How to Share Medical Data Without Revealing Anything?

Imagine each hospital has a medical puzzle, but the pieces are secret for privacy. Today, researchers have found a clever way to combine these puzzles without showing individual pieces. Instead of exchanging raw patient data, each hospital sends only a 'blurred photo' of its data, a much smaller representation that doesn't reveal personal information. This technique, called dimensionality reduction, allows scientists to combine information and create accurate survival curves, just as if they had seen all the data together. The method works both when data is split between hospitals (horizontal) and when each hospital has different information about the same patient (vertical). Tests on seven medical datasets show that this approach is even more reliable than analyses conducted by a single center. It's a step forward for collaborative medical research, which can now rely on powerful tools without compromising privacy.

Technically, the method is based on dimensionality reduction to extract essential features from the covariate matrix. These low-dimensional features are then exchanged between institutions. Analysts reconstruct an aggregated dataset on which they apply propensity score matching, a statistical technique that balances patient characteristics to simulate a controlled experiment. Subsequently, Kaplan-Meier curves are estimated, a non-parametric tool for analyzing time to an event (e.g., healing or death). The key innovation is that the matching and estimation process occurs on the aggregated data, not the raw data, ensuring privacy. Experimental results, both on simulated datasets and real ones like MIMIC-III or TCGA, show that curves obtained with this distributed method are more stable and accurate than those calculated individually by each center, overcoming the problem of limited and biased samples typical of single-site analyses.