Previously submitted to: JMIR Formative Research (no longer under consideration since Aug 08, 2018)
Date Submitted: Jul 18, 2017
(closed for review but you can still tweet)
Warning: This is an author submission that is not peer-reviewed or edited. Preprints - unless they show as "accepted" - should not be relied on to guide clinical practice or health-related behavior and should not be reported in news media as established information.
Automatic Health Assessment of Smart Home Residents via Unsupervised Learning from Ambient Sensor Data: An Observational Study
Background:
Existing approaches to health assessment for smart home residents mostly rely on the activities recognized by supervised learning methods, which require well-annotated sensor data. Obtaining such labeled data is not a feasible task, especially for senior adults.
Objective:
Developing an unsupervised learning method to health assessment for smart home residents without the need of labeled data.
Methods:
Ambient sensors are deployed in participants’ home to collect activity data. Home resident’s health is assessed via a two-step approach. In the first step, the clusters of sensor data are discovered using K-means or Gaussian Mixture Models (GMM). In the second step, the time spent on each data cluster is used to compute a health score that represents the physical health of the resident. As we could not obtain ground-truth labels, we conduct this project as an observational study.
Results:
We test the first step of our approach (sensor data clustering) on the Aruba dataset, which was acquired from a testbed occupied by a senior resident for more than a year. In this step, GMM achieves its best performance of 0.75 (????1_score) and 0.79 (Rand_index) and outperforms K-means. Qualitative results obtained on our smart home dataset also demonstrate the correlation between the discovered sensor data clusters with resident’s daily routine. The health score derived in the second step of our approach is highly associated with the health events happened to one of our participants.
Conclusions:
The experimental results demonstrate that the sensor data clusters discovered by standard clustering methods are highly correlated with the ground-truth activities or daily routine. Between the two tested methods, GMM outperforms K-means in all metrics. The health score derived by our two-step unsupervised approach is highly related to the health events happened to the resident.
ClinicalTrial:
The CSIRO Health and Medical Research Human Research Ethics Committee (#12/17).
Citation
