An intelligent system for prognosis of noncommunicable diseases’ risk factors

Telematics and Informatics xxx (xxxx) xxx–xxx

Contents lists available at ScienceDirect

Telematics and Informatics journal homepage: www.elsevier.com/locate/tele

An intelligent system for prognosis of noncommunicable diseases’ risk factors Fábio Pittolia, Henrique Damasceno Viannaa, Jorge Luis Victória Barbosaa, Emerson Butzena, Mari Ângela Gaedkeb, Juvenal Soares Dias da Costab, ⁎ Renan Belarmino Scherer dos Santosa, a

Applied Computing Graduate Program (PIPCA), University of Vale do Rio dos Sinos (UNISINOS), 950, Unisinos Av., São Leopoldo, Rio Grande do Sul, Brazil b Collective Health Graduate Program, University of Vale do Rio dos Sinos (UNISINOS), 950, Unisinos Av., São Leopoldo, Rio Grande do Sul, Brazil

A R T IC LE I N F O

ABS TRA CT

Keywords: Ubiquitous computing Mobile computing Bayesian Networks Chronic disease

Noncommunicable diseases are the main reason to the rise of diseases incidence in the developed world. The management and prevention of these diseases can be done by controlling the behavioral and biological risk factors which are related to them. ChronicPrediction is an intelligent system for noncommunicable diseases care which determines in real time the impact on risk factors due to actions taken by users. Based on impact information, the system presents on users’ smartphones strategic messages to help in their treatment. ChronicPrediction applies Bayesian Networks (BNs) which use risk factors for mapping the causes of noncommunicable diseases worsening. The support to multiple chronic diseases and the integrated use of multiple BNs based on risk factors are the main contributions of this work and differentiate the proposed system from related work. We have built a functional prototype that allowed us to conduct two experiments. The first one successfully tested the main functionalities provided by ChronicPrediction to support BNs based on risk factors and the sending of messages to users’ smartphones. The evaluation involved the building of a BN for predicting coronary artery disease made with real world data obtained in a prospective cohort study. The study involved 302 patients from a hospital localized in southern Brazil. The second experiment assessed the ChronicPrediction support to multiple BNs at same time. The test involved the previous BN and another from a thirty part research work to map risk factors of diabetes. The results were encouraging and show potential for implementing ChronicPrediction in real-life situations.

1. Introduction Noncommunicable diseases (NCDs) are the main reason to the rise of diseases incidence in the developed world. The NCDs incidence has raised with high rates due the population aging and the increased longevity of people with many chronic conditions. These diseases create economic impacts in individuals, families, health systems and general society, once they affect people in their productive years, reducing the productivity and the ability of gain at domestic level (Puoane et al., 2008). The main NCDs include cardiovascular diseases, cancers, chronic respiratory diseases and diabetes. Visual impairment and blindness, hearing impairment and deafness, oral diseases and genetic disorders are other chronic conditions also classified as NCDs. ⁎

Corresponding author. E-mail addresses: [email protected] (J.L. Victória Barbosa), [email protected] (E. Butzen), [email protected] (M.Â. Gaedke), [email protected] (J.S. Dias da Costa), [email protected] (R.B. Scherer dos Santos). https://doi.org/10.1016/j.tele.2018.02.005 Received 7 November 2017; Received in revised form 14 February 2018; Accepted 15 February 2018 0736-5853/ © 2018 Published by Elsevier Ltd.

Please cite this article as: Pittoli, F., Telematics and Informatics (2018), https://doi.org/10.1016/j.tele.2018.02.005

Telematics and Informatics xxx (xxxx) xxx–xxx

F. Pittoli et al.

The causes of most NCDs are well established and known, which are mainly related to behavioral and biological risk factors. Among the most present risk factors are unhealthy diets, including the excessive ingestion of calories, lack of physical activity and smoking habits (WHO, 2005). Despite the clinical differences between each chronic condition, these diseases put patients and families to face the same needs, for example, change their behavior to tackle with the social impact of symptoms and interact with medical care all the time. In turn, the medical assistance should ensure that patients will receive the required treatment, information and support needed to control their diseases, minimize symptoms and manage their health (Wagner and Groves, 2002). Most people with NCDs struggle with physical, psychological and social needs of their diseases without much aid from health professionals. Frequently, the received help cannot give the right support nor comply with the people needs for an effective health management. This is further aggravated if the NCD needs a continuous treatment (Wagner et al., 2001; Bodenheimer et al., 2002). Hence, it is strategic that NCD patients have quick and straight access about their situation regardless of the time or location. In that sense, the rise of mobile devices with Internet access such as smartphones, offers great potential for easing the control and continuous tracking of diseases by the patients, since most people carry their smartphones everywhere. Furthermore, with ubiquitous Internet access they can get specialized aid whenever is needed (Google, 2012). Weiser (1991) called ubiquitous computing (UbiComp) the massive and integrated use of computers and sensors that interact seamlessly to create an always present computational platform. This high density of distributed computational devices generates a massive quantity of data that can be used to anticipate situations faced by NCD patients (Pejovic and Musolesi, 2015). For instance, the physical activity practice is related to acute myocardial infarction (Ôunpuu et al., 2015). Data collected by sensors embedded in smartphones may be used to monitor physical activity of chronic patients helping to anticipate an infarct. In this sense, we propose ChronicPrediction, a system for ubiquitous prognosis of NCD risk factors. ChronicPrediction uses Bayesian Networks (BNs) created from the mapping of causal relationship of risk factors. The BNs consider the behavioral data generated by the patients, while they use their smartphones. In this way, ChronicPrediction ubiquitously helps patients to be aware of their treatment, assessing whether the actions taken are helping to improve their health conditions or not. The main strength of ChronicPrediction resides in its capacity to deal with multiple types of NCDs, and its architecture that allows a near real time prediction of NCDs risks factors based on patients’ behaviors. On the other hand, the support of multiple NCDs by ChronicPrediction is not entirely automated. It needs the collaboration of health specialists that are responsible for designing BNs for each NCD. This article is organized in six sections. Section 2 presents an introduction to UbiComp and NCDs care. Section 3 addresses works that apply, in some manner, prediction in NCDs care. Section 4 describes the ChronicPrediction. Implementation aspects, assessment and results are covered in Section 5. Section 6 presents final remarks and directions for future works. 2. Background: ubiquitous computing, u-Health and NCDs Circulatory and respiratory diseases, cancers and diabetes, are considered the NCDs with most prevalence in a global level (WHO, 2005; WHO, 2008; Brazilian Health Ministry, 2008). These diseases belong to the group of chronic conditions that involves a large spectrum of health problems, such as long term mental disorders, permanent transmissible infections and continuous physical impairment (WHO, 2002). In general terms, these conditions have the same characteristics as they require “life style changes” and “long term health management”. Moreover, they are caused by lifestyles and risk behaviors, such as unhealthy diets, smoking habits, sedentary lifestyle, alcoholism, high-risk sexual behaviors and social stress. In addition to those behaviors factors, age and biological factors can influence the incidence of chronic conditions. The management, prevention and control of NCDs can be done through the care of behavioral and biological risk factors related to those diseases. According to World Health Organization (WHO, 2005), “if these risk factors were eliminated, at least 80% of all heart disease, stroke and type 2 diabetes would be prevented; over 40% of cancer would be prevented”. Moreover, about 25 years ago, Weiser (1991) introduced the concept of Ubiquitous Computing, or UbiComp, predicting a world where computing devices would be present in objects, environments and human beings themselves. These devices would interact naturally with the users without being noticed. Ten years after, Mahadev Satyanarayanan reinforced the concept through an article that would become a classic (Satyanarayanan, 2001). Context-Aware Computing has been considered a strategic research topic to support the UbiComp (Hoareau and Satoh, 2009). Context is any information that can be used to identify the situation of an entity (i.e. a person, a place or an object) and that has importance in the interaction between user and application (Dey et al., 2001). Among the goals of context-aware computing are the acquisition and usage of information of the physical world to then, select, configure and offer a variety of services in a meaningful way. Context-aware systems are fundamentally interested in acquiring context information (for example, by sensors) and understand this context (for example, by merging sensed perception for a particular context). This kind of system is able to fit its operations for the actual context without an explicit user intervention, and by that, raise its usability and effectiveness by taking into account the environment context (Baldauf et al., 2007). Context history, or trail (Barbosa et al., 2016), has been recognized as the collection of users’ past contexts. According to Ciaramella et al. (2010), context history is an important piece of information to recognize the user’s status. When recorded for a long time, such histories can offer the opportunity to infer users’ actions, and by doing so, reinforce the services offered by computational systems. In addition, prediction based on recorded contexts is an actual challenge for the context history subject (Barbosa et al., 2013; König et al., 2013; da Rosa et al., 2016). UbiComp has found application in a diverse range of knowledge areas, such as, health (Vianna and Barbosa, 2014), commerce (Barbosa et al., 2016), competence management (Rosa et al., 2015), learning (Barbosa et al., 2013; Barbosa et al., 2014; Abech et al., 2

Telematics and Informatics xxx (xxxx) xxx–xxx

F. Pittoli et al.

2016), logistics (Oliveira et al., 2013; Oliveira et al., 2015), accessibility (Tavares et al., 2015) and games (Oliveira et al., 2008). The UbiComp applied in the health field is also called u-Health. In general, u-Health applications are centered in hospital routine management (Orwat et al., 2010), patients monitoring (Agoulmine et al., 2011), or well-being (Buttussi and Chittaro, 2010). u-Health hospital applications focus on the enhancement of hospital staff routines such as medicine administration and prescription, medical team collaboration, and management of conferences, surgeries and emergencies rooms. Patients monitoring applications are commonly used for aiding the care of elder and chronic patients remotely. The well-being class encompasses those applications oriented to personal health management, such as diet management, personal training assistants and games oriented to the enhancement of well-being. 3. Prediction in NCDs care Today, ubiquitous systems used for NCDs support and care, prioritize patients monitoring and alerts generation. Others needed features for helping in the continuous treatment of these diseases, such as training patients’ or integrating them with health agents are rare in this type of systems. Another important aspect is that few ubiquitous systems give support to decision-making and recommendation, even specific systems for NCDs management and control. As the NCDs care must be done continuously, patients must be aware on the progress of their treatment and if the actions done by them daily are helping to improve health. For example, if the food consumed will aid in lowering their blood glucose levels, or if the amount of physical activities are enough to better control the cholesterol ratio. The approach based on prediction has begun to be analyzed by some researchers. However, these studies are oriented to specific diseases or risk factors (Wu et al., 2015; Suchánek et al., 2014; van der Heijden et al., 2013; Guo et al., 2012; Park and Cho, 2012), and not oriented to design an architecture or model to support multiple NCDs, in order to recommend actions for improving the patients’ treatment. This section reviews six works that use predictive models for the care of NCDs. The aim is to assess how they adhere to the notion ubiquitous support of NCDs. The works are compared in how they support health/well-being, risk factors analysis, multiple NCDs, daily activities assistance and mobile devices. We also assessed which technique they use, if they are context aware, and if they store patient history data for later use. Maxwell et al. (2017) compared how different deep learning architectures perform when classifying comorbidities. They used a medical center dataset in which each row represented a patient with results from different examination types and the classification of six chronic diseases. Seven deep learning architectures were tested to classify the presence of hypertension, diabetes or fatty liver. Deep Neural Network and Random Forest powered by Random k-Labelsets (Tsoumakas and Vlahavas, 2007) performed better. Wu et al. (2015) presented the design and results related to their artificial neural network (ANN) for predicting systolic blood pressure based on correlated factors, such as, age, gender, serum cholesterol (high-density lipoprotein, or HDL cholesterol; lowdensity lipoprotein, or LDL cholesterol; and triglycerides), fasting blood sugar and electrocardiograph signal. In their experiment, the authors have found the set of factors that better fit for systolic blood pressure prediction for the designed ANN. Suchánek et al. (2014) proposed a model for automatic learning of diagnoses. The model aims to generate new diagnoses based on a diagnosis history base. The history takes into account the symptoms reported by patients and past diagnosis done by physicians. A probabilistic model is then used to infer new diagnosis from the diagnosis history base. Yet, van der Heijden et al. (2013) presented their chronic care architecture and, in particular, their probabilistic model for detecting exacerbation in cases of chronic obstructive pulmonary disease (COPD). The authors define exacerbations as “acute events of worsening of COPD-related health status”. The model uses a Bayesian Network (BN) which was designed in conjunction with pulmonologists of the Radboud University Nijmegen Medical Centre. This BN was also trained and validated with data sets from COPD researches. A decision making system for helping middle aged people predict diabetes type 2 was presented by Guo et al. (2012). A Naive Bayes network was projected and built to support the prediction of diabetes type 2. Using the leave-one-out method, the authors obtained a prediction precision improvement in relation to other BNs used for predicting diabetes. Finally, Park and Cho (2012) proposed a model to predict metabolic syndrome. Metabolic syndrome is a set of risk factors that contributes to abdominal weight, insulin resistance, dyslipidemia and hypertension (Park and Cho, 2012). Their predictive model uses a BN and its design was aided by medical specialists. According to authors, the proposed method has a better prediction performance than other predictive methods studied. Table 1 summarizes the comparison among the related work. The item Support to health/well-being indicates if the work has any type of health support or assistance for its users, in order to improve their quality of life and management of their routine. The Risk factors analysis identifies if the work considers the features that influence the aggravation of some determined NCD. The prediction technique used is described by Used technique. The definition if the compared work can be applied for the caring of more than one NCD is shown in Support for multiple NCDs. The Daily activities assistance indicates if the work gives continuous support to users so they can monitor their health variables and habits. The definition if the work uses some type of context information to improve users’ experience is shown in Context awareness. The Context history data (trails) points if context history is used as an artifact for predicting future contexts (da Rosa et al., 2016). Finally, Support for mobile devices describes the usage of the technique in mobile devices. Nearly all works support context history data, except the proposals of Guo et al. (2012) and Park and Cho (2012). Furthermore, almost all related work are focused on predictive models, so the Daily activities assistance feature is not supported by them, except the system of van der Heijden et al. (2013), which supports this feature partially. The system has some kind of activity assistance, but those activities are more focused on particular aspects of the monitored disease, and not in those aspects that may lead to improve the 3

Telematics and Informatics xxx (xxxx) xxx–xxx

F. Pittoli et al.

Table 1 Comparison among the related work. Support type

Maxwell et al. (2017)

Wu et al. (2015)

Suchánek et al. (2014)

van der Heijden et al. (2013)

Guo et al. (2012)

Park and Cho (2012)

Support to health/wellbeing Risk factors analysis Used technique Support for multiple NCDs Daily activities assistance Context awareness Context history data (trails) Support for mobile devices

No

No

No

Yes

Yes

Yes

No Various Yes No No Yes No

No RNA No No No Yes No

No BN No No Yes No

Yes BN Yes Partially Yes Yes Yes

Yes BN No No No No No

Yes BN No No No No No

patients’ well-being, like diet or physical activities control. In addition, it is not clear if patients can manage activities, or create activities for managing other diseases or aspects of their own diseases. 4. ChronicPrediction ChronicPrediction is an intelligent system for monitoring NCDs risk factors. Monitoring can be understood by the prediction of improvement or worsening of such risk factors. The system uses the service infrastructure and knowledge representation defined by the U’Ductor (Vianna et al., 2017), which is a model for ubiquitous care of NCDs. The model was designed to provide services that are common in the prevention and treatment of such diseases. U’Ductor acts as a middleware for connecting NCDs care applications, i.e. it works as a “service bus” for NCDs care applications (Bhadoria et al., 2017a; Bhadoria et al., 2017b). So, U’Ductor tracks activities, notifying clients with useful information, and searching for nearby resources and people. ChronicPrediction enhances U’Ductor by coupling an engine that uses BNs to enable the prediction of future trends in the risk factors. This section is organized in two subsections. The first synthesizes the operation of the system and how it integrates with the U’Ductor, while the second details its components and how they work. 4.1. ChronicPrediction overview The ChronicPrediction works as an U’Ductor executable module which main goal is to monitor risk factors that influence NCDs. Furthermore, the system works along with the ChronicDuctor, which is a personal assistant for supporting NCDs care, communication between patients community and health organizations, and the searching for health care resources that are located nearby (Vianna and Barbosa, 2014). Fig. 1 shows the U’Ductor middleware stack, and how ChronicPrediction fits in U’Ductor as an executable module, running together with ChronicDuctor. The other parts of the middleware are responsible for the features of shared resources, context, access control, messaging, localization and peering with other U’Ductor nodes. Fig. 2 shows the dynamics between ChronicDuctor and ChronicPediction. ChronicDuctor is responsible for managing the patients’ care plan. In the plan is described the activities which the patients must follow to obtain success in their treatment. The care plan activities are scheduled by the ChronicDuctor (Fig. 2, step 1), so in the right time ChronicDuctor can notify the patients about activities that they must accomplish. Such activities can be the intake of a medicine, a meal, the practice of a physical activity or the

Fig. 1. U’Ductor Middleware.

4

Telematics and Informatics xxx (xxxx) xxx–xxx

F. Pittoli et al.

Fig. 2. ChronicPrediction model overview.

reading of a vital sign (step 2). After notifying the patients, the ChronicDuctor awaits for patients to insert the actions they took in the activity (step 3). For example, if the activity was to measure their blood pressure, ChronicDuctor will offer a form, so the patients can fill with their last blood pressure reading. The information entered by the patients in the ChronicDuctor are sent to ChronicPrediction (step 4). ChronicPrediction uses these information to generate predictions to aid patients in improving the quality of their NCDs treatment (step 5 and 6). Giving an example, for patients suffering from diabetes, ChronicPrediction will use the information entered to suggest them to increase the frequency which they practice physical activities or that they prioritize the consumption of food with low fat and low sugar.

4.2. ChronicPrediction architecture The ChronicPrediction architecture has three main process. These are shown in Fig. 3 as the external rectangles numbered from 1 to 3. The context acquisition process (rectangle number 1) encompasses the use by patients of U’Ductor applications such as ChronicDuctor. The U’Ductor context component implements a publish/subscribe pattern by which applications can listen to

Fig. 3. ChronicPrediction architecture model.

5

Telematics and Informatics xxx (xxxx) xxx–xxx

F. Pittoli et al.

contextual information provided by other applications or add new contextual information. Each contextual data has a type and the value is an instance of the type definition (Vianna et al., 2017). In the U’Ductor prototype contextual data are expressed using Resource Description Framework (RDF).1 That use generates trails, i.e. information about patient’s actions and visited contexts, which are then stored in the Historical trails repository. In the selection and data processing process (rectangle number 2) the global contexts subscriber acquires context data from patients’ smartphones by using the publish/subscribe pattern (Coulouris et al., 2011). These subscribed context data from the smartphones are stored in the unified contexts repository, which centralizes the storage of context data acquired by the global subscriber. These context data will then be queried by the trails selection module. The trails selection module queries the unified contexts repository for the most applicable data related to NCDs supported by the system. That’s because different NCDs types require different BNs which, by their turn, require inferring different types of information. Finally, the processing trails module creates the discretized data set that will be used for the BNs training, i.e., it does the transformation of continuous data into discrete data. The discretized data are then stored in the discretized data repository. In the inference and prediction process (rectangle number 3) specialists from the medical or nutritional area, or anyone that has the required knowledge, can create and train the BNs that will be later used for inference. Here, the generation and training Bayesian Networks module controls the BNs’ features of generation, modeling and training. The BN modeling accepts the reasoning models of cause and effect (Pearl, 2000), diagnoses’ inference (Agre, 1997) and inter-casual inference (Kjrulff and Madsen, 2007). After the BNs generation, they will be trained and calibrated with the discretized data generated by the processing trails module and stored in the Bayesian Networks repository. Finally, the Bayesian Networks inference module does predictions using the data generated by patients. Such data can be informed on forms by patients, or automatically collected by ChronicDuctor from sensors present in the users’ smartphones. Also, this module is responsible for selecting the BNs that are related to the data inputted by patients, doing multiple predictions. The generated predictions are interpreted, and the results of the interpretations are sent to users as feedbacks that indicate if the actions, habits and attitudes of users are in conformance with what is expected for their treatments. The Bayesian Networks repository and the Bayesian Networks inference module form the main pieces of ChronicPrediction, as they are the engine behind the capacity of the system in dealing with multiple NCDs. 5. ChronicPrediction evaluation As a way to evaluate the system, we built a BN for predicting coronary artery disease (CAD) and a prototype of the proposed architecture, which was functionally tested in two experiments. The BN training and the first experiment were done with the use of real data that came from the prospective cohort project of a large philanthropic hospital from Porto Alegre, Brazil. The second experiment was done to test the capacity of the system in dealing with multiple NCDs using multiple BNs. This section will describe the ChronicPrediction prototype implementation, the BN implementation aspects for CAD prediction, the prospective cohort project from which came the data for the evaluation and the two experiments that were applied. In both experiments, scenarios were used to simulate possible decisions done by patients on their care activities, and so assess the consistency of ChronicPrediction’s inferences. 5.1. ChronicPrediction’s implementation aspects The mobile application used in the functional test was developed as a U’Ductor middleware’s executable module (Vianna et al., 2017). The services of the global contexts subscriber and of the Bayesian Networks inference module were developed following the REST architectural style (Fielding and Taylor, 2000) with aid of Jersey2 package available for Java. The BNs were created with the aid of the GeNIe3 tool, and its inference process was realized with the jSMILE4 library. Both allow to create, edit and save graphical models for using in probabilistic reasoning and decision making under uncertainty. 5.2. Location of study and data description The data used for generating the BN as also to evaluate ChronicPrediction came from a prospective cohort study ran by a hospital located at Porto Alegre City, Rio Grande do Sul State, Brazil. It has 400 beds, more than 1700 employees, and a built area of 50,000 m2. In addition, it has diagnoses resources of cutting edge technology. The prospective cohort study (Franken et al., 2012) included patients with 30 years or older, of both genders, egressed from the Vascular Medicine Institute, residents at the Rio Grande do Sul State, and diagnosed with Acute Coronary Syndrome (ACS) including unstable angina, acute myocardial infarction, i.e., (AMI), without elevation of the ST segment and AMI with elevation of the ST segment. Patients admitted and discharged between may 2009 to may 2011 and matching the inclusion criteria were added to the 1

https://www.w3.org/RDF/. http://Jersey.java.net. 3 http://genie.sis.pitt.edu/. 4 http://genie.sis.pitt.edu/wiki/Introduction_to_jSMILE. 2

6

Telematics and Informatics xxx (xxxx) xxx–xxx

F. Pittoli et al.

study. In the end, 302 patients diagnosed with ACS where included in the study. Three models of standardized and precoded questionnaires were used for the data acquisition of the cohort study: one baseline questionnaire and two follow-up questionnaires, applied at six months and one year after discharge. In the baseline questionnaire was listed problems and diagnoses that led to patient’s hospitalization, allowing the identification of those with ACS (in all its clinical forms). This questionnaire was applied by interviewers directly to patients during hospitalization. Physical examination data and laboratory information were collected from the patients’ health records. The study logistics ensured that all information were available to interviewers. By that questionnaire was also collected data about reported morbidities and socioeconomic, demographic and behavioral profiles. Data about medicine use was implicit and refers to the medicines that the patients were taking at time of hospitalization. So, information about prescribed medicines at discharge were not acquired. Therefore, a search was realized in every patient’s health record of the study, aiming to collect data about prescribed medicines at the patients’ last day of hospitalization. The follow-up questionnaires were realized by telephone interviews, which is a reliable and frequently used method for collecting data from epidemiological studies (Frey and Oishi, 1995; McColl et al., 2002; Nelson et al., 2003). The first follow-up questionnaire was applied six months after patients’ discharge, and a last questionnaire was applied one year after discharge. These tools were used to look for information about occurrence of death, number of medical visits in the period, hospital readmission, occurrence of complications, interventions, life style, medicines usage and results of laboratory tests. After the data collection, the questionnaires were coded by the interviewers and revised by the field supervisor. After the questionnaires were ready, they were forwarded for double data entry processing, and finally, their correction.

5.3. Bayesian network for predicting coronary artery disease The control of risk factors can reduce coronary diseases and their mortality. This control can explain why such mortality decayed in some countries and raised in others (Piegas et al., 2003). For the creation of a BN for CAD prediction was needed to identify the main risk factors that cause CAD and then, define in which way those risk factors are related with each other. The Framingham Score (Lotufo, 2013; Polanczyk, 2005) was used as a scientific base for the acquisition of risk factors. This score measures the likelihood of occurrence of coronary disease by the use of variables as age, gender, systolic blood pressure, cholesterol ratio, smoking habits and diabetes. In addition to the Framingham Score, we consulted four medical and nutritional specialists from Collective Health Graduate Program from University of Vale do Rio dos Sinos (Unisinos), about the risk factors implication in CAD. Thus, the following risk factors were defined as the main involved with CAD: Smoking; Systolic Blood Pressure; Diastolic Blood Pressure; Total Cholesterol; HDL Cholesterol; Blood Sugar; Gender; Weight; Waist Measurement; Physical Activity Practice; NCDs Family History; and Age. The intervals of values used to classify the risk factors into discrete values are listed in Table 2. This classification was based in the medical literature (Cacoub et al., 2011; Santos Filho and Martinez, 2002; Santos et al., 2002; Sposito et al., 2013). After this classification, the casual and effect relations used in the CAD BN were defined by the medical and nutritional specialists. The designed BN is shown in the Fig. 4. A proper data set is required if realistic results are wanted and, for that reason, a data set containing real data was used in the BN training. Moreover, the usage of real data gives the BN model the ability to make coherent predictions and diagnoses. For this purpose, we used the data from (Gaedke, 2013) which were derived from the “Prospective Cohort Study of Patients with Acute Coronary Syndrome” realized at the Vascular Medicine Institute of the large philanthropic hospital from Porto Alegre, Brazil. Then we applied statistical validation tests on the data from the Prospective Cohort Study, to check if the data show similarity with the conditions presented in real world. For this purpose, we assessed the agreement of the data with the conclusions presented by the Framingham Heart Study by doing a statistical validation and an external validity assessment.

Table 2 CAD risk factors and their values intervals. Risk factor

Systolic Blood Pressure Diastolic Blood Pressure Total Cholesterol Blood Sugar Weight Waist Measurement Age HDL Cholesterol

Values intervals Normal

High

< 140 mmHg < 90 mmHg < 200 mg/dl < 110 mg/dl < 25 value of BMI < 102 cm (Men) < 88 cm (Women) < 45 years (Men) < 55 years (Women) Low < 40 mg/dl

> =140 mmHg > =90 mmHg > =200 mg/dl > =110 mg/dl > =25 value of BMI > =102 cm (Men) > =88 cm (Women) > =45 years (Men) > =55 years (Women) Normal > =40 mg/dl

7

Reference Cacoub et al. (2011) Cacoub et al. (2011) Cacoub et al. (2011) Cacoub et al. (2011) Santos et al. (2002) Santos et al. (2002) Santos Filho and Martinez (2002) Sposito et al. (2013)

Telematics and Informatics xxx (xxxx) xxx–xxx

F. Pittoli et al.

Fig. 4. Bayesian network for predicting coronary artery disease.

5.3.1. Statistical validity As a way to ensure the data set statistical validity a chi-squared test and a significance test for logistic regression were done. Both tests were done with aid of the R language,5 that is suited for statistical computation, and provides a large set of techniques and statistical tests. The data set used for the tests had the following variables: Smoking, Systolic Blood Pressure, Diastolic Blood Pressure, Total Cholesterol, HDL Cholesterol, Blood Sugar, Gender, Body Mass Index (BMI), Weight, Waist Measurement, Leisure, Age and CAD. The variables Systolic Blood Pressure, Diastolic Blood Pressure, Total Cholesterol, HDL Cholesterol, Weight, Waist Measurement and Age were assumed to be dichotomous, once they were previously discretized by the rules defined in Table 2. The variables Smoking, Gender, Leisure and CAD are naturally dichotomous. Finally, BMI was the only continuous variable in the data set. Before beginning the validation we removed 41 records from a total of 1,005 from the Cohort data set because they had inconsistencies in their fill – i.e. the value of one or more variables were not set or had some invalid value. The chi-squared test was run in the dichotomous variables as way to test their independence in relation to the dependent variable (Cook and Campbell, 1979). In this way, we assumed the variables Smoking, Systolic Blood Pressure, Diastolic Blood Pressure, Total Cholesterol, HDL Cholesterol, Blood Sugar, Gender, Weight, Waist Measurement, Leisure and Age as independent dichotomous variables and CAD as a dependent variable. With exception of the variable Gender, it was indicated a dependence between the dichotomous variables, with a p-value < 0.05 (i.e. p-value < 0.05). The test run on the Gender variable resulted in p-value equals to 0.62 (i.e. p-value = 0.62). A significance test for logistic regression was run to assess the dependence of the variable BMI in relation to the CAD variable (Cook and Campbell, 1979). Such test is most suited when there is a dichotomous variable (CAD) and a quantitative variable (BMI). The significance test for logistic regression resulted in a p-value lower than 0.05 (i.e. p-value < 0.05), what indicates that the variable BMI is statistically significant over the CAD variable. 5.3.2. External validity The external validity should indicate if the proposed model has condition to be used in broader contexts (Farias et al., 2014), i.e. if the BN for CAD can be applied in other situations, not only for patients of philanthropic hospital. In this way, we need to reinforce that the proposed BN model is based in risk relations defined by the Framingham Heart Study. This is an ongoing study, which has begun in 1948, and showed the possibility to establish the risk of myocardial infarction based on age, gender, systolic blood pressure, 5

https://www.r-project.org.

8

Telematics and Informatics xxx (xxxx) xxx–xxx

F. Pittoli et al.

Table 3 Simulation results. Scenario

Expected result

Interaction number

Activity

Measurement Last

New

P(CAD) Before

After Insertion

Subject 1

Improvement

1

2 Subject 2

Improvement

1 2

3

Subject 3

Worsening

1 2

Subject 4

Worsening

1

2

30 min walk? None Yes High (99%) High (99%) Feedback: The daily practice of exercises is a key component to acquire a better health and contributes to your well being. Your treatment for Coronary Artery Disease is not being properly conducted. Follow your care plan carefully and see your doctor. Weight in 83 kg 64 kg High (99%) Normal (68%) Feedback: Congratulations! You are following your treatment and care plan properly. Keep the good work! Smoking? Yes No High (97%) Normal (74%) Feedback: None. Measuring blood 150/100 mmHg 150/100 pressure Feedback: High blood pressure is one of the biggest risk factors for cardiovascular disease, increasing risk o circulatory problems which can be fatal.Your treatment for Coronary Artery Disease is not being properly conducted. Follow your care plan carefully and see your doctor. Measuring blood 150/100 mmHg 130/80 mmHg Normal (74%) Normal (74%) pressure Feedback: Congratulations! Your treatment for Coronary Artery Disease is being properly conducted and your care plan is being followed correctly. Smoking? No Yes Normal (97%) High (87%) Feedback: None. Weight in 132.27 lb 165.34 lb High (87%) High (87%) Feedback: Have the ideal weight contributes to a better health and well-being. Your treatment for Coronary Artery Disease is not being properly conducted. Follow your care plan carefully and see your doctor. Measuring blood sugar 98 mg/dl 118 mg/dl Normal (97%) High (55%) Feedback: Keeping your blood sugar level lower than 99 mg/dl is crucial to your health. Higher levels indicate risks for the incidence of other noncommunicable diseases. Your treatment for Coronary Artery Diseases is not being properly conducted. Follow your care plan carefully and see your doctor. Measuring cholesterol 195 mg/dl 218 mg/dl High (55%) High (75%) Feedback: A high cholesterol is one of the main risks for the incidence of heart diseases. Your treat for Coronary Artery Disease is not being properly conducted. Follow your care plan carefully and see your doctor.

cholesterol ratio, smoking and diabetes (Lotufo, 2013). Thus, since the statistical tests demonstrated the same relations that are presented in the Framingham Heart Study, we can state that the BN model has external validity. Finally, through the statistical assessment we guaranteed that the data set used for BN training includes the risk relations for CAD, established by the Framingham Heart Study, in other words, it is suited for training a BN which predicts Coronary Artery Disease. The fact that the Gender variable is independent from the CAD variable was expected, which demonstrates the existence of a balance in the population in relation to the gender. Furthermore, as presented in the Framingham Heart Study, gender is not a straight indicator of CAD risk, but a conditional indicator of this kind of risk. In other words, the CAD risk will be greater or lower according to gender, if the other variables indicate the same. Once the results of tests are in accordance with Framingham Heart Study, we believe that the data set is statistically valid. Next section will describe the study that originated the data for the BN training with better details. 5.4. ChronicPrediction’s usage simulation with prospective cohort study data With aims of evaluating the relevance of the ChronicPrediction use in the care of NCDs, four scenarios were designed for functionally test the system. Four individuals were selected from the prospective cohort study data provided by the Unisinos Collective Health Graduate Program. By the time of evaluation, the selected individuals were identified as “Subject 1”, “Subject 2”, “Subject 3” and “Subject 4”. The four patients were selected according to the diversity of their profiles in relation to the variables age, gender, BMI, physical activity practice, blood sugar, blood pressure, cholesterol and NCDs history. For example, considering the age characteristic, we selected one young individual, one middle age individual and two elders. Table 3 show the scenarios used in simulation. Each scenario was designed to test the use of ChronicPrediction as a tool for predicting changes of CAD risk in near real-time, and so, helping patients to acquire a better lifestyle. In this way, the scenarios Subject 1 and Subject 2 show behaviors which lower risks of CAD (column Expected Result = Improvement). Whereas the scenarios Subject 3 and Subject 4 address the opposite, i.e., behaviors that increase the risks (column Expected Result = Worsening). A care plan was created for each patient. The care plan describes the activities and goals that the patient must accomplish, and also defines mechanisms and tools that the patient can use to accomplish these activities. In these plans were included those activities that were straight related with CAD risk factors but that are also related with other NCDs. Such activities, if are followed and accomplished, tend to bring great gains for the subjects’ health. Another important facet is that the proposed activities belong to the 9

Telematics and Informatics xxx (xxxx) xxx–xxx

F. Pittoli et al.

modifiable risk factors group, and thus, may change during the subject treatment. After the definition of the care plan, ChronicPrediction usage simulation was run for each defined subject. First, the subject’s care plan was scheduled in the ChronicDuctor. Thus, the ChronicDuctor could notify the patients about care plan activities they must accomplish at the planned moment. Each interaction with application is identified by column “Interaction number”, and is composed by a activity notification, data input and a feedback. Along with each activity notification is shown a form where the patients must input information about the activity. For example, a “weight in” activity will bring a field for weight insertion. The inserted activity data for each subject is shown in the column “Measurement”, sub-column “New” of Table 3. The activity data inserted in ChronicDuctor is used by ChronicPrediction to infer the CAD likelihood in near real-time. After the inference, ChronicPrediction will send a feedback message for counseling the patients if they did not reach their planned goal. Otherwise, a congratulation message will be given to the patients. The CAD likelihood and the feedback messages resulting from the simulation are shown in the columns “P(CAD)” and “Feedback” of Table 3. The individual in charge of the simulation was responsible for executing the following activities in each scenario: create the care plan; execute the prototype; and input the data for the related activities in ChronicDuctor. Next paragraphs describe the results obtained in the four simulations starting by the scenario Subject 1. When the simulation starts, ChronicDuctor reminds the Subject 1 about a planned 30 min of walk activity. The Subject 1 checks this activity as done and, by doing this, the ChronicPrediction is activated by this new context information. So, ChronicPrediction uses this information to assess the CAD risk probability of the patient, which yet is high (see Table 3, Interaction number 1). Then, ChronicPrediction shows a feedback message informing about her CAD risk (Fig. 5(a)). After a while, ChronicPrediction notifies Subject 1 about a new planned activity. This activity asks Subject 1 to inform her weight. Once the weight is inputted, ChronicPrediction uses this new context information to assess Subject’s 1 CAD risk probability again. Due to her new weight, Subject 1 CAD risk probability has decreased (see Table 3, Interaction number 2), which makes ChronicPrediciton to show her a congratulation message (Fig. 5(b)). When the second simulation has started a change in the Subject 2 profile was made indicating that she has quit smoking. In consequence of that, the P(CAD) changed from a high risk of 97%, to a normal risk of 74% (see Table 3, Interaction number 1). After, the subject receives a notification about an activity to monitor blood pressure, and then ChronicDuctor is used to record a new blood pressure measurement. The subject then records a new measure with a value of 150/100 mmHg (same as the previous) and so, maintaining the P(CAD) with a normal risk of 74% (see Table 3, Interaction number 2). Then, the ChronicPrediction sends a message to Subject 2 (Fig. 5(c)) alerting her that she did not reach an acceptable value for this vital sign. When the ChronicDuctor notifies the Subject 2 about a new monitor blood pressure activity, the value 130/80 mmHg is recorded. This value is different from the previous, even so, the P(CAD) did not change. However, the subject has reached a more acceptable level for blood pressure measurement, and so, the ChronicPrediction sends her a congratulation message (Fig. 5(d)). Before starting Subject 3 scenario, the individual in charge of the simulation set the profile of this subject to smoker (see Table 3, Interaction number 1). When the simulation started, the ChronicDuctor requests Subject 3 for weight insertion. Once the weight is inserted, ChronicPrediction is activated to assess the CAD risk probability of Subject 3. In this time, the inferred probability for the subject is already high, once it was aggravated since Subject 3 became a smoker. Therefore, ChronicPrediction has sent a feedback message which is shown in Fig. 5(e). Finally, the Subject 4 scenario simulation starts and ChronicDuctor requests the subject to insert her last blood sugar measurement. The inserted measurement is higher than the last (see Table 3, Interaction number 1), what made the ChronicPrediction generate the feedback presented in Fig. 5(g). Following the simulation, ChronicDuctor notifies Subject 4 of a new activity. At this time the subject needs to insert a new cholesterol measurement. Again, after the insertion of the data, ChronicPrediction notifies the subject of her condition (Fig. 5(h)), once the CAD risk probability of the subject increased due to her new cholesterol measurement, that has increased too (see Table 3, Interaction number 2).

5.5. ChronicPrediction’s usage with multiple NCDs In this second evaluation, we tested the capacity of the system in dealing with two NCDs using different BNs. For this, we designed a scenario where a patient suffers from CAD and diabetes type 2. The BN used to infer CAD was the same of the first evaluation, while we used the BN from Guo et al. (2012) to infer diabetes. To train the diabetes BN we used the Pima Indians Diabetes Data Set (Knowler et al., 1990). Table 4 shows the patient’s profile and the desirable health values. The health parameters age, gender, height, smoker, physical activity and total cholesterol were inserted into the prototype before the beginning of the test. Once the test started, the prototype asks the patient’s systolic and diastolic blood pressure (Fig. 6). Based on the inserted blood pressure values, the ChronicPrediction has sent two feedbacks to the patient. These are shown in Fig. 7(a) and (b). Both feedbacks were given due to patient’s hypertension risk. Fig. 7(a) was the feedback given based on diabetes, while Fig. 7(b) was based on CAD. After a while, the prototype asked the patient her weight. So, the value 171.96 lb was inserted. Due to patient’s overweight risk, the ChronicPrediction has sent two more feedbacks to the patient. One related to the implications of overweight in diabetes (Fig. 7)) and the other related to CAD (Fig. 7(d)). This section showed the capability of ChronicPrediction to deal with more than one disease simultaneously. Though the test demonstrates the functionality with two diseases, we can state that same is applicable to more diseases, as far as the system has the BN correspondent to each disease. 10

Telematics and Informatics xxx (xxxx) xxx–xxx

F. Pittoli et al.

(a) Feedback sent to (b) Feedback sent to (c) Feedback sent to Subject 1 after inform- Subject 1 after weight Subject 2 after the first ing about 30 min walk insertion blood pressure insertion activity completion

(d) Feedback sent to (e) Feedback sent to (f) Feedback sent to Subject 2 after the Subject 3 after weight Subject 4 after blood second blood pressure insertion sugar insertion insertion

(g) Feedback sent to Subject 4 after cholesterol insertion Fig. 5. Notifications sent to subjects after data insertion.

6. Conclusions The U’Ductor model (Vianna et al., 2017) showed a ubiquitous architecture to support the care of NCDs. A prototype of the model and an application that runs over this prototype were built and evaluated by ten chronic patients. These patients used the application and answered to a questionnaire based in technology acceptance model (TAM) (Davis, 1989). The results showed a good acceptance of the proposed model, since a general agreement of 95% was obtained in perceived usefulness acceptance assessment. The ChronicPrediction extends the U’Ductor model by adding an architecture that measures the effects of the actions taken by its users. This feature enables the U’Ductor to anticipate health situations of its users, and so, allowing them to modify their lifestyles in advance, preventing future health complications. 11

Telematics and Informatics xxx (xxxx) xxx–xxx

F. Pittoli et al.

Table 4 Patient’s data. Health parameter Age Gender Height Weight BMI Smoker Physical Activity Blood sugar Systolic Blood Pressure Diastolic Blood Pressure Total Cholesterol HDL Cholesterol NCD Family History NCD

Health value 35 Female 5.41 ft 171.96 lb 28.65 No Yes 98 mg/dl 150 mmHg 100 mmHg 120 mg/dl 65 mg/dl No CAD Diabetes

Desirable value < 55 for women and < 45 for men – – – < 25 No Yes < 110 mg/dl < 140 mmHg < 90 mmHg < 200 mg/dl > 60 mg/dl No – –

Fig. 6. Systolic and diastolic blood pressure form.

As seen in Section 3, ChronicPrediction was designed to provide support to health/well-being, risk factors analysis, multiple NCDs, daily activities assistance, context-aware, context history data and mobile devices. The support to health/well-being is done continuously by the assessment of patients behaviors in order to identify the improvement or worsening of their health. Section 5.4 and 5.5 explained how ChronicPrediction supports risk factors analysis. Section 5.5 has also presented how support to multiple NCDs is reached by ChronicPrediction. Furthermore, the use of the U’Ductor model empowers ChronicPrediction to offer patients daily activities assistance, as they can manage their care plan, and also enables them to search for resources and people that may help in their treatment. The near-real time notifications done by ChronicPrediction is made possible by the context-aware architecture of U’Ductor model, which automatically transmits contexts change notifications to ChronicPrediction’s modules. Context history data were used for purpose of training the BNs. We simulated this feature by using datasets from the prospective cohort project (Gaedke, 2013) and the Pima Indians Diabetes Data Set (Knowler et al., 1990). This process was detailed in 5.3, 5.2 and 5.5. As ChronicPrediction was developed as an android application, it naturally gives support to mobile devices. The main strength of ChronicPrediction resides in its capacity to deal with multiple types of NCDs, as in its architecture that allows a near real time prediction of NCDs risks factors based on patients behaviors. This is a novel contribution that was not found in any of the related work. On the other hand, the support to multiple NCDs by ChronicPrediction is not entirely automated. It needs the 12

Telematics and Informatics xxx (xxxx) xxx–xxx

F. Pittoli et al.

(a) Hypertension feedback for a (b) Hypertension feedback for a patient with diabetes patient with CAD

(c) Overweight feedback for a pa- (d) Overweight feedback for a patient with CAD tient with diabetes Fig. 7. Multiple disease feedback.

collaboration of health specialists that are responsible for designing BNs for each NCD. Even with this weakness, ChronicPrediction is ahead from the compared works as it is deployed as a mobile platform, it is a context-aware platform and it has support for daily activities assistance. These features were not present in the compared works. Another contribution of this work was the BN that calculates CAD probability. This BN was created in partnership with health specialists from the Unisinos Collective Health Graduate Program. To compute the CAD likelihood, the BN uses data from risk factors recorded by the patients in the ChronicDuctor. The BN was trained with real data from a prospective cohort project of a large philanthropic hospital from Porto Alegre, Brazil (Gaedke, 2013). Two tests were made to evaluate ChronicPrediction. In the first test we could assess the ChronicPrediction system dynamics and the CAD BN throughout four scenarios fed with real data from the prospective cohort project. These scenarios were configured to demonstrate behaviors that could cause the increasing or lowering of CAD risks. The second test was done to demonstrate how ChronicPrediction can deal with different diseases. For this, we created a scenario with a patient that suffers from CAD and diabetes type 2. In this case, the system was configured to use the CAD BN to infer CAD risk and, to infer diabetes type 2 risk, we used the BN from Guo et al. (2012) trained with the Pima Indians Diabetes Data Set. By the tests results we concluded that ChronicPrediction is able to maintain users aware of those behaviors that might alleviate or aggravate their diseases. Although the simulations could be used to prove the feasibility of the model, we still need to evaluate effectiveness and hazards of its use. In this sense, we intend to run an effectiveness evaluation of the system by patients that suffer from CAD. The evaluation will be conducted jointly with the Collective Health Graduate Program of Unisinos. In this evaluation, CAD patients will be recruited and assigned to different groups. The individuals from one group will use ChronicPrediction as their personal health assistant. The other group will receive normal medical guidance. After one year the results of each group will be compared. The work could also be enriched with an evaluation of the technology acceptance. Upon execution, the users would answer an assessment questionnaire based on the Technology Acceptance Model (TAM). TAM measures the satisfaction through perceived usefulness and perceived ease of use. The TAM model has been considered a standard to evaluate the acceptance of new technologies (Marangunić and Granić, 2015). Furthermore, a long-term experiment with patients would support a more complete assessment of the effectiveness of the model and would bear further studies based on user behavior, such as the influence of the health notifications in engagement of patients in the continuity of their care routine. Also, a hazard analysis may be conducted to identify possible risks and side effects that may occur while using the proposed approach. In addition, we want to address how to automatically or semi automatically generate models for predicting NCDs risk worsening, once this is one of the weakness of this work. Furthermore, it is also relevant to investigate the use of other predictions models besides BNs. Also, ChronicPrediction may be improved in providing a way for health professionals customize the recommendation messages, as they currently are very specific.

13

Telematics and Informatics xxx (xxxx) xxx–xxx

F. Pittoli et al.

References Abech, M., da Costa, C.A., Barbosa, J.L.V., Rigo, S.J., da Rosa Righi, R., 2016. A model for learning objects adaptation in light of mobile and context-aware computing. Personal Ubiquitous Comput. 20 (2), 167–184. Agoulmine, N., Deen, M., Lee, J.-S., Meyyappan, M., 2011. U-health smart home. IEEE Nanatechnol. Mag. 5 (3), 6–11. http://dx.doi.org/10.1109/MNANO.2011. 941951. URL: < http://ieeexplore.ieee.org/lpdocs/epic03/wrapper.htm?arnumber=5993590 > . Agre, G., 1997. Diagnostic bayesian networks. Comput. Artif. Intell. 16 (1), 47–67. Baldauf, M., Dustdar, S., Rosenberg, F., 2007. A survey on context-aware systems. Int. J. Ad Hoc Ubiquitous Comput. 2 (4), 263–277. http://dx.doi.org/10.1504/ IJAHUC.2007.014070. Barbosa, D.N.F., Barbosa, J.L.V., Bassani, P.B.S., Rosa, J., Martins, M., Nino, C., 2013. Content management in a ubiquitous learning environment. Int. J. Learn. Technol. 46, 24–35. http://dx.doi.org/10.1504/IJCAT.2013.051385. Barbosa, J.L.V., Barbosa, D.N.F., Oliveira, J.M.D., Rabello, S.A., Jr, 2014. A decentralized infrastructure for ubiquitous learning environments. j-jucs 20 (12), 1649–1669. Barbosa, J.L.V., Martins, C., Franco, L.K., Barbosa, D.N.F., 2016. Trailtrade: a model for trail-aware commerce support. Comput. Ind. 80, 43–53. Bhadoria, R.S., Chaudhari, G.S., Tomar, N., Singh, S., 2017. Exploring Enterprise Service Bus in the Service-Oriented Architecture Paradigm. IGI Global. Bhadoria, R.S., Chaudhari, N.S., Tomar, G.S., 2017b. The performance metric for enterprise service bus (esb) in soa system: theoretical underpinnings and empirical illustrations for information processing. Inf. Syst. 65 (Suppl. C), 158–171. http://dx.doi.org/10.1016/j.is.2016.12.005. URL: < http://www.sciencedirect.com/ science/article/pii/S0306437915301952 > . Bodenheimer, T., Wagner, E.H., Grumbach, K., 2002. Improving primary care for patients with chronic illness. J. Am. Med. Assoc. 288 (15), 1909–1914. Brazilian Health Ministry, 2008. Diretrizes e recomendações cuidado integral de doenças crônicas não-transmissíveis: Promoção da saúde, vigilância, prevenção e assistência,. Buttussi, F., Chittaro, L., 2010. Smarter phones for healthier lifestyles: an adaptive fitness game. Pervasive Comput. IEEE 9 (4), 51–57. http://dx.doi.org/10.1109/ MPRV.2010.52. Cacoub, P.P., Zeymer, U., Limbourg, T., Baumgartner, I., Poldermans, D., Rother, J., Bhatt, D.L., Steg, P.G., 2011. Effects of adherence to guidelines for the control of major cardiovascular risk factors on outcomes in the reduction of atherothrombosis for continued health (reach) registry Europe. Heart 97 (8), 660–667. Ciaramella, A., Cimino, M.G.C.A., Lazzerini, B., Marcelloni, F., 2010. Using context history to personalize a resource recommender via a genetic algorithm. In: 2010 10th International Conference on Intelligent Systems Design and Applications, IEEE. IEEE, pp. 965–970. Cook, T., Campbell, D., 1979. Quasi-experimentation: Design & Analysis Issues for Field Settings, Houghton Mifflin. URL: < https://books.google.com.br/books?id= BFNqAAAAMAAJ > . Coulouris, G., Dollimore, J., Kindberg, T., Blair, G., 2011. Distributed Systems: Concepts and Design, fifth ed. Addison-Wesley Publishing Company, USA. da Rosa, J.H., Barbosa, J.L., Ribeiro, G.D., 2016. Oracon: an adaptive model for context prediction. Expert Syst. Appl. 45, 56–70. http://dx.doi.org/10.1016/j.eswa. 2015.09.016. URL: < http://www.sciencedirect.com/science/article/pii/S0957417415006302 > . Davis, F.D., 1989. Perceived usefulness, perceived ease of use, and user acceptance of information technology. MIS Q. 319–340. Dey, A.K., Abowd, G.D., Salber, D., 2001. A conceptual framework and a toolkit for supporting the rapid prototyping of context-aware applications. Hum. Comput. Interact. 16 (2), 97–166. http://dx.doi.org/10.1207/S15327051HCI1623402. Farias, K., Garcia, A., Lucena, C., 2014. Effects of stability on model composition effort: an exploratory study. Softw. Syst. Model. 13 (4), 1473–1494. http://dx.doi. org/10.1007/s10270-012-0308-2. Fielding, R.T., Taylor, R.N., 2000. Principled design of the modern web architecture. In: Proceedings of the 22nd International Conference on Software Engineering, ICSE ’00. ACM, New York, NY, USA, pp. 407–416. http://dx.doi.org/10.1145/337180.337228. Franken, D.L., Olinto, M.T.A., Paniz, V.M.V., Henn, R.L., Junqueira, L.D., da Silveira, F.G., Roman, V.R., Manenti, E.R.F., Dias da Costa, J.S., 2012. Behavioral changes after cardiovascular events: a cohort study. Int. J. Cardiol. 161 (2), 115–117. http://dx.doi.org/10.1016/j.ijcard.2012.06.033. Frey, J., Oishi, S., 1995. How to Conduct Interviews by Telephone and in Person, How to Conduct Interviews by Telephone and in Person. Sage Publications. URL: < https://books.google.com.br/books?id=31IfAQAAIAAJ > . Gaedke, M.A., 2013. Uso de medicamentos de prevenção secundária após síndrome coronariana aguda (Master’s thesis). Universidade do Vale do Rio dos Sinos, Programa de Pós-graduação em Saúde Coletiva, São Leopoldo, RS. Google, May 2012. Our Mobile Planet. < http://www.mmaglobal.com/files/BrazilEnglish.pdf > . Guo, Y., Bai, G., Hu, Y., 2012. Using bayes network for prediction of type-2 diabetes. In: Internet Technology And Secured Transactions, 2012 International Conferece For. IEEE, pp. 471–472. Hoareau, C., Satoh, I., 2009. Modeling and processing information for context-aware computing: a survey. New Gener. Comput. 27 (3), 177–196. http://dx.doi.org/10. 1007/s00354-009-0060-5. Kjrulff, U., Madsen, A., 2007. Bayesian Networks and Influence Diagrams: A Guide to Construction and Analysis, Information Science and Statistics. Springer. URL: < http://books.google.com.br/books?id=BB4lL6BK7IC > . Knowler, W.C., Pettitt, D.J., Saad, M.F., Bennett, P.H., 1990. Diabetes mellitus in the pima indians: incidence, risk factors and pathogenesis. Diab. Metab. Rev. 6 (1), 1–27. König, I., Klein, B.N., David, K., 2013. On the stability of context prediction. In: Proceedings of the 2013 ACM Conference on Pervasive and Ubiquitous Computing Adjunct Publication. ACM, pp. 471–480. Lotufo, P.A., 2013. O escore de risco de framingham para doenças cardiovasculares. Rev. Med. 87 (4), 232–237. Marangunić, N., Granić, A., 2015. Technology acceptance model: a literature review from 1986 to 2013. Univ. Access Inf. Soc. 14 (1), 81–95. http://dx.doi.org/10. 1007/s10209-014-0348-1. Maxwell, A., Li, R., Yang, B., Weng, H., Ou, A., Hong, H., Zhou, Z., Gong, P., Zhang, C., 2017. Deep learning architectures for multi-label classification of intelligent health risk prediction. BMC Bioinf. 18 (14), 523. McColl, E., Jacoby, A., Thomas, A., Soutter, A., Bamford, A., Steen, A., Thomas, A., Harvey, A., Garratt, A., Bond, A., Jan. 2002. Design and use of questionnaires: a review of best practice applicable to surveys of health service staff and patients. Nelson, D.E., Powell-Griner, E., Town, M., Kovar, M.G., 2003. A comparison of national estimates from the national health interview survey and the behavioral risk factor surveillance system. Am. J. Public Health 93, 1335–1341. Oliveira, R.R., Noguez, F.C., Costa, C.A., Barbosa, J.L., Prado, M.P., 2013. Swtrack: an intelligent model for cargo tracking based on off-the-shelf mobile devices. Expert Syst. Appl. 40 (6), 2023–2031. http://dx.doi.org/10.1016/j.eswa.2012.10.021. URL: < http://www.sciencedirect.com/science/article/pii/ S0957417412011359 > . Oliveira, R.R., Cardoso, I.M., Barbosa, J.L., da Costa, C.A., Prado, M.P., 2015. An intelligent model for logistics management based on geofencing algorithms and RFID technology. Expert Syst. Appl. 42 (15–16), 6082–6097. http://dx.doi.org/10.1016/j.eswa.2015.04.001. URL: < http://www.sciencedirect.com/science/article/ pii/S0957417415002316 > . Orwat, C., Rashid, A., Holtmann, C., Wölk, M., Scheermesser, M., Kosow, H., Graefe, A., 2010. Adopting pervasive computing for routine use in healthcare. IEEE Pervasive Comput. 9 (2), 64–71. Ôunpuu, S., Negassa, A., Yusuf, S., 2015. INTER-HEART: a global study of risk factors for acute myocardial infarction. Am. Heart J. 141 (5), 711–721. http://dx.doi. org/10.1067/mhj.2001.114974. Park, H.-S., Cho, S.-B., 2012. Evolutionary attribute ordering in bayesian networks for predicting the metabolic syndrome. Expert Syst. Appl. 39 (4), 4240–4249. Pearl, J., 2000. Causality: Models, Reasoning and Inference, vol. 29 Cambridge Univ Press. Pejovic, V., Musolesi, M., 2015. Anticipatory mobile computing: a survey of the state of the art and research challenges. ACM Comput. Surv. 47 (3). http://dx.doi.org/ 10.1145/2693843. 47:1–47:29.

14

Telematics and Informatics xxx (xxxx) xxx–xxx

F. Pittoli et al.

Piegas, L.S., Avezum, A., Pereira, J.C.R., Neto, J.A. M.R., Hoepfner, C., Farran, J.A., Ramos, R.F., Timerman, A., Esteves, J.P., 2003. Risk factors for myocardial infarction in brazil. Am. Heart J. 146 (2), 331–338. Polanczyk, C.A., et al., 2005. Fatores de risco cardiovascular no brasil: os próximos 50 anos. Arq. Bras. Cardiol. 84 (3), 199–201. Puoane, T., Tsolekile, L., Sanders, D., Parker, W., et al., 2008. Chronic non-communicable diseases. South African Health Rev. 1, 73–87. Rosa, J.H., Barbosa, J.L.V., Kich, M., Brito, L., 2015. A multi-temporal context-aware system for competences management. Int. J. Artif. Intell. Educ. 25 (4), 455–492. http://dx.doi.org/10.1007/s40593-015-0047-y. Santos, R.D., Timerman, S., Spósito, A.C., Halpern, A., Segal, A., Ribeiro, A.B., Garrido, A., Mady, C., Fernandes, F., Lorenzi Filho, G., et al., 2002. Diretrizes para cardiologia clínica e funcor da sociedade brasileira de cardiologia. Arq. Bras. Cardiol. 78 (Suppl. 1), 1–13. Santos Filho, R.D., Martinez, T.L., 2002. Fatores de risco para doença cardiovascular: velhos e novos fatores de risco, velhos problemas!. Arquivos Bras. Endocrinol. Metabol. 46 (3), 212–214. Satyanarayanan, M., 2001. Pervasive computing: vision and challenges. IEEE Pers. Commun. 8, 10–17. Segatto, W., Herzer, E., Mazzotti, C.L., Bittencourt, J.A.R., Barbosa, J., 2008. Mobio threat: a mobile game based on the integration of wireless technologies. Comput. Entertain. 6 (3), 14. http://dx.doi.org/10.1145/1394021.1394032. 39(1–39). Sposito, A.C., Caramelli, B., Fonseca, F.A., Bertolami, M.C., Afiune Neto, A.A., Souza, A.D., Lottenberg, A.M.P., Chacra, A.P., Faludi, A.A., Loures-Vale, A.A., et al., 2013. V diretriz brasileira sobre dislipidemias e prevenção da aterosclerose: Departamento de aterosclerose da sociedade brasileira de cardiologia. Arq. Bras. Cardiol. 101, 2–19. Suchánek, P., Marecki, F., Bucki, R., 2014. Self-learning bayesian networks in diagnosis. Proc. Comput. Sci. 35, 1426–1435. http://dx.doi.org/10.1016/j.procs.2014. 08.200. URL: < http://www.sciencedirect.com/science/article/pii/S187705091401165X > . Tavares, J., Barbosa, J., Cardoso, I., Costa, C., Yamin, A., Real, R., 2015. Hefestos: an intelligent system applied to ubiquitous accessibility. Univ. Access Inf. Soc. 1–19. http://dx.doi.org/10.1007/s10209-015-0423-2. Tsoumakas, G., Vlahavas, I., 2007. Random k-Labelsets: An Ensemble Method for Multilabel Classification. Springer Berlin Heidelberg, Berlin, Heidelberg, pp. 406–417. van der Heijden, M., Lucas, P.J., Lijnse, B., Heijdra, Y.F., Schermer, T.R., 2013. An autonomous mobile system for the management of COPD. J. Biomed. Inform. 46 (3), 458–469. http://dx.doi.org/10.1016/j.jbi.2013.03.003. URL: < http://www.sciencedirect.com/science/article/pii/S1532046413000373 > . Vianna, H., Barbosa, J.L.V., 2014. A model for ubiquitous care of non-communicable diseases. Biomed. Health Inf. IEEE J.(99). http://dx.doi.org/10.1109/JBHI.2013. 2292860. 1–1. Vianna, H.D., Barbosa, J.L.V., Pittoli, F., 2017. In the pursuit of hygge software. IEEE Softw. 34 (6), 48–52. Vianna, H.D., Pittoli, F., Marques, E.B., Barbosa, J.L.V., 2017. Exploring Enterprise Service Bus in the Service-Oriented Architecture Paradigm, IGI Global. Ch. Towards a Middleware Based on SOA for Ubiquitous Care of Non-Communicable Diseases. Voigtmann, C., Schutte, C., Wacker, A., David, K., 2013. A new approach for distributed and collaborative context prediction. In: Pervasive Computing and Communications Workshops (PERCOM Workshops), 2013 IEEE International Conference on. IEEE, pp. 20–24. Wagner, E.H., Groves, T., 2002. Care for chronic diseases. BMJ 325 (7370), 913–914. http://dx.doi.org/10.1136/bmj.325.7370.913. Wagner, E.H., Austin, B.T., Davis, C., Hindmarsh, M., Schaefer, J., Bonomi, A., 2001. Improving chronic illness care: translating evidence into action. Health Affairs 20 (6), 64–78. Weiser, M., 1991. The computer for the 21st century. Sci. Am. 265 (3), 94–104. WHO, 2002. Innovative Care for Chronic Conditions: Building Blocks for Actions: Global Report. < http://apps.who.int/iris/bitstream/10665/42500/1/ WHONMCCCH02.01.pdf > . WHO and others, 2005. Preventing Chronic Diseases: A Vital Investment: Who Global Report. < http://www.who.int/chp/chronicdiseasereport/fullreport.pdf > . WHO and others, 2008. < http://www.who.int/dietphysicalactivity/Indicators > . Wu, T., Kwong, E.-Y., Pang, G.-H., 2015. Bio-medical application on predicting systolic blood pressure using neural networks. In: Big Data Computing Service and Applications (BigDataService), 2015 IEEE First International Conference on. pp. 456–461. doi:10.1109/BigDataService.2015.54.

15