A lightweight unsupervised approach for adverse health detection and digital biomarker discovery in people living with dementia

Abstract Background Sensor‐based remote health monitoring of persons living with dementia (PLwD) can be used to gain insights into their health and monitor the progression of their condition, with minimal intrusion. This helps minimize preventable hospital admissions, while allowing researchers to improve their understanding of dementia. Existing approaches for detecting activity and behavioural anomalies in PLwD are challenged by noise in data, lack of annotated datasets, multivariate data, scalability, data drift and explainability. Method We propose and evaluate a solution based on the Matrix Profile, an exact, ultra‐fast distance‐based anomaly detection algorithm, specifically the Contextual Matrix Profile (CMP), to detect anomalies that may indicate unusual activity and onset of UTI. Daily household movement data collected via passive infrared (PIR) sensors are used to generate CMPs from location‐wise sensor counts, duration and change in hourly movement patterns. We create CMP‐based multivariate anomaly detection models to generate a single daily normalized anomaly score for each patient. We discover digital biomarkers of anomalies and evaluate our method vs. three state‐of‐the‐art algorithms. Result CMP‐based models yield up to 85% recall with only a 5% alert rate, when evaluated on a subset of 9363 days from 15 participant households with 41 clinically validated incidences of urinary tract infections (UTI) and hospitalization, collected by the UK Dementia Research Institute between August 2019 and July 2021. Our multidimensional CMP model offers the best balance of recall vs. anomalies raised, with excellent generalisation. We discover that bathroom early AM activity (midnight to 6 am) is the prime cross‐patient digital biomarker of anomalies. This validates findings in literature that unusual bathroom activity is a clinically significant feature in UTI for dementia. We also demonstrate a cross‐patient view of anomaly patterns. Conclusion We address the need for anomaly detection and scoring using multivariate time series sensor data in remote health monitoring. The CMP allows configurability, ability to denoise and detect patterns, and explainability to clinical practitioners. With higher sensitivity, fewer alerts and better overall performance than state‐of‐the‐art methods, and the ability to discover digital biomarkers of anomalies, the CMP is a clinically meaningful unsupervised anomaly detection technique for dementia and beyond.