Indirect recognition of predefined human activities

sensors Article Indirect Recognition of Predefined Human Activities Ojan Majidzadeh Gorjani * , Antonino Proto , Jan Vanus and Petr Bilik Department of Cybernetics and Biomedical Engineering, Faculty of Electrical Engineering and Computer Science, VSB—Technical University of Ostrava, 70833 Ostrava-Poruba, Czech Republic; [email protected] (A.P.); [email protected] (J.V.); petr[email protected] (P.B.) *Correspondence: [email protected] Received: 31 July 2020; Accepted: 25 August 2020; Published: 26 August 2020   Abstract: The work investigates the application of artificial neural networks and logistic regression for the recognition of activities performed by room occupants. KNX (Konnex) standard-based devices were selected for smart home automation and data collection. The obtained data from these devices (Humidity, CO 2 , temperature) were used in combination with two wearable gadgets to classify specific activities performed by the room occupant. The obtained classifications can benefit the occupant by monitoring the wellbeing of elderly residents and providing optimal air quality and temperature by utilizing heating, ventilation, and air conditioning control. The obtained results yield accurate classification. Keywords: deep learning; logistic regression; activity recognition; prediction; classification; artificial neural network; smart homes; intelligent buildings 1. Introduction The availability of various affordable and cost-effective technologies for automation drives the rapid increase in smart homes. Such technologies provide the possibility of monitoring and tracking events such as unauthorized entry detection, the status of doors and windows, and presence monitoring. An increase in the number of sensors and integration with the Internet of Things (IoT) within smart homes creates new possibilities for improving the daily life of the residents, such as monitoring the activity and well-being of disabled people or seniors [1]. In recent years the health care and assisted living has gained much attention among researchers. In a case study, Panagopoulos et al. [ 2 ] presented a usability assessment of “Heart Around”, an integrated homecare solution incorporating communication functionalities, as well as health monitoring and emergency response features. Loukatos et al. [ 3 ] investigated educationally fruitful speech-based methods to assist people with special needs to care for potted plants. Wiljer et al. [ 4 ] suggested improving health care by developing an artificial intelligence-enabled healthcare practice. Many of the works in the field of activity recognition are emphasizing fall detections [ 5 – 7 ]. Sadreazami et al. [ 5 ] proposed using the StandoffRadar and a time series-based method for detecting fall incidents in human daily activities. A time was obtained by summing all the range bins corresponding to the ultra-wideband radar return signals. Ahamed et al. [ 6 ] investigated accelerometer-based fall detection, the Feed Forward Neural Network and Long Short-Term Memory based on deep learning networks, applied to detect falls. Dhiraj et al. [ 7 ] proposed two vision-based solutions, one using convolutional neural networks in 3D-mode and another using a hybrid approach by combining convolutional neural networks and long short-term memory networks using 360-degree videos for human fall detection. On a larger scale, Hsuseh et al. [ 8 ] adopted deep learning techniques to learn the long-term dependencies from videos for human behavior recognition in a multi-view framework detection. Sensors 2020,20, 4829; doi:10.3390/s20174829 www.mdpi.com/journal/sensors Sensors 2020,20, 4829 2 of 19 Often, camera-based solutions create concerns regarding security and privacy. Therefore, several works are based on indirect occupancy monitoring. Szczurek et al. [ 9 ] investigated occupancy determination based on time series of CO 2 concentration, temperature and relative humidity. There are works [ 1 ] monitoring the daily living activities in smart home care using CO 2 concentration. Vanus et al. [ 10 ] designed an indirect method for human presence monitoring in an intelligent building. Vanus et al. [ 11 ] used the IBM SPSS modeler tool and neural networks for CO 2 prediction within smart home care. Vanus et al. [ 12 ] compared neural networks, random trees, and linear regression for the purpose of indirect occupancy recognition in intelligent buildings. This paper proposes to employ an identical KNX-based setup building on the above contributions with a significant difference in expanding the occupancy monitoring to activity recognition. The indirect recognition of human activity is one of the most highly anticipated research topics. Albert et al. [ 13 ] used mobile phones for activity recognition in Parkinson’s patients. Nweke et al. [ 14 ] reviewed deep learning algorithms for human activity recognition using mobile and wearable sensor networks. Lara et al. [ 15 ] reviewed human activity recognition using wearable sensors and Yousefi et al. [ 16 ] reviewed behavior recognition using Wi-Fi channel state information. Minarno et al. [ 17 ] compared the performance Logistic Regression and Support Vector Machine to recognize activities such as laying, standing sitting, walking, walking upstairs or downstairs. Kwapisz et al. [ 18 ] proposed using logistic regression and multilayer perceptron with data obtained from cell phone accelerometers to recognize similar human activities. In a similar study, Bayat et al. [ 19 ] proposed using accelerometer data from smartphones to recognize more complex human activities such as running and dancing. Trost et al. [ 20 ] compared results obtained from the hip and wrist-worn accelerometer data for the recognition of seven classes of activities. This study is aimed at taking the data analysis within smart homes beyond occupancy monitoring and fall detection. Although there are a few available works in the field of activity recognition, this study targets new types of recognizable activities beyond common walking, running, and climbing stairs. The proposed method employs KNX standard-based devices to obtain room air quality data (Humidity, CO 2 , temperature) and combines the obtained data with two wearable gadgets that provide movement-related data. KNX-based devices were selected due to properties such as cost-effectiveness, compatibility and wide availability within locations such as smart homes, office buildings, shopping centers, medical facilities, and industrial locations. Initially, logistic regression-based models are developed (using IBM SPSS statistic 26) to classify the obtained datasets. Logistic regression is one of the most used methods in the field of activity recognition. Therefore, it provides a good reference for the evaluation of the method using artificial neural networks. Ultimately, the article proposes to use artificial neural networks and the obtained datasets to classify few types of human daily activities such as relaxing, eating, cleaning, exercising using a stationary bike and using a computer. IBM SPSS statistic 26 and IBM SPSS modeler 18 were selected as suitable data analysis platforms to develop required logistic regression and artificial neural network predictive models. In addition to monitoring the wellbeing of elderly residents, the obtained predictions can benefit the occupant by providing optimal air quality and temperature by utilizing heating, ventilation, and air conditioning control. The obtained results yield highly accurate prediction accuracies. 2. Materials and Methods The proposed method contains three main stages of data collection, pre-processing, and predictive analytics (Figure 1). In the first stage, the KNX devices were employed to monitor the air quality of the room in terms of room temperature (C), humidity level (%), CO 2 Concentration level (ppm). The movements of the room occupant were monitored using two individual wearable gadgets based on the Inertial Measurement Unit (IMU). After data synchronization and dealing with the missing data, predictive analytics were applied. Figure 1shows the application of logistic regression using IBM SPSS statistics 26. A separate predictive model with binary output was dedicated to each type of activity classes, where 0 represents false and 1 represents true. Since logistical regression is commonly used Sensors 2020,20, 4829 3 of 19 in this particular field of research, it provides a good benchmark or reference point for the evaluation of the artificial neural network-based method. Figure 2shows the application of artificial neural networks using IBM SPSS modeler 18. It can be observed that in the second approach a single output was used to determine the outcome of the predictive model. Figure 1. Block diagram of the proposed method using logistic regression. Figure 2. Block diagram of the proposed method using artificial neural networks. 2.1. Data Collection The data collection was performed in laboratory EB312 at the new Faculty of Electrical Engineering and Computer Science building of the VSB Technical University of Ostrava. The data collection was performed on the 19 July 2019 (08:28:00 to 10:31:00) and 26 July 2019 (08:09:00 to 10:10:00). The activities were performed by a single occupant present in the room. The performed activities were divided into five classes that are described in Table 1. These classes can simulate part of the daily activities performed in a single occupant room. Table 1. Description of activity categories. Activity Class Description Class 1 Relaxing with minimal movements Class 2 using the computer for checking emails and web surfing Class 3 Preparing tea and sandwich—eating breakfast Class 4 Cleaning the room by wiping the Tables and vacuum cleaning Class 5 Exercising using stationary bicycle Sensors 2020,20, 4829 4 of 19 2.1.1. KNX Technology A KNX (Konnex) setup was used to monitor the experiment’s room. In general, KNX is an open standard (EN 50090 [ 21 ], ISO/IEC 14543 [ 22 ]) for commercial and domestic building automation in a variety of locations such as office buildings, shopping centers, medical facilities, and industrial locations. It can be used to control functions such as heating, cooling, ventilation, energy management, and lighting control. The KNX bus system is a decentralized system with multi-master communication. KNX modules are commonly commissioned using the Engineering Tool Software (ETS). In addition to ETS, a .net-based software was developed [ 12 ] to ensure the connection of KNX-based devices and IBM cloud storage technology, which enables the communication between IBM Watson IoT platform and KNX smart installation. The measurements of CO 2 accumulation, indoor temperature, and humidity were performed using the MTN6005-0001 module. The measuring range of this device is listed in Table 2. Table 2. List of measured parameters and their unit. Sensor Unit Range CO2ppm 300 to 9999 Temperature 1 0 to +40 Relative humidity sensor % 20 to 100 2.1.2. Wearable Gadgets Two wearable gadgets were used to monitor the experimenter’s movements [ 23 , 24 ]. One was worn on the right hand and the other on the right leg (Figure 3). The wearable gadgets were based on the new generation of the Inertial Measurement Unit (IMU), developed by x-io Technologies, UK. The IMU is a compact data acquisition platform that combines diverse onboard sensors (as displayed in Table 3), and it is largely used for the evaluation of gait variability [24,25]. As regards this study, it comprises an 8-channel analog input, and an SD-card to store the data. The analog input of the IMU is equipped with a 10-bit AD-converter that allows us to acquire and convert the signals from a variety of modules. Table 3shows the measured parameters and their units. Figure 3. Inertial Measurement Unit (IMU) worn on a leg. Table 3. List of measured parameters using wearable gadgets. Parameter Unit Gyroscope X, Y, Z deg/s Accelerometer X, Y, Z g Magnetometer X, Y, Z µT Barometer hPa Sensors 2020,20, 4829 5 of 19 2.2. Pre-Processing The wearable gadgets are using an approximated data collection rate of 30 to 60 samples per second and the KNX-based data collection rate is between 1 to 10 samples per minute. This large difference creates a database synchronization problem. Therefore, the data collected from KNX devices had been expanded to match the fast rates of the wearable gadget. A .Net-based script was used to perform data synchronization. Missing data could result in algorithm failure or decrease the accuracy of the analysis. Therefore, IBM SPSS software tool automatically removes all of the records with missing data from the analysis. Using the IBM SPSS software tool time-related variables were removed and correct variable types were assigned to each parameter (continuous and binary). 2.3. Predictive Analytics Predictive modeling is the general concept of building a model that uses big data to develop models capable of making reliable predictions. In general, these models are based on variables (also known as predictors) that are most likely to influence the outcome [ 26 ]. Predictive models are widely applied in various applications such as weather forecasting [ 27 – 29 ], Bayesian spam filters [ 30 – 33 ], business [ 34 – 37 ], and fraud detection [ 38 – 40 ]. Predictive models typically include a machine learning algorithm that learns certain properties from a training dataset. The learning process can be applied using supervised learning [ 41 , 42 ], unsupervised learning [ 42 ], semi-supervised learning [ 43 ], active learning. In the purposed method, supervised learning was employed by presenting a set of solved (labeled) examples to the model for training. Once the model is established, a pattern between the predictors and the outcome could solve similar predictions on its own. 2.3.1. Logistic Regression Regression is one the oldest and often used algorithms in machine learning with a supervised learning strategy [ 44 , 45 ]. Linear Regression and Logistic Regression are the two famous types of regression. In general, Linear Regression is used for solving Regression problems whereas Logistic Regression is used for solving the Classification problems such as predicting the categorical dependent variable with the help of independent variables or where the probabilities between two classes are required [45]. Logistic regression is used in various fields, including machine learning, most medical fields, and social sciences [ 46 – 50 ]. The weighted sum of inputs passes through the logistic function Equation (1) that can map values in between 0 and 1. The logistic function is a sigmoid function [ 51 ] and the curve obtained is called a sigmoid curve or S-curve (Figure 4). Figure 4. Example of the logistic function. The output of the binary logistic regression model can be only binary (either 0 or 1). Outputs with more than two values are modeled by multinomial logistic regression and if the multiple categories are ordered, by ordinal logistic regression. The logistic regression is not a classifier by itself, it simply provides a probability of output in terms of input. However, it can be used to make a classifier, for instance by choosing a cutoffvalue and classifying inputs with probability greater than the cutoff Sensors 2020,20, 4829 6 of 19 as 1 and below the cutoffas 0; this is a common way to make a binary classifier. The general equation of logistic regression is provided by Equation (2). f(x)=1 1+e−k(x−x0)(1) y=1 1+e−(β0+β1x1+β2x2+β3x3+···+βnxn)(2) Regression models can be created using multiple algorithms, these algorithms specify how independent variables are entered into the model [ 52 – 56 ]. The common algorisms are Enter (Regression) [ 56 , 57 ], Stepwise [ 58 ], Backward Elimination [ 59 ] and Forward Selection [ 60 , 61 ]. The Hosmer–Lemeshow test and Omnibus test are some of the most common statistical tests used to examine the goodness of fit for logistic regression. It compares the observed event rates and expected event rates in subgroups of the model population. The test mainly identifies subgroups as the deciles of fitted risk values. Well calibrated models are the models with similar expected and observed event rates in their subgroups. The expected probability of success is given by the equation for the logistic regression model. In general, the Hosmer–Lemeshow test is useful to determine if the lack of fit (poor prediction) is significant but it does not properly take overfitting into account. The omnibus test is a likelihood-ratio chi-square test of the current model versus the null (in this case, the intercept) model. Generally, the significance value of less than 0.05 indicates that the current model outperforms the null model. The odds ratio is often used to quantify the strength of the association between two events. In logistic regression, the odds ratio shows the amount of increase in the output variable with every unit increase in a specific input variable. The odds ratio for a continuous independent variable can be defined as Equation (3). This exponential relationship provides an interpretation for β1 where the odds is multiplied by eβ1 for every 1-unit increase in x [ 62 ]. If a, b, c and d can are cells in a 2 × 2 contingency table then formula 4 describes odds ratio for a binary independent variable. odds ratio(OR)=p(x+1) p(x)=eβ0+β1(x+1) eβ0+β1(x)=eβ1(3) odds ratio(OR)=ad bc (4) 2.3.2. Artificial Neural Network Due to their power flexibility and ease of use, artificial neural networks are widely used [ 63 – 69 ]. Artificial neural networks obtain their knowledge from the learning process and then use interneuron connection strengths (known as synaptic weights) to store the obtained knowledge [ 70 , 71 ]. One of the most used classes of artificial neural networks is a multilayer perceptron which is a feedforward neural network that belongs to deep learning. Deep learning utilizes a hierarchical level of artificial neural networks to carry out the process of machine learning. Unlike traditional programs, the hierarchical function of deep learning systems enables machines to process data with a nonlinear approach. The multilayer perceptron utilizes backpropagation for training [ 72 – 74 ]. Due to its multiple layers and nonlinear activation, a multilayer perceptron can distinguish data that are not linearly separable [75]. In deep learning, in addition to input and output, layers of the neural network contain multiple hidden layers and each can contain multiple neurons. The first layer of the neural network processes a raw data input like the amount of the transaction and passes it on to the next layer as output. The second layer processes the previous layer’s information by including additional information. This continues across all levels of the neural network. Each layer of its neural network builds on its previous layer. The multilayer perceptron artificial neural network with two hidden layers was chosen as a suitable deep learning method for this article (Figure 5). Sensors 2020,20, 4829 7 of 19 Figure 5. Example of the developed multilayer perceptron artificial neural network model with 24 neurons input layer, eight neurons in the first hidden layer, four neurons in the second hidden layer, five neurons output layer. The multilayer perceptron artificial neural network was implemented in the IBM SPSS Modeler 18 software. The IBM SPSS modeler algorithm guide mathematically describes its multilayer perceptron model as followings [76]: Input layer: j0=p units, a0:j , . . . , a0:j0 , with a0:j=xj , where j is the number of neurons in the layer and X is the input. ith hidden layer: ji units, ai:1 , . . . , ai:ji , with a1:k=γi(Ci:k) and Ci:k=Pji−1 j=0ωI:j1 , kai−1:j , where ai−1:0 = 1, γi is the activation function for the layer I, and ωI:j1 is weight leading from layer i − 1. At this layer, the model uses hyperbolic tangent as an activation function provided by γ(C)=tanh(c)ec−e−c ec+e−c. Output layer: jI=R units, aI:1 , . . . , aI:JI , with aI:k=γI(CI:k) and CI:k=PJ1 J=0ωI:j , kai−1:j , where ai−1:0 =1. The SoftMax function (γ(Ck)=eck Pj∈Γhecj) is used as an activation function. To evaluate the performance of predictive modeled three methods of splitting, partitioning, and scoring is commonly used. In the partitioning method, the datasets are randomly divided into training, testing, and validation partitions where models are trained, tested and evaluated using different segments of the dataset. Partitioning is mostly recommended for very large datasets. The scoring method uses entirely different datasets for training and evaluation. One dataset is solely used for training and a separate dataset for evaluation. Therefore, it provides a better indication of the real accuracy of the models. 3. Implementation and Results This section discusses the implementation and results of the classifications performed by logistic regression and artificial neural networks (multilayer perceptron). The logistic regression is a commonly used classification method in the field of activity recognition. Therefore, it can provide a good comparison point for the main purposed method using artificial neural networks. 3.1. Linear Regression The obtained data from measurements performed on the 19 July 2019 (dataset A) and 26 July 2019 (dataset B) were analyzed using IBM SPSS Statistics 26 software tool. Since IBM SPSS Modeler Sensors 2020,20, 4829 8 of 19 18 does not natively include logistic regression, IBM SPSS Statistics 26 software was used to perform the logistic regression analysis. In the first stage, the datasets A and B were individually classified. The logistical regression models were developed using enter configuration, classification cutoffof “0.5” and a maximum of 20 iterations. The goodness of fit describes how well a statistical model fits a set of observations. Hosmer and Lemeshow and omnibus tests were used to determine the goodness of fit. For a good fit, the Hosmer & Lemeshow test significance value should be more than 0.05 and omnibus should have a significance value less than 0.05. These conditions were satisfied with large margins across all models. Tables 4and 5 show the accuracy of classification for data obtained from measurements of dataset A (total of 296,188 records) and dataset B (total of 290,174 records). The result shows that all models obtained classification accuracy above 91.2%. In analysis performs on the measurement interval of dataset A (Table 4), the activity Class 3 shows almost complete accuracy (only two wrong predictions in 296,188 records) and activity Class 4 shows the lowest accuracy (97.4%). Table 4. Classification table using dataset A (19 July 2019 interval). Class Observed Predicted Percentage Correct Overall Accuracy 0 1 Class 1 0 273,204 2758 99.9% 98.9% 1 422 19,804 97.9% Class 2 0 213,908 3764 98.3% 97.4% 1 3877 74,639 95.1% Class 3 0 220,320 0 100.0% 100.0% 1 2 75,866 100.0% Class 4 0 243,244 5327 97.9% 95.4% 1 8209 39,408 82.8% Class 5 0 223,644 1096 99.5% 99.3% 1 869 70,579 98.8% Table 5. Classification table using dataset B (26 July 2019 interval). Activity Observed Predicted Percentage Correct Overall Accuracy 0 1 Class 1 0 270,378 867 99.7% 99.5% 1 640 18,829 96.7% Class 2 0 213,182 3791 98.3% 97.0% 1 5068 68,673 93.1% Class 3 0 216,883 95 100% 99.9% 1 51 73,685 99.9% Class 4 0 235,594 9914 96.0% 91.2% 1 15,720 29,486 65.2% Class 5 0 212,940 1653 99.2% 98.9% 1 1535 74,586 98.05 Similar characteristics can be observed from the dataset B (Table 5) result where Class 3 yields the highest accuracy (99.9%) and Class 4 the lowest (91.2%). Table 1indicates that Class 4 is dedicated to cleaning activities such as wiping tables and vacuum cleaning. Therefore, the lower accuracy could be the direct result of less consistent movement during this activity class. Using a stationary bicycle (Class 5 activity) is a high energy activity, and on the contrary, relaxing with minimal movements (Class 1) is a low energy activity. Regardless of energy levels, both of these activities provide consistent movements that directly translate to a more recognizable pattern within data. This can be easily Sensors 2020,20, 4829 9 of 19 observed within the classification results (99.3% and 98.9% for Class 5 and 98.9 and 99.5% for Class 1). Summing up the classification accuracy resulted in 97.8% of correctly classified records. Table 6shows the odds ratio of different parameters in developed models. The odds ratio shows the amount of increase in the output variable with every unit increase in a specific input variable. Simply, the output variable is more associated with changes in parameters with a larger absolute value of the odds ratio. Table 6shows consistent odds ratios for gyroscope and magnetometer (both devices and across all three axes), CO 2 for all models. Therefore, it affects all models with a similar significance. By comparing each model with its alternative interval, it can be observed that except for Model 5, the temperature has a similar range on both datasets. However, models 1, 2, and 4 share the very high odds ratio and model 3 shows a null effect. Table 6also shows that this null effect is also shared with the models based on the dataset A. It is also apparent that KNX-based data do not influence the recognition of Class 3 activity. On the other hand, the accelerometer y-axis (both devices) share similar large odds Table 6shows the odds ratio of different parameters in developed models. The odds ratio shows the amount of increase in the output variable with every unit increase in a specific input variable. Simply, the output variable is more associated with changes in parameters with a larger absolute value of the odds ratio. Table 6shows consistent odds ratios for gyroscope and magnetometer (both devices and across all three axes), CO 2 for all models. Therefore, it affects all models with a similar significance. By comparing each model with its alternative interval, it can be observed that except for Model 5, the temperature has a similar range on both datasets. However, models 1, 2, and 4 share the very high odds ratio and model 3 shows a null effect. Table 6also shows that this null effect is also shared with the models based on the dataset A. It is also apparent that KNX-based data do not influence the recognition of Class 3 activity. On the other hand, the accelerometer y-axis (both devices) share similar large odds ratio across both models, and in the case of exercising using a stationary bicycle (Class 5 activity), this large effect can be observed on the X-axis of the leg accelerometer. With few exceptions, the odds ratio of both datasets remains within a similar range, this indicates the consistency of the analysis. Overall, it can be observed that activity Class 1 is mainly affected by temperature and the Activity Classes 2, 4, 5 are mostly affected by temperature and are accelerometer-based. The obtained conclusions from the odds ratio were verified by regression weights, the test of significance, and Wald statistic. In the last stage of the analysis, the developed models were further evaluated by alternative datasets (scoring), resulting in a significant drop in the prediction accuracy (up to 50% decrease). This indicated is an indication of overfitting. Although Hosmer and Lemeshow and omnibus tests are a good indication for the goodness of fit, they do not detect overfitting. Across both models, and in the case of exercising using a stationary bicycle (Class 5 activity), this large effect can be observed on the X-axis of the leg accelerometer. With few exceptions, the odds ratio of both datasets remains within a similar range, this indicates the consistency of the analysis. Overall, it can be observed that activity Class 1 is mainly affected by temperature and the Activity Classes 2, 4, 5 are mostly affected by temperature and are accelerometer-based. The obtained conclusions from the odds ratio were verified by regression weights, the test of significance, and Wald statistic. In the last stage of the analysis, the developed models were further evaluated by alternative datasets (scoring), resulting in a significant drop in the prediction accuracy (up to 50% decrease). This indicated is an indication of overfitting. Although Hosmer and Lemeshow and omnibus tests are a good indication for the goodness of fit, they do not detect overfitting. Sensors 2020,20, 4829 16 of 19 Author Contributions: Conceptualization, O.M.G.; A.P. and J.V.; Data curation, O.M.G. and A.P.; Formal analysis, O.M.G.; Funding acquisition, P.B.; Investigation, O.M.G.; Methodology, O.M.G.; Project administration, J.V. and P.B.; Resources, O.M.G.; A.P. and P.B.; Software, O.M.G.; Supervision, O.M.G.; J.V. and P.B.; Validation, O.M.G.; Visualization, O.M.G.; Writing—original draft, O.M.G. and A.P.; Writing—review & editing, O.M.G. All authors have read and agreed to the published version of the manuscript. Funding: This work was funded by the European Regional Development Fund in the Research Centre of Advanced Mechatronic Systems project, project number CZ.02.1.01/0.0/0.0/16019/0000867 within the Operational Programme Research, Development and Education. Acknowledgments: This work was supported by the European Regional Development Fund in the Research Centre of Advanced Mechatronic Systems project, project number CZ.02.1.01/0.0/0.0/16019/0000867 within the Operational Programme Research, Development, and Education. This work was supported by the Student Grant System of VSB Technical University of Ostrava, grant number SP2020/151. Conflicts of Interest: The authors declare no conflict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to publish the results. References 1. Vanus, J.; Belesova, J.; Martinek, R.; Nedoma, J.; Fajkus, M.; Bilik, P.; Zidek, J. Monitoring of the daily living activities in smart home care. Hum. Cent. Comput. Inf. Sci. 2017,7, 30. [CrossRef] 2. Panagopoulos, C.; Menychtas, A.; Tsanakas, P.; Maglogiannis, I. Increasing Usability of Homecare Applications for Older Adults: A Case Study. Designs 2019,3, 23. [CrossRef] 3. Loukatos, D.; Arvanitis, K.G.; Armonis, N. Investigating Educationally Fruitful Speech-Based Methods to Assist People with Special Needs to Care Potted Plants. In International Conference on Human Interaction and Emerging Technologies; Springer: Cham, Switzerland, 2019; pp. 157–162. 4. Wiljer, D.; Hakim, Z. Developing an artificial intelligence–enabled health care practice: Rewiring health care professions for better care. J. Med. Imaging Radiat. Sci. 2019,50, S8–S14. [CrossRef] [PubMed] 5. Sadreazami, H.; Bolic, M.; Rajan, S. Fall Detection Using StandoffRadar-Based Sensing and Deep Convolutional Neural Network. IEEE Trans. Circuits Syst. II Express Briefs 2020,67, 197–201. [CrossRef] 6. Ahamed, F.; Shahrestani, S.; Cheung, H. Intelligent Fall Detection with Wearable IoT. Adv. Intell. Syst. Comput. 2020,993, 391–401. [CrossRef] 7. Dhiraj; Manekar, R.; Saurav, S.; Maiti, S.; Singh, S.; Chaudhury, S.; Neeraj; Kumar, R.; Chaudhary, K. Activity Recognition for Indoor Fall Detection in 360-Degree Videos Using Deep Learning Techniques. Adv. Intell. Syst. Comput. 2020,1024, 417–429. [CrossRef] 8. Hsueh, Y.-L.; Lie, W.-N.; Guo, G.-Y. Human Behavior Recognition from Multiview Videos. Inf. Sci. 2020 ,517, 275–296. [CrossRef] 9. Szczurek, A.; Maciejewska, M.; Pietrucha, T. Occupancy determination based on time series of CO2 concentration, temperature and relative humidity. Energy Build. 2017,147, 142–154. [CrossRef] 10. Vanus, J.; Machac, J.; Martinek, R.; Bilik, P.; Zidek, J.; Nedoma, J.; Fajkus, M. The design of an indirect method for the human presence monitoring in the intelligent building. Hum. Cent. Comput. Inf. Sci. 2018 ,8, 28. [CrossRef] 11. Vanus, J.; Kubicek, J.; Gorjani, O.M.; Koziorek, J. Using the IBMSPSS SWTool withWavelet Transformation for CO2 Prediction within IoT in Smart Home Care. Sensors 2019,19, 1407. [CrossRef] 12. Vanus, J.M.; Gorjani, O.; Bilik, P. Novel Proposal for Prediction of CO2 Course and Occupancy Recognition in Intelligent Buildings within IoT. Energies 2019,12, 4541. [CrossRef] 13. Albert, M.V.; Toledo, S.; Shapiro, M.; Koerding, K. Using mobile phones for activity recognition in Parkinson’s patients. Front. Neurol. 2012,3, 158. [CrossRef] [PubMed] 14. Nweke, H.F.; Teh, Y.W.; Al-Garadi, M.A.; Alo, U.R. Deep learning algorithms for human activity recognition using mobile and wearable sensor networks: State of the art and research challenges. Expert Syst. Appl. 2018 , 105, 233–261. [CrossRef] 15. Lara, O.D.; Labrador, M.A. A survey on human activity recognition using wearable sensors. IEEE Commun. Surv. Tutor. 2012,15, 1192–1209. [CrossRef] 16. Yousefi, S.; Narui, H.; Dayal, S.; Ermon, S.; Valaee, S. A survey on behavior recognition using wifi channel state information. IEEE Commun. Mag. 2017,55, 98–104. [CrossRef] Sensors 2020,20, 4829 17 of 19 17. Minarno, A.E.; Kusuma, W.A.; Wibowo, H. Performance Comparisson Activity Recognition using Logistic Regression and Support Vector Machine. In Proceedings of the 2020 3rd International Conference on Intelligent Autonomous Systems (ICoIAS), Singapore, 26–29 February 2020; pp. 19–24. 18. Kwapisz, J.R.; Weiss, G.M.; Moore, S.A. Activity recognition using cell phone accelerometers. ACM SIGKDD Explor. Newsl. 2011,12, 74–82. [CrossRef] 19. Bayat, A.; Pomplun, M.; Tran, D.A. A study on human activity recognition using accelerometer data from smartphones. Procedia Comput. Sci. 2014,34, 450–457. [CrossRef] 20. Trost, S.G.; Zheng, Y.; Wong, W.K. Machine learning for activity recognition: Hip versus wrist data. Physiol. Meas. 2014,35, 2183. [CrossRef] 21. European Committee for Standards. Home and Building Electronic System (HBES); Cenelec EN50090; European Committee for Standards: Brussels, Belgium, 2005. 22. International Organization for Standardization. K.N.X. Standard ISO/IEC14543-3; International Organization for Standardization: Geneva, Switzerland, 2006. 23. NGIMU X-Io Technologies. Available online: https://x-io.co.uk/ngimu/(accessed on 13 August 2020). 24. Rantalainen, T.; Karavirta, L.; Pirkola, H.; Rantanen, T.; Linnamo, V. Gait Variability UsingWaistand Ankle-Worn Inertial Measurement Units in Healthy Older Adults. Sensors 2020,20, 2858. [CrossRef] 25. Fida, B.; Bernabucci, I.; Bibbo, D.; Conforto, S.; Schmid, M. Pre-processing effect on the accuracy of event-based activity segmentation and classification through inertial sensors. Sensors 2015,15, 23095–23109. [CrossRef] 26. Alankar, B.; Yousf, N.; Ahsaan, S.U. Predictive Analytics for Weather Forecasting using Back Propagation and Resilient Back Propagation Neural Networks. In Advances in Intelligent Systems and Computing; Springer: Berlin/Heidelberg, Germany, 2020; Volume 1030, pp. 99–115. [CrossRef] 27. Poornima, S.; Pushpalatha, M. Prediction of rainfall using intensified LSTM based recurrent Neural Network with Weighted Linear Units. Atmosphere 2019,10, 668. [CrossRef] 28. Pooja, S.B.; Siva Balan, R.V. Linear program boosting classification with remote sensed big data for weather forecasting. Int. J. Adv. Trends Comput. Sci. Eng. 2019,8, 1405–1415. [CrossRef] 29. Yang, Y.; Hu, R.; Sun, G.; Qiu, C. Chinese Spam Data Filter Model in Mobile Internet. In Proceedings of the International Conference on Advanced Communication Technology ICACT, Xi’an, China, 17–20 February 2019; pp. 339–343. [CrossRef] 30. Maguluri, L.P.; Ragupathy, R.; Buddi, S.R.K.; Ponugoti, V.; Kalimil, T.S. Adaptive Prediction of Spam Emails Using Bayesian Inference. In Proceedings of the 3rd International Conference on Computing Methodologies and Communication ICCMC, Erode, India, 27–29 March 2019; pp. 628–632. [CrossRef] 31. Mansourbeigi, S.M.-H. Stochastic Methods to Find Maximum Likelihood for Spam E-mail Classification. Adv. Intell. Syst. Comput. 2019,927, 623–632. [CrossRef] 32. Mallik, R.; Sahoo, A.K. A novel approach to spam filtering using semantic based naive bayesian classifier in text analytics. Adv. Intell. Syst. Comput. 2019,813, 301–309. [CrossRef] 33. Van Nguyen, T.; Zhou, L.; Chong, A.Y.L.; Li, B.; Pu, X. Predicting customer demand for remanufactured products: A data-mining approach. Eur. J. Oper. Res. 2019,281, 543–558. [CrossRef] 34. Liu, C.; Lei, Z.; Morley, D.; Abourizk, S.M. Dynamic, Data-Driven Decision-Support Approach for Construction Equipment Acquisition and Disposal. J. Comput. Civ. Eng. 2020,34, 34. [CrossRef] 35. Wang, J.; Lai, X.; Zhang, S.; Wang, W.M.; Chen, J. Predicting customer absence for automobile 4S shops: A lifecycle perspective. Eng. Appl. Artif. Intell. 2020,89. [CrossRef] 36. Park,G.; Song,M.Predictingperformancesinbusinessprocessesusingdeepneuralnetworks. Decis. Support. Syst . 2020,129, 113191. [CrossRef] 37. Sarno, R.; Sinaga, F.; Sungkono, K.R. Anomaly detection in business processes using process mining and fuzzy association rule learning. J. Big Data 2020,7, 5. [CrossRef] 38. Matos, T.; Macedo, J.A.; Lettich, F.; Monteiro, J.M.; Renso, C.; Perego, R.; Nardini, F.M. Leveraging feature selection to detect potential tax fraudsters. Expert Syst. Appl. 2020,145, 113128. [CrossRef] 39. Shi, C.; Li, X.; Lv, J.; Yin, J.; Mumtaz, I. Robust geodesic based outlier detection for class imbalance problem. Pattern Recognit. Lett. 2020,131, 428–434. [CrossRef] 40. Han, J.; Pei, J.; Kamber, M. Data Mining: Concepts and Techniques; Elsevier: Waltham, MA, USA, 2011. 41. Kachalsky, I.; Zakirzyanov, I.; Ulyantsev, V. Applying Reinforcement Learning and Supervised Learning Techniques to Play Hearthstone. In Proceedings of the 2017 16th IEEE International Conference on Machine Learning and Applications (ICMLA), Cancun, Mexico, 18–21 December 2017; pp. 1145–1148. Sensors 2020,20, 4829 18 of 19 42. Nijhawan, R.; Srivastava, I.; Shukla, P. Land cover classification using super-vised and unsupervised learning techniques. In Proceedings of the 2017 International Conference on Computational Intelligence in Data Science (ICCIDS), Chennai, India, 2–3 June 2017; pp. 1–6. 43. Liu, Q.; Liao, X.; Carin, L. Semi-Supervised Life-Long Learning with Application to Sensing. In Proceedings of the 2007 2nd IEEE International Workshop on Computational Advances in Multi-Sensor Adaptive Processing, St. Thomas, VI, USA, 12–14 December 2007; pp. 1–4. 44. Yan, X. Linear Regression Analysis: Theory and Computing; World Scientific: Hackensack, NJ, USA, 2009; pp. 1–2. ISBN 9789812834119. 45. Rencher, A.C.; Christensen, W.F. Chapter 10, Multivariate regression—Section 10.1 Introduction Methods of Multivariate Analysis. In Wiley Series in Probability and Statistics, 3rd ed.; John Wiley & Sons: Hoboken, NJ, USA, 2012; Volume 709, p. 19. ISBN 9781118391679. 46. Tolles, J.; Meurer, W.J. Logistic regression: Relating patient characteristics to outcomes. JAMA 2016 ,316, 533–534. [CrossRef] [PubMed] 47. Boyd, C.R.; Tolson, M.A.; Copes, W.S. Evaluating trauma care: The TRISS method Trauma Score and the Injury Severity Score. J. Trauma 1987,27, 370–378. [CrossRef] [PubMed] 48. Le Gall, J.R.; Lemeshow, S.; Saulnier, F. A new simplified acute physiology score (SAPS II) based on a European/North American multicenter study. JAMA 1993,270, 2957–2963. [CrossRef] [PubMed] 49. Berry, M.J.; Linoff, G.S. Data Mining Techniques: For Marketing, Sales, and Customer Relationship Management; John Wiley & Sons: Hoboken, NJ, USA, 2004. 50. Truett, J.; Cornfield, J.; Kannel, W. A multivariate analysis of the risk of coronary heart disease in Framingham. J. Chronic Dis. 1967,20, 511–524. [CrossRef] 51. Cybenko, G. Approximation by superpositions of a sigmoidal function. Math. Control. Signals Syst. 1989 ,2, 303–314. [CrossRef] 52. Freedman, D.A. Statistical Models: Theory and Practice; Cambridge University Press: Cambridge, UK, 2009. 53. Efroymson, M.A. Multiple Regression Analysis. Mathematical Methods for Digital Computers; Ralston, A., Wilf, H.S., Eds.; Wiley: New York, NY, USA, 1960. 54. Hocking, R.R. The Analysis and Selection of Variables in Linear Regression. Biometrics 1976 ,32, 1–49. [CrossRef] 55. Draper, N.; Smith, H. Applied Regression Analysis, 2nd ed.; John Wiley & Sons, Inc.: New York, NY, USA, 1981. 56. SAS Institute. SAS/STAT User’s Guide, 4th ed.; Version 6; SAS Institute Inc.: Cary, NC, USA, 1989; Volume 2. 57. Knecht, W.R. Pilot Willingness to Take offinto Marginal Weather, Part II: Antecedent Overfitting with forward Stepwise Logistic Regression. In Technical Report DOT/FAA/AM-O5/15; Federal Aviation Administration: Oklahoma City, OK, USA, 2005. 58. Flom, P.L.; Cassell, D.L. Stopping Stepwise: Why Stepwise and Similar Selection Methods are Bad, and What You Should Use; NESUG: Corvallis, OR, USA, 2007. 59. Myers, R.H.; Myers, R.H. Classical and Modern Regression with Applications; Duxbury Press: Belmont, CA, USA, 1990; Volume 2. 60. Bendel, R.B.; Afifi, A.A. Comparison of stopping rules in forward “stepwise” regression. J. Am. Stat. Assoc. 1977,72, 46–53. 61. Kubinyi, H. Evolutionary variable selection in regression and PLS analyses. J. Chemom. 1996 ,10, 119–133. [CrossRef] 62. Szumilas, M. Explaining odds ratios. J. Can. Acad. Child. Adolesc. Psychiatry 2010,19, 227. 63. Moosavi, S.R.; Wood, D.A.; Ahmadi, M.A.; Choubineh, A. ANN-Based Prediction of Laboratory-Scale Performance of CO2-Foam Flooding for Improving Oil Recovery. Nat. Resour. Res. 2019 ,28, 1619–1637. [CrossRef] 64. Zarei, T.; Behyad, R. Predicting the water production of a solar seawater greenhouse desalination unit using multi-layer perceptron model. Sol. Energy 2019,177, 595–603. [CrossRef] 65. Yilmaz, I.; Kaynar, O. Multiple regression, ANN (RBF, MLP) and ANFIS models for prediction of swell potential of clayey soils. Expert Syst. Appl. 2011,38, 5958–5966. [CrossRef] 66. Heidari, E.; Sobati, M.A.; Movahedirad, S. Accurate prediction of nanofluid viscosity using a multilayer perceptron artificial neural network (MLP-ANN). Chemom. Intell. Lab. Syst. 2016,155, 73–85. [CrossRef] 67. Abdi-Khanghah, M.; Bemani, A.; Naserzadeh, Z.; Zhang, Z. Prediction of solubility of N-alkanes in supercritical CO2 using RBF-ANN and MLP-ANN. J. CO2Util. 2018,25, 108–119. [CrossRef] Sensors 2020,20, 4829 19 of 19 68. Ahmed, S.A.; Dey, S.; Sarma, K.K. Image texture classification using Artificial Neural Network (ANN). In Proceedings of the 2011 2nd National Conference on Emerging Trends and Applications in Computer Science, Shillong, India, 4–5 March 2011; pp. 1–4. 69. Zarei, F.; Baghban, A. Phase behavior modelling of asphaltene precipitation utilizing MLP-ANN approach. Pet. Sci. Technol. 2017,35, 2009–2015. [CrossRef] 70. Behrang, M.; Assareh, E.; Ghanbarzadeh, A.; Noghrehabadi, A. The potential of different artificial neural network (ANN) techniques in daily global solar radiation modeling based on meteorological data. Sol. Energy 2010,84, 1468–1480. [CrossRef] 71. Haykin, S. Neural Networks: A Comprehensive Foundation, 2nd ed.; Macmillan College Publishing: New York, NY, USA, 1998. 72. Ripley, B.D. Pattern Recognition and Neural Networks; Cambridge University Press (CUP): Cambridge, UK, 1996. 73. Hastie, T.; Tibshirani, R.; Friedman, J. The Elements of Statistical Learning: Data Mining, Inference, and Prediction; Springer: New York, NY, USA, 2009. 74. Rosenblatt, F. Principles of Neurodynamics. Perceptrons and the Theory of Brain Mechanisms; Defense Technical Information Center (DTIC): Baer Fort, VA, USA, 1961. 75. Rumelhart, D.E.; Hinton, G.E.; Williams, R.J. Learning Internal Representations by Error Propagation; The MIT Press: Cambridge, MA, USA, 1985. 76. IBM SPSS Modeler 18 Algorithms Guide. Available online: http://public.dhe.ibm.com/software/analytics/ spss/documentation/modeler/18.0/en/AlgorithmsGuide.pdf (accessed on 1 November 2019). © 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).