Artificial Neural Network-Driven Predictive Modeling for Early Lung Cancer Risk Assessment

Abstract: We develop a framework that incorporates clinical information, smoking history and computed tomography (CT) derived radiomics into an artificial neural network (ANN) that can predict early lung cancer risk. We create a multimodal dataset by combining institutional medical record data from LIDC-IDRI images, we extract radiomic features from the images including nodule size, texture entropy, nodule edge sharpness, etc., and we normalize our data through proper imputation and outlier removal techniques and reduce dimensionality of all our extracted data through Principal Component Analysis (PCA). We use a patient split on training data to prevent overfitting in our model and measure performance with several metrics (AUC, Sensitivity, Specificity, Error Inspection through ROC Curve and Confusion Matrix). We pair our predictions with SHAP/LIME based explanations at a case level so that the physician or clinician can identify what variables contributed to their patients' risk scores and assist in developing appropriate thresholds for clinical evaluation. Overall, the combination of our prediction and explanation results provide evidence of the benefits of multimodal ANN risk assessments as well as demonstrate the importance of a transparent and appropriately governed deployment strategy.

Keywords: Lung cancer; risk prediction; artificial neural networks; radiomics; CT imaging; biomarkers; smoking exposure; SHAP; LIME; ROC analysis; confusion matrix; clinical decision support.