Infrared spectroscopy with statistical analysis and machine learning for cancer risk assessment in inflammatory bowel disease patients.

Patients with inflammatory bowel disease (IBD) undergo regular colonoscopic surveillance due to their elevated risk of developing colorectal cancer (CRC). However, current clinical, endoscopic and histopathological risk stratification methods can be limited in sensitivity and objectivity, highlighting the need for complementary molecular approaches. In this study, Attenuated Total Reflectance-Fourier Transform Infrared Spectroscopy (ATR-FTIR) was used to acquire mid-infrared (MIR) spectra from IBD-associated endoscopic biopsy samples (30 patients analysed; 10 who developed dysplastic pre-cancerous lesions and 20 who did not during long-term follow-up). Samples were stratified according to baseline clinical CRC risk (high/low) and subsequent pre-cancerous lesion development, enabling assessment of molecular signatures associated with future cancer risk rather than solely current disease status. Spectral variations were analysed using chemometric methods and machine learning classifiers. Principal component analysis (PCA) was performed to evaluate spectral separation, with the first three components (PC1, PC2 and PC3) accounting for 76.8%, 19.2% and 2.7% of the total variance, respectively. Hierarchical clustering analysis (HCA) further explored similarity patterns across defined spectral regions. Among the evaluated models, PLS-based classifiers achieved a balanced accuracy of 0.61-0.70 under repeated nested cross-validation; however, a permutation test indicated that this performance was not significantly above chance (