A Machine Learning Based Decision Support System for Healthcare Triage Applications

Machine learning (ML) models are useful for developing clinical decision support systems (CDSSs). However, the flexibility of ML models leads to uncertainty and inconsistency, which requires a balance of various factors. This study aims to explore the development and application of ML-based predictive models in the context of a self-triage CDSS. Specifically, we build ML models to predict and assign medical specialists for patients based on the input symptoms. Two ML classifiers, namely a deep neural network (DNN) and a K-nearest neighbour (KNN) model, are developed. Firstly, data preprocessing is performed to clean and transform patient related data. Feature selection is then applied to optimize computational resources by identifying only the relevant features (symptoms). Hyper-parameter tuning is also carried out to improve the model performance. Then, the optimized features and hyper-parameters are fed into the DNN and KNN models. These models are trained and evaluated on datasets containing a large number of diseases and medical specialists. The results indicate that the DNN model outperforms the KNN counterpart in terms of accuracy and F1-score. Implications of ML-based prediction models are analysed and discussed using a use case pertaining to patient self-triage and appointment scheduling.