Sleep Apnea Detection in Fog Based Ambient Assisted Living System Ace Dimitrievskia, Natasa Koceskab, Eftim Zdravevski a, Petre Lameskia, Betim Cicoc, Saso Koceskib and Vladimir Trajkovika a Faculty of Computer Science and Engineering, Ss.Cyril and Methodius University, Skopje, N. Macedonia b Faculty of Computer Science, University Goce Delcev Stip, N.Macedonia c MetropolitanTirana University, Tirana, Albania Abstract Ambient Assisted Living environments use different sensors and actuators to enable their end- users to live in their preferred environments. Unlike smart homes, where a target audience is usually a family unit, standard Ambient Assisted Living end users are care receivers and care providers. This article describes an approach based on the fog computing paradigm to detect sleep apnea in an Ambient Assisted Living context unobtrusively. The edge nodes process and detect local activities of daily living events and have direct control of the local environment. The fog nodes are used to further process and transmit data. The cloud is used for more complex and anonymous data computation. This research shows that sensors, which are unobtrusive and do not interfere with users' daily routines, can be successfully used for pattern observation. Keywords Ambient Assisted Living (AAL), Fog computing, Cloud computing, Personal health care. 1. Introduction obtained with unobtrusive sensing can give a more detailed picture of the care receivers' health and personal habits. In that way, Advancements in cloud computing and the technology directly impacts elderly and Internet of Things (IoT) have had a positive disabled people’s ability to remain at home and impact on pervasive computing and can live more independent lives [3][4]. AAL is also improve Ambient Assisted Living (AAL) addressing the growing cost of traditional solutions. Fog computing is a newer discipline health care. Advances in AAL's research that brings an opportunity to fill in some gaps provide tools and methods for improving the and improve many aspects of cloud-based AAL health of the elderly and people with systems, mainly by increasing user privacy if disabilities. On the other hand, Enhanced used correctly [1], [2]. Living Environment (ELE) is a field that Technology for monitoring, assisting, and provides resources for personal health for the improving personal health has improved general population. Although AAL and ELE considerably with affordable wearable and address different target audiences, both fields unobtrusive sensors, cloud computing, and benefit from similar technology [5]. improved Internet connectivity. The presence A typical AAL goal is to enable care and rapid growth of the Internet of Things (IoT) providers to have technology-enabled paradigm has also impacted how people continuous monitoring of care receivers. It monitor their health. Most current wearable reduces care costs on the one hand and devices can monitor heart rate and physical increases care efficiency on the other hand activity. More appliances come with Internet [6][7]. Cloud paradigm fits well for this connection capability, and smart sensors are scenario as data can be aggregated and analyzed becoming increasingly common. The data in a centralized location. An interface for care Proceedings of RTA-CSIT 2021, May 2021, Tirana, Albania EMAIL: ace.dimitrievski@gmail.com (A. 1); natasa.koceska@ugd.edu.mk (A. 2); eftim.zdravevski@finki.ukim.mk (A. 3) petre.lameski@finki.ukim.mk (A. 4) bcico@umt.edu.al (A. 5) saso.koceski@ugd.edu.mk (A. 6) trvlado@finki.ukim.mk (A. 7 © 2021 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). CEUR Wor Pr ks hop oceedi ngs ht I tp: // ceur - SSN1613- ws .or 0073 g CEUR Workshop Proceedings (CEUR-WS.org) providers can be provided from the cloud using bandwidth requirement and the need for the web and mobile devices. Network reliability real-time cloud communication. and demand for real-time processing of risk • Cloud downtime or connectivity issues factors and different privacy concerns require can be a problem in the case of AAL [12]. some local data processing. Fog computing While many large cloud providers have addresses these problems by its definition. multiple availability zones, the cost of A single device, individual, or group of having high availability of the cloud is unobtrusive sensors present in the ELE can higher. Edge nodes can more easily be provide input on a limited set of health aspects. clustered, allowing for the high availability Smartwatches and health trackers can track of fog computing. body temperature, heart rate, walking or In this paper, we identify multiple benefits running; environmental sensors can detect of fog computing in the typical AAL scenario temperature, humidity, fall detection, and and propose an architecture that would make movement within the home. A more holistic them possible. The AAL fog-based architecture picture of these devices and sensors can be is described in Section 2 of this paper. The provided if connected to the cloud, where all the proposed architecture benefits are illustrated data is analyzed for more robust data processing with the experiment presented in Section 3, and techniques. By using data from many users, Section 4 concludes the paper. machine learning (ML) algorithms can learn and predict health hazards and find correlations 2. Fog based AAL architecture between the environment and human health [8][9]. While IoT cloud-based computing benefits Fog computing adds a layer to the cloud are visible both in research and daily use, there computing architecture. However, it should not are many drawbacks when it comes to personal be interpreted as an extension of the cloud. Fog health care data clouds. The most significant computing spans to adjacent physical locations. ones are the following: It supports online analytics and various • The lack of security of IoT devices and communications networks in performing distributed computing [13]. There are four companies' un-proper practices that gather logical layers of Fog computing. and abuse personal data have made consumers more proactive in protecting Data is generated by sensors that can be their data [10]. There is a potential of wearable or body sensors and peripheral or environmental sensors on the first layer. Data targeted advertisement to identify personal health details, with a possibility that future can also be generated from external sources such are: social networks, clinical center employers could refuse potential employees because of their health risks or personal information systems, or medical databases. habits. Insurance companies can purchase Data collected by the sensors can include vital signs, personal habits, or environmental factors. personal data and use it to deny coverage or increase premiums. Protections against External data sources can provide different these practices vary and can be loose in information, including medical check results, some jurisdictions. Even when such medical databases for diagnostics, and similar. protections exist, the legal expenses can be The fog layer gathers the sensor data, high, and the case can be challenging to processes them, and passes either processed or prove. Fog computing can have a role in data portions of raw data to the cloud. The devices directly connected to the sensors are called edge protection by moving some data analysis to the edge nodes and anonymizing the cloud's nodes. Aside from collecting data, they can take data. action with the user. Each LAN environment can have one or more edge nodes, depending on • Personal healthcare and AAL systems the application requirements and scale. In can generate a significant quantity of data elderly care facilities, data for multiple tenants [11]. Some ALE scenarios, such as fall could be processed on the same edge nodes. detection, require having an immediate These actions can include providing feedback reaction of the system by triggering an alarm to the person to take their medicine or to start to the care provider. Data pre-processing on exercising. They can directly interact with the edge nodes can significantly reduce the environment, such as: activating the humidifier or regulating the room temperature, controlling support several features such as acting as a electrical appliances, and cutting-off for water, repository to temporarily store sensors’ and gas, and electricity in case of an emergency. users’ information and bring intelligence by The fog network usually has a more limited enhancing data fusion, aggregation, and capacity than the cloud for data computation interpretation techniques. It is essential to and cannot do complex machine learning and provide preliminary local processing of sensors' feature extraction. data, which is the primary role of a smart e- However, fog nodes could be able to run health gateway. Smart e-health gateway can algorithms developed by machine learning. As tackle many challenges in ubiquitous healthcare the machine learning system improved and systems such as energy efficiency, scalability, evolved, regular updates could be pushed down interoperability, and reliability issues [17]. to the fog network to improve sensor data Due to the privacy concerns and the patterns. Using this methodology, ADL technical aspects for scalability and detecting ML could receive continuous data interoperability, it is crucial to identify and and improve the detection rate. Events that take trace the data flow in the system. Sensor data a brief time, such as when a person falls, can be originates when sensors acquire measurement detected by the fog nodes using the latest ML from the physical world. This measurement is model improved in the cloud. represented by an electrical signal transferred to The cloud layer assembles and processes the a controller that would interpret the signal. data from multiple sources and creates machine Some sensors are manufactured to include the learning models. The feature extraction is done electronic circuits to digitalize the reading, and at this layer as well. Data from the fog and some are even Internet-connected, enabling external sources is collected and processed by them to upload the data to a remote system the data fusion component [14]. The output is directly. The sensor data is then passed to the an improved knowledge base. The service layer local processing nodes. These nodes are part of uses this knowledge base in turn. the fog and can communicate to other layers of The service layer is the product of the the fog. The data on these edge nodes is system. Knowledge obtained by analyzing the processed for local events detection. data is used for services, including creating Only the edge nodes or smart e-health customized recommendations for diet and gateways should be able to get unfiltered raw exercise, improving diagnostics systems, sensor data. The data that is passed on to other providing updates to the health providers, and layers of the fog is pre-processed [18]. From adding additional information in medical this point, the data can be split into multiple databases. processing paths depending on the desired The critical features that should be satisfied function. Data with person-identifying by the system include security, privacy, high properties can only be passed to the fog areas availability, and interoperability. Security and used for healthcare provider usage in a privacy [15] can be addressed by implementing compliant way with local regulations for best practices to protect the network and the handling medical data. Data used for science data. Redundancy and automatic fail-over are research can also contain medical data, but needed to provide high availability, primarily personal identifiers should be stripped or when the health care recipient’s life depends on hashed. Other service types might require the assisted living system. The increased aggregated data that does not expose the user’s complexity requires ensuring connected and medical conditions. It, for example, can include inter-operable components by using the average time spent outdoors. Such data can frameworks intended to ensure mutual be correlated with local weather to determine compatibility [16]. the best time to organize group activities for the In fog computing, the nodes nearest to the community's senior members. Some data might devices are named edge nodes. In healthcare be of the type that the person would like to share systems, these nodes represent smart e-health on social media or other platforms. It might gateways. They act as a bridge for medical include exercise data such as walking, hiking, sensors to cloud computing platforms. The or riding a bike. main requirement of a gateway is to support Each of the services dealing with user data various wireless protocols and inter-device is logically independent and can be hosted on communication. Its role can be extended to separate cloud platforms. The health provider service is independent of social media or information will not cause significant medical research databases. The separation of improvements in the algorithms [20]. the cloud can be implemented by separation on • Data encryption is used to protect data any level in the fog network. As the data is as it passes through the network. Data passed between layers of the fog network, encryption can be full or partial. For several processing types can occur. Data example, a gateway node would encrypt processing tasks mostly would take place on the sensor readings and meta-data of the person. edge nodes or smart e-health gateways as the However, the personal information would gateways would directly interface with the be encrypted so that only the healthcare sensor network and receive raw sensor data. At provider’s network would have the this layer, we can identify the following types decryption key. The sensor data would be of tasks: encrypted so the fog nodes would decrypt Data filtering is used to filter noise, invalid and, without person-identifying meta-data, sensor readings, and redundant information that pass it onto cloud instances to do statistical does not contribute to the desired information analysis or machine learning. This method the system should induce. Sensor data contains will reduce data duplication in the network valuable information. However, they also carry as the same information will not have to be non-deterministic errors such as motion transmitted twice from the gateway to artifacts, data corruption issues, and unwanted different fog nodes. signals that are also significantly uploaded to • Error code correction can be used to increasing storage requirements and power ensure validity during transmission. The fog consumption. Fog computing could play an network can rely on various data essential role in increasing efficiency and transmission techniques and technologies to reduce storage requirements for medical big pass on the information; sometimes, the data solutions [19]. network protocol would have a built-in Anonymizing of data strips or replacing feature to ensure valid transmission. When person-identifying information from data this is not the case, the fog nodes would have packets. When there is a requirement to to ensure the data's validity by identifying separate patient/customer data, personal and correcting transmission errors. The information is replaced with unique identifiers. same applies to the data traveling from the This data can be passed on to the fog nodes for sensors to the gateway, as many sensors do added security using an enterprise service bus not have a buffer memory and cannot re- (ESB). transmit data. Error code correction will be Data fusion automatically transforms used to identify faulty readings and discard information from different sources and points them (because having gaps in the data is into a representation that provides practical often better than having inaccurate data). support for automated decision-making. Applying data fusion in gateways provides several advantages: reduced data ambiguity, 3. Experiment extended coverage in space and time, robustness and reliability, and increased data A common usage of sensor networks is to quality. After data is fused, only final results are train machine learning models and enable transmitted through the network so the network different end-user actions. Depending on the bandwidth can be efficiently utilized, and the number of sensors used, the number of features system can be more energy-efficient [15]. extracted from the sensor data, and the data Data processing that can be done on any generation rate, generating the model will most layer of the fog network includes: likely be done in the cloud due to the resources • Data compression is used to reduce the demand and a potential need to use data from amount of bandwidth required to transmit other locations. On the other hand, the the sensor network's information. implementation of the model can and should be Compression can be lossy or lossless. Lossy done on edge. As an example, we will consider compression can be acceptable in many a data flow model to detect sleep apnea using cases, especially if the sensor data's noninvasive sensors, illustrated in Figure 1. resolution is too high. Besides, the extra Figure 1: Data processing for unobtrusive sleep apnea detection The sensor readings from multiple care The first phase is to pre-process the data by recipients are collected. In scenarios of multiple identifying body movements. As described in occupants, such as in a hospice, edge nodes [21], sleep apnea is accompanied by body or leg retain personal or identifiable information, movement, which noninvasive sensors can which is then stripped by the edge. detect. We have used multiple PIR sensors and The sensor readings from multiple care piezoelectric-based sensors placed under the recipients are collected. In scenarios of multiple mattress (see Figures 2 and 3). occupants, such as in a hospice, edge nodes retain personal or identifiable information, which is then stripped by the edge. Figure3: Sensor for movement detection on the bed under the mattress The strong correlation between the two sensor types, shown in the diagram of recorded Figure 2: Floor plan and sensor layout sensor data over 8 hours, is presented in Figure 4. When motion is detected, the data from instances from the dataset. The sampling is multiple noninvasive sensors is processed on random but consistent while growing a single the edge node. The local machine learning tree. The multiple decision trees are trained on model is run, and the possible occurrence of the training data independently. sleep apnea is diagnosed. Periodical sessions The tree branching is performed by finding with invasive sensors or medical professionals' the best split from the features on each node. observations can be carried out to label the data During classification, trees vote for the class, set [22]. and the majority class is eventually predicted. Like RF, the Extremely Randomized Trees (ERT) algorithm [26] also generates trees' ensembles. ERT chooses the split from the attributes randomly, unlike RF. As a result, the number of calculations per node is decreased, thus increasing the training speed. Both algorithms provide excellent classification performance and can train models on extensive Figure 4: Movement in bed over 8 hours of datasets very fast. continuous sleeping Both ERT and RF provide feature importance estimates, a property used for After anonymization of the data, it is feature ranking and discarding of low- packaged and sent to the cloud for additional importance features during the feature selection processing. The data model on the cloud side is phase. We have used the feature importance run to verify the outcome for the received data. estimates when training an ERT classifier due If the model present in the cloud makes positive to its better speed than RF. detection for the received data and if the data Additionally, we have also used the Support was previously labeled with a negative result by Vector Machines (SVM) classifier [27] with the edge node, then the updated model is sent Gaussian kernel. Even though SVMs are much back to the edge node, which in turn processes slower algorithms as the dimensionality of data the data against the updated model. increases, they are compelling, especially after Sensor data that does not suggest strong parameter tuning [28]. Whenever we used negative results are marked for further labeling SVMs, the datasets were normalized so that the if additional data such as monitoring from training dataset will have a mean and standard medical equipment or video that can be deviation of 0 and 1, respectively. The RF and analyzed by a trained professional is available. ERT parameters were the default per their Such feedback is periodically included to build implementation in the [29] library. We did not the cloud model continuously. notice any significant gain by tuning their parameters (i.e., number of features per tree). 3.1. Classification algorithms Both ERT and RF classifiers were trained using 100 trees, which was appropriate for this size dataset. Using fewer trees improved the speed This section explains the classification while offering slightly worse classification algorithms used for feature ranking and performance. This library was used for the other construction classification models. The classification algorithms as well. accuracy was used for the comparison of various classification models throughout the system. One of the classification algorithms 3.2. Feature extraction used in our experiments is logistic regression [23]. For small datasets, it is straightforward The measurements from sensors can detect and provides easily interpretable models. atomic actions or states. More complex actions Moreover, it is a lightweight algorithm, which are depending on the context, which recent can be useful if the system is deployed on measurements can determine. Therefore, the hardware with limited resources. data needs to be first adequately segmented, and Random Forest (RF) [24] is an effective then feature extraction performed [30]. This algorithm that creates an ensemble of decision study additionally discusses the window size trees [25] by randomly sampling training impact on activity recognition. Generally, lower sensor frequencies entail longer The system evaluates different feature sets windows. It is considered during our by building classification models using the experiments by using different window lengths training dataset and evaluating them with the and analyzing the accuracy depending on them. validation dataset. The test set is not utilized at The segmentation into windows, step 1 on this stage at all. Thus, only the feature set that Figure 5, was performed, thus excluding the results in the best classification accuracy is border intervals when the activity changes from retained. To summarize, the purpose of this step one activity to another. is to significantly reduce the feature set size by Segmenting of streaming data into windows discarding features with low importance or high is performed in step 2 in Figure 5. Step 2 data drift sensitivity. extracts the following types of features (the The system evaluates different feature sets number of measurements within one window is by building classification models using the denoted by n.): training dataset and evaluating them with the • Basic statistics results in 14 features validation dataset. The test set is not utilized at per time series. this stage at all. Thus, only the feature set that • Equal-width histogram calculated with results in the best classification accuracy is [log ! 𝑛+1] intervals, based on the Sturges retained. To summarize, the purpose of this step rule [31]. It results in 5 to 8 features when is to significantly reduce the feature set size by the window length varies from 5s (25 discarding features with low importance or high measurements) to 20s (100 measurements). data drift sensitivity. • Quantile-based features: first quartile, After the feature set is reduced, step 5 uses median, third quartile, interquartile ranges, the training and validation sets to perform and other percentiles (5, 10, 20, 30, 40, 60, parameter tuning for the SVM. 70, 80, 90, 95), also used in [32]. From one- Finally, step 6 evaluates different classifiers time series, it generates 14 features. by building classification models with the • Auto-correlation of the measurements training and validation dataset's union and within one sliding window [33]. Let τ denote evaluating it using the independent test set. the amount of shift, and its domain is defined as τ ∈ [1, ë n û] • For exponentially increasing values of τ in that range, classical autocorrelation and Pearson correlation are calculated. Additionally, it calculates both correlations using the first and second half of measurements within one sliding window. This results in 3 to 4 τ values when the window length varies from 5s (25 measurements) to 20s (100 measurements). • Pearson correlations between pairs of time series; For five-time series, this results in 9 features. • Linear and quadratic fit coefficients; There are two linear fit and three quadratic fit coefficients, yielding five features in total per time series. • As a result of step 2, 250 to 270 features are generated depending on the window length. In step 3 performs feature importance and drift sensitivity estimation is done. Next, step 4 performs coarse-grained feature selection, which tests a set of Figure 5: Feature extraction, selection, and thresholds used to discard features with low classification flow importance or high drift sensitivity. 4. Results and evaluation Skopje, Macedonia and is supported by the networking activities provided by the ICT COST Actions IC1303 AAPELE and CA16226 The duration of our experiment was 8 hours. SHELD-ON. The sampling rate was set to 10Hz, thereby We also acknowledge Microsoft Azure's measuring ten values from each sensor every support for research through a grant providing second. We divided the dataset into three computational resources for this work. different subsets: training, validation, and testing. The training subset consisted of the first 45% records for each action, and the validation 7. References subset consisted of the next 25% records. The remaining 30% of records belonged to the test [1] M. Aazam and E.-N. Huh, “Fog subset. When performing parameter tuning for computing and smart gateway based SVMs and making feature selection, the communication for cloud of things,” in training set was used to build models, and the Future Internet of Things and Cloud validation set was used to evaluate their (FiCloud), 2014 International Conference performance. Once this phase was completed, on. IEEE, 2014, pp. 464–470. the final evaluation was performed only with [2] T. N. Gia, et al., “Fog computing in the best feature set decided after the feature health- care internet of things: A case screening and using the most optimal study on ECG feature extraction,” in parameters. The union of the training and Computer and Information Technology; validation sets was used to build classification Ubiquitous Computing and models for making final predictions. The test Communications; Dependable, set was used for building predictions and the Autonomic and Secure Computing; performance evaluation. Pervasive Intelligence and Computing (CIT/IUCC/DASC/PICOM), 2015 IEEE 5. Conclusion International Conference on. IEEE, 2015, pp. 356–363. [3] V. Stantchev, et al., “Smart items, fog As personal health becomes pervasive and and cloud computing as enablers of the data generated by it increases in volume, fog servitization in healthcare,” Sensors & computing offers a solution for many critical Transducers, vol. 185, no. 2, p. 121, 2015. challenges. The added flexibility of the fog [4] P. Maresova, et al., "Technological architecture enables better placement of Solutions for Older People with computing and network resources. Smarter data Alzheimer's Disease" Current Alzheimer flow could protect personal data, bandwidth Research, Volume 15, No. 9 April 2018, cost could be reduced, and more scalable, DOI:10.2174/1567205015666180427124 secure, and interoperable systems can be 547, pp. 975-983 designed. This paper illustrates those benefits [5] R. Goleva, et al.,, “AAL and ELE Platform by providing an experimental illustration of Architecture” in book: Ambient Assisted typical AAL service provided by fog-based Living and Enhanced Living health care ELE. Environments: Principles, Technologies By using simple hardware, the AAL data and Control, , Elsevier, 2016. was streamed to a cloud-based system, where it [6] L. Gu, D. Zeng, S. Guo, A. Barnawi, was fused. Using a systematic and automated and Y. Xiang, “Cost-efficient resource feature extraction and selection process, we management in fog computing supported could extract robust and reliable features that medical cyber-physical system,” IEEE facilitated building powerful classification Transactions on Emerging Topics in models. Computing, vol. 5, no. 1, pp. 108–119, 2017. 6. Acknowledgements [7] O. Kotevska, et al. "Towards a Patient- Centered Collaborative Health Care This work was partially financed by the System Model," International Journal of Faculty of Computer Science and Engineering Computer Theory and Engineering , vol. at the Ss. Cyril and Methodius University, 4, no. 6, November 2012, DOI: 10.7763/IJCTE.2012.V4.631, pp. 1025- [17] A.M. Rahmani, et al, “Smart e-health 1029. gateway: Bringing intelligence to internet- [8] L. Xu, et al."What Clinics Are Expecting of-things based ubiquitous healthcare From Data Scientists? A Review on Data- systems,” in Consumer Communications Oriented Studies Through Qualitative and and Networking Conference (CCNC), Quantitative Approaches." In IEEE 2015 12th Annual IEEE. IEEE, 2015, pp. Access, Volume 7, 826–834. January2019,DOI:10.1109/ACCESS.201 [18] A. Atanasov, et al. “Testbed Environment 8.2885586, pp. 641-654. for Wireless Sensor and Actuator [9] I. Kulev, et al. “Recommendation Network”, in Proc of the 5th International Algorithm Based on Collaborative Conference on Systems and Network Filtering and its Application in Communications – ICSNC 2010, CD Healthcare”, The 10th Conference for publication, Nice, France, August 2010. Informatics and Information Technology [19] H. Dubey, et al, “Fog data: Enhancing (CIIT 2013), Bitola, R. Macedonia, April telehealth big data through fog 18-21, 2013. computing,” in Proceedings of the ASE [10] M. Barhamgi, et al., “Enabling end-users BigData & SocialInformatics 2015. to protect their privacy,” in Proceedings of ACM, 2015, p. 14. the 2017 ACM on Asia Conference on [20] E. Zdravevski, et al. Improving activity Computer and Communications Security. recognition accuracy in ambient-assisted ACM, 2017, pp. 905–907. living systems by automated feature [11] E. Vlahu-Gjorgievska, et al., "Connected- engineering. IEEE Access, 5,2017, pp. Health Algorithm: Development and 5262-5280. Evaluation." Journal of Medical Systems, [21] V. K. Somers, et al.,“Sympathetic neural vol. 40, no. 4, 2016, DOI: mechanisms in obstructive sleep apnea.” 10.1007/s10916-016-0466-9, pp. 1-7. Journal of Clinical Investigation, vol. 96, [12] A. M. Rahmani, et al., “Exploiting smart no. 4, 1995, p. 1897. e-health gateways at the edge of [22] J.-C. Vazquez, et al., “Automated analysis healthcare internet-of-things: a fog of digital oximetry in the diagnosis of computing approach,” Future Generation obstructive sleep apnoea,” Thorax, vol. Computer Systems, vol. 78, pp. 641–658, 55, no. 4, 2000, pp. 302–307. 2018. [23] D. W. Hosmer Jr, S. Lemeshow, and R. X. [13] Y. Shi, et al., “The fog computing service Sturdivant, Applied logistic regression. for healthcare,” in Future Information and John Wiley & Sons, 2013, vol. 398. Communication Technologies for [24] L. Breiman, “Random forests,” Machine Ubiquitous HealthCare (Ubi-HealthTech), Learning, vol. 45, no. 1, 2001, DOI: 2015 2nd International Symposium on. 10.1023/A:1010933404324, pp. 5–32. IEEE, 2015, pp. 1–5. [25] J. R. Quinlan, “Induction of decision [14] A. Zdravevska, et al, “Cloud-based trees,” Machine learning, vol. 1, no. 1, recognition of complex activities for 1986, pp. 81–106. ambient assisted living in smart homes [26] P. Geurts, D. Ernst, and L. Wehenkel, with noninvasive sensors,” in Smart “Extremely randomized trees,” Machine Technologies, IEEE EUROCON 2017- learning, vol. 63, no. 1, 2006, pp. 3–42. 17th International Conference on. IEEE, [27] C. Cortes and V. Vapnik, “Support-vector 2017, pp. 769–774. networks,” Machine Learning, vol. 20, no. [15] A. Alrawais, A. Alhothaily, C. Hu, and X. 3, 1995. DOI: 10.1007/BF00994018, pp. Cheng, “Fog computing for the internet of 273–297. things: Security and privacy issues,” IEEE [28] P. Lameski, E. Zdravevski, R. Mingo, and Internet Computing, vol. 21, no. 2, pp. 34– A. Kulakov, “Svm parameter tuning with 42, 2017. grid search and its impact on reduction of [16] M. Memon, et al, “Ambient assisted model overfitting,” in Rough Sets, Fuzzy living healthcare frameworks, platforms, Sets, Data Mining, and Granular standards, and quality attributes,” Sensors, Computing. Springer, 2015, pp. 464–474. vol. 14, no. 3, pp. 4312–4341, 2014. [29] F. Pedregosa, et al., “Scikit-learn: Machine learning in Python,” Journal of Machine Learning Research, vol. 12, 2011, pp. 2825–2830. [30] O. Banos, et al., “Window size impact in human activity recognition,” Sensors, vol. 14, no. 4, 2014. DOI: 10.3390/s140406474, pp. 6474–6499. [31] H. A. Sturges, “The choice of a class interval,” Journal of the American Statistical Association, vol. 21, no. 153, 1926, pp. 65–66. [32] P. Siirtola and J. R¨oning, “Recognizing human activities user independently on smartphones based on accelerometer data,” International Journal of Artificial Intelligence and Interactive Multimedia, vol. 1, no. 5, 2012, pp. 38–45. [33] H. Mart´ın, A. M. Bernardos, J. Iglesias, and J. R. Casar, “Activity logging using lightweight classification techniques in mobile devices,” Personal and Ubiquitous Computing, vol. 17, no. 4, 2013. doi: 10.1007/s00779-012-0515-4, pp. 675– 695.