Advanced ML Superior to TNM Staging in Predicting Pancreatic Cancer Survival

SAN ANTONIO — Accurate prognostic prediction is crucial for treatment planning in patients with pancreatic cancer (PC). Advanced machine learning (ML) techniques are superior to the existing TNM staging system for accurately predicting the survival of patients with PC, according to a study presented at the American College of Gastroenterology’s Annual Scientific Meeting & Postgraduate Course in San Antonio, Texas, on October 28, 2019.

To evaluate ML-based analytic models, researchers used data from the Surveillance, Epidemiology, End-Results (SEER) database to identify patients aged ≥18 years with histologically confirmed PC diagnosed between 2004 and 2015 (N=42,673). Included in the analysis were demographic, socioeconomic, and clinical variables (American Joint Committee on Cancer [AJCC] tumor stage, site, grade, size of tumor, treatment received, and survival in months). Classification tree-based, bayesian net, neural net, support vector, and K -nearest neighbor classifier prediction analytic models were built and trained using supervised learning algorithms to predict 1-year survival probability. Using only baseline variables (A model), separate models were built and included treatment-related variables (surgery, chemotherapy, and radiotherapy [B model]).

Multivariate Cox regression analysis identified 10 independent prognostic variables: year of diagnosis, tumor stage, tumor grade, presence of metastasis, age, tumor size, marital status, insurance status, US region, and socioeconomic factors. These were entered into the ML models. Using AUROC (Area Under the Receiver Operating Characteristics) of the A and B models, 1-year survival probabilities were 0.804 (95% confidence interval [CI], 0.800-0.808) and 0.832 (95% CI, 0.828-0.836), respectively, which were both significantly higher survival probabilities predicted by the TNM staging system (AUROC=0.696; 95% CI, 0.691-0.700; P <.0001).

The authors concluded that traditional prognostic evaluation methods, ie, TNM staging, do not incorporate certain clinical and socioeconomic factors that impact patient survival. A more accurate and comprehensive tool is needed; advanced ML techniques may serve as an effective tool for prognostic evaluation of PC in clinical settings.

Reference

Das A, Ngamruengphong S. Machine learning based predictive models are more accurate than TNM staging in predicting survival in patients with pancreatic cancer. Paper presented at: Annual Scientific Meeting & Postgraduate Course; October 28, 2019; San Antonio, Texas. Abstract P0942