36 HEALTH 37 How AI-Powered Knowledge Aggregation Could Help Find A Cure For Alzheimer’s Privacy-preserving federated learning could overcome a big stumbling block: data-sharing friction Computer illustration of a broken brain’s neural network represented by lines and dots. Some areas are not connected, depicting dementia and Alzheimer’s disease. Over the last two decades the biotech Careful evaluations of key research in- «It is time to try industry has conducted over 156 sights driven by data, most notably the clinical trials and spent hundreds of recent announcement of the revival of new approaches billions of dollars on the discovery and a drug developed by U.S. drug com- to stop this How AI Could Help amplify or muffle these signals,” says Berkowitz. “This testing of these drugs, and yet, not pany Biogen called Aducanumab, offer methodology can remove noise, improve confidence, and one clinically impactful drug has been some hope. “This possible break- slow-moving The new federation will harvest insights from AIs that run uncover new areas of exploration without ever sharing brought to market in that time. through raises many questions,” says locally on each dataset, explains Berkowitz. After local sensitive and proprietary raw data. “ Berkowitz, “One of the most pressing epidemic.» analysis, generated knowledge- not data- are aggregated There is no cure and no treatment to is whether there are more meaningful from each AI into a combined set of findings that could By iterating a few times, the AI that has aggregated slow down a disease, which in the U.S. insights hidden in the failed drug trial both uncover previously weak signals and inform further knowledge can dig into these signals and gain greater alone is expected to impact 16 million data.” among them are privacy concerns and areas of inquiry and treatment. clarity. “In this system, the consistency of data becomes people by 2050, generating $1.2 tril- competition issues. less of an issue. We aggregate the generated knowledge, lion in medical and hospice care costs. One of the most valuable assets “What we are talking about is privacy-preserving federated not the data. Noise can be filtered out, and we can turn created by these trials is the data col- Berkowitz will be in Davos during the learning,” he says. “Now we can say to drug companies the data from two decades of research into knowledge “The economic impact of this disease lected. While there has been progress, annual meeting of the World Econo- ‘you do not need to share your data.’ An AI can sit on your that, when aggregated, will lead to clinically meaningful is mind-boggling,” says Barak Berko- the vast majority of this data has yet mic Forum January 21-24 to recruit local server and analyze the data and simply share the insights,” Berkowitz says. witz, co-founder of this project. He is to be shared or, more importantly, potential partners in a new federation insights that come from that analysis.” also the Director of Operations and aggregated, Berkowitz says. Data ag- that seeks to overcome these hurdles The federation will work on encrypted data eliminating Strategy at the MIT Media Lab where gregation could turn one hundred fifty by removing data sharing from the Sharing research insights is a principle that has been used critical privacy concerns; the AI will focus on local data he came in contact with the power relatively small data sets into a truly equation. The federation is seeking for the critical advancement of science for decades, at servers, so raw data never leaves its owner’s site; aggre- of federated learning systems. It is massive data asset of more than one firms that have conducted clinical scientific conferences and congresses, leading to new gate insights to amplify weak signals and identify new estimated that in the U.S. more than 16 hundred and fifty thousand subjects. trials focused on Alzheimer’s and understandings of disease and new areas of inquiry. “It is correlations and correct for small-sampling bias. million Americans provide unpaid care “There is a robust scientific basis those that specialize in helping artifi- not just listening to reports, but the discussions that they for people with Alzheimer’s or other for believing this aggregated data cial intelligence (AI) filter out “noise”, stimulate, that make congresses so valuable,” says Ber- Members of the federation gain value in three critical dementias. That is almost 10% of the could provide meaningful insights i.e. extraneous information. MIT’s kowitz. In the same way, new technologies enable AIs to ways, says Berkowitz. The federation will share insights working age population. “Alzheimer’s into these trials, reasons for failures, Connected Science Group will serve share their insights with other AIs in a secure mode. “No AI related solely to a specific trial on a proprietary basis with is a growing global catastrophe tied missed efficacy, and possibly clinically as a data trustee to the federation, sees raw data, so it can not be shared. Instead, the AIs col- the data creator. It will share generated insights around to aging in a world that is aging and meaningful treatments,” he says. also which means it will be responsible for lect correlations and other data signals they may ascertain, the data creator. It will share generated insights around the scale of the catastrophe is right up significant forces slowing or stopping making sure the AI is not doing things and these signals then create knowledge. The knowledge obvious key questions with all federation members. And there with climate change,” he says. such in-depth data sharing. Chief or passing data it should not. is then aggregated from multiple AIs and is combined to all Federation members can interrogate the aggregated