I am a pediatric cardiologist and have cared for children with heart disease for the past three decades. In addition, I have an educational background in business and finance as well as healthcare administration and global health – I gained a Masters Degree in Public Health from UCLA and taught Global Health there after I completed the program.
“Sharing knowledge is not about giving people something, or getting something from them. That is only valid for information sharing. Sharing knowledge occurs when people are genuinely interested in helping one another develop new capacities for action; it is about creating learning processes.”
Peter Senge, American systems scientist
This is a very timely article from Nature Medicine on the application of federated learning in the training process of artificial intelligence models with medical image and other data from 20 institutions around the world; this is achieved while maintaining data anonymity to obviate the need for data sharing. The study, aptly named Electronic Medical Record Chest X-ray AI Model, or EXAM, utilizes vital signs (oxygen saturation, blood pressures, and respiratory rate), laboratory data (CBC, D-dimer, lactate, COVID PCR test, etc), and chest X-rays to predict the future oxygen requirements of patients with COVID-19 with the same neural network, which aggregated the updated model weights from the 20 centers. The outcomes were: 24-hour and 72-hour oxygen device, death, and time of death. Unsurprisingly, the EXAM model achieved an area under the curve (AUC) of > 0.92 while providing a 16% improvement in AUC and a 38% increase in generalizability (compared to single site data models). In addition, when this model was used in 3 hospitals independent from this original cohort, it showed 95% sensitivity and 88% specificity in predicting ventilator need in 24 hours from arrival in the emergency room.
The EXAM model remains limited by the quality of the training data like any other machine learning model, but is particularly useful for centers with small datasets that do not provide adequate substrate for artificial intelligence. This pioneering study showed the feasibility of real-world rapid science collaboration without data exchange and ushers in a new era of collaboration amongst centers for concomitant data protection and robust insights. Additional approaches utilizing automated hyper parameter searching, neural architecture search, and other automated machine learning methodologies will further improve this federated learning model and collaborative clinical science concept. The authors deserve high praise for executing a large study with many centers in federated learning, and this teamwork approach can be an important roadmap for artificial intelligence in medicine in the future.