Using traditional and deep machine learning methods on predicting triage level in an emergency room

This research addresses the challenge of predicting triage levels in emergency room using traditional and deep machine learning methods in Turkey. The complexity of the local healthcare system and the Turkish language poses challenges for such predictive modeling. We employed a comprehensive methodology using various machine learning models, text embedding techniques, and sequential data analysis. We explored the performance of traditional machine learning models like Logistic Regression, Random Forest, and XGBoost, with the deep models like simple Neural Network model and a CNN-LSTM-Attention model. We evaluated these models using a dataset consisting of structured and unstructured data from emergency department visits at Ödemiş Devlet Hastanesi in İzmir. The unstructured data, particularly the complaints, were processed using various text embedding methods, including Bag of Words, Word2Vec, and BERT-based embeddings.. An important aspect of the study was the incorporation of patient history into the models. This included data on previous visits and diagnosed diseases, which were modeled using sequential methods like Recurrent Neural Networks and Long Short-Term Memory networks. The study also examined the effectiveness of the CNN-LSTM-Attention model, which combined different inputs. Also, we provided a new performance measure, customized F1 score, which is calculated by giving different weights for each error type. As a result, XGBoost with Word2vec emerged as the most effective traditional model, especially when using a comprehensive set of features. Deep models, combined various inputs, showed better performance in certain scenarios. As conclusion with future suggestions , emphasizing the need for more sophisticated models for several inputs to handle the complex patterns in emergency room data.