Deep neural networks and association with inpatient survival for patients admitted to the hospital with colon cancer undergoing open colectomy.

Abstract

e15537 Background: During hospitalization for colectomy patients with colon cancer may be at risk for mortality due to several reasons including underlying cancer, other comorbidities or age. Our aim was to develop a machine learning algorithm to predict end of admission mortality in patients undergoing colectomy for their cancer. We hypothesized that machine learning could be applied to patients undergoing open colectomy as part of their colon cancer treatment. Methods: All adult patients (> 18 years) for the National in-patient Services (NIS) database 2014 was used to develop our algorithm. We extracted all patients with a diagnosis of cancer using the ICD-9 codes and all patients that underwent colectomy. We conducted a multivariate analysis to look at the risk factors that dictate mortality in these patients. We also developed a linear regression model and a deep learning algorithm to predict mortality in these patients. Results: We identified a total of 4120 patients that underwent open colectomy with colon cancer in the NIS 2014. We observed that several clinical and lab-based parameters were statistically significant for multivariate analyses while others were not. With most significant being all patients refined diagnosis related groups (APRDRG) risk mortality (HR = 6.20, 95%CI = 4.10-9.36), APRDRG severity (HR = 4.78, 95%CI = 3.27-7.00), chronic anemia (HR = 0.61, 95%CI = 0.40-0.94), coagulation disorders (HR = 2.32, 95%CI = 1.31-4.10), chronic electrolytes disorders (HR = 2.12, 95%CI = 1.39-3.24), neurological disorders (HR = 1.92, 95%CI = 1.11-3.32), underweight (low BMI) (HR = 2.40, 95%CI = 1.05-5.48), Hyperkalemia (HR = 2.40, 95%CI = 1.28-4.50), Acidosis (HR = 4.04, 95%CI = 2.64-6.19). In the machine learning analysis, we found out that our proposed DNN outperformed the RF with test set accuracy of 91.1%, sensitivity of 88.5%, specificity 91.2%, PPV of 24.7, NPV of 99.5% and AUROC of 0.968 [95%CI = 0.08-.014]. Conclusions: Our novel DNN model outperformed RF classifier models. The model is easy to implement, user friendly and with good accuracy. However, further external validation of the model is required.