Pancreatic Cancer Survival: Analysis of Risk Factors Using Classical and Deep Learning Methods
Statistical Methods and Data Analysis, Yalçın Tahtalı (ed) Tokat Gaziosmanpaşa University,Lütfi Bayyurt (ed) Tokat Gaziosmanpaşa University,Samet Hasan Abacı (ed) Ondokuz Mayıs University, Editör, Özgür Yayınları, Gaziantep, ss.165-184, 2026
- Yayın Türü: Kitapta Bölüm / Araştırma Kitabı
- Basım Tarihi: 2026
- Doi Numarası: 10.58830/ozgur.pub1404
- Yayınevi: Özgür Yayınları
- Basıldığı Şehir: Gaziantep
- Sayfa Sayıları: ss.165-184
- Editörler: Yalçın Tahtalı (ed) Tokat Gaziosmanpaşa University,Lütfi Bayyurt (ed) Tokat Gaziosmanpaşa University,Samet Hasan Abacı (ed) Ondokuz Mayıs University, Editör
- Dokuz Eylül Üniversitesi Adresli: Evet
Özet
Due to its aggressive clinical course and complex prognostic profile, pancreatic cancer remains one of the most challenging types of cancer to predict survival for. Although the traditional Cox Proportional Hazards (CPH) model is considered the gold standard, frequent violations of the "proportional hazards" assumption in pancreatic cancer datasets limit its predictive accuracy. This study presents a comprehensive comparison of classical statistical methods with modern machine learning and deep learning architectures using a large-scale dataset of 96,298 patients obtained from the SEER database. The study applied Kaplan-Meier, Time-Dependent Cox, and Parametric Accelerated Failure Time (AFT) models, as well as deep learning frameworks such as Random Survival Forests (RSF), DeepSurv, DeepHit, and DeepAFT. Model performance was rigorously evaluated using the C-Index, Integrated Brier Score (IBS), and time-dependent AUC metrics. The findings indicate that the Log-Logistic AFT model is the most effective method in terms of overall discriminatory power, with a C-Index of 0.8120. This suggests that pancreatic cancer survival data are well described by the Log-Logistic distribution. Additionally, in long-term projections (60 and 100 months), the RSF model achieved the highest sensitivity (AUC = 0.9550), while the DeepAFT model reduced prediction error (IBS = 0.0713). Variable importance analysis identified treatment method, histological type, and stage as key survival determinants; in particular, it was found that the combination of surgery and therapy extended survival time by 7.19 times. These results suggest that parametric models and machine learning tailored to the nature of the data in this field may offer superior reliability and interpretability compared to complex deep learning architectures.