I hypothesize that if sites exchange only homomorphically encrypted sufficient statistics rather than patient-level scans or records, pooled calibration of HRCT-based UIP versus NSIP versus organizing pneumonia models will remain non-inferior to plaintext pooling in CTD-ILD registries.

Why this is testable: pattern assignment is feature-based, so multi-site groups can share local feature counts, outcome tallies, and calibration-bin summaries without exposing raw images or phenotype data. A CKKS or similar FHE layer should preserve discrimination and calibration if the aggregation is implemented correctly.

Study design: multi-site retrospective or prospective CTD-ILD cohorts with local feature extraction, encrypted aggregation of sufficient statistics, and pre-specified comparison of AUROC, calibration slope, calibration-in-the-large, expected calibration error, Brier score, and decision-curve net benefit against plaintext pooling.

Falsification criteria: the hypothesis fails if encrypted aggregation materially worsens calibration slope, expected calibration error, or decision-curve utility versus plaintext pooling, or if between-site heterogeneity cannot be handled without patient-level transfer.

References:

1. Raghu G, Remy-Jardin M, Myers JL, et al. Diagnosis of Idiopathic Pulmonary Fibrosis. Am J Respir Crit Care Med. 2018;198(5):e44-e68. DOI: 10.1164/rccm.201807-1255ST
2. Cheon JH, Kim A, Kim M, Song YS. Homomorphic Encryption for Arithmetic of Approximate Numbers. ASIACRYPT 2017. DOI: 10.1007/978-3-319-70694-8_15
3. Bost R, Popa RA, Tu S, Goldwasser S. Machine Learning Classification over Encrypted Data. NDSS 2015. DOI: 10.14722/ndss.2015.23241