2.1 Loneliness Predictive Models

Assessing loneliness in older adults is challenging due to the negative desirability biases associated with being lonely. However, some studies have objectively detected loneliness. In the work of Austin et al. [¹⁹], authors proposed a system to measure loneliness by assessing in home multiple objective behavioral measures. Elderly movements were captured using wireless motion sensors. Open and close door events were captured using contact sensors installed on the doors. Call activities were captured using phone monitors. Computer behavior was captured using software. Data was collected from 16 older adults who agreed to have the sensor platform installed in their own homes for up to 8 months. Loneliness was assessed using the University of California Los Angeles (UCLA) Loneliness Scale [²⁷]. The results showed that loneliness is significantly associated with time out-of-home and number of computer sessions.

In the work of Badal et al. [¹⁷], authors assessed loneliness employing natural language processing techniques in transcribed speech text of older adults. Participants completed semi-structured qualitative interviews regarding the experience of loneliness and a quantitative self-report UCLA loneliness scale [²⁷] to assess loneliness. Lonely individuals showed sadness and longer responses regarding loneliness questions. Women were more likely to be lonely. Men used more fearful and joyful words in their responses. The resulting predictive models predicted qualitative loneliness with 94% precision and quantitative loneliness with 76% precision.

In the work of Sanchez et al. [¹⁸], authors developed four predictive models to determine each aspect of loneliness: family loneliness, spousal loneliness, social loneliness, and existential crisis, according to [²⁸]. Data was collected from 100 older adults who answered a questionnaire regarding their daily activities. Predictive models were developed by implementing a range of classifier algorithms. The best predictive models were implemented in a mobile application called “¡Vive!” which can collect data and predicting loneliness. Mobile application was evaluated with 12 participants. Results showed that the mobile application obtained an accuracy of 91.67% for family loneliness, an accuracy of 83.34% for spousal loneliness, an accuracy of 66.67% for social loneliness, and an accuracy of 83.34% for existential crisis.

2.2 Social Isolation Predictive Models

In the work of Campos et al. [²⁴], authors presented a prediction model of social isolation in older adults through ambient intelligence and social networking sites. They identified attributes correlated with social isolation.

Data was collected from 144 older adults. They applied a survey to collect demographic data and the Lubben scale [²⁹] to determine social isolation level. The prediction model obtained an accuracy score of 85% and a type II error rate of 15%. To evaluate the model, 8 older adults were monitored for one month using their mobile phones, four wireless IP cameras, a wireless router, an internet connection, and their Facebook message history.

A comparison between Lubben scale and predictive model was carried out, obtaining a prediction accuracy of 87% and a type II error rate of 15%.

2.3 Loneliness and Social Isolation Predictive Models

In the work of Goonawardene et al. [²³], authors have detected social isolation in older adults through a sensor-based monitoring system which comprises passive infrared motion sensors installed in every room of the apartment and a door contact sensor. Each sensor reports if motion has been detected within its coverage area.

They extracted different features to examine the association with elderly’s social isolation level: going out, sleep, time spent in the living room, and activity level in the kitchen. Results showed that goingout behavior, daytime napping, and time spent in the living room are associated with different social isolation dimensions.

Further, elderly who perceived themselves as socially lonely tend to take more naps during the daytime.

In the work of Akgul et al. [²¹], authors presented an expert system which uses the artificial neural network and linear regression models to estimate emotional loneliness and social loneliness. Data was collected from 809 older adults through the Loneliness Scale for Elderly [³⁰] using face-to-face interview technique. The results showed RMSE values of 0.1994 for social loneliness and 0.1883 for emotional loneliness with an artificial neural network model and RMSE values of 0. 3553 for social loneliness and 0.3837 for emotional loneliness with linear regression model. The artificial neural network model performed better than the linear regression model.

4.1 Data Collection

The understanding of data is the core of the first phase of the proposed methodology. Two databases were considered in this phase: loneliness [²] and social isolation [⁵].

The loneliness database was created through a questionnaire applied to 100 older adults, including both men and women between 60 and 90 years of age (69 ± 7.165). The participants' older adults fulfill the following characteristics: they have full mental and physical capacities, they use a cellular phone, they had no difficulty understanding the questions, they signed the informed consent, and they were willing to take part in the research.

Older adults with physical, sensory, or cognitive disability were excluded because they were not part of the scope of the study. The questionnaire comprises two parts.

The first one gathered the level of loneliness in the older adult through the ESTE-R scale [²⁸] which is targeted to Spanish speakers. This scale includes different dimensions of loneliness and groups them into four factors: family loneliness, spousal loneliness, social loneliness, and existential crisis. The scale consists of 34 items and the response options are: never, rarely, sometimes, regularly, or always. The Cronbach’s Alpha is 0.915, which indicates an excellent internal consistency. The scale results reflect the subject’s loneliness level and are valid for one month before the test is applied.

The loneliness database is composed of 100 information records on older adults organized into 5 sections (see https://docs.google.com/ document/d/1QIYIMCwqL7zbX9Hbn7e1A5T_m65 _IxTo). The first section is made up of 5 demographic questions. The next segment represents 15 questions about older adult cell phone information. The next segment contains 7 questions related to the times that the older adult is away from his house.

The next segment represents the information of the activities that the elderly adult performs at home, this segment is represented by 4 questions. In addition, the time that the elderly person spends in each of the rooms of their home is also analyzed through 4 questions. Finally, the level of loneliness of the older adult detected through the ESTE-R [²⁸] loneliness survey is also collected

The data of the loneliness levels obtained from the questionnaires are the following: 17 people with loneliness, 12 people with principles of loneliness and 71 people without loneliness.

The isolation database was created through a questionnaire applied to 144 older adults. The sample included both men and women between 60 and 89 years of age (68.2 ±8.9), all of them with following characteristics: they have full cognitive and physical capacities without mobility impairment, they use a cellular phone and can use it to make calls or send text messages, they had no difficulty understanding the questions, they signed the informed consent, and they were willing to take part of the research. The older adults were surveyed about their social interactions over the last 30 days.

The questionnaire consists of two parts. The first one is the Spanish version of the Lubben Social Network Scale (LSNS-6) [²⁹], which was used to determine the level of social isolation of older adults. The second part of the questionnaire collected data concerning demographic information as well as social interaction activities. The questions formulated by the LSNS-6 request information about the frequency of social interactions during one month prior to the interview, which is often difficult to remember accurately.

The social isolation database is composed of 144 records of information on older adults organized into 4 sections¹. The first section is made up of 8 demographic questions. The next segment represents 17 questions related to the use of the cell phone of the elderly, such as: communication activities with friends and family. The next segment contains information about communication activities through social networks (8 questions). The next segment is made up of information on mobility activities of the elderly inside and outside the home. This segment is represented by 9 questions. Finally, the level of social isolation detected through the Lubben of social isolation survey [²⁹] is also collected.

The social isolation database consists of 44 fields and 144 records of older adult participants. They were classified as follows: 48 people with high isolation, 93 people with low isolation and 3 people with no isolation. Having little information on people with no social isolation, inconsistent classifications were generated in the results, so it was decided to discard this class.

4.2 Data Processing

Data processing is the conversion of data into a usable and desired form called dataset. Raw data is highly susceptible to noise, missing values, and inconsistency. The quality of data affects the data mining results. To improve the quality of the data and, consequently, of the mining results, raw data is preprocessed to improve the efficiency and ease of the mining process.

In this research work, the quality of raw data was improved by using a preprocessing procedure that includes a) data transformation, b) training and test data c) class balancing and d) the selection of relevant attributes to improve the miner efficiency. As a result of these processes, two preprocessed datasets prepared to perform the attribute union analysis and generate behavior models were obtained. Following, the data preprocessing procedures are presented in detail.

4.3 Predictive Model Construction

Several experiments were developed to generate the predictive model of loneliness and social isolation proposed in this research work. The experiments were conducted considering six of the top ten data mining algorithms identified by the IEEE International Conference on Data Mining [³⁵]: C4.5, CART, Support Vector Machine, AdaBoost with Random Forest classifier, k Nearest Neighbor, and Naive Bayes.

The other four algorithms were not considered because they are focusing on clustering and because they have characteristics that were not adapted to the problem addressed in this research work. Two metrics were used to determine the performance of the six algorithms: Accuracy and F1 Score.

4.4 Model Combination

We have constructed two new models for loneliness and social isolation using only common attributes. Data training and evaluation were carried out using the same algorithms and metrics of the previous steps. Results are shown in Table 3. For the case of loneliness, the best model was AdaBoost-Random Forest algorithm with an accuracy and an F-measure of 0.7. For the case of social isolation, the best model was k-Nearest-Neighbor algorithm with an accuracy of 0.8 and an F-measure of 0.794.

These models have the same attributes, so we test original social isolation data using loneliness model and vice versa. This process is carried out with the objective of defining new classes that allow combining the models into one.