E209: Machine Learning to Predict Early Death Despite Pancreaticoduodenectomy

ePoster Abstract Author(s)

Clayton T. Marcinak, MD

Resident
Division of Surgical Oncology, Department of Surgery, School of Medicine and Public Health, University of Wisconsin–Madison, Madison, WI, USA
Madison, Wisconsin, United States

Submitter(s)

Clayton T. Marcinak, MD

Resident
Division of Surgical Oncology, Department of Surgery, School of Medicine and Public Health, University of Wisconsin–Madison, Madison, WI, USA
Madison, Wisconsin, United States

Author(s)

Clayton T. Marcinak, MD

Resident
Division of Surgical Oncology, Department of Surgery, School of Medicine and Public Health, University of Wisconsin–Madison, Madison, WI, USA
Madison, Wisconsin, United States
KA

Kaleem Ahmed, MD, MSAI

Research Specialist
Wisconsin Surgical Outcomes Research Program, Department of Surgery, University of Wisconsin–Madison, Madison, WI, USA, United States
SI

Sheriff Issaka, BS

Data Scientist
Division of Surgical Oncology, Department of Surgery, University of Wisconsin–Madison, United States
SZ

Syed Nabeel Zafar, MD, MPH

Surgical Oncologist
University of Wisconsin School of Medical and Public Health, United States

Introduction:

Pancreaticoduodenectomy (PD) is the only surgical option for right-sided pancreatic ductal adenocarcinoma (PDAC). However, over 25% of patients undergoing PD die within one year of diagnosis. These patients carry all the risks for significant postoperative morbidity with no survival advantage when compared to non-surgical treatment options. We aimed to determine if machine learning (ML) models have superior accuracy to traditional models at predicting futile surgery in patients with PDAC.

Methods:

We analyzed data from patients in the National Cancer Database undergoing PD for PDAC between 2004 and 2020. PD was defined as futile if the patient died within 12 months of cancer diagnosis. We trained predictive models using 80% of the dataset and 16 preoperative input variables including demographics, clinical and pathologic factors, and neoadjuvant treatment status. Models included logistic regression (LR), decision tree (DT), random forest (RF), and gradient boosting (GB) classifiers. A grid search and cross-validation were used for model optimization. Models were tested on a 20% hold out dataset using area under the receiver operating characteristic curve (AUC).

Results:

A total of 66,331 patients were included in the analysis. Of these, 34,260 (51.7%) were men, with a median age of 67 years (IQR, 59 to 74 years). The median follow-up time for the cohort was 75.8 months (IQR, 44.1 to 119.5 months). A total of 16,768 (25.2%) patients met the criteria for futile surgery. Of all trained models, GB and LR outperformed the other models with AUC scores of 0.6936 and 0.6858, respectively (Figure). Variables associated with the highest odds of futile PD were poor tumor differentiation, and tumor size greater than 4 cm. Variables associated with the lowest odds of futile PD were receipt of neoadjuvant systemic therapy, tumor size ≤ 2 cm, and well differentiation.

Conclusions:

Despite using a limited clinical data set, we demonstrate the ability of statistical models to predict the odds of futile PD with decent accuracy. ML models did not significantly outperform multivariable LR in this analysis using 16 input variables. Similar analyses on more granular datasets holds promise to more accurately identify patients who may not benefit from a complex operation. These findings have important implications for shared decision-making and optimizing care for patients with PDAC.

Learning Objectives:

Define operative futility in the cohort.
List the factors associated with the highest risk of early death after pancreaticoduodenectomy.
Describe the performance of various machine learning models to predict futile surgery.