TIHM: An open dataset for remote healthcare monitoring in dementia

Abstract

Dementia is a progressive condition that affects cognitive and functional abilities. There is a need for reliable and continuous health monitoring of People Living with Dementia (PLWD) to improve their quality of life and support their independent living. Healthcare services often focus on addressing and treating already established health conditions that affect PLWD. Managing these conditions continuously can inform better decision-making earlier for higher-quality care management for PLWD. The Technology Integrated Health Management (TIHM) project developed a new digital platform to routinely collect longitudinal, observational, and measurement data, within the home and apply machine learning and analytical models for the detection and prediction of adverse health events affecting the well-being of PLWD. This work describes the TIHM dataset collected during the second phase (i.e., feasibility study) of the TIHM project. The data was collected from homes of 56 PLWD and associated with events and clinical observations (daily activity, physiological monitoring, and labels for health-related conditions). The study recorded an average of 50 days of data per participant, totalling 2803 days.

Fig. 1

Demonstration of a residential setting equipped with PIR sensors for in-home activity monitoring and other sensors for sleep and physiology monitoring in the TIHM project. PIR and door sensors are included in each room of the house. An under-the-mattress sensor is used for sleep and in-out-of-bed monitoring. Connected devices which are operated manually are also used in the setting to acquire physiology data.

Fig. 2

(a) An overview of the number of movements per location, per day. The bar chart on the right shows the total number of movements that occurred at each location over the given period. The large drop on the 14th of June 2019 is caused by a technical failure in the data collection server. (b) An overview of the total number of participants joining the study within the timeline, the increasing trend of activities (a) corresponds to an increase in the number of households and participants recruited (b). On average, each participant was involved in the study for about 50 days.

Fig. 3

(a) In-home activities of PLWD with frequent episodes of neuropsychiatric symptoms. Activity patterns are irregular and there are no consistent habitual patterns of daily activities. (b) In-home activities of a PLWD with no neuropsychiatric symptoms. Habitual patterns are identifiable, and activity can be inferred.

Fig. 4

Visualisation of multi-source data for a participant. It is shown the daily physiology data (blood pressure, body weight, temperature) for a participant, aligned with the alerts generated in the dataset.

Fig. 5

Performance of baseline models for classifying Agitation alerts using daily activity and physiology information. (a) Demonstration of cross-validation in the experiments with a 5-fold cross-validation as demonstrated and each test set consists of data from a 7-day period. (b) Average sensitivity and specificity of the baseline models across 5-fold cross-validation, with error bars indicating the standard deviation for each model.

Fig. 6

The feature importance learned by the Logistic Regression model. The SHAP value of each feature represents its impact on the model output regarding a given input. The violin plot illustrates the distribution of SHAP values for each feature, which are estimated by test samples during the cross-validation. The colour indicates whether the raw value of a feature is high or low.