Explainable and Adaptive AI For Predicting and Profiling Patients Visiting NHS Emergency Departments: An Early Exploratory Study
Abstract
This project is an early exploratory study investigating the health, lifestyle and demographic factors that lead to an increased risk of hospitalisation, and whether traditional ML and DL models are able to use these factors to predict hospital length of stay. Insights from professionals within the NHS were gathered using a short survey, and findings from this were used to generate a synthetic dataset containing 8 features, with two labels: short or long stays. The chosen features were Age, Gender, Health Conditions, Mental Health, Smoking, Alcohol Consumption, Socio-economic Status and Exercise. Synthetic data was chosen as a basis for experimentation due to availability of sensitive data, and for this early proof-of-concept stage, the flexibility of synthetic data is beneficial.
Data generation was carried out using Python, and four models were tested: Random Forest, KNN, SVM and Neural Network. The Random Forest performed best with a F1 score of 0.81, with SVM performing worst with a F1 score of 0.61. Each model was analysed using F1 score and confusion matrices. None of the models performed poorly, demonstrating that machine learning is a viable approach for predicting length of stay and supporting effective allocation of resources in NHS emergency departments. Limitations of this work include the lack of correlation between features in the synthetic dataset, which does not fully reflect empirical data. Future work will develop a more sophisticated model to predict and profile patients visiting the ED, using data from Hywel Dda NHS Health Board. This work serves as a starting point, with considerations for which models should be chosen, societal factors leading to increased patient risk, and the potential benefits of AI within a healthcare setting.
Download Meg's Thesis