Enhancing Asthma Attack Predictions with Neural Network Architectures in Electronic Health Record Data

Stratifying asthma patients based on their risk of attacks is crucial to effective asthma management. However, the potential of neural network (NN) methods to develop models for asthma attack prediction has been largely unexplored. We systematically investigated the performance of various NN models for asthma attack prediction over 12 months using the Optimum Patient Care Research Database (OPCRD), a UK-wide primary care database. Our study included data from 550,478 patients aged 18–80 with current asthma. We evaluated the performance using two distinct approaches: A fully Connected Network (FCN) with manually extracted features and a combined FCN and Long-Short Term Memory (LSTM) leveraging manually extracted features and a sequence of raw clinical codes. We also investigated the impact of the number of unique clinical codes in the sequence. Our findings reveal the promising potential of the combined FCN and LSTM model, utilising the top 30% of the common clinical codes in the dataset, which achieved the highest performance with an area under the curve (AUC) of 0.785. This study underscores the significant potential of NN methodologies in advancing clinical risk prediction tasks using routinely collected clinical data, offering a hopeful outlook for future advancements in asthma management.