Late Breaking Abstract - Recurrent Neural Networks And Environmental Data For Copd Exacerbation Prediction

Introduction: TelEPOC is a telemedicine program developed for COPD patients with frequent exacerbations. Patients submit the following variables each morning: heart rate, temperature, oxygen saturation, respiratory rate, steps walked and a questionnaire about cough, sputum, dyspnea and general health status. The current dataset is composed of 3.5 years of data, where 111 patients submitted 80303 questionnaire responses. Objectives: We trained a Recurrent Neural Network (RNN) with the goal of predicting whether a patient will suffer an exacerbation within the next 3 days. Besides, we augmented the dataset by including 99 variables which capture daily environmental conditions existing in the place where each patient resides. Methods: We divided our dataset into 3 different splits: train, validation and test. We then performed an hyperparameter search in order to find the best model. We repeated this process with 4 different random splits of the data and report the mean score obtained on the test set. Results: The best RNN trained with the TelEPOC dataset achieved an Area under the ROC curve (AUROC) of 0.94 and an Area under Precision-Recall (AUPRC) of 0.84. The AUPRC achieved by random predictions was 0.14. Adding the environmental data to that dataset resulted in no improvement of the scores. We trained an additional model using just the environmental data, achieving an AUROC of 0.74 and an AUPRC of 0.53. This shows that environmental variables do contain information about future exacerbations, but such information is probably already contained in the TelEPOC dataset. The most relevant environmental variables for these predictions where: rain, atmospheric pressure, and NO2, SO2 and NH3 levels.