Early detection of lung cancer in smokers using miRNA profiles and a hybrid deep learning framework.

Abstract

e13682 Background: Lung cancer remains the primary cause of cancer-related mortality globally, with early diagnosis being crucial for improving patient survival rates. Conventional diagnostic approaches are often inadequate for detecting early-stage malignancies, underscoring the necessity for minimally-invasive and reliable biomarkers, particularly within high-risk populations such as smokers. Methods: We conducted an in-depth analysis of small-RNA sequencing data from the GSE93284 dataset in the GEO, including miRNA expression profiles of bronchial brushings of 347 smokers who underwent bronchoscopy for suspected lung cancer, 194 diagnosed with lung cancer within a year, and the rest confirmed with benign diseases. The dataset was preprocessed, tailored to the model's loss function requirements. Mutual information-based feature selection was employed to retain the top 200 miRNAs, significantly reducing dimensionality while preserving predictive information. Our model architecture integrated a hybrid wide and deep neural network enhanced with feature-wise self-attention mechanisms, designed to capture complex miRNA interactions. To address class imbalance, we implemented Focal Loss with class-specific weighting and augmented the training data using a Conditional Variational Autoencoder, which synthesized additional miRNA profiles conditioned on class labels. Furthermore, we incorporated Gradient Boosted Trees into an ensemble framework, combining predictions via a weighted averaging approach. The dataset was stratified into training, validation, and test subsets to ensure robust evaluation. Results: The ensemble model achieved an accuracy of 0.85 on the test set, with a sensitivity of 0.95 and precision of 0.83, resulting in an F1-score of 0.89. The ROC-AUC was calculated at 0.81, reflecting strong discriminative capability. Optimal threshold determination using Youden’s J statistic yielded a threshold of 0.47, effectively balancing sensitivity and specificity. Confusion matrix analysis revealed high true positive rates alongside manageable false positive rates, affirming the model's efficacy for clinical screening applications. Conclusions: Our deep learning framework, leveraging miRNA expression profiles and advanced augmentation and ensemble techniques, demonstrates high sensitivity and accuracy for the early detection of lung cancer in smokers. This approach holds significant promise for enhancing clinical screening and improving patient prognoses.