Privacy-Preserving Clinical Analytics with Threshold Homomorphic Encryption: Insights from Hematologic Toxicity During Craniospinal Irradiation

The increasing need for safeguarding medical data privacy, particularly in research involving sensitive patient information, requires innovative solutions. This work proposes a conceptual architecture that uses (threshold) homomorphic encryption for statistical analysis on encrypted medical data shared between different institutions. By utilizing this type of encryption, sensitive patient data is kept secure throughout the analysis process, minimizing the risk of re-identification. The data is encrypted locally before being processed on secure computation servers, ensuring privacy while enabling several statistical analysis. Performing computation on encrypted data is expensive and this has led to widespread skepticism regarding its practicality. Our work shows that, with recent advances and careful design and engineering the technology can indeed be harnessed to facilitate medical research. This method aligns with key data protection regulations and lays the groundwork for more privacy-preserving collaborative research in the medical field. This paper presents a conceptual model rather than a full platform implementation. We securely replicate, on encrypted patient-level data from pediatric craniospinal irradiation, three routine statistics—Pearson’s correlation, Wilcoxon rank-sum, and χ 2 —in a three-party threshold-HE workflow. Across tested sizes, χ 2 ≤ 0 . 5 % error, Pearson’s r ≈0–5.7% (≈5% at n =512), Wilcoxon’s z ≈13.5–21.9%, with millisecond-scale runtimes ( ≈ 5 0 – 8 5 0 m s ). We also outline a concise systems blueprint for cross-institution analytics (Fig. [1](#fig-cap-0001)).