Machine Learning-based Prediction of the Likelihood of Colorectal Cancer Using miRNA Expression

[Background] Colorectal cancer (CRC) comprises 10% of all cancer diagnoses, making it the third most diagnosed cancer globally. Despite its prevalence, most current methods for identifying CRC lack sensitivity and consistency while being invasive and costly. Thus, this study aimed to develop artificial neural network (ANN) models that could accurately detect CRC using miRNA expressions in tissue and plasma samples. [Methods] The study used miRNA expression profiles of formalin-fixed paraffin-embedded tissue and plasma samples obtained from CRC patients and healthy controls. ANNs were trained to discriminate between CRC patients from healthy controls using the relative expression of miR-21-5p, miR-196b-5p, miR- 135b-5p, miR-92a-3p, miR-29a-3p, and miR-197-3p in colorectal tissues and blood plasma. Multivariate logistic regression (MLR) and decision tree (DT) models were used to compare the performance of the ANN models. [Results] The ANNs achieved an accuracy of 98.5 and 88.2%, a sensitivity of 90.9 and 80.4%, a specificity of 92.6 and 84.7%, and an area under the ROC curve of 0.92 and 0.83 for the plasma and tissue samples, respectively. Moreover, sensitivity analysis of the ANN models showed that miR-135b-5p and miR-92a-3p had the greatest influence in distinguishing CRC from healthy plasma and malignant from neoplasm-free colorectal tissues, respectively. However, only miR-135b-5p was significantly downregulated in both CRC plasma and malignant colorectal tissue samples. Results from the MLR and DT models support the results from the ANN sensitivity analysis. [Conclusion] Our results show that the trained ANNs were able to accurately and confidently detect CRC using the considered six miRNA expression levels in colorectal tissue and plasma samples, providing an accurate, rapid, and less-invasive approach to diagnosing CRC.