471343 Emerging Methodologies for Environmental Exposure Assessment: Coupling Personal Sensor Data and Agent Based Modelling (ABM)
A trial multi-sensor campaign took place in Thessaloniki, Greece, where 30 participants carried a series of devices such as: (a) a temperature logger, to detect changes between indoor/outdoor conditions, (b) a fitness tracker, to capture motion and intensity of activity, (c) a GPS device, to track coordinates and speed along with (d) Moves, a smartphone application that enables tracking of location and activity. Additionally, a time – location – activity diary was filled out on paper each day. Aiming to substitute cumbersome time logs and to overcome the major drawback of the GPS-type sensors, that fail to differentiate whether an individual stay indoors, outdoors or in transit, the capability of predicting the type of location based solely on sensor data was explored using an Artificial Neural Network (ANN) model. A multi-layer perceptron network was utilized, which is a feed-forward artificial neural network model that maps sets of input data onto a set of appropriate output. Specifically, the independent variables that fed the ANN input layer consisted of (a) the personal temperature, Temp, derived from the wearable temperature sensor, (b) the observed outdoor temperature, Tempout, derived from a central meteorological station of Thessaloniki, (c) the ratio Temp/ Tempout, and (d) personal speed, derived from the GPS devices wore by the participants. Moreover, information on day light was also included as an input variable, transformed into a categorical element (day or night). The initial database was divided into training and validation set (85% and 15% of the total record entries, respectively) and the models developed from the training set were tested using the validation set. Results illustrate that the ANN model performs very well in predicting the various locations, especially the indoor ones, which also comprise the vast majority of registries of the training set, since most of the time of the daily activity is spent in indoor locations.
As a next step, using Monte Carlo analysis, distributions of participants’ movement and activities were extrapolated to a larger population. These data were used as input to a spatially explicit ABM computational platform in order to model trajectories of individual citizens. ABM is a simulation technique where a system -in this case a city- is modelled as a collection of individual heterogeneous actors, the agents. It is a stochastic approach that enables the observation and prediction of emerged behaviours. Every entity taking part in this system is agentified, it is therefore considered as an individual agent (the system is composed of roads, buildings, vehicles and human agents). Specifically, imported road and building shapefiles of a building block resolution contain spatial information on the capacity of a street and land use information (e.g.: type of building) respectively. The extrapolated observations and distributions on spatiotemporal behaviours, derived from the sensors campaign, together with population data from local authorities were transformed into coding lines that define and shape the agents’ world. Particular emphasis is also given in the case of in-model incorporation of socio-economic status (SES) data. SES variables (e.g., information on educational level, income, occupational status) can explain differences in external exposure because of different prevalence of specific preferences and decision-making in various groups of population. Different human agents based on different age, sex or income will follow different rules, will express different behaviours and this could lead to a different exposure profile. Moreover, knowledge of human agent characteristics by other human agents provide a signal that acts to enable or prevent interaction from occurring. Data on lifestyle/behaviour patterns (e.g., timetables for various activities per gender and age group) and SES information, derived from censuses in the city of Thessaloniki, were implemented in the human agents’ population. Survey outputs were associated with human agent behavioural rules, with the aim to model representative to real world conditions.
When the model is initialized, human agents are clustered in age groups, depending on their age and are randomly allocated to a residential place, which serves as their house for the whole simulation. Human agents’ characteristics provide capabilities or constraints on the agents’ behavioural rules. The reference point could be, for example, income that can influence their preferences and decision-making. Based on their age and SES characteristics, and of course based on the distance between point of departure and their targeted destination, human agents will choose different means of transportation. In the same way, different human agents will follow a different sequence and types of activities. For example, children and adults are programmed to move from a household to an assigned school or office whereas human agents that belong to the elderly will follow a different sequence of activities. At the end of a run (a single run corresponds to a typical day), human agents’ trajectories, derived by the coded routine, are captured as points and exported as a GIS shapefile, a layer which can then be superposed onto high spatial resolution urban air quality maps of hourly PM10 concentration of pollutants, for the city of Thessaloniki. Personal exposure, expressed as inhalation-adjusted exposure to air pollutants is then evaluated by assigning pollutant concentrations to a human agent based on his/her coordinates, activities and the corresponding inhalation rate.
By estimating the daily time-activity patterns (predicted by the coupled sensors-ANN-ABM platform) of vulnerable subgroups of population, we were able to estimate their personal exposure and intake dose per body weight. On average, personal exposure results were between 10 and 20% more accurate than the equivalent estimate using ambient air concentration of PM as exposure proxy.
An individual exposure model was developed, where societal dynamics are being explicitly taken into account. The model can feed into a population-based exposure assessment without imposing prior bias, but rather basing its estimations onto emerging properties of the agents’ system behaviour. The establishment of an approach with the capacity for aggregation and analysis at various levels of population size, together with the integration of SES indicators can lead to an exposure assessment model that can be useful especially for vulnerable groups of population, such as children, the elderly and people with low SES. The dynamic nature of intake dose assessment at the individual level allows for the derivation of guidance regarding behavioural options that limit exposure to high levels of pollution. This study represents the first step of the development of a methodological approach towards a refined external exposure assessment and urban air pollution and climate change mitigation management. The model can be further used as a means for estimating and comparing the probable effects of different public health strategies prior to implementation, therefore reducing the time and expense required to identify effective policies.