What question did this study set out to answer?

To validate a privacy-preserving architecture for surgical-robot digital twin applications across multiple hypotheses.

June 2, 2026Open Access

Pre-Registered Validation of Bilateral Composition III Privacy-Preserving Architecture for Surgical-Robot Patient-Pair Digital Twins-- A 78-Study Campaign

Key Points

To validate a privacy-preserving architecture for surgical-robot digital twin applications across multiple hypotheses.
Conducted a pre-registered empirical validation campaign with 78 studies under deterministic protocols and zero post-hoc adjustments.
Evaluated approximately 425 hypotheses comparing the Composition III architecture to other privacy mechanisms.
Deployed on real biological kinematic data from KIMORE rehabilitation studies.
Composition III demonstrated superior performance compared to Laplace and Gaussian mechanisms (AUC 0.468 vs 0.591 and 0.533, p<0.001).
Mutual information between output and patient identity was determined to be 0.000 bits, ensuring privacy.
The architecture scales effectively for cohorts up to 1000 patients, confirming its utility in multi-arm configurations.

Abstract

This deposit contains the complete pre-registration, runner code, summary verdicts, per-study reports, figures, and arXiv-format manuscript for a 78-study pre-registered empirical validation campaign of a bilateral Composition III privacy-preserving architecture for surgical-robot patient-pair digital-twin deployments. The campaign comprises approximately 425 pre-registered hypotheses across six batches, executed under continuous deterministic protocol (PYTHONHASHSEED=0) with zero post-hoc threshold adjustments. Approximately 75% of pre-registered hypotheses are SUPPORTED; remaining outcomes are honestly disclosed as bounded-negative findings that characterize deployment boundaries. The architecture extends the companion Paper 8 (10.5281/zenodo.20466035) from single-channel-group musculoskeletal state to bilateral patient-and-robot state. Key empirical findings: (i) Composition III strictly dominates the Laplace and Gaussian differential-privacy mechanisms at every tested privacy budget (patient AUC 0.468 vs Laplace 0.591 and Gaussian 0.533 at ε=1.0, with task classification accuracy 1.000 vs 0.375 and 0.271 respectively); (ii) mutual information between protected output and patient identity equals 0.000 bits; (iii) the formal Rényi-DP bound is approximately 40× looser than the empirically-audited ε; (iv) GDPR Article 17 right-to-be-forgotten is O(1)-computable with statistical indistinguishability from a never-trained model; (v) the architecture scales cleanly to 1000-patient cohorts and to multi-arm da Vinci configurations; (vi) fresh execution on real KIMORE rehabilitation skeleton data confirms architecture function on real biological human kinematic data. The deposit includes 78 pre-registration files, runner Python scripts, JSON summary verdicts, 8 figures, a comprehensive POC summary document (78 detailed addenda), three Phase XXII synthesis documents, and both v1 and v2 of the manuscript (v2 is current; v1 retained for transparency). All studies reproduce bit-identically across macOS, Linux, and Python 3.9-3.11. KIMORE data (Capecci et al. 2019) and CMU Motion Capture Database real-data anchors are referenced; native JIGSAWS surgical-robot data validation remains future work pending Data Use Agreement approval. Released under CC-BY 4.0.

Read Full Paperexternally

Mark Helpful

Bookmark

Relay

View Full Paper