Predictive machine learning models for survival outcomes in patients with pancreatic cancer.

Abstract

625 Background: Identifying prognostic markers is essential for accurately assessing risk levels in pancreatic cancer patients, leading to more personalized treatment approaches. In light of this, our study endeavors to build a survival prediction model for pancreatic cancer patients using various machine learning (ML) techniques, taking into account both clinical and inflammatory factors. Methods: The research included 311 pancreatic cancer patients on a prospective basis. Their data was sourced from the Clinical Data Warehouse of National Cheng Kung University Hospital. For the purpose of predicting 6 and 12-month survival, patients were randomly allocated into training, validation, and test groups in an 8:1:1 ratio. We utilized machine learning techniques and assessed them on the QOCA Aim (Quanta Omni Cloud Care for AI Medical Platform, version 2.0) platform. The model's effectiveness was gauged using the area under the receiver operating characteristic curve (AUC). Results: The cohort exhibited a mean age of 63.5 years with a standard deviation of 11. The median overall survival was documented at 16.8 months (95% confidence interval [CI] = 15.6–18). Analytical evaluation indicated that the optimal model for prognostication of 6-month survival was the balanced bagging classifier, yielding an AUC of 93.7%. Conversely, for the 12-month survival prediction, the k-nearest neighbors (kNN) classifier was superior, demonstrating an AUC of 86.1%. Subsequent variable importance analysis elucidated that salient predictors encompassed C-Reactive protein (CRP), surgical intervention, carbohydrate antigen (CA)19-9, age at diagnosis, and the neutrophil–lymphocyte ratio (NLR). Conclusions: The findings indicated that machine learning can be employed to develop high-efficacy predictive models for survival in pancreatic cancer patients. Such models offer promise for enhanced risk categorization and personalized prognostic assessment in this patient cohort.