Interpretable multi-label prediction of institution-defined IgA nephropathy clinical patterns using routine cross-sectional clinical and laboratory data

Abstract

IgA nephropathy (IgAN) clinical heterogeneity leads to inconsistent documentation of multiple, institution-specific pattern labels, creating an imbalanced multi-label phenotyping challenge. We developed an interpretable hybrid model, the Bidirectional Long Short-Term Memory-Convolutional Neural Network with Class-Specific Attention (BiLSTM-CNN-CAT), using routine cross-sectional data from 500 biopsy-confirmed IgAN patients. The model integrates BiLSTM and 1D-CNN feature extractors with class-specific attention for label-dependent representation. Training employed class-weighted binary cross-entropy with validation-based threshold calibration. Evaluation used an independent hold-out test set and five-fold cross-validation, with label-wise micro metrics, per-label ROC analysis, computational-complexity reporting and exploratory paired comparisons. The model achieved a micro-F1 of 0.989 and a micro-averaged ROC-AUC of 0.980 on the independent test set, with stable cross-validation performance. Ablation studies supported the contributions of the hybrid architecture and attention mechanism, while subgroup and missingness analyses provided preliminary checks of robustness. BiLSTM-CNN-CAT provides an interpretable baseline framework for standardising multi-label IgAN phenotyping from routine data. The findings should be considered exploratory until externally validated in multicentre cohorts.