographics, FL can enable disruptive innovations for the future.

FL can also be used to advance academic research. For instance, in 2020, the American College of Radiology, Diagnosticos da America, Partners HealthCare, Ohio State University and Stanford Medicine used Federated Learning to developed better predictive models to assess breast tissue density for mammograms (4). The study showed that the FL-generated models outperformed those trained on a single institute’s data and were more generalizable. Similarly, the HealthChain project aims to develop and deploy a FL framework across four hospitals in France (5). This solution generates common models that can predict treatment response for breast cancer and melanoma patients. This can further help oncologists to ascertain the most effective way to treat each patient based on their microscopic anatomy or dermoscopy images.

Further, collating data using ‘FAIR’ principles (Findable Accessible Interoperable Reusable) make the data readable and understandable to a great extent from autonomous algorithms that can be complex [2]. For instance, representing data in the form universally standardized vocabularies such as Uniform Resource Identifier (URI) and publicly searchable ontologies, can help many in this field of research, overcome language barriers in clinical data accessed from global clinical sites.

Strengthening FL by combining with FAIR principles can lead to meaningful clinical insights from big data on a global scale.

Federated Learning can create a huge impact on various stakeholders such as clinicians, patients, hospitals, AI researchers and healthcare providers. Despite the advantages of FL, researchers and AI developers must pay careful attention on the study design, selection of clinical protocols, data heterogeneity and data quality to alleviate model bias. Federated Learning is a promising concept to secure accurate, safe and unbiased data models. By enabling multiple parties to train collaboratively without the need to exchange or centralize data sets, FL addresses issues related to sensitive medical data. Federated Learning will have a great impact on precision medicine and holds the potential to improve patient care globally.

References

Nicola Rieke et.al. “The future of digital health with federated learning”, Nature Partner Journal, Digital Medicine (2020) 3:119 ; https://doi.org/10.1038/s41746-020-00323-1

Kairouz, P. et al. Advances and open problems in federated learning. arXiv preprint arXiv:1912.04977 (2019).

Petros Kalendralis et.al, “FAIR-compliant clinical, radiomics and DICOM metadata of RIDER, interobserver, Lung1 and head-Neck1 TCIA collections”, Volume47, Issue11, November 2020, Medical Physics.

https://blogs.nvidia.com/blog/2020/04/15/federated-learning-mammogram-assessment/

https://www.substra.ai/en/healthchain-project