Clinically relevant and explainable feature extraction from remote monitoring data for the prediction of urinary tract infections

Abstract

AbstractBackgroundA frequent cause for hospitalisation in people living with dementia (PLWD) is urinary tract infections (UTIs) (Dufour et al., 2015). Early detection aids in avoiding unplanned hospitalisations and machine learning and connected sensors enable development of risk analysis models based on in‐home monitoring data. This work makes use of data from the ongoing Minder study at the UK Dementia Research Institute, including in‐door movement, appliance use, physiology, sleep, and environmental information. Previous work (Li et al., 2020) shows complex models can achieve high accuracy when detecting UTIs, but interpretability and generalisability remain an open issue. This work evaluates how clinically interpretable features and simple models perform.MethodThe engineered features reflect visible symptoms of UTIs in older adults and are daily aggregated and pre‐processed information from PLWD’s homes containing passive sensors. 14 activity, sleep and physiological features were engineered relating to bathroom visits, entropy rate, physiological readings, and sleep behaviour. Raw data consisted of frequencies of the activity sensor firings. Their use in predicting cases of UTIs was compared by evaluating multilayer perceptron models on labelled data (643 negatives and 311 positives, from 39 PLWD), using 5‐fold cross‐validation (80%/20% train/test split), with 5 repeats. Prior to training and testing, the data was z‐normalised using unlabelled data. When evaluating, recall is most essential because reducing false positive rates is crucial in our setting.ResultThe highest performing model achieved the best accuracy, recall, precision and F1 score (85.6%, 69.9%, 81.5%, 74.8%) on combinations of engineered and raw features; though not significantly higher than with raw features alone (83.2%, 63.5%, 80.5%, 70.8%). Full results can be seen in the attached figure.ConclusionOur engineered features do not significantly improve results but add interpretability and generalisability to predictions. The models' hidden layers have a high capacity for learning patterns in the raw data, however, may not be clinically interpretable and we hypothesise that engineered features improve generalisation. Engineered features are homogeneous across the cohort and new data, and so create a way of representing the data which is not single device dependent. To verify this, we will continue our data collection and will conduct further experiments.