Developing and validating a new neural network model for early identification of high‐risk acute pancreatitis patients

Abstract

Abstract Introduction Existing research studies on the use of machine‐learning algorithms for predicting mortality in acute pancreatitis patients is limited and characterized by low methodological quality. Furthermore, those risk‐scoring systems have demonstrated suboptimal performance. Our objective was to develop a high‐performing, robust neural network (NN) model that can identify acute pancreatitis patients who are at risk of 30‐day mortality early. Methods Adult patients with confirmed diagnosis of acute pancreatitis were included in our study. Patients with preexisting diagnoses of pancreatic cancer, iatrogenic endoscopic pancreatitis, abdominal trauma, exacerbations of chronic pancreatitis, or pregnancy were excluded. The measurements of the variables were obtained within the first hour after their admission to the emergency department. The patients were divided into development and validation cohorts with a ratio of 75 to 25 and a NN model was derived in the derivation cohort after oversampling. For the NN model, multilayer perceptron was used. 5‐fold cross‐validation was performed for internal validation and the final model was externally validated. DeLong's test was used to compare the area under the curves (AUC). Results There were 279 and 93 patients in the development and validation cohorts, respectively. In the out of sample validation, the model had an AUC of 0.936 (95% confidence interval (CI) = 0.864–0.999), sensitivity of 100% (95% CI = 47.8–100), specificity of 80.7% (95% CI = 70.9–88.3), and negative likelihood ratio of ≈0. Conclusion Our NN model demonstrated excellent performance in both the derivation and validation phases for the early identification of acute pancreatitis patients at high mortality risk. Notably, it was able to identify all patients who had 30‐day mortality. Our study offers several key advantages, including a transparent methodology, credible validation, and the provision of a calculator to assess the model's performance across different populations and settings.

Keywords

• acute pancreatitis
• machine learning
• mortality
• neural networks
• prediction