Privacy-preserving AI for heart data keeps 97–99.9% accuracy but can cost 12x compute

An Istanbul Technical University dissertation splits the ECG data lifecycle into three trust boundaries, each with its own protection. On ~21,800 ECG recordings, homomorphic encryption plus federated learning added 30% overhead in the simplest setup and up to 1,200% when sites held different data about the same patients, while accuracy stayed at 97–99.9%.

• Tested on ~21,800 ECG recordings split across simulated hospitals
• Encryption overhead: 30% with distinct patients, ~450% with distinct features
• Worst-case tangled data arrangement pushed compute costs to 1,200%
• Diagnostic accuracy held between 97% and 99.9%