There are circumstances in which patients require immediate medical care. Usually, hospitals have a specialized unit ready to take care of emergencies such as: strokes, heart attacks, sepsis, severe wounds, and other medical conditions that are also classifies as emergencies. This research compared models based on artificial neural networks (ANN) to predict the diagnosis of patients received at the emergency department (ED) of a public hospital.

The research was carried out following a classic 4-phase methodology: analysis, design, development, and validation. During the analysis, patient records collected by ED personnel during 2020 were thoroughly reviewed and preprocessed. During the design, some of the records were selected and divided into several subsets following specific criteria. Each subset was a set of independent variables used as input to build a prediction model. Technical considerations relate to the artificial neural networks (ANN) such as the number of hidden layers and the number of epochs, batch size, size of the training and test set, and size of the validation dataset were also defined at these point. The phases of construction and the validation are carried out entirely using the WEKA 3.9.6. Numerous experiments, trial and error adjustments, and replications were necessary to produce the results shown in this article.

For the purposes is the present work around 15,000 records were considered. The complete dataset was divided in two parts: 80% for training and test, and the remaining 20% for validation. The approach to predict the patient diagnose considered the construction of ANNs with of independent variables.

During the construction several configurations of ANNs were tried to achieve better results. Only the records allocated for training and test dataset is used during the construction. The performance of the proposed models was measured in terms of the rate of correct predictions made with records allocated for validation.

The experimental result revealed that, depending on the number of diagnoses, the values of the target class, the proposed models were able to predict correctly between 55% and 85% of the cases. In conclusion, ANN-based models can help predict ED patients’ diagnosis with a reasonable degree of certainty. However, the success of the model depends greatly on the numbers of values that the target class has.