Hybrid Forecasting Framework for Risk Analysis in Cystic Fibrosis

For efficient clinical decision support and prompt intervention in the setting of cystic fibrosis (CF), it is imperative to properly forecast fast decreases in lung function. However, the variability in lung function deterioration and the limitations of current forecasting methods, particularly those relying on Forced Expiratory Volume in 1 second (FEV1) criteria, pose significant challenges. Traditional survival models often fail to incorporate longitudinal biomarker and risk factor observations, leading to inadequate mortality predictions. To address these limitations, this work introduce a Hybrid Forecasting Framework utilizing deep learning algorithms to enhance survival prediction for CF patients. This proposed approach, termed the Temporal Fusion Transformer (TFT) for Competing Risk Prediction, integrates longitudinal data including FEV1 %, Lung Allocation Score, age, gender, weight, and pancreatic sufficiency to dynamically update survival predictions for multiple competing risks. Employing D-SMOTE for data balancing and multi-head attention to capture dynamic temporal connections, the TFT model effectively manages complex temporal relationships and dependencies. This advanced framework surpasses traditional methodologies, offering clinicians a powerful tool for making informed decisions regarding the timing of lung transplantation for CF patients. The TFT for Competing Risk Prediction significantly outperforms current survival models, with an accuracy, precision, and Mean Absolute Error (MAE) of 97%, 91%, and 0.2, respectively. This presents a highly accurate and precise method for predicting rapid declines in lung function in patients with CF.