Cross-Device Risk Calibration for Surface-Enhanced Raman Spectroscopy-Based Urologic Disease Classification

Surface-enhanced Raman spectroscopy (SERS) combined with deep learning offers a noninvasive route for molecular screening of urologic diseases. Its broader analytical utility, however, is constrained by two coupled challenges: overconfident predictions near interclass boundaries and cross-device/acquisition-condition distribution shift. We therefore tested whether risk-calibrated patient-level SERS analysis could identify reportable outputs under the protocol-linked composite shift observed in the external data set. Internally, a derivative-guided dual-stream classifier with soft-risk retain/defer triage separated a trusted reportable subset from high-risk samples sent to manual review and kept accuracy above 95% among retained spectra. Externally, the workflow used guarded patient-level calibration-path selection followed by support-dependent few-shot covariance alignment to recover covered known-class outputs under a cross-device/acquisition-condition composite shift. Across the four covered external classes, patient accuracy recovered to 85.82%, and accuracy among retained patients reached 89.95%. When prostate cancer was added to the external test stream as unknown input, a front-end dual-evidence gate increased the unknown patient defer rate from 4.80% to 78.40% while maintaining 90.07% accuracy among gate-passed known-class patients. The evaluation therefore goes beyond closed-set accuracy, testing internal release/defer control, few-shot calibrated recovery of covered external classes, and prereport deferral of unsupported inputs.