Automated Computational Detection of Disease Activity in ANCA-Associated Glomerulonephritis Using Raman Spectroscopy: A Pilot Study

Abstract

Biospectroscopy offers the ability to simultaneously identify key biochemical changes in tissue associated with a given pathological state to facilitate biomarker extraction and automated detection of key lesions. Herein, we evaluated the application of machine learning in conjunction with Raman spectroscopy as an innovative low-cost technique for the automated computational detection of disease activity in anti-neutrophil cytoplasmic autoantibody (ANCA)-associated glomerulonephritis (AAGN). Consecutive patients with active AAGN and those in disease remission were recruited from a single UK centre. In those with active disease, renal biopsy samples were collected together with a paired urine sample. Urine samples were collected immediately prior to biopsy. Amongst those in remission at the time of recruitment, archived renal tissue samples representative of biopsies taken during an active disease period were obtained. In total, twenty-eight tissue samples were included in the analysis. Following supervised classification according to recorded histological data, spectral data from unstained tissue samples were able to discriminate disease activity with a high degree of accuracy on blind predictive modelling: F-score 95% for >25% interstitial fibrosis and tubular atrophy (sensitivity 100%, specificity 90%, area under ROC 0.98), 100% for necrotising glomerular lesions (sensitivity 100%, specificity 100%, area under ROC 1) and 100% for interstitial infiltrate (sensitivity 100%, specificity 100%, area under ROC 0.97). Corresponding spectrochemical changes in paired urine samples were limited. Future larger study is required, inclusive of assigned variables according to novel non-invasive biomarkers as well as the application of forward feature extraction algorithms to predict clinical outcomes based on spectral features.