aeon: a toolkit for machine learning with time series

In recent years, machine learning frameworks such as scikit-learn have become essential infrastructure of modern data science. They have become the principal tool for practitioners and central components in scientific, commercial and industrial applications. But despite the ubiquity of time series data, until recently, no such framework exists for machine learning with time series. In 2019, sktime was conceived to fill this gap and it has become an established toolkit and software component for time series analysis used world-wide by academics and industry alike.
It is an easy-to-use, flexible and modular framework for a wide range of time series machine learning tasks. Techniques for learning from time series have been developed in a range of disciplines, including: statistics; machine learning; signal processing; econometrics; and finance. sktime aims to link these communities by providing a unified interface for related time series tasks such as forecasting, classification, clustering, regression, annotation, anomaly detection and segmentation. It provides scikit-learn compatible algorithms and gives easy access to implementations of state of the art algorithms not accessible in other packages. This project will allow sktime to continue to sustain and grow its operations by providing dedicated maintenance resource, enhancing the functionality and increasing engagement with scientific and industrial stakeholders. We wish to broaden the functionality of sktime to include new areas of active machine learning research and deepen our user base to reach new communities of researchers. Our aim is to link theory and practice by making it easier and faster for state of the art time series algorithms to be applied to real world problems of genuine scientific interest. To demonstrate this potential we will collaborate with domain experts on two applications. The first relates to predicting the early onset of dementia using electroencephalography (EEG). EEG are time series that record electrical activity in the brain using a series electrodes placed on the scalp. The equipment is relatively cheap and portable. If we could use it to screen for early onset dementia it could make a huge difference to the outcomes for many patients. However, the accuracy needed for clinical use is very hard to achieve. We will collaborate with experts in Cambridge who have clinical data and see if the state of the art predictive models can outperform traditional approaches. The second application involves analysing data generated from intensive care monitoring of children in Great Ormond Street Hospital (GOSH). Intensive care patients are continually monitored for vital body functions (heart rate, blood pressure, breathing rate, etc). Increasingly, this time series data is captured and can be mined to improve clinical practice. We will collaborate with a research team already working with GOSH to explore whether sktime can be used to decrease the time it takes to analyse this data.
This research may lead to insights that improve clinical practice by answering questions such as "when is the best time to remove the tube that is helping a patient breathe?". It will also help us reach our broader goal to speed up the discovery and dissemination of best practice. Data sharing between hospitals is, quite sensibly, difficult and time consuming. We wish to develop a new user base of hospital data scientists willing to share their research findings and code rather than their data. So, for example, if we discover something interesting in the GOSH data, we would like to rapidly share this finding and the code that verifies it in our data. This code sharing via sktime will dramatically reduce the time taken to test hypotheses on different observational data sets and give greater confidence in finding verified on independent groups of patients conducted transparently by different researchers.