Occupancy and Daily Activity Event Modelling in Smart Homes for Older Adults with Mild Cognitive Impairment or Dementia Flávia Dias Casagrande Evi Zouganeli Department of Electronic Engineering, OsloMet - Oslo Metropolitan University, Norway. {flavia.casagrande,evi.zouganeli}@oslomet.no Abstract In this paper we present event anticipation and prediction of sensor data in a smart home environment with a limited number of sensors. Data is collected from a real home with one resident. We apply two state-of-the-art Markov- based prediction algorithms - Active LeZi and SPEED - and analyse their performance with respect to a number of parameters, including the size of the training and testing set, the size of the prediction window, and the number of sensors. The model is built based on a training dataset and subsequently tested on a separate test dataset. An accu- racy of 75% is achieved when using SPEED while 53% is achieved when using Active LeZi. Keywords: smart home, prediction models, sensor data, occupancy modelling, event modelling 1 Introduction We present results from the Assisted Living project, an interdisciplinary project that aims to develop assisted liv- ing technology (ALT) to support older adults with mild cognitive impairment or dementia (MCI/D) live a safe and independent life at home. The project is carried out by experts in the field of nursing and occupational therapy, ethics, and technology (Zouganeli et al., 2017). MCI and dementia involve cognitive decline, which can affect at- tention, concentration, memory, comprehension, reason- ing, and problem solving. Smart homes can potentially include a number of intelligent functions that can pro- vide valuable support to older adults with MCI/D, such as prompting support e.g. in order to assist or encourage, diagnosis support tools, as well as prediction, anticipation and prevention of hazardous situations. Activity recogni- tion and prediction is a prerequisite and a necessary tool for achieving the majority of these. We present our first results on prediction of binary sen- sor data in a smart home environment. Several algorithms have been reported in the literature for this purpose. How- ever, to the extent of our knowledge, such prediction algo- rithms have not yet been tested in a real home, nor have they been proven to be accurate enough to be implemented in real homes. In addition, there is no comprehensive study comparing the different available algorithms or pro- viding guidelines as to which application areas they are best suited for. In this paper we apply two algorithms on data from a real home, compare their performance, and shed some light regarding their application areas. 2 Related Work Data prediction algorithms have been extensively re- searched on in the literature (Wu et al., 2017). Event or activity prediction can for example lead to an improved operation of automation functions (e.g. turn on the heater sufficient time prior to the person arriving at home); facili- tate useful prompting systems (e.g. prompt the resident in case the predicted next activity is not performed) (Holder and Cook, 2013); or detect changes/ anomalies in certain behaviour patterns (e.g. movement, everyday habits, etc.) and hence assist to indicate the onset or the progress of a condition (Riboni et al., 2016). The Active LeZi (ALZ) algorithm has been extensively applied for prediction on sequential data (Gopalratnam and Cook, 2007). The algo- rithm was tested on the Mavlab testbed dataset and was shown to achieve a 47% accuracy. Some of the ideas of ALZ have been used in the implementation of a new algo- rithm, the sequence prediction via enhanced episode dis- covery (SPEED)(Alam et al., 2012). SPEED was tested on the same dataset as ALZ and achieved an accuracy of 88.3% when the same dataset was used both for training and for testing. These algorithms are based on Markov models, where at any given point in time the next state depends solely on the previous one (Rabiner and Juang, 1986). Hence, the most probable next event can be esti- mated based on the current state. Besides probabilistic algorithms, neural networks have also been used for event prediction. A root square mean error (RMSE) of 0.05 using Echo State Network (ESN) and Non-linear Autoregressive Network (NARX) was re- ported by using a number of input/output configurations (Lotfi et al., 2012; Mahmoud et al., 2013). Other relevant research includes prediction of the time when a certain ac- tivity will happen using decision trees (Minor and Cook, 2016) or time series (Moutacalli et al., 2015). Prediction of the next activity as well as the time, location, and day it would occur has also been reported (Nazerfard and Cook, 2015). In this paper, we use the Active LeZi and SPEED algo- rithms for the prediction of the next sensor to be activat- ed/ deactivated in an event sequence obtained from a real home with one resident. https://doi.org/10.3384/ecp18153236 236 Proceedings of The 59th Conference on Simulation and Modelling (SIMS 59), 26-28 September 2018, Oslo Metropolitan University, Norway