Abstract
<title>Abstract</title> <p>Lung cancer remains one of the leading causes of cancer-related mortality worldwide, underscoring the need for computational tools that can support prognosis and treatment decision-making. This study presents an integrative survival-analysis pipeline that combines structured clinical variables with unstructured clinical notes embedded using Bidirectional Encoder Representations from Transformers (BERT). A curated cohort of 500 patient records was drawn from a publicly available repository of 23,657 cases to evaluate three modelling approaches: Cox proportional hazards regression, Random Forest, and DeepHit. Kaplan–Meier estimation showed no statistically significant differences in survival across treatment groups (log-rank p = 0.774). The Cox model achieved a concordance index of 0.566, identifying tumor location (upper lobe, HR = 1.25) and disease stage (Stage III, HR = 0.77; Stage IV, HR = 0.76) as significant predictors. The Random Forest regressor achieved the best predictive accuracy (MSE = 1226.97), with several BERT-derived text features ranking among the most important predictors, demonstrating that free-text notes carry prognostic information beyond structured fields. In contrast, the DeepHit neural survival model underperformed (C-index = 0.485), reflecting the limitations of deep learning on modest-sized cohorts. A Windows Presentation Foundation (WPF) C# graphical interface was developed to surface Kaplan–Meier curves, patient-specific risk scores, feature-importance plots, and GPT-2 narrative summaries. For a representative patient, surgery was recommended as the most beneficial treatment, although uniformly high confidence across options highlighted the need for probability calibration. These findings show that classical regression remains robust for modest datasets, while ensemble methods benefit from text features, and deep learning requires larger cohorts to be effective. The study demonstrates the translational potential of integrating transformer-based NLP with survival modelling and provides a clinician-facing interface to support decision-making.</p>