Updated: Sep 29
AI Developers at Massachusetts General Hospital created an AI model that can detect undiagnosed brain aneurysms.
Brain aneurysms are common in the population, and aneurysm rupture can lead to devastating consequences.
To improve and validate the performance of the AI model, the developers are training the model across a federated network of seven hospitals around the globe - using federated learning, so they never have to transfer patient data.
The enabling infrastructure was established in just weeks due to the ease of creating federating learning environments with the Rhino Health platform; setting-up future collaborations between network members will be possible in a matter of minutes.
Tackling The Problem
Medical researchers and clinicians have long worked to more quickly and more accurately identify intracranial aneurysms, which have an estimated prevalence of 3-6% in the general population (Vlak, et al 2011). Accurate and early diagnosis is key to preventing the devastating neurological consequences of aneurysm rupture. Artificial intelligence (AI) solutions hold great promise in aiding radiologist detection of aneurysms using computed tomography angiography (CTA). But - as is often the case with AI and machine learning - the performance is correlated to the volume and diversity of data used for training the models. Models are too often trained on data from a single site and their performance degrades when tested on patients from different hospitals, due to data differences (e.g., patient demographics, scanner types). Thus, without training AI models on large volumes of data from diverse patient populations (in other words, data that better represents today’s real-world patient populations), model performance and health equity suffers. Traditionally, access to these broader datasets has been the primary constraint in the development and deployment of new AI models, as small volumes of relatively homogenous patient data sit in isolated silos at health systems around the globe.
To break this data access constraint, researchers at Massachusetts General Hospital (MGH) teamed-up with Rhino Health, using our ‘Federated Learning as a Platform’ solution to quickly and easily set up collaborations with six other institutions around the world to validate and better generalize their locally well-performing algorithm - without moving any data.
Reimagining What’s Possible
As Director of Clinical Engagement at Rhino Health, I identify high-impact collaboration opportunities that have real clinical value and help bring them to life. This global research collaboration was led by MGH’s Quanzheng Li, PhD, Dufan Wu, PhD, and Daniel Montes, MD. It aims to solve the false-negative rate in aneurysm detection - which means more missed aneurysms, especially small ones - when radiologists are stuck in the traditionally difficult, time-consuming process of CTA interpretation when a patient presents with symptoms (Hochberg, et al 2011).
Six institutions on four continents participated:the University of Cambridge School of Medicine (in the UK) led by Tomasz Matys, MD, PhD, Lahey Hospital and Medical Center (in the U.S.) led by Christoph Wald, MD, DASA (a leading integrated healthcare provider in Brazil) led by Felipe Kitamura, MD, PhD, the Medical Data Analytics Laboratory (MeDA Lab) at National Taiwan University led by Weichung Wang, PhD, Seoul National University Hospital (in South Korea) led by Young-Gon Kim, PhD, and Assuta Medical Center (in Israel) led by Michal Guindy, MD.
Remarkably, this truly global research collaboration is taking place without any transfer of the underlying data. Rather, the MGH-developed algorithm is being trained at each participating site locally - with key learnings from across these disparate datasets being used to inform and refine the global model.
Up and Running - and Learning - Quickly
To understand more about how MGH quickly set up this global collaboration powered by the Rhino Health Platform, watch this video.
The MGH model is being validated and retrained at each of the participating sites as they join the project and collect data. Researchers are closely monitoring performance of the model after applying this federated learning approach; the hypothesis, of course, is that model performance will improve as the project will use more diverse datasets than any one institution could provide on its own. The timeline is aggressive, and the federated learning infrastructure was set up in just a few weeks. Now that each site has joined the network, a similar research effort can now be conducted in a matter of days!
What’s Next, Why It Matters
The goal, in this case, is to ultimately improve patient outcomes by reducing the occurrence of brain aneurysm rupture, which typically happens in 15-30% of cases and often leads to high morbidity (e.g., neurological disability) and mortality (Muehlschlegel 2018). This will be addressed by running the new, improved, model over large amounts of data from each participating site.
All participating sites and researchers have joined the growing global network of healthcare AI innovators connected via the Rhino Health Platform. Federated learning allows collaboration without data sharing, thus ensuring security and data privacy. The Rhino Health platform enables federated learning, streamlined by NVIDIA’s FLARE SDK, and includes other features that make collaborations easy.
Rhino Health is collaborating with institutions around the globe to advance federated learning, including MGH, Cambridge University, Lahey Hospital, DASA, National Taiwan University, Seoul National University Hospital, and Assuta Medical Centers. If you are interested in joining this collaboration, feel free to reach out. I’m at email@example.com.
Thanks to the team of collaborators driving this important work forward, including: At MGH - Aoxiao Zhong. At Cambridge - Joshua Kaggie, PhD, Annabel Sorby-Adams, PhD, Abhishekh Ashok, MBBS PhD, and Fiona Gilbert, MD. At Lahey - Emanuele Orru', MD, and Adam Medina. At DASA - Paulo Eduardo de Aguiar Kuriki, MD, Gustavo Pinto, MD, Suely F, MD, and Vitor de Lima Lavor. At the MeDA Lab - Chien-Hua Huang, MD, Che-Yu Hsu, MD, Po-Ting Chen, MD, Chun-Hao Chang, Hsin-Han Tsai, and Shih-Chieh Ma. At SNUH - Inpyeong Hwang, MD, Daseul Park, BS, and Seung Yeon Cho, BS. At Assuta - Daniel Rabina, Anatoly Budylev, Judith Luckman, MD, Miel Dabush Kasa, and Ofir Orpaz.
Vlak, M. H.; Algra, A.; Brandenburg, R.; and Rinkel, G. J. 2011. Prevalence of unruptured intracranial aneurysms, with emphasis on sex, age, comorbidity, country, and time period: a systematic review and meta-analysis. The Lancet Neurology 10(7):626–636.
Hochberg, A. R.; Rojas, R.; Thomas, A. J.; Reddy, A. S.; and Bhadelia, R. A. 2011. Accuracy of on-call resident interpretation of ct angiography for intracranial aneurysm in subarachnoid hemorrhage. American Journal of Roentgenology 197(6):1436–1441.
Muehlschlegel, S. 2018. Subarachnoid hemorrhage. CONTINUUM: Lifelong Learning in Neurology 24(6):1623–1657.
Yang, J.; Xie, M.; Hu, C.; Alwalid, O.; Xu, Y.; Liu, J.; Jin, T.; Li, C.; Tu, D.; Liu, X.; et al. 2021. Deep learning for detecting cerebral aneurysms with ct angiography. Radiology 298(1):155–163