-
Working PhD Project Title:
Exploring methods for localization and identification of behaviour changes indicative of Dementia in the CUBOId household
-
Academic Background
MSc Physics, Imperial College, 2011-2015
General Profile:
I graduated in 2015 with a first class Msci Physics degree from Imperial College London. I then worked at Goldman Sachs Investment Bank where I used statistical techniques to track performance and make predictions on several different project initiatives. I joined the Interactive Artificial Intelligence CDT in September 2020 and am still developing my research interests. I am currently interested in research at the intersection of AI / Neuroscience and Natural Language Processing / Computer Vision.
Research Project Summary:
Two broad research goals were addressed in the summer project, the first being to explore methods for accurate indoor localisation and the second to develop methods to identify changes in patient behaviour within a household. Synthetic data, SPHERE Challenge data and CUBOId data were used. Hidden Markov Models, KNN and Random Forest models were implemented across the three data sets showing that received signal strength (RSS) data alone could effectively be used for room level localisation. Accuracy scores of over 80 % were achieved across all data sets. Distance metrics to measure change in data distributions were defined on synthetic data, with difference in accuracy performance of a classifier was found to be most appropriate for use on real data. This distance metric was implemented on SPHERE Challenge data and CUBOId data from two homes with the aim of identifying changes in a dementia patients' behaviour over time. Results on CUBOId data were uninformative, limitations were acknowledged, and changes proposed for future work that could lead to improved results.
For the PhD some direct extensions to this distance analysis work will be explored. Data used will be extended from RSSI only to include accelerometer data with environmental sensors also being used to aid labelling on unsupervised daily living CUBOId data. Furthermore, engineering of additional features such as hour of day that would help identify behavioural changes will be explored.
Additional modelling techniques for localisation will also be explored in the PhD. With regards to accuracy of indoor localisation, a joint modelling approach should be explored for the CUBOId household. This is motivated by the hypothesis that there exist patterns between the locations of two individuals who live together in a house due to their interactions. One would therefore expect that the novel joint modelling approach would outperform existing methods if patterns between the two individuals' locations were successfully learnt. This joint approach will likely be explored via several modelling techniques that have performed well in the literature for localisation. Theoretically an LSTM network for example could be trained to simultaneously predict the location of both individuals taking the sensor data from both as features. Conditional Random Fields (CRFs) have also been shown to perform well on the localisation task. W.Yang , R. Poyiadzi et al. have previously implemented a CRF for localisation in the CUBOId household. Future work could extend this CRF to incorporate both participants jointly.
A knowledge driven approach could also be explored, extending previous work done by W.Yang, R. Poyiadzi et al. that showed incorporating knowledge of daily living into the model could improve performance on the localisation task within the home. Knowledge can range from knowing there is an increased probability to be in the bedroom at night to knowing that if someone is in the bathroom that there is a high probability that they are the only person in the bathroom. (people seldom use the bathroom together!)
Finally, the CUBOId project provides time series data collected from smart home sensors. The machine learning methods developed and trialled on this data set may also be applied to other domains that use time series data as a part of the PhD.
