Machine learning based reference assisted signal recovery under anatomical domain shift in simulated radiofrequency fracture sensing
Year: 2026
Type: article
Abstract: Healthy-reference information may improve radiofrequency fracture-sensing representations under anatomical domain shift, but its benefit must be separated from noise dependence and threshold transfer. We evaluated 5,580 transmitter acquisitions from 1,395 physical configurations across five idealised locations. Twenty independent noise repetitions underwent fixed feature extraction and source-only LightGBM fitting, calibration and threshold selection. All-stage mean macro AUROC was 0.7718. For early healing, independent same-SNR referencing increased mean macro AUROC from 0.7935 to 0.9725; the matched difference was 0.1790 (multiplicity-adjusted 99.5% H2 interval 0.1343–0.2072), conditional on fixed anatomy and fitted models. All six matched endpoint contrasts favoured reference assistance. Equal-budget comparison favoured same-SNR over cross-SNR referencing; a second same-SNR reference provided no clear additional early-healing benefit. Early-healing pooled AUROC was 0.8567; locked sensitivity-targeted thresholds transferred with mean sensitivity 82.83% and specificity 20.00%. Saved-model replay on ten selected configurations across twelve numerical/noise settings changed 1,266 of 9,600 Youden and 539 sensitivity-targeted decision instances relative to baseline. These repeated instances are not independent patients or new cohort performance. The framework separates reference-information benefit, noise variation and decision transfer. These simulation-only findings do not estimate human diagnostic accuracy or electromagnetic continuum accuracy.
Source: Scientific Reports
Institutions: University of Bradford, British University Vietnam
Language: English
Subfield: Orthopedics and Sports Medicine
Field: Medicine
Domain: Health Sciences
Keywords: fracture healing, LightGBM, Youden index
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7236163619
