Speaker: Presentation: Date: Session: Time: Place: Joss Langford Event based classification of human behaviour Tuesday 16th September 2014 (4) Business Applications of R 1330 ‐ 1500 hrs Tower Suite 3 Event based classification of human behaviour: strategic commercial use of open source tools GENEActiv • Activinsights specialises in the science of behavioural measurement with the aim of benefiting people’s overall health • GENEActiv wrist-worn accelerometers are the device of choice for studies at many of the world’s leading universities, hospitals and professional sports bodies • The original open raw data accelerometer – giving uncompressed data in SI units © Activinsights Limited 2014 Strategic – Commercial - Open Background: • GENEActiv is a general purpose instrument – specialisation comes at the software & analysis stage • Our customers often produce large data sets – up to 10,000 subjects at 0.5Gb each • GENEActiv is largely used in research (academic & commercial) Open source tools allow us to: • Support a huge range of applications without significant business risk and overhead • Collaborate with academics and other commercial partners • Provide referenced open releases of tools via CRAN © Activinsights Limited 2014 Impact Many studies and most large national studies now use raw data in SI units – often GENEActiv & GENEAread Examples: • General behaviour • Sports • Sleep • Visualisation • Algorithms American Journal of Epidemiology AssociationSleep Between Questionnaire- and Journal Accelerometer-Assessed Physical Activity: The Role of Estimates Sociodemographic Factors Sleep Using Journal of Applied Physiology Microelectromechanical Systems (MEMS) Auto-calibration of accelerometer data CRAN: for GGIR free-living physical activity assessment using local gravity and temperature: an evaluation on four continents CRAN: GENEAsphere CRAN: GGIR © Activinsights Limited 2014 Event-based classification • While we have been successful moving the industry to SI units – we still have work to do in the area of fixed epoch summaries • Large time-series data sets are not intuitive for human handling • Nearly all our customers use fixed epoch summaries – but variable length event-based summaries have many advantages • Is it a classification or a changepoint problem? • There is a natural granularity of everyday life © Activinsights Limited 2014 Coelition • Coelition is an open standards organisation that is developing a standard to help classify everyday living • The Classification of Everyday Living (COEL) is a hierarchical taxonomy with 32 top level clusters - currently coding over 7,000 behavioural events • The hierarchical structure allows information from different sources to be recorded in the same framework © Activinsights Limited 2014 Raw data view of a morning Accelerometer © Activinsights Limited 2014 Event-based view Accelerometer Diary child woke during night got up sleep Simple phone app bought coffee drove to station train to London phone screen on location 1 location 2 location 3 © Activinsights Limited 2014 Future challenges • Using ‘GENEAclassify’ as a starting point, build our abilities to intelligently analyse raw data accelerometry • Develop better deployment tools e.g. file handling, parallelisation and interfacing • Develop new tools for the analysis of event-based data e.g. sequence analysis © Activinsights Limited 2014 Acknowledgements & information Zhou Fang, Biomathematics & Statistics Scotland Séverine Sabia, University College London Vincent van Hees, Newcastle University John Bardon, University of Regina Alex Rowlands, University of South Australia Matt Reed, Coelition joss@activinsights.co.uk © Activinsights Limited 2014