Predicting hospital admissions and its cost due to respiratory diseases in Brazil using Machine Learning Time Series Forecasting

Abstract

Background: Respiratory diseases (RD) impose an immense health burden and over 1 billion people suffer from acute or chronic RD. Artificial Intelligence (AI) can improve the quality of healthcare, with the potential of assisting in the medical diagnosis of many diseases and reducing unnecessary hospitalizations and costs. This article aims to develop a Machine Learning (ML) model to predict the healthcare resources utilization (HCRU) and costs associated to RD hospitalizations in the Brazilian public health system (SUS). Methods: Data were extracted from three public databases: Hospital Information System (SIH), “e-saúde” database and Meteorological Database, in the city of Curitiba, between 2017 and 2019. All analyzes considered the number of hospitalizations per day. The outcomes predicted by ML were the cost and the number of hospitalizations in the next seven days after a RD claim. The models were created by data mining process. Different algorithms were tested by the model building process up to five times. The best model for the seven-day cost and utilization forecasts was defined according to mean absolute percentage error (MAPE), mean absolute error (MAE), root mean squared error (RMSE). The SHAP method was used to analyze the interpretability of the best selected model. Results: There were, on average, 315.41 hospitalizations and 97,596 primary care services for RD per week in the city of Curitiba between 2017 and 2019, with an average cost of 246,390.30 US dollars (R$ 549,332.87). The Recurrent Neural Network (RNN) methods (LSTM and GRU) presented the best results for forecasting costs and HCRU. LSTM model outperformed all other algorithms in both models with a RMSE of 0.07 and 0.04 respectively. The most impacting variables in the model (SHAP analysis) were the meteorological ones. However, the forward to specialist, type of attendance and medical specialty on the ambulatorial records were also important. High average temperatures support the model to make a prediction of a smaller number of hospitalization days for that period. Conclusion: The prediction model used was robust enough to predict information about hospitalization and costs related to RD, demonstration its applicability as a tool to optimize resources allocation and health promotion strategies.