Antibiotic Consumption Forecasting using a Combinatorial Convolutional Neural Network with Long Short-Term Memory Model

Document Type : Research Paper


1 Department of Information Technology Management, Faculty of Management and Accounting, Qazvin Branch, Islamic Azad University, Qazvin, Iran

2 Faculty of Electrical, Biomedical and Mechatronics Engineering, Qazvin Branch, Islamic Azad University, Qazvin, Iran



Background and Objectives: In recent years, medicine supply chain management has become more significant, especially after the Covid-19 pandemic. The most important issue is supply chain cost control. Medicine costs include nearly 30% of hospital expenses. If the drug inventory is not properly managed, it will lead to issues such as the lack of inventory of certain drugs, provision of excess inventory, increased costs, and, finally, patient dissatisfaction.
Method: In this study, an attempt has been made to predict and manage the pharmaceutical needs of hospitals using artificial neural networks and deep learning algorithms. The prescription and consumption information of one of the general hospitals in Hamedan city from 2013 until 2017 was extracted from the HIS databases. Since most drugs have different and specific characteristics and it is impossible to create the same prediction model for all drugs, the employed dataset is limited to the category of antibiotics. As a case study, the accuracy of the predictive model is evaluated, especially for cefazolin. We use a deep model to analyze the medical data time series efficiently. This model consists of two parts, a Convolutional Neural Network and a Long Short-Term Memory network (LSTM), which can sufficiently recognize the change history in time series prediction applications. The proposed model with many adjustable parameters in convolutional neural networks will bring good performance to overcome the complexities of the learning problem.
Results: Using deep learning in the training process can increase robustness by reducing the effects of complexity and uncertainty in medical data. Eventually, the prediction evaluation results and analytical criteria such as forecasting error and convergence speed and some statistical tests like R^2, MAE, and RMSE were presented. The average forecasting error for the proposed method is 0.028, and the measured values for RMSE, MAE, and R^2 are 0.095, 0.081, and 0.788, respectively.
Conclusion: A comprehensive comparison between some other predictive methods and the implemented model shows the outperformance of the proposed approach. Additionally, the evaluation results indicate the efficiency of the proposed approach.