Machine learning models for curative and palliative oesophageal cancer treatment pathway prediction

Thavanesan, Navamayooran, Parfitt, Charlotte, Bodala, Indu, Walters, Zoë, Ramchurn, Sarvapali, Underwood, Timothy and Vigneswaran, Ganesh (2024) Machine learning models for curative and palliative oesophageal cancer treatment pathway prediction. European Journal of Surgical Oncology, 50 (1), [107152]. (doi:10.1016/j.ejso.2023.107152).

Abstract

Introduction: Oesophageal Cancer Multidisciplinary Teams (OC MDTs) operate under significant caseload pressures. This risks variability of decision-making which may influence patient outcomes. Machine Learning (ML) offers the ability to streamline and standardise decision-making by learning from historic treatment decisions to prediction treatment for new patients. We present internally validated ML models designed to predict OC MDT treatment decisions for curative and palliative OC patients.

Methods: four ML algorithms (multinomial logistic regression (MLR), random forests (RF), extreme gradient boost (XGB) and decision tree (DT)) were trained using nested cross-validation on a cohort of 938 OC cases from a single tertiary unit over a 12-year period. The models classified predicted treatments into one of: Surgery (S), Neoadjuvant Chemotherapy (NACT) + S, Neoadjuvant Chemoradiotherapy (NACRT) + S, Endoscopic or Palliative treatment. Performance was assessed on Area Under the Curve (AUC).

Results: across algorithms, all models performed strongly with mean AUC for Surgery = 0.849±0.026, NACT +S = 0.884±0.008, NACRT +S = 0.834±0.035, Endoscopic = 0.923±0.067 and Palliative = 0.963±0.033. MLR and XGB models performed most successfully (AUC 0.915±0.051 and 0.911±0.051 respectively). Models were integrated into a web-application to allow clinicians a user-friendly interface.

Conclusion: this study is the first successful use of ML to predict both curative and palliative MDT treatment decisions for OC patients, readily integrated into a user-friendly interface for real-time treatment pathway prediction. This offers significant potential to streamline MDT caseload, focus discussion on complex cases and provide an simple interface for such decision-support tools.

This record has no associated files available for download.

Contributors

Author: Navamayooran Thavanesan

Author: Zoë Walters

Author: Sarvapali Ramchurn

Author: Timothy Underwood

Author: Ganesh Vigneswaran

Download statistics