Abstract
BACKGROUND: Accurate forecasting of lung cancer incidence is crucial for early prevention, effective medical resource allocation, and evidence-based policymaking. OBJECTIVE: This study proposes a novel deep learning framework-PSOA-LSTM-that integrates Particle Swarm Optimization (PSO) with an attention-based Long Short-Term Memory (LSTM) network to enhance the precision of lung cancer incidence prediction. METHODS: Using the Global Burden of Disease 2019 (GBD 2019) dataset, the model predicts age- and gender-specific lung cancer incidence trends for the next 5 years. The proposed model was compared against traditional models including ARIMA, standard LSTM, Support Vector Regression (SVR), and Random Forest (RF). RESULTS: The PSOA-LSTM model achieved superior performance across five key evaluation metrics: mean squared error (MSE) = 0.023, coefficient of determination (R (2)) = 0.97, mean absolute error (MAE) = 0.152, normalized root mean squared error (NRMSE) = 0.025, and mean absolute percentage error (MAPE) = 0.38%. Visualization results across 12 age groups and both genders further validated the model's ability to capture temporal trends and reduce prediction error, demonstrating enhanced generalization and robustness. CONCLUSION: The proposed PSOA-LSTM model outperforms benchmark models in predicting lung cancer incidence across demographic segments, offering a reliable decision-support tool for public health surveillance, early warning systems, and health policy formulation.