Research Article
Classifying persons with dementia from control subjects when ascending and descending stairs based on a single pelvis-mounted sensor
@INPROCEEDINGS{10.4108/eai.16-5-2016.2263971, author={Catherine Holloway and Ian McCarthy and Tatsuto Suzuki and Keir Yong and Biao Yang and Amelia Carton and Nadia Bianchi-Berthouze and Nick Tyler and Sebastian Crutch}, title={Classifying persons with dementia from control subjects when ascending and descending stairs based on a single pelvis-mounted sensor }, proceedings={Pervasive Health Workshop on Affective Interaction with Virtual Assistants within the Healthcare Context'}, publisher={ACM}, proceedings_a={AIVAHC2016}, year={2016}, month={6}, keywords={dementia pattern recognition in the wild}, doi={10.4108/eai.16-5-2016.2263971} }- Catherine Holloway
Ian McCarthy
Tatsuto Suzuki
Keir Yong
Biao Yang
Amelia Carton
Nadia Bianchi-Berthouze
Nick Tyler
Sebastian Crutch
Year: 2016
Classifying persons with dementia from control subjects when ascending and descending stairs based on a single pelvis-mounted sensor
AIVAHC2016
EAI
DOI: 10.4108/eai.16-5-2016.2263971
Abstract
As part of a larger program of work to understand how people with dementia navigate their environment and use visual cues we present data which uses a single sensor – an IMU placed on the pelvis – to classify people into two groups on the basis of hesitancy when ascending/descending stairs: individuals with dementia vs age-matched controls. The classification was conducted on data collected from 34 people (14 controls; 20 people with dementia, comprising 10 with typical Alzheimer’s disease [tAD] and 10 with posterior cortical atrophy [PCA]) walking up a set of 4 steps. Attributes used to discriminate those with and without dementia were the mean and root mean square values of: resultant acceleration, roll, pitch and yaw. Each person’s data was allocated to one of two datasets (N=17, N=17). A weighted nearest neighbor classifier was trained on each dataset in turn and subsequently used on the remaining dataset. Overall accuracy of the classifier was 0.67, with a precision of 0.62 and recall of 0.47.
