How Does Federated Learning Work In Healthcare?
THE TECH
A central organization such as a laboratory or research hospital begins by developing a machine learning model. This is an AI tool that can identify patterns in patient data to make decisions and predictions.
A model can analyze the DNA of patients all suffering from the same illness, for example, to identify their genetic similarities. Such insights can be used to find disease treatments or help identify high-risk individuals for proactive monitoring.
The more data the model analyzes, the more accurate it becomes. Micah Sheller, staff research scientist at Intel’s security and privacy research labs, says a lack of data diversity is a major challenge organizations face when training AI models for healthcare research.
“You want more data from more technicians, from different machines, from different geographies over different patient populations, taken from different types of doctors,” he says.
Researchers Design An AI Model
To pinpoint the gene that predisposes certain space explorers to the disease without risking exposure of their private information, the researchers used federated learning—allowing an AI model to train on data spread across multiple hospitals and laboratories, without directly accessing it.
Because only a few hundred people have ever traveled to space, the team needed more data on which to train the AI models. They used data from mice that had been exposed to comparable levels of radiation in lab experiments, hoping to find a genetic overlap.
Separate models trained on the human and mouse datasets then had their insights aggregated—and this optimized model was ultimately able to identify a gene that may contribute to the development of the disease.
By better understanding the disease, we can better tailor our response. This discovery could lead to genetic screening to identify individuals at high risk of getting cancer or enable the formulation of targeted therapies.
AI industry experts from Intel teamed up with researchers from the Frontier Development Lab to find out why cosmic radiation causes cancer in some astronauts and not others.
Intel worked with a university on a joint research project to identify malignant brain tumors. By training and evaluating its AI models on secure datasets from 71 institutions across six continents, researchers were able to improve brain tumor detection by 33% compared to models trained on the largest public data set at the time.
Better Brain Tumor Detection
Federated learning is a paradigm shift in healthcare.
By securely unlocking the answers hidden in siloed data, this method accelerates medical research and enhances treatment efficacy, touching lives from the clinic to the cosmos.
The collaboration of organizations, safeguarded by technology like Intel SGX, is not just an improvement—it’s a revolution. We’re entering a new era of healthcare, one where data-driven insights have the potential to save millions of lives for generations to come.
AN AI MODEL
STEP 2:
A SECURE SAFEROOM
STEP 3:
INSIGHTS ON THE MOVE
Traditionally, an organization would need to collate patient data from multiple sources, such as hospitals, to feed its model. This process is often unfeasible due to patient privacy laws.
With OpenFL—an open-source framework from Intel—instead of bringing confidential patient data into the lab, the lab allows multiple institutions to train a shared model, contributing insights without releasing any of the data itself.
Intel Software Guard Extensions (SGX) provide a secure enclave within the computers at each organization where this training occurs. It acts like a digital lockbox, shielding the data and the AI model from unauthorized access.
It Learns In A Computer’s Secure Safe Room
After each institution completes its local model training, it sends back the model updates—not the raw data—to the lab. Think of these updates as insights or lessons learned from each hospital’s unique patient dataset.
The central research lab then combines all of these insights. The result is a more intelligent model that benefits from diverse data points without compromising patient privacy.
“It’s a really nice way to learn from these vast, diverse datasets because you don’t actually have to collect the data,” says Sheller. “You can leave the data where it is.”
The AI Brings Insights (Not Data) Back Home
STEP 1:
How Did It Find A Cancer Gene With The Help Of Astronauts?
SEE IT IN ACTION
Three More Healthcare Discoveries
Made Possible By Federated Learning
THE FUTURE
Healthcare organizations are using federated learning to train AI models to accurately identify high breast density in mammograms, a risk factor for developing breast cancer. This automated detection could lead to focused, proactive monitoring of high-risk patients.
More Accurate Mammograms
Researchers deployed AI models early in the Covid-19 pandemic in a New York City hospital system to analyze health data and predict how the disease would likely progress in new patients. By employing such tools in future pandemics, doctors could identify the highest-risk patients more quickly than traditional methods allow.
Targeted Pandemic Responses
