Privacy-Preserving Artificial Intelligence Methods in Healthcare

After decades of turbulent research and development with multiple fundamental setbacks during so-called AI winters, artificial intelligence is currently starting to proliferate everyday applications. Made possible by remarkable advances in computing capabilities of modern hardware and exponentially increasing masses of available data, self-learning models match or even outperform human experts in tasks across many domains. However, the technology’s transition into medicine and healthcare is noticeably slowed down by a range of factors. As medical data is highly sensitive and thus worthy of utmost protection, concerns around patient privacy are among the most challenging hurdles for amassing large datasets which are essential for artificial intelligence models. To address them in the current digital age, privacy-preserving methods are required that go beyond simply removing names from datasets.

In this thesis, techniques showing potential in both academia as well as industry are presented and grouped into three categories: information obfuscation, insight prevention and data synthesisation. Furthermore, to test their practicability, experiments with methods of all three groups have been conducted as part of this thesis. For information obfuscation, privacy-preserving record linkage was investigated and applied in an operational registry for heart failure patients. To circumvent the need to see data directly (insight prevention), novel decentral learning schemes were introduced. To test data synthesisation, biosignals were synthesised with generative adversarial networks. Additionally, studies to highlight the usefulness of artificial intelligence in healthcare have been published. This thesis’ results indicate not only that satisfying methods exist in theoretical literature work, but also demonstrated that they are viable in real-world applications. Methods like the ones investigated and presented in this thesis aim to facilitate the transition of artificial intelligence into health applications by increasing the amount of knowledge – not necessarily data – models can access in a privacy-preserving manner and thus enabling the development more capable digital tools. In times of overburdened healthcare systems and ageing populations in many industrialised regions, such assistive aids could help healthcare professionals to provide better and ultimately more efficient care for their patients in the future.