A lightweight dual-representation cross-attention framework for urinary FTIR classification of lupus nephritis and subtype discrimination.

BACKGROUND: LN subtype stratification is important for treatment and prognosis but relies on invasive renal biopsy, limiting repeated assessment. Urinary FTIR spectroscopy offers a label-free, repeatable alternative, yet remains insufficient for interpretable discrimination of clinically important subtypes, especially PLN and mixed LN. We therefore developed a compact, data-efficient, and interpretable urinary FTIR model for LN screening and PLN-mixed LN stratification. RESULTS: First-morning urine spectra (500-4000 cm-1) were collected and represented as synchronized raw and preprocessed signals. A lightweight Spectral Interaction Attention Former (SIA-Forme) was developed to encode patch-wise spectral information and fuse the two representations through bidirectional cross-attention. Under repeated stratified evaluation, the model achieved an ROC-AUC of 0.988 ± 0.013 for LN versus healthy controls (129 LN, 113 controls), and an accuracy of 86.0 ± 5.1% with an AUC of 0.875 ± 0.079 for PLN versus mixed LN discrimination (61 vs 48). The proposed method outperformed conventional chemometric models and several deep-learning baselines under the same protocol. Complementary interpretability analyses, including Welch's t-test with FDR correction, DeepSHAP, and spectral band masking, consistently highlighted the 1500-1700 cm-1 and 3300-3500 cm-1 regions as major contributors to subtype discrimination. SIGNIFICANCE: The results show that dual-representation fusion of urinary FTIR spectra can improve classification and interpretability in small-cohort LN analysis. The proposed framework provides a compact strategy for urinary spectral measurement and pattern analysis, with potential value for noninvasive LN screening and subtype-related assessment.