through wearable devices and other sensing technologies, geo-located online surveys,
open/public smart city datasets…), on individual and collective (population/cluster)
levels. Overall well-being and its main domains (vitality, supportive relationships,
stress levels…) are all significant factors affecting the onset and exacerbation of the
stated chronic diseases which are becoming more and more widespread and progressing in urban environments, and overall resilience of citizens and urban communities is increasingly important against other pertaining global and sustainability
challenges, like climate change.
The proposed and deployed elementary statistical model presented in this paper is
to be the basis for interpretation and contextualization of changes to well-being, and a
performance benchmark for evaluation and comparison of more complex and advanced
well-being models of the aimed predictive analytics and final intelligent system (incorporating machine learning methods) in subsequent development, supporting the
PULSE PHOs (Public Health Observatories established for the relevant policy making
and execution in smart cities).
2 Conceptual Background and Approach
The activity and vital/health parameters data measured mostly unobtrusively by
wearable devices (wristbands, smartwatches) have particular significance for behaviour
analysis and change recognition in PULSE, as these are the input data streams with
highest volume, acquisition “velocity”, and temporal resolution/granularity of all the
various data collected in the Project, and therefore practically the most (and only)
suitable data comprising the sufficiently continuous and non-sparse time series over
months, to properly derive or construct the behavioural patterns and analyze behaviour
changes. Recent studies performed by the stated major wearable device manufacturers
over billions of records of temporal measurements data [1, 2], as well as the experiences from projects like the just concluded City4Age (www.city4ageproject.eu) [3, 4],
show the significance and general predictive ability of the measured main vital/health
and activity parameters (walking, climbing stairs, physical activity/exercise, heart rate
data, consumed calories…) for overall health and physiological/physical well-being
assessment. The additional complementary socio-demographic, health, lifestyle/habits
and environmental data in less frequent temporal resolution, ingested from the
open/public datasets or manual “obtrusive” inputs, are combined to cross-check, adjust
and improve integrity of the recognized behaviour changes derived from the main timeseries data acquired through the wearable devices.
We adopt a combined knowledge- and data-driven approach in detection and
characterization of relevant behaviours that denote significant variations in well-being,
with multi-level hierarchical model topology and range/threshold based computational
rules as basic primary formal knowledge structures, and statistical analytics as baseline
(and performant) data-driven detection methods.
The complexity of human behaviours is commonly represented through multi-level
hierarchical structured models, decomposed to more granular “units” like activities and
action events [5, 6], with multiple variables from behavioural, physiological and
environmental domains of well-being known to additionally increase complexity and
156
V. Urošević et al.
open/public smart city datasets…), on individual and collective (population/cluster)
levels. Overall well-being and its main domains (vitality, supportive relationships,
stress levels…) are all significant factors affecting the onset and exacerbation of the
stated chronic diseases which are becoming more and more widespread and progressing in urban environments, and overall resilience of citizens and urban communities is increasingly important against other pertaining global and sustainability
challenges, like climate change.
The proposed and deployed elementary statistical model presented in this paper is
to be the basis for interpretation and contextualization of changes to well-being, and a
performance benchmark for evaluation and comparison of more complex and advanced
well-being models of the aimed predictive analytics and final intelligent system (incorporating machine learning methods) in subsequent development, supporting the
PULSE PHOs (Public Health Observatories established for the relevant policy making
and execution in smart cities).
2 Conceptual Background and Approach
The activity and vital/health parameters data measured mostly unobtrusively by
wearable devices (wristbands, smartwatches) have particular significance for behaviour
analysis and change recognition in PULSE, as these are the input data streams with
highest volume, acquisition “velocity”, and temporal resolution/granularity of all the
various data collected in the Project, and therefore practically the most (and only)
suitable data comprising the sufficiently continuous and non-sparse time series over
months, to properly derive or construct the behavioural patterns and analyze behaviour
changes. Recent studies performed by the stated major wearable device manufacturers
over billions of records of temporal measurements data [1, 2], as well as the experiences from projects like the just concluded City4Age (www.city4ageproject.eu) [3, 4],
show the significance and general predictive ability of the measured main vital/health
and activity parameters (walking, climbing stairs, physical activity/exercise, heart rate
data, consumed calories…) for overall health and physiological/physical well-being
assessment. The additional complementary socio-demographic, health, lifestyle/habits
and environmental data in less frequent temporal resolution, ingested from the
open/public datasets or manual “obtrusive” inputs, are combined to cross-check, adjust
and improve integrity of the recognized behaviour changes derived from the main timeseries data acquired through the wearable devices.
We adopt a combined knowledge- and data-driven approach in detection and
characterization of relevant behaviours that denote significant variations in well-being,
with multi-level hierarchical model topology and range/threshold based computational
rules as basic primary formal knowledge structures, and statistical analytics as baseline
(and performant) data-driven detection methods.
The complexity of human behaviours is commonly represented through multi-level
hierarchical structured models, decomposed to more granular “units” like activities and
action events [5, 6], with multiple variables from behavioural, physiological and
environmental domains of well-being known to additionally increase complexity and
156
V. Urošević et al.