Abstract
BACKGROUND: Managing chronic viral infections like Hepatitis C virus (HCV) often requires expensive healthcare resources and highly qualified personnel, making efficient diagnostic methods essential. Despite remarkable therapeutic advancements for the treatment of HCV, several challenges remain, such as improved fast diagnostic procedures allowing universal screening. OBJECTIVE: We propose a novel approach that combines Near-Infrared Spectroscopy (NIRS) and clinical data with machine learning (ML) to improve Hepatitis C Virus (HCV) detection in serum samples. METHODS: NIRS offers a fast, non-destructive, and residue-free alternative to traditional diagnostic methods, while ML models enable feature selection and predictive analysis. We applied L1-regularized Logistic Regression (L1-LR) to identify the most informative wavelengths for HCV detection within the 1,000-2,500 nm range, and then integrated these spectral features with routine clinical markers using a Random Forest (RF) model. Our dataset comprised 137 serum samples from 38 patients, each represented by a NIRS spectrum and clinical data from blood tests. RESULTS: After preprocessing with Standard Normal Variate (SNV) correction and downsampling, the best-performing RF model, which combined NIRS features and clinical data, achieved an accuracy of 72.2% and an AUC-ROC of 0.850, outperforming models using only clinical or spectral data. Feature importance analysis highlighted specific wavelengths near 1,150 nm, 1,410 nm, and 1,927 nm, associated with water molecular states and liver function biomarkers (GPT, GOT, GGT), reinforcing the biological relevance of this approach. CONCLUSIONS: These findings suggest that integrating NIRS and clinical data through machine learning enhances HCV diagnostic capabilities, offering a scalable and non-invasive alternative for early detection and risk assessment.