Develop predictive model for direct treatment cost of acute coronary syndrome using a neural network algorithm

Tóm tắt

Background: Acute coronary syndrome (ACS) accounts for half the global economic burden. Current models to predict the ACS treatment cost have low accuracy and high complexity. This study aimed to build a more accurate predictive model using a neural network algorithm. Objectives: 1) Survey the cost of treating ACS at research hospitals. 2) Analyze factors associated with total direct cost of treating ACS at research hospitals. 3) Build and assess a model that predicts the total direct cost of treating ACS at research hospitals. Subjects and methods: A cross-sectional descriptive analysis was conducted based on the electronic medical records of 496 ACS inpatients at Cho Ray and Bach Mai hospitals. Factors associated with the total direct cost were used as inputs to build the neural network model. The grid search tool and k-fold cross-validation were used to select the best set of hyperparameters.Results: Mean total direct cost per ACS patient per course was 75,443,006±52,443,599 VND. Gender, health insurance type, course duration, health status at discharge, and number of comorbidities influenced the cost and were used as model inputs. Regarding the best set of hyperparameters, the distribution was Laplace, the transfer function was rectifier with dropout, the loss function was Absolute, the number of neurons in each hidden layer was 40, the number of hidden layers was 2, the lasso value was 1.0E-5, the ridge value was 1.0E-3, and the rho value was 0.999. The training set root mean squared error (RMSE) (25,091,949 VND) was smaller than those of the validation and test sets (33,025,969 and 29,202,777 VND, respectively); the difference between total predicted and actual cost was not significant, indicating that the optimization and regularization criteria were reached. Conclusions: The predictive model has relatively high accuracy and may be applicable in real-world settings. The model should be continuously enhanced to improve predictions and expanded to other patient groups based on big medical data.

Abstract

Background: Acute coronary syndrome (ACS) accounts for half the global economic burden. Current models to predict the ACS treatment cost have low accuracy and high complexity. This study aimed to build a more accurate predictive model using a neural network algorithm. Objectives: 1) Survey the cost of treating ACS at research hospitals. 2) Analyze factors associated with total direct cost of treating ACS at research hospitals. 3) Build and assess a model that predicts the total direct cost of treating ACS at research hospitals. Subjects and methods: A cross-sectional descriptive analysis was conducted based on the electronic medical records of 496 ACS inpatients at Cho Ray and Bach Mai hospitals. Factors associated with the total direct cost were used as inputs to build the neural network model. The grid search tool and k-fold cross-validation were used to select the best set of hyperparameters.Results: Mean total direct cost per ACS patient per course was 75,443,006±52,443,599 VND. Gender, health insurance type, course duration, health status at discharge, and number of comorbidities influenced the cost and were used as model inputs. Regarding the best set of hyperparameters, the distribution was Laplace, the transfer function was rectifier with dropout, the loss function was Absolute, the number of neurons in each hidden layer was 40, the number of hidden layers was 2, the lasso value was 1.0E-5, the ridge value was 1.0E-3, and the rho value was 0.999. The training set root mean squared error (RMSE) (25,091,949 VND) was smaller than those of the validation and test sets (33,025,969 and 29,202,777 VND, respectively); the difference between total predicted and actual cost was not significant, indicating that the optimization and regularization criteria were reached. Conclusions: The predictive model has relatively high accuracy and may be applicable in real-world settings. The model should be continuously enhanced to improve predictions and expanded to other patient groups based on big medical data.