A Semi-Supervised Machine Learning Model for Risk Analysis of Urinary Tract Infections in People with Dementia

Abstract

Background: Urinary Tract Infections (UTIs) are common conditions in the elderly population and one of the key causes of hospital admissions in people with dementia. One of the key challenges in dealing with UTIs is that the condition often remains undiagnosed until it advances and shows severe symptoms which can lead to hospitalisation.     

Method: This study presents a new system for detecting UTI in people with dementia. The system relies on low-cost in-home sensory monitoring combined with a novel machine learning algorithm. Environmental and physiological monitoring devices provide continuous data collection from homes, and the proposed machine learning model is trained to extract and identify the risk of UTIs. In conditions such as UTI, collecting sufficient training data to train machine learning models is a challenging task. The limited access to labelled data and examples creates the risk that learning models can become overfitted or biased. We propose a semi-supervised learning model to process a large unlabelled data set in combination with a smaller labelled dataset. We use an unsupervised learning technique which trains a model by extracting the underlying features from a set of unlabelled data and use the probabilistic neural network to estimate the density distribution. We use the unsupervised model for feature extraction and use these features from the smaller dataset to train a supervised classifier.  The proposed approach is tested in real-world application and within a healthcare monitoring system with 110 participants.   

Finding: The proposed system can record the environmental and clinical data in an integrated framework. This provides information for the proposed machine learning model to identify the risks of UTIs by processing the patterns and changes in vital signs and activity patterns. The evaluation results show that the proposed model can detect the risk of UITs with an F1-score of 84%, which outperforms the baseline methods by 10% on average.      

Interpretation: The proposed model uses continuous environmental monitoring data and vital signs to detect early symptoms and the risk of UTIs. This is a first study using deep learning and semi-supervised models to analyse in-home sensory data in remote healthcare monitoring scenarios. The model provides a new approach to analyse and interpret continuous observation and measurement data for dementia care. 

Funding: The UK Dementia Research Institute, Medical Research Council (MRC), Alzheimer's Society and Alzheimer's Research UK.

Declaration of Interests: The authors declare no competing interests.

Ethics Approval Statement: The study protocol for this work has been reviewed and approved by the South East Coast Surrey NHS Research Ethics Committee and all the research has been performed in accordance with relevant guidelines and regulations. We obtained informed consent from all the study participants.