Machine learning pipeline with custom grid search for colorectal Raman spectroscopy data

Colorectal cancer remains a major health burden, and its early detection is crucial for effective treatment. This study investigates the use of a handheld Raman spectrometer in combination with machine learning to classify colorectal tissue samples collected during colonoscopy. A dataset of 330 spectra from 155 participants was preprocessed using a standardized pipeline, and multiple classification models were trained to distinguish between healthy and pathological tissue. Due to the strong class imbalance and limited data size, a custom grid search approach was implemented to optimize both model hyperparameters and preprocessing parameters. Unlike standard GridSearchCV, our method prioritized balanced accuracy on the test set to reduce bias toward the dominant class. Among the tested classifiers, the Decision Tree (DT) and Support Vector Classifier (SVC) achieved the highest balanced accuracy (71.77% for DT and 70.77% for SVC), outperforming models trained using traditional methods. These results demonstrate the potential of Raman spectroscopy as a rapid, non-destructive screening tool and highlight the importance of tailored model selection strategies in biomedical applications. While this study is based on a limited dataset, it serves as a promising step toward more robust classification models and supports the feasibility of this approach for future clinical validation.