Published in AAAI Health Intelligence Workshop (pending publication in Computation Intelligence Springer Book Series), 2023
We use Electroencephalography (EEG) data to detect early stage Parkinson’s Disease. Firstly, we present a novel representation for EEG data, a 7-variate series of band power coefficients, which enables the use of (previously inaccessible) time series classification methods. Our approach achieves over 90% accuracy, recall and precision which rivals state of the art methods. This is particularly impressive given the early stage Parkinson’s Disease participants and limited optimisation of the model thus far. Secondly we present a framework for determining the importance of individual brain regions in Parkinson’s Disease classification. We find that across different EEG data types, it is the Prefrontal brain region that has the most predictive power for the presence of Parkinson’s Disease
We completed a submission to the AAAI-2023 Hackathon. Details of our submission will be made available after the competition has closed.
The SPHERE challenge presents an activity recognition task based on multimodal data collected from Smart Home sensors. While aiming to advance the safety of vulnerable people, the intrusive nature of the SPHERE technology renders transparency a particularly important feature of any SPHERE solution. Accordingly, current researchers present an interactive visualization dashboard that offers clear insights into the collected data, while giving healthcare professionals an opportunity to evaluate an algorithm’s predictions and assess sensor activity at the proximate time of an identified incident. Two phases of research are presented: 1) data modelling, 2) dashboard development. After curating multiple datasets that attempted to address key pre-processing challenges (e.g. pertaining to missing data points and an imbalanced class distribution), the data modelling phase trained and tested a series of machine learning models. Using Brier score as the selected evaluation metric, a multilayer perceptron, trained on a subset of selected features, was identified as the top performer with a score of .205. The dashboard development phase then visualised a sample recording sequence, displaying multiple data features for each second of the sequence. Visualizations included: bounding box positions, accelerometer data, the patient’s precise location within the smart home, and the predicted label distribution associated with the given timestep. The resulting dashboard offered insightful information, however there are potential improvements such as integrating a recent history of actions to give a clearer outline of data trends. Its efficacy can be further enhanced by optimising model performance, for example through a more sophisticated weighted re-sampling procedure and a greater focus on recall to emphasise minority classes.