Pre-Registered Multi-Dataset Validation of Three-Way Composition III Privacy-Preserving Architecture for Musculoskeletal-Kinematic Clinical Digital Twins (Paper 8 Reproducibility Archive)

This reproducibility archive accompanies the Paper 8 manuscript "Pre-Registered Multi-Dataset Validation of Three-Way Composition III Privacy-Preserving Architecture for Musculoskeletal-Kinematic Clinical Digital Twins. " Background. Clinical digital twin (DT) systems for musculoskeletal rehabilitation increasingly aggregate longitudinal kinematic data across multiple sites and patient populations. These deployments require simultaneous guarantees of (a) federated coefficient learning utility, (b) membership-inference privacy against realistic adversaries, (c) byte-budgeted transmission for Internet-of-Things radio links, and (d) deployment-time robustness to anatomical, populational, and protocol heterogeneity. Methods. We present a pre-registered empirical validation campaign spanning 21 studies and 91 hypotheses for the three-way Composition III architecture (per-subject phase randomization, cohort mean aggregation, federated AR (1) coefficient learning) applied to musculoskeletal-kinematic clinical DT deployments. Each study's decision rules were frozen before runner execution. The campaign covers foundational properties; Composition III privacy and utility under homogeneous, moderately heterogeneous, and extremely heterogeneous (50: 1 imbalance, 50x sensor noise differential) deployments; multi-classifier shadow-attacker bounds across LogisticRegression, RandomForest, and GradientBoosting families; differential-privacy noise calibration with non-monotone trade-off characterization; multi-period longitudinal observation up to T=10 periods; cross-anatomy generalization from N=3 (knee) to N=23 (hand-MANO model) ; and three real-world dataset validations using CMU Motion Capture Database (walking) and KIMORE Rehabilitation Dataset (rehabilitation exercises with 44 healthy controls plus 34 subjects with low-back pain, Parkinson's disease, and post-stroke conditions). Results. 71 of 91 pre-registered hypotheses supported (78%). All 20 non-supported outcomes are substantively interpretable: 3 metric-choice issues resolved by follow-up studies, 4 deployment-condition-dependent disclosures (including a non-monotone DP privacy-utility trade-off, an anatomy-specific sigmaDP scaling law, and a heterogeneity x longitudinal compound-leakage interaction), and 13 honest bounded-negatives identifying empirical boundaries of the recommended deployment configuration. Three findings stand out: (i) the realistic-attacker bound generalizes within +/- 0. 06 across synthetic and two real datasets (maximum observed 0. 60 against multi-classifier shadow attackers) ; (ii) patient-vs-healthy subgroup analysis on KIMORE yields identical oracle accuracy across populations (delta = 0. 0000) ; (iii) anatomy-specific DP calibration scaling sigmaDP proportional to 1/sqrt (N) validated for N >= 5 with explicit knee N=3 outlier disclosure. Conclusions. Composition III is a viable privacy-preserving federated learning architecture for musculoskeletal-kinematic clinical digital twins. Empirical validation across synthetic data, two real datasets, multiple anatomies, three classifier families, three observation horizons, and mixed healthy/patient populations provides comprehensive grounding for clinical deployment. Archive contents. 21 frozen pre-registrations, 21 deterministic Python runners, 21 study reports, 21 verdict summary JSON files, 21 raw CSV data files, source code for the spiral-domain encoder and CMU MoCap and KIMORE parsers, 6 manuscript figures with generators, and the manuscript itself (Markdown and Word formats). Raw third-party data files are not redistributed per their respective licensing terms; download instructions and subject IDs are documented. Companion datasets. CMU Motion Capture Database (CC-BY 3. 0, http: //mocap. cs. cmu. edu/, funded by NSF EIA-0196217). KIMORE Rehabilitation Dataset (CC-BY 4. 0, Capecci et al. 2019 IEEE TNSRE 27 (7): 1436-1448), segmented version distributed via the EGCN repository. Reproducibility. Deterministic execution under PYTHONHASHSEED=0; numpy/scikit-learn/matplotlib versions specified in source README. See source/READMEREPRODUCE. md for full reproduction instructions.