Skip to main content

Activity-driven detection of cognitive impairment using deep learning.

Arifoglu, D., 2019. Activity-driven detection of cognitive impairment using deep learning. Doctoral Thesis (Doctoral). Bournemouth University.

Full text available as:

[img] PDF
ARIFOGLU, Damla_Ph.D._2019.pdf
Restricted to Repository staff only until 1 October 2021.



While life expectancy is on the rise all over the world, more people face health related problems such as cognitive decline. Cognitive impairment is a collective name for progressive brain syndromes which affect memory, cognition, behaviour and emotion. People suffering from cognitive impairment may lose their abilities to perform daily life activities and they get dependent on their caregivers. Although some medications can slow the progress of the disease, currently there is no way to stop its development. Sufferers may require special needs which increase the cost of care. Thus, detecting the indicators of cognitive decline before it gets worse would be very crucial. Current assessment methods mostly rely on queries from questionnaires or in-person examinations, which depend on recall of events that may poorly represent a person’s typical state. The aim in this thesis is to adapt deep learning techniques for analysing daily activities of elderly people and detecting abnormalities in the activity patterns. Recent studies suggest that indicators of cognitive decline can be observed in daily life activity patterns. The spatio-temporal and hierarchical relationship of activities and their intrinsic structures are important in the context of cognitive decline analysis. Existing studies treat each activity as an atomic unit and fail to capture the relationship among sub-activities. Also, existing studies rely on fixed length features to model activities, ignoring the granular level information coming from raw sensor activations. Moreover, there exists no daily activity dataset representing the behaviour of dementia sufferers because producing such datasets requires time and adequate experimental environment. Given these challenges, the present thesis addresses the following research questions: How can we cope with the scarcity of dataset reflecting on cognitive status of elderly people? How can activities be modelled taking into account their spatio-temporal neighbourhood and hierarchical information? How can we represent raw data to encode the granular level details? These research questions are addressed in the following way. Firstly, two methods are proposed to cope with the scarcity of data: (i) synthetic data generation and (ii) transfer learning adoption. Secondly, the activity recognition problem is emulated (i) as a sequence labelling problem to model spatio-temporal patterns. (ii) as a hierarchical learning problem to model sub-activities. (iii) as a graph labelling problem to encode granular level details. Thirdly, raw sensor measurements stemming from sequential data are used to model sensor activation relationships. The proposed methods are also compared against the state-of-art methods. The preliminary results obtained indicate that pro- posed data simulation and transfer learning approaches are useful to cope with the scarcity of data reflecting cognitive status of elderly people. Moreover, experiments show that the proposed deep learning methods are promising to detect abnormalities in the context of cognitive decline. Proposed methods are not only promising to detect abnormal behaviour at a fine-grained level, but some of them can also model activities hierarchically by taking sub-activities into account and then can detect abnormal behaviour occurring at granular levels.

Item Type:Thesis (Doctoral)
Additional Information:If you feel that this work infringes your copyright please contact the BURO Manager.
Uncontrolled Keywords:sensor-based activity recognition; abnormal behaviour detection; dementia; elderly care; deep learning
Group:Faculty of Science & Technology
ID Code:32750
Deposited By: Unnamed user with email symplectic@symplectic
Deposited On:10 Sep 2019 09:06
Last Modified:10 Sep 2019 09:06


Downloads per month over past year

More statistics for this item...
Repository Staff Only -