Introduction

The annual global societal cost of dementia was US $1313.4 billion in 2019, representing an enormous economic burden []. With the global dementia population projected to increase from 57.4 million in 2019 to 152.8 million by 2050 [], this burden is expected to rise substantially []. To alleviate it, it is crucial to identify risk factors and implement preventive interventions at the most effective times across the dementia spectrum [].

Social isolation is one of the risk factors for dementia among older adults [-]. It is a complex concept that primarily refers to the objective aspect of low frequency of social interactions but is also closely related to the subjective experience of loneliness []. While social interaction refers to the frequency of contact with others [], loneliness is a subjective and negative feeling caused by a lack of social networks []. Experiencing social isolation has been associated with reduced gray matter volume in the memory-related hippocampus [,], potentially increasing the risk of developing dementia. As a modifiable risk factor for dementia [-], it is essential to identify and address aspects of social isolation among at-risk individuals through early detection.

Early detection of social isolation in at-risk groups for dementia, such as individuals with subjective cognitive decline (SCD) and mild cognitive impairment (MCI), may enhance the effectiveness of dementia prevention efforts. SCD and MCI represent the preclinical and prodromal stages of dementia, respectively []. SCD refers to a sustained, self-perceived decline in cognitive abilities compared with a previous level, without objective cognitive impairment []. By contrast, MCI differs from SCD in that it involves objective cognitive impairment []. Compared with individuals with normal cognition, the risk of dementia is more than twice as high for those with SCD [], and over 23 times higher for those with MCI, indicating that SCD and MCI are vulnerable groups in the progression to dementia []. However, both SCD and MCI represent stages where recovery of cognitive function and delayed progression are possible through prompt identification of high-risk individuals and reduction of modifiable risk factors [,]. Thus, the SCD and MCI stages may represent an optimal window to implement interventions aimed at improving social interaction and preventing loneliness—both modifiable risk factors for dementia []. Additionally, with an estimated 315 million people globally affected by SCD and 69 million by MCI in 2020 [], addressing social isolation at these stages through early detection could have significant public health implications for dementia prevention.

Ecological momentary assessment (EMA) enables real-time self-reported data collection in everyday environments, reducing recall bias and allowing for more accurate measurement []. Previous studies have primarily relied on retrospective methods to assess social interaction and loneliness [,]. However, these methods are subject to recall bias and typically involve single-time assessments, which may not adequately capture actual experiences. This limitation is especially pronounced in populations with memory impairments, such as individuals with SCD and MCI []. EMA offers several advantages over traditional measurement methods. Recent studies have evaluated EMA as a more time-sensitive tool for capturing the characteristics of social isolation [], highlighting its potential to provide a more nuanced understanding of this phenomenon []. These strengths underscore EMA’s suitability as a reliable method for assessing social interaction and loneliness, particularly in cognitively vulnerable populations.

Additionally, actigraphy continuously and noninvasively records data on activity and sleep in real time during everyday activities, minimizing recall bias and enabling objective measurement [-]. Given these advantages, actigraphy has proven useful in predicting social isolation []. However, a retrospective study that used regression models to explore the relationship between social interaction, loneliness, and sleep among older adults in the United States [] had limitations in capturing sleep characteristics. This was due to the use of only 3 sleep variables—total sleep time (TST), sleep efficiency, and wake after sleep onset (WASO)—measured via actigraphy over a relatively short period (3 days). The previous study has limited applicability to older adults with cognitive impairment, as this population was not included. In addition, although increased physical activity has the potential to promote contact with others—thereby enhancing social interactions and reducing loneliness—prior research has primarily focused on sleep-related variables, failing to fully consider other factors associated with social isolation [,]. Therefore, applying objectively assessed sleep and physical activity data collected through real-time assessment methods would not only enhance data reliability but also provide a more comprehensive understanding of social isolation.

Machine learning (ML) is a powerful tool for identifying patterns and predicting outcomes in large data sets, including those collected over extended periods []. ML is particularly valuable for processing actigraphy data, which includes metrics such as physical activity and sleep, and shows significant promise for predictive purposes [,,]. Given its capacity to handle large data sets, applying ML to mobile EMA data, actigraphy-based sleep and physical activity data, and survey data represents a novel approach to accurately predict social isolation among at-risk groups. Therefore, the objective of our study was to develop and validate models to explore factors related to 2 aspects of social isolation—social interaction frequency and levels of loneliness—among older adults in the predementia stage.

MethodsDesign

This study used a prospective observational design to build models exploring factors associated with social interaction frequency and levels of loneliness among older adults with SCD or MCI.

Ethics Considerations

Ethical approval for this study was obtained from the Yonsei University Health System Severance Hospital Institutional Review Board (approval number 4-2022-0637). All participants voluntarily provided written informed consent before enrollment and were free to withdraw at any time without penalty. Anonymized participant data were collected and stored on an encrypted server. This manuscript uses data from the first wave of a larger 3-wave study, for which the study protocol has been previously published [].

Participants and SettingsInclusion and Exclusion Criteria

Older adults with SCD or MCI were recruited from the Dementia Relief Center and a community service center in Seoul, Korea. The Dementia Relief Center houses an outpatient clinic specializing in the early diagnosis and care management of dementia, whereas the community service center serves as a space for older adults to engage in recreational and wellness programs, regardless of dementia diagnosis. The common inclusion criteria for study participants were (1) age over 65 years, (2) ability to use a smartphone, and (3) ability to respond to momentary questionnaires via a mobile app. The recruitment criteria were specifically tailored to the characteristics of SCD and MCI. The inclusion criteria for each group are listed in the following sections.

SCD Group

All participants with SCD answered “yes” to the question, “Do you think your memory has worsened compared to before?” Considering the different characteristics of the Dementia Relief Center and the community service center, tailored recruitment criteria were applied for each setting. Participants recruited from the Dementia Relief Center had no history of MCI or dementia, as reported by a nurse. Those recruited from the community service center self-reported no history of MCI or dementia and had a score of 24 or higher on the Korean Mini-Mental State Examination, second edition (K-MMSE-2) [-].

MCI Group

The MCI group consisted of participants who had been clinically diagnosed with MCI by a physician at the Dementia Care Center. To ensure their cognitive function was not consistent with severe cognitive impairment, individuals with MCI were additionally screened using the K-MMSE-2 and were required to score 18 or higher.

Participants were excluded if they met any of the following criteria: (1) illiteracy; (2) diagnosis of neurological diseases such as epilepsy, stroke, Parkinson disease, or other forms of brain damage; (3) diagnosis of psychiatric disorders such as schizophrenia or bipolar disorder; or (4) undergoing treatment for critical illnesses such as chemotherapy, having severe cardiovascular disease, or a history of substance abuse (including narcotics or alcohol) within the past 3 years.

Procedure

Participant recruitment and data collection were conducted through in-person visits from October 2022 to November 2023. These activities were carried out by a registered nurse enrolled in a master’s program (SY) and undergraduate nursing students, including CK. A master’s-level RN (DH), serving as the research coordinator, trained the research team in recruitment and data collection procedures according to the study protocol. Recruitment was primarily conducted by RNs, who verified participants’ eligibility, explained the purpose of the study, and obtained informed consent. Undergraduate nursing students, serving as research assistants, supported the process by installing the mobile EMA app on participants’ phones and registering their information in the actigraphy program. After the 2-week study period, the research team met with participants in person to collect actigraphy and survey data and to remove the EMA application from their smartphones.

EMA Features

Participants recorded their social interaction frequency and levels of loneliness in real time using a mobile EMA app on their smartphones. These variables were assessed 4 times daily, based on the Time-Use Survey from Statistics Korea []: night (9 PM-9 AM), morning (9 AM-2 PM), afternoon (2 PM-6 PM), and evening (6 PM-9 PM).

Social interaction frequency was measured using a 5-point Likert scale (0=no contact to 4=four or more times), based on the frequency of interpersonal contact [], including in-person meetings, phone calls, and video calls. Although validation studies of single-item EMA measures of social interaction frequency in older adults are limited, we assessed the reliability of this item using the intraclass correlation coefficient (ICC), which was 0.33, indicating fair agreement [].

Levels of loneliness, as a subjective feeling, were assessed using a 5-point Likert scale ranging from “not lonely at all” to “extremely lonely.” A previous study using a similar single-item EMA approach to assess momentary loneliness among older adults reported an ICC of 0.75 []. In our study, the ICC for the loneliness item was 0.56, indicating moderate agreement [] and supporting the reliability of this measure for repeated assessments.

Survey Data

We utilized a self-reported survey to collect demographic and health-related data. Demographic variables included sex, age, educational level, marital status, household type, subjective economic status, previous employment status, and duration of previous employment. Health-related variables included medical history, vision impairment, hearing impairment, subjective health status, number of medications, cognitive status, functional status, and psychological status.

Cognitive Status and Functional Status

In this study, the K-MMSE-2 [,] was used to assess cognitive functioning. Scores range from 0 to 30, with lower scores indicating poorer cognitive function. We also used the Korean version of the Subjective Cognitive Decline Questionnaire (SCD-Q) to measure self-perceived cognitive decline []. Scores range from 0 to 24, with higher scores indicating a greater subjective perception of cognitive decline [].

We utilized the Korean Instrumental Activities of Daily Living (K-IADL) to assess functional status. Among the 10 items in the K-IADL, 3 items are scored on a scale of 1-3 points, while the remaining items are scored on a scale of 1-4 points. Higher scores indicate a greater level of dependency []. Participants were categorized into 2 groups: those reporting dependency on 1 or more items were classified as dependent, while those independent on all items were classified as independent []. We also used the Frailty Phenotype Questionnaire to assess frailty among community-dwelling older adults. Scores range from 0 to 5, with 0 indicating robust status, 1-2 indicating prefrailty, and 3-5 indicating frailty [].

Psychological Status

We used the Korean version of the Short Form Geriatric Depression Scale (SGDS-K) to measure depression []. The SGDS-K ranges from 0 to 15, with higher scores indicating greater depressive symptoms. Additionally, we used the Korean version of the Geriatric Anxiety Inventory (K-GAI) [] to assess anxiety levels. The K-GAI ranges from 0 to 20, with higher scores indicating higher levels of anxiety. We also utilized the Korean version of the Mild Behavioral Impairment Checklist (MBI-C) [] to assess symptoms of mild behavioral impairment (MBI). MBI is a concept encompassing changes in motivation, affect, impulse control, social appropriateness, and perception or thought content [], and is associated with the transition from a nondementia state to dementia []. The Korean version of the MBI-C comprises domains corresponding to the 5 core components of MBI. The presence of MBI was determined by any score greater than 0, assessed from both global and domain-specific perspectives. Additionally, MBI severity was evaluated using summed scores, which were calculated separately as global total scores and domain-specific total scores.

Actigraphy

Participants wore a wrist-worn Actiwatch (Actiwatch Spectrum PRO; Philips Respironics) continuously for 2 weeks (24 hours a day). They were instructed to wear the device throughout the study period, except during unavoidable situations such as showering or when experiencing discomfort. The Actiwatch is a reliable and effective device for measuring sleep and physical activity [] and has been widely used in studies involving older adults [,]. We used the Actiwatch Spectrum PRO, which is equipped with an accelerometer that samples data 32 times per second. It was configured with a 30-second epoch length and a threshold value of 40 to collect sleep and physical activity data. Raw data were exported using Actiware software (version 6.1.2.1; Philips Respironics). The Actiware algorithm is commonly used for actigraphy data extraction and is well-regarded for its performance [].

If activity counts were below the threshold value of 40, they were categorized as sleep, while counts exceeding 40 were classified as wake []. Additionally, immobile and mobile states were defined by Actiware based on the level of activity generated during each 30-second epoch. Activity was classified as mobile if the activity count was 2 or more within the epoch, and immobile if the activity count was less than 2 [].

Sleeping Features

We generated 5 sleep features: TST (ie, the sum of the lengths of sleep bouts), number of sleep bouts (ie, the count of continuous sleep bouts), WASO (ie, the count of awakenings during sleep bouts, where sleep/wake is scored as 1), sleep efficiency (ie, TST divided by the sum of TST and WASO, multiplied by 100), and fragmentation index (ie, the sum of percent mobile and percent 1-minute immobile bouts, divided by the number of immobile bouts for the given interval). Among these 5 features, TST was classified as an indicator of sleep quantity, whereas the number of sleep bouts, WASO, sleep efficiency, and fragmentation index were classified as sleep quality indicators.

Physical Activity Features

We derived 3 physical activity features: total activity counts (the sum of all activity counts), average mobile bouts (mobile time divided by the number of mobile bouts), and immobile time (the sum of intervals with activity counts of 2 or fewer). Total activity counts and average mobile bouts were classified as indicators of physical movement, whereas immobile time (minutes) was categorized as an indicator of sedentary behavior.

Data Processing

For survey data preprocessing, we standardized continuous variables using a Min-Max scaler and applied one-hot encoding to categorical variables. For actigraphy data, temporal pattern characteristics were extracted using an autoencoder. An autoencoder is a neural network composed of an encoder and a decoder, which enables automatic feature learning from unlabeled data []. Eight actigraphy features were extracted across the 4 designated periods.

Because of challenges in completing all 4 mobile EMA assessments on the first and last days of the study—resulting from variations in participants’ start and end times—we used data collected from the second to the thirteenth day. This approach ensured consistency and reliability in the analysis and enabled precise time-based analysis and pattern identification. Participants were allowed to correct their responses if they initially submitted incorrect entries via the mobile EMA. In cases of duplicate responses at the same time point, the most recent entry was considered the corrected value and used for analysis. For missing responses, values within a 1-hour window were used to fill in the gaps.

The mobile EMA scores for social interaction frequency and loneliness levels are inherently subjective, with individual differences in interpreting neutral points. These variations made it challenging to determine whether a participant exhibited low social interaction frequency or high levels of loneliness based solely on changes in EMA scores over the study period. To address this issue, we applied Deep Embedding Clustering (DEC) [], a deep learning–based unsupervised learning method, to group participants into a small number of clusters using dissimilarity-based analysis of their EMA responses (). A detailed description of the DEC model architecture, hyperparameter settings, and training procedure is provided in . DEC processes 48 EMA responses each for social interaction frequency and loneliness, embedding them into vectors via an autoencoder. These compressed vectors were then used for clustering.

‎Figure 1. Deep embedding used for clustering data into 3 clusters.

The optimal number of clusters was determined by testing cluster counts ranging from 2 to 10, using 2 methods: (1) identifying the point where the silhouette coefficient is maximized [], and (2) applying the elbow method, which involves selecting the point where the within-cluster sum of squares sharply declines []. Priority was given to the silhouette coefficient, while the elbow method was used as a supplementary approach to confirm the selection. Based on the results presented in , 3 clusters were identified as optimal for social interaction levels, as this was the point where the silhouette coefficient reached its maximum. For the level of loneliness, although the silhouette coefficient was maximized at 2 clusters, the difference between 2 and 3 clusters was minimal. Moreover, 3 clusters corresponded to the point where the within-cluster sum of squares showed a sharp decline. Based on this, 3 clusters were selected as the optimal number. Consequently, we determined that 3 clusters were optimal for both social interaction frequency and loneliness levels.

The derived clusters were labeled S1, S2, and S3 for social interaction frequency, and L1, L2, and L3 for loneliness. To explore factors associated with low social interaction frequency and high levels of loneliness—the primary objectives of this study—these clusters were reclassified into binary groups. For this reclassification, we analyzed the mean EMA scores across the 4 daily time points during the study period () and examined the dispersion of participants’ embedding vectors generated by the Autoencoder (). Based on these evaluations, we redefined the 3 clusters into 2 binary groups to classify participants as having either low or high social interaction frequency and loneliness.

‎Figure 2. The classification of social interaction frequency and levels of loneliness of the 3 clusters. (A) Average graph of mobile EMA responses for the social interaction frequency of the 3 clusters. (B) Average graph of mobile EMA responses for the loneliness levels of 3 clusters. EMA: ecological momentary assessment. Exploratory Modeling

We used leave-some-out cross-validation to construct the training and validation data sets. The complete data set of 99 samples was divided into 10 folds using stratified sampling, separately for low social interaction frequency and high levels of loneliness. Each fold was used once for validation, and the performance metrics from all 10 iterations were aggregated for a comprehensive evaluation. Subsequently, we applied supervised learning models, including logistic regression, random forest (RF), Gradient Boosting Machine (GBM), and Extreme Gradient Boosting (XGBoost), to identify factors associated with social interaction frequency and loneliness.

The relative importance of the variables associated with the models was assessed using the Permutation Feature Importance algorithm [] ().

‎Figure 3. The sequential steps involved data processing and exploratory modeling. EMA: Ecological Momentary Assessment. Statistical Analysis

We conducted descriptive analyses to examine the general characteristics of the study population. Continuous variables are reported as means with SDs, and categorical variables are presented as frequencies and percentages. To assess differences in general characteristics between the 2 groups for social interaction frequency and level of loneliness, chi-square tests and independent 2-tailed t tests were performed. The ML models were evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, precision, specificity, and F1-score. Descriptive analyses were conducted using SAS version 9.4 (SAS Institute Inc), and a 2-sided P value of <.05 was considered statistically significant. ML analyses were conducted using Python version 3.9 (Python Software Foundation).

ResultsOverview

A total of 145 participants were initially recruited for this study. Of these, 18 participants withdrew consent, and 1 participant was excluded due to device loss. Additionally, 27 individuals were excluded for having mobile EMA records covering fewer than 13 days, in order to ensure sufficient data for the 12-day analysis period and maintain data completeness. As a result, the final analysis included 99 participants—67 with SCD and 32 with MCI.

Social Interaction and Loneliness Group Categorization

Based on the results of the DEC, social interaction frequency was categorized into 2 groups: the low group, which had below-average social interaction frequency (<average; S1), and the high group, which had above-average frequency (>average; S2 and S3). Similarly, levels of loneliness were divided into 2 groups: the low group, with below-average loneliness (<average; L1), and the high group, with above-average loneliness (≥average; L2 and L3; ).

As shown in , visual inspection confirmed that the 2 groups were appropriately classified. For social interaction frequency, 43 participants were classified into the low group (S1), while 10 and 46 participants were classified into the high group (S2 and S3, respectively). For loneliness, 62 participants were classified into the low group (L1), while 21 and 16 participants were classified into the high group (L2 and L3, respectively).

General Characteristics of Participants

Of the 99 participants, 62 (63%) were female. The mean age of all participants was 76.8 (SD 6.0) years. The mean age was 78.4 (SD 5.7) years in the low social interaction frequency group and 75.6 (SD 5.9) years in the high social interaction frequency group. presents the baseline characteristics according to levels of social interaction. The low social interaction frequency group included 43 (43%) individuals. Compared with the high social interaction group, this group had significantly higher scores on the Korean version of the MBI-C (mean 9.9, SD 10.1 vs mean 5.1, SD 6.2; P=.01).

Table 1. Baseline characteristics of participants by social interaction frequency (N=99).VariablesLow (n=43)High (n=56)P valuePredementia stage, n (%)

.63
SCDa28 (65)39 (70)

MCIb15 (35)17 (30)
Sex, n (%)

.004
Men23 (53)14 (25)

Women20 (47)42 (75)
Age (years), n (%)

.13
<703 (7)12 (21)

70-746 (14)11 (20)

75-7917 (40)14 (25)

≥8017 (40)19 (34)
Educational level, n (%)

.66
≤Elementary school15 (35)14 (25)

Middle school8 (19)9 (16)

High school9 (21)15 (27)

≥College11 (26)18 (32)
Marital status, n (%)

.45
Married30 (70)38 (68)

Widowed12 (28)13 (23)

Single1 (2)5 (9)
Type of household, n (%)

.03
Alone11 (26)9 (16)

With spouse30 (70)34 (61)

With others2 (5)13 (23)
Subjective economic status, n (%)

.25
Low10 (23)8 (14)

Moderate30 (70)39 (70)

High3 (7)9 (16)
Previous employment status, n (%)c

.73
Yes30 (70)42 (75)

No8 (19)10 (18)
Vision impairment (yes), n (%)10 (23)13 (23)>.99Hearing impairment (yes), n (%)10 (23)7 (13).16Subjective health status, n (%)c

.18
Poor11 (26)8 (14)

Fair9 (21)5 (9)

Good9 (21)18 (32)

Very good11 (26)15 (27)

Excellent3 (7)9 (16)
Number of medications, n (%)c

.35
≤212 (28)25 (45)

3-515 (35)17 (30)

≥614 (33)12 (21)
Cognitive function (K-MMSE-2d), mean (SD)27.5 (1.8)28.1 (1.6).10Subjective Cognitive Decline (SCD-Qe), mean (SD)8.1 (6.1)5.8 (5.4).05IADL disability (K-IADLf score>10), n (%)12 (28)11 (20).33Frailty (FPQg), n (%)

.07
Robust20 (47)36 (64)

Prefrailty15 (35)17 (30)

Frailty8 (19)3 (5)
Depression (SGDS-Kh), mean (SD)5.6 (1.9)5.1 (1.9).22Anxiety (K-GAIi), mean (SD)4.8 (5.7)3.5 (4.6).19Mild Behavioral Impairment (MBI-Cj), mean (SD)9.9 (10.1)5.1 (6.2).01

aSCD: subjective cognitive decline.

bMCI: mild cognitive impairment.

cSome missing data.

dK-MMSE-2: Korean Mini-Mental State Examination, second edition.

eSCD-Q: Subjective Cognitive Decline Questionnaire.

fK-IADL: Korean Instrumental Activities of Daily Living.

gFPQ: Frailty Phenotype Questionnaire.

hSGDS-K: Korean version of the Short Form-Geriatric Depression Scale.

iK-GAI: Korean version of the Geriatric Anxiety Inventory.

jMBI-C: Mild Behavioral Impairment Checklist.

presents the baseline characteristics based on levels of loneliness. The high loneliness group consisted of 37 (37%) individuals, while the low loneliness group included 62 (63%) individuals. The mean age in the low loneliness group was 77.1 (SD 5.9) years, compared with 76.4 (SD 6.1) years in the high loneliness group.

Table 2. Baseline characteristics of participants by levels of loneliness (N=99).VariablesLow (n=62)High (n=37)P valuePredementia stage, n (%)

.36
SCDa44 (71)23 (62)

MCIb18 (29)14 (38)
Sex, n (%)

.43
Men25 (40)12 (32)

Women37 (60)25 (68)
Age (years), n (%)

.77
<709 (15)6 (16)

70-7410 (16)7 (19)

75-7918 (29)13 (35)

≥8025 (40)11 (30)
Educational level, n (%)

.48
≤Elementary school17 (27)12 (32)

Middle school9 (15)8 (22)

High school18 (29)6 (16)

≥College18 (29)11 (30)
Marital status, n (%)

.83
Married43 (69)25 (68)

Widowed16 (26)9 (24)

Single3 (5)3 (8)
Type of household, n (%)

.04
Alone9 (15)11 (30)

With spouse40 (65)24 (65)

With others13 (21)2 (5)
Subjective economic status, n (%)

<.001
Low3 (5)15 (41)

Moderate52 (84)17 (46)

High7 (11)5 (14)
Previous employment status, n (%)c

.49
Yes46 (74)26 (70)

No12 (19)6 (16)
Vision impairment (yes), n (%)13 (21)10 (27).49Hearing impairment (yes), n (%)11 (18)6 (16).85Subjective health status, n (%)c

.60
Poor11 (18)8 (22)

Fair8 (13)6 (16)

Good15 (24)12 (32)

Very good17 (27)9 (24)

Excellent10 (16)2 (5)
Number of medications, n (%)c

.85
≤224 (39)13 (35)

3-518 (29)14 (38)

≥617 (27)9 (24)
Cognitive function (K-MMSE-2d), mean (SD)28.0 (1.6)27.5 (1.8).10Subjective Cognitive Decline (SCD-Qe), mean (SD)6.3 (5.6)7.6 (6.2).30IADL disability (K-IADLf score > 10), n (%)15 (24)8 (22).77Frailty (FPQg), n (%)

.26
Robust39 (63)17 (46)

Prefrailty17 (27)15 (41)

Frailty6 (10)5 (14)
Depression (SGDS-Kh), mean (SD)5.1 (1.5)5.6 (2.5).19Anxiety (K-GAIi), mean (SD)3.2 (4.7)5.5 (5.6).03Mild Behavioral Impairment (MBI-Cj), mean (SD)6.0 (7.0)9.1 (10.5).11

aSCD: subjective cognitive decline.

bMCI: mild cognitive impairment.

cSome missing data.

dK-MMSE-2: Korean Mini-Mental State Examination, second edition.

eSCD-Q: Subjective Cognitive Decline Questionnaire.

fK-IADL: Korean Instrumental Activities of Daily Living.

gFPQ: Frailty Phenotype Questionnaire.

hSGDS-K: Korean version of the Short Form-Geriatric Depression Scale.

iK-GAI: Korean version of the Geriatric Anxiety Inventory.

jMBI-C: Mild Behavioral Impairment Checklist.

Performance Comparison for Exploratory Models

and present the performance of the models exploring factors associated with low social interaction frequency and high levels of loneliness. Among the models, the RF model demonstrated the highest performance in identifying factors related to low social interaction frequency, while the GBM model showed the best performance in identifying factors related to high levels of loneliness. Specifically, the RF model achieved an AUC of 0.935, with macro and micro F1-scores of 0.824 and 0.828, respectively. By contrast, the GBM model yielded an AUC of 0.887, with a macro F1-score of 0.867 and a micro F1-score of 0.871.

Table 3. Performance comparison of the exploratory models for low social interaction frequency and high levels of lonelinessa.Exploration goal and modelAUCbAccuracyPrecisionSpecificityF1-scoreMacrocMicrodLow social interaction frequency






GBMe0.9090.8500.8370.8570.8290.828LRf0.8500.7800.8370.7320.7630.766RFg0.9350.8490.8370.8570.8240.828XGBoosth0.9070.8400.8140.8570.8140.814High levels of loneliness






GBM0.8870.8380.8710.7840.8670.871LR0.8040.7790.8390.6760.8250.825RF0.9090.8180.8710.7300.8540.857XGBoost0.8580.7980.8390.7300.8380.839

aThe italicized values indicate the machine learning models with the best performance within each category (ie, low social interaction frequency and high levels of loneliness).

bAUC: area under the receiver operator characteristic curve.

cAverage F1-score for 10 folds.

dCalculated as the sum of the confusion matrix of folds.

eGBM: Gradient Boosting Machine.

fLR: logistic regression.

gRF: random forest.

hXGBoost: Extreme Gradient Boosting.

‎Figure 4. Receiver operator characteristic (ROC) curve for each model’s analysis of low social interaction frequency and high levels of loneliness. (A) ROC curve of the low social interaction frequency exploratory model. (B) ROC curve of the high levels of loneliness exploratory model. GBM: Gradient Boosting Machine; LR: logistic regression; RF: random forest; XGB: Extreme Gradient Boosting. Feature Importance

summarizes the key features identified by the RF model, which exhibited the best performance in identifying factors associated with low social interaction frequency. Among all features, physical movement emerged as the most influential variable, showing the highest feature importance. Given the relatively higher importance of actigraphy-derived features compared with demographic and health-related variables, the feature importance values for the survey-based data are presented separately for clarity (see ). Among the demographic and health-related variables, the total score across the MBI-C domains demonstrated the highest feature importance.

Table 4. The feature importance for exploring factors related to low social interaction frequency, as analyzed by 4 groups.VariablesFeature importancePhysical movement5.223Sleep quantity2.128Sleep quality0.795Sedentary behavior0.751Sum of MBI-Ca0.103Impulse dyscontrol domain in MBI-C0.084Sum of K-GAIb0.053Living with offspring0.048Sum of SCD-Qc0.044Sex0.035

aMBI-C: Mild Behavioral Impairment Checklist.

bK-GAI: Korean version of the Geriatric Anxiety Inventory.

cSCD-Q: Subjective Cognitive Decline Questionnaire.

presents the representative influential features associated with a high level of loneliness, as identified by the GBM model, which demonstrated the best performance for this outcome. Among all features, actigraphy data emerged as the most influential, with sleep quality showing the highest feature importance. Given the relatively greater influence of actigraphy-derived variables compared with survey-based data, the feature importance values for the survey data are provided separately for clarity (see ). When actigraphy features were excluded, the total score on the SGDS-K emerged as the most influential survey-based predictor.

Table 5. The feature importance for exploring factors related to high levels of loneliness, as analyzed by 4 groups.