How amalgamated learning could scale medical AI

AI shows tremendous promise in finding new patterns embedded in the mountains of data. However, some data differ in different silos for technical, ethical, and commercial reasons. A promising new AI and machine learning technique called amalgamated learning can help eliminate these silos, find new cures for diseases, prevent fraud and improve industrial equipment. It can also provide a way to create digital twins from incompatible forms of data.

At the Imec Future Summit conference, Roel Wuyts gave an exclusive interview with VentureBeat detailing how integrated learning works and how it compares to related technologies such as federated learning and homomorphic encryption. Wuyts is the manager of ExaScale Life Lab at Imec, a cross-industry scientific collaboration in Europe and a professor at Katholieke Universiteit Leuven in Belgium.

He is leading a team focused on exploring different approaches to AI scaling in different participants to improve semiconductor manufacturing, medical research and other areas.

“We want to do population-based data-based analyzes to look at novel markers that no one has seen before,” Wuts said. Going.” At the population level privacy can help improve computing.

Maintaining the privacy of medical research

At one end of the spectrum, new computing technologies such as homomorphic encryption allow multiple participants to share data to collaborate on new AI models with high confidence. However, it also adds a lot of computational overhead. Older implementations were apparently about ten-thousand times slower than comparable algorithms, and now researchers are slowing them down a thousand-fold. Wuyts said it is still not practical for large-scale population research.

At the other end of the spectrum, federated learning techniques allow various participants to update the machine learning model locally without sending sensitive data to others. In this case, only updates of the model are shared with others. This type of approach is more efficient than homomorphic encryption. His team has found ways to predict atrial fibrillation by applying federated learning to multiple hospitals.

Arterial fibrillation is an irregular heart rhythm that can cause blood clots in the heart. The hope is that better medical data and newer smart watches could provide better warning signs to help reduce these risks. But hospitals face various ethical and privacy issues in sharing such data at the population level. His team has already seen some promising initial results in this collaboration. Along the way, he predicts that we can all take advantage of the data we collect from our neighbors’ smartwatches.

Federated learning limits

However, there are some challenges in federal education. For starters, all hospitals or healthcare companies involved must use the same model and techniques. This could be a problem if the hospital hopes to commercialize the new AI model.

“In some cases, they are not willing to share the data or models they develop because it could provide a competitive advantage,” Wuyts said.

The second problem is that all that data needs to be generalized. This is not a big issue in areas like heart research, where there is a consensus on how and what to measure. However, it can be more problematic because when teams have different processes for collecting and analyzing data, teams try to bring in more data from new sources. Wuyts noted that even in areas such as genomics research, each hospital may differ in how they collect data, which affects the results of the study.

Another issue is how doctors code various diseases. For example, in some of their research, they discovered regional differences in how doctors in different healthcare systems would record similar illnesses in the healthcare system. This can result from different types of compensation for different diseases treated using similar approaches.

Integrated education

His team has recently begun experimenting with integrated education for large-scale cancer research. Like federated learning, it is much faster than homomorphic encryption and does not require participants to share data. Another advantage is that it supports multiple models, so participants do not have to share the intellectual property baked into it. This could encourage cross-industry medical research by competitors that improves outcomes for everyone while protecting commercial interests.

This technique seems to work even when each participant encodes the data a little differently. The main thing is that the technique takes advantage of the differences discovered in each local data set. As a result, everyone can learn from the experience of others, even if their own hospital data collection processes are different, as long as these processes are internally consistent. “We don’t think we need to generalize data from all parties to train the local model,” Wuyts said.

One concern is that this unified learning makes it difficult to overcome prejudice or to find out how the model has reached a definite conclusion compared to traditional approaches. As a result, they are focusing on using more explanatory AI techniques that allow them to identify and audit various factors that may affect the results.

“You need to create a complete stack of tools to check and log what’s going on, so people can see,” Wuts said. “They’re focusing on more explanatory models so that if something goes wrong, people can investigate and point out what went wrong.”

Another benefit is that integrated learning will also help individuals customize their digital twins, even if their local set points are slightly different for things like temperature or other important signals. For example, some individuals are more likely to get hotter than others. It is more important to observe the changes taking place in each individual than the global set point in the entire population.

“If we can capture the true signal, it’s more interesting than showing the raw value,” Wuyts said.

Similar Posts

Leave a Reply

Your email address will not be published.