Novelty detection algorithms for behavioural analysis in healthcare

One of the key challenges in digital health is the understanding of human behaviour. To this end, using activity recognition from home sensors based on machine learning has been proven successful in recent years. The standard features computed from raw data using signal processing techniques include both time domain features & frequency domain features, as well as features based on expert knowledge. These activities (e.g.sleeping, exercising, cleaning) provide a good proxy for the characterisation of the patient's lifestyle over time. In many of these cases, the objective will be not just to recognise a particular behaviour but the detection of anomalous behaviour, in which case the deviation from standard behavioural patterns is of interest.
The aim of this project is to automatically identify changes in the lifestyle that could be used as evidence of the presence or progress of a disease. Many aspects of healthcare make the application of out-of-the-box novelty detection techniques challenging. The algorithms aim to distinguish between the positive class
& everything else. The definition of the region that covers all possible normal behaviours is difficult. Also, the boundary separating the positive instances from anomalies is imprecise & thus observations that lie close to it can be easily misclassified. Moreover, it is relatively easy to compile a dataset with patients with a common disease, but this is not the same for the negative class. Patients cannot be assumed to be negative instances unless they have been tested. These are the motivation for the 4 phases of the PhD summarised below. In the novelty detection literature changes are assumed to be abrupt and sparse, and (the observed time series is often univariate or of fixed dimensions. In contrast, healthcare data calls for new novelty detection methodologies, which are able to detect gradual and smooth changes, handle multiple heterogeneous data streams, & perform online analysis as the monitoring of the patients is on-going, without causing
any false alarm. This will also require the design of new retrieval models adapted to the particular signal characteristics, e.g. uneven-sampling, missing data, non-Gaussianity, presence of non-stationary processes, or non-i.i.d. relations. Additionally, a 2nd challenge of this project will be to incorporate an additional level
of abstraction to analyse complex sequences of activities and the interplay between them in a rigorous manner. These activities are usually organised in a hierarchical fashion, from basic posture/ambulation to Activities of Daily Living. Thus, new hierarchical novelty detection algorithms that account for this inherent structure will be designed in such a way that patterns can be categorised as novelties at a given level of the hierarchy, not necessarily at all of them. In 3rd place, the project will need to advance algorithms that derive meaning from the time series of activities, moving from a simple detection of individual activities to a proper understanding of sequences of activities. These will be specifically tailored to the behavioural patterns of interest to clinicians as not all novelties may be as informative or useful for each clinician. In order to account for this, the project will incorporate recent results from exploratory data analysis, where users are typically interested in visualisations that highlight surprising information & patterns. That is, users are interested in finding patterns that complement or contradict their prior expectations, rather than those that confirm them. The project will study methods that take into account the clinician's prior expectations in order to return the most relevant novelties. Finally, a key property of digital health's data is the large-scale nature of the data. In this case, data management becomes more challenging due to the increasing availability of new data sources.