| 초록 |
Objectives: Glomerulonephritis is a primary renal disease typically diagnosed by renal biopsy. However, alternative diagnostic approaches based on clinical data are needed for patients who are unable to undergo biopsy. Electrophoretic patterns are known to vary among different types of glomerulonephritis. Therefore, we aimed to investigate whether electrophoresis-based image analysis could aid in disease classification. Methods: We retrospectively reviewed data from 206 patients diagnosed with glomerulonephritis by renal biopsy between 2010 and 2019. A total of 124 patients were included in the final analysis: 63 with IgA nephropathy, 15 with minimal change disease, 29 with membranous nephropathy, and 17 with focal segmental glomerulosclerosis. Graphic images of six protein fraction zones obtained from serum or urine electrophoresis were analyzed using artificial intelligence–based pattern recognition to generate disease-specific profiles. Diagnostic performance, including recall, accuracy, F1 score, area under the receiver operating characteristic curve (AUC), and average precision (AP), was evaluated for each subtype. Results: An initial multiclass classification model using 65 extracted features—including protein fractions, fractional areas, and intrinsic curve characteristics sampled across 256 intervals—failed to accurately discriminate among the four subtypes of glomerulonephritis. However, in a binary classification model based on urine electrophoresis distinguishing IgA nephropathy from other glomerular diseases, diagnostic performance was favorable. The model demonstrated a high recall rate, with an AUC of 0.903, an F1 score of 0.823, and an AP of 0.870 (Table 1). Conclusion: Urine electrophoresis image patterns may help distinguish IgA nephropathy from other forms of glomerulonephritis. This approach may serve as a useful ancillary diagnostic tool for patients in whom renal biopsy is not feasible. |