This paper is in the following e-collection/theme issue:

Peer-to-Peer Support and Online Communities (633) Dementia and Cognitive Decline (478) Supporting Informal Care and Caregivers (291)

Published on in Vol 25 (2023)

Preprints (earlier versions) of this paper are available at https://preprints.jmir.org/preprint/48193, first published .
Examining Online Behaviors of Adult-Child and Spousal Caregivers for People Living With Alzheimer Disease or Related Dementias: Comparative Study in an Open Online Community

Examining Online Behaviors of Adult-Child and Spousal Caregivers for People Living With Alzheimer Disease or Related Dementias: Comparative Study in an Open Online Community

Examining Online Behaviors of Adult-Child and Spousal Caregivers for People Living With Alzheimer Disease or Related Dementias: Comparative Study in an Open Online Community

Authors of this article:

Congning Ni1 Author Orcid Image ;   Qingyuan Song1 Author Orcid Image ;   Bradley Malin1, 2, 3, 4 Author Orcid Image ;   Lijun Song5 Author Orcid Image ;   Patricia Commiskey6 Author Orcid Image ;   Lauren Stratton7 Author Orcid Image ;   Zhijun Yin1, 2, 3 Author Orcid Image
Article Authors Cited by Tweetations (3) Metrics

    Original Paper

    1Department of Computer Science, Vanderbilt University, Nashville, TN, United States

    2Department of Biomedical Informatics, Vanderbilt University Medical Center, Nashville, TN, United States

    3Center for Genetic Privacy & Identity in Community Settings, Vanderbilt University Medical Center, Nashville, TN, United States

    4Department of Biostatistics, Vanderbilt University Medical Center, Nashville, TN, United States

    5Department of Sociology, Vanderbilt University, Nashville, TN, United States

    6Department of Neurology, Vanderbilt University Medical Center, Nashville, TN, United States

    7Care and Support, Alzheimer’s Association, Chicago, IL, United States

    *these authors contributed equally

    Corresponding Author:

    Zhijun Yin, PhD

    Department of Biomedical Informatics

    Vanderbilt University Medical Center

    2525 West End Avenue

    Nashville, TN, 37203

    United States

    Phone: 1 6159363690

    Email: zhijun.yin@vumc.org


    Background: Alzheimer disease or related dementias (ADRD) are severe neurological disorders that impair the thinking and memory skills of older adults. Most persons living with dementia receive care at home from their family members or other unpaid informal caregivers; this results in significant mental, physical, and financial challenges for these caregivers. To combat these challenges, many informal ADRD caregivers seek social support in online environments. Although research examining online caregiving discussions is growing, few investigations have distinguished caregivers according to their kin relationships with persons living with dementias. Various studies have suggested that caregivers in different relationships experience distinct caregiving challenges and support needs.

    Objective: This study aims to examine and compare the online behaviors of adult-child and spousal caregivers, the 2 largest groups of informal ADRD caregivers, in an open online community.

    Methods: We collected posts from ALZConnected, an online community managed by the Alzheimer’s Association. To gain insights into online behaviors, we first applied structural topic modeling to identify topics and topic prevalence between adult-child and spousal caregivers. Next, we applied VADER (Valence Aware Dictionary for Sentiment Reasoning) and LIWC (Linguistic Inquiry and Word Count) to evaluate sentiment changes in the online posts over time for both types of caregivers. We further built machine learning models to distinguish the posts of each caregiver type and evaluated them in terms of precision, recall, F1-score, and area under the precision-recall curve. Finally, we applied the best prediction model to compare the temporal trend of relationship-predicting capacities in posts between the 2 types of caregivers.

    Results: Our analysis showed that the number of posts from both types of caregivers followed a long-tailed distribution, indicating that most caregivers in this online community were infrequent users. In comparison with adult-child caregivers, spousal caregivers tended to be more active in the community, publishing more posts and engaging in discussions on a wider range of caregiving topics. Spousal caregivers also exhibited slower growth in positive emotional communication over time. The best machine learning model for predicting adult-child, spousal, or other caregivers achieved an area under the precision-recall curve of 81.3%. The subsequent trend analysis showed that it became more difficult to predict adult-child caregiver posts than spousal caregiver posts over time. This suggests that adult-child and spousal caregivers might gradually shift their discussions from questions that are more directly related to their own experiences and needs to questions that are more general and applicable to other types of caregivers.

    Conclusions: Our findings suggest that it is important for researchers and community organizers to consider the heterogeneity of caregiving experiences and subsequent online behaviors among different types of caregivers when tailoring online peer support to meet the specific needs of each caregiver group.

    J Med Internet Res 2023;25:e48193

    doi:10.2196/48193

    Keywords

    Alzheimer disease or related dementias; informal caregivers; adult-child caregivers; spousal caregivers; online community; sentiment analysis; topic modeling; text classification 


    Background

    Dementia is a clinical syndrome that severely impairs a person’s memory, language, and reasoning abilities [1]. Alzheimer disease is the most common cause of dementia, and approximately 6.5 million Americans aged ≥65 years live with Alzheimer disease or related dementias (ADRD) [2]. Owing to a lack of effective treatments, 80% of persons living with dementia are cared for at home by their informal caregivers (eg, family members, friends, or other unpaid caregivers) [3]. This results in tremendous physical, financial, and emotional challenges for these caregivers [4,5], making them known as invisible second patients [6]. It has further been shown that informal ADRD caregivers in different kin relationships (eg, adult children, spouses, or children-in-law, which we refer to as relationships in this paper for simplicity) differ in their caregiving experience, physical and mental health, and support needs [7,8].

    Existing Literature and Motivation

    To date, research on how to support informal ADRD caregivers has focused primarily on the support provided by health professionals [9] (eg, cognitive-behavioral therapy [10], benefit-finding interventions [11], and educational program [12]). Although professional support can improve caregiving skills and emotional well-being, it is challenging to implement it on a large scale because of limitations in workforce availability and financial support [13-15]. At the same time, an increasing body of theoretical and empirical studies indicates that peer-based social support from offline social networks has a positive impact on informal ADRD caregivers [16,17]. For example, the source-need matching theory [18-21] suggests that social support, including informational and emotional support, from caregivers who share similar social backgrounds or caregiving experiences, is more beneficial for stress coping than the support from those who do not. Notably, research has shown that informal ADRD caregivers with shared caregiving experiences can help each other with unique informational and emotional support [22]. However, similar to professional services, offline peer support also faces scalability issues. For example, offline informal ADRD caregivers often report small, high-quality networks of only 4 people on average. By contrast, those with larger networks tend to report experiencing less role strain and better well-being [23].

    The integration of the internet into daily life has enabled caregivers to increasingly discuss health-related topics on online platforms [20], including Twitter [24], Facebook [25], and other online communities [26]. In recent research, higher emotional stress and financial hardship of caregivers were found to be the most significant emerging factors associated with increased health-related internet use [27]. Participation in online peer discussions can reduce depressive symptoms [28], improve quality of life [29], and reduce feelings of loneliness [30]. Caregivers seek support and are willing to share experiences and practical information that they believe will assist other caregivers in online environments [31]. In addition, numerous online communities offer the added advantage of facilitating asynchronous discussions and connecting caregivers without the limitations of time and location. These advantages enable ADRD caregivers to receive sufficient social support while fulfilling the time and space obligations of providing care [32]. However, the studies to date that analyzed online caregiving discussions have paid limited attention to the nature of the relationships between informal ADRD caregivers and persons living with dementia. In this respect, it should be recognized that spouses and adult children constitute the 2 largest types of primary informal ADRD caregivers. Spouses regard caregiving as part of their marital obligations, whereas adult children consider such tasks as an important change in their lifestyle [33]. Moreover, spouses tend to be less selective about whether to take on caregiving responsibilities and are at a higher risk of developing health complications compared with adult-child caregivers [34]. By contrast, adult children tend to experience greater family conflict or work disruptions than spouses [33,35]. Given that there are likely differences in caregiving experiences, characterizing the online behaviors of different types of caregivers by relationships can help design customized solutions to online social support that can efficiently assist each type of informal ADRD caregiver.

    Most online platforms (eg, Reddit) do not establish specific forums based on each type of caregiver relationship, such that all caregivers discuss topics within the same forum. Even in online communities with separate forums for different types of caregiver relationships (eg, Talking Points managed by the Alzheimer’s Society in the United Kingdom), not every caregiver discloses their relationship with persons living with dementia publicly, and any caregiver can publish their posts in any forum that they believe is appropriate. Considering the massive number of online sources that may contain ADRD caregiving discussions, it is crucial for researchers to develop machine learning algorithms that can efficiently identify the relationship between caregivers and persons living with dementia based on their online posts, which will be a valuable social determinant that can be applied to online social support-based research. Various machine learning algorithms have been developed to extract relevant information from massive, noisy online data to facilitate biomedical research [36], including, but not limited to, predicting mental health status [37], disseminating study information during the COVID-19 pandemic [38], and learning the needs of patients with ovarian cancer and their caregivers [39]. However, most studies in the ADRD caregiving field have either identified online ADRD caregiving discussions [26] or summarized caregiving challenges [35]. Few studies have differentiated caregivers’ relationships with persons living with dementia.

    Objectives

    In this study, we aimed to conduct a comprehensive comparison of the online behaviors of the 2 distinct groups of informal ADRD caregivers: adult-child and spousal caregivers. To achieve this goal, we conducted topic analysis, sentiment analysis, and caregiver relationship classification using the large-scale retrospective data collected from ALZConnected [40], an open online community powered by the Alzheimer’s Association for any person affected by ADRD in North America. Our study not only enriches our understanding of the online behaviors of informal ADRD caregivers in online communities but also highlights the potential for improving online support systems, thereby contributing to the advancement of targeted support for informal ADRD caregivers.


    Ethical Considerations

    This study was exempted from human participants research by the Institutional Review Board of Vanderbilt University Medical Center (221732). The online posts are publicly accessible through the ALZConnected online community. In this paper, we only present results on an aggregated level, and any quotes presented are carefully rephrased to maintain the privacy of the corresponding users [41].

    Data Collection

    ALZConnected is a free online community for caregivers and persons living with dementia to post questions, offer and receive support, and make public and private groups around dementia-related topics and issues. ALZConnected allows individuals to read online discussions without registering on the platform. However, to use all the functions in ALZConnected, including publishing a post, individuals must register to become a member on the platform. During registration, a user is asked to provide information about their relationship with persons living with dementia. Registered users may choose to make such information visible to other logged-in online users. ALZConnected has 10 forums, each organizing online discussions into a collection of topic threads. Each thread is initialized with the first post as the top post and can contain many other posts called comments. We collected data from ALZConnected using a web crawler that we developed based on the BeautifulSoup v4.11 python package. The data contain online user interactions that transpired between October 2011 and August 2022. We maintained all users who self-reported as ADRD caregivers and excluded administrators and moderators.

    For this study, we extracted all the posts published in three main caregiver forums: (1) Caregivers Forum, (2) Spouse or Partner Caregiver Forum (referred to as Spouses Forum for simplicity), and (3) Caregivers Who Have Lost Someone Forum (referred to as Lost Forum). Although it is unlikely that a caregiver is active online every day, in this study, we define the active days of a user as the number of days between the dates when the user published the first and the most recent posts to improve the readability of the paper. In our analysis, we focused primarily on adult-child and spousal caregivers; all other relationships were categorized as other caregivers.

    Activity Statistics

    We characterized caregivers’ online activities along several dimensions: (1) posts and relationships, (2) views and comments, and (3) posts and active days. To do so, we first identified all the self-reported relationships and the corresponding post volumes to provide a big picture with respect to all the online caregivers. Second, ALZConnected displays the number of views and comments for each topic thread publicly, which enables us to compare these 2 statistics on a log scale separately for adult-child and spousal caregivers. Reading and posting are the 2 types of essential activities in online communities that can benefit caregivers in different ways. Third, we compared the distributions of the online posts and active days on a log scale separately for adult children and spouses to gain insights into online activities for both short-term and long-term online caregivers. To avoid the logarithm of 0, we incorporated a weak uniform prior in the form of a small pseudo count. Specifically, this was accomplished by adding 1 to each count.

    Structural Topic Modeling

    We applied structural topic modeling (STM) v1.3.6 [42], an R package developed by researchers and contributors, to top posts to investigate the topic prevalence for adult children and spousal caregivers. In general, a caregiver initiates a topic thread by asking a question or sharing their experience in the top post, whereas other caregivers make comments in the response below. Therefore, focusing on the top posts instead of comments enables an examination of the challenges communicated by these caregivers. Furthermore, focusing on top posts can reduce the bias induced by super-active users who published substantially more posts than others. STM allows for the integration of post-level metadata (eg, authorship) into the topic modeling process. In contrast to standard topic modeling methods such as latent Dirichlet allocation, STM can provide better intuition into the dynamics of social representations through a comparative view [43-45].

    In our model training, we incorporated caregiver relationships as a binary meta-variable. This variable indicates whether a top post was composed by an adult child or a spouse. More specifically, we extracted the top posts from all the topic threads, removed stop words and special symbols, and dropped the words that appeared <10 times in the data set. To determine the optimal number of topics in STM, we relied on the measures of exclusivity, which refers to the distinctiveness of the words with the highest frequencies in the topic, and semantic coherence [46], which quantifies how the words in a topic frequently co-occur together in general contexts [42]. We chose the optimal topic number from a predefined list of [5, 10, 15, 20, 25, 30] by finding the one that made a trade-off between the 2 metrics.

    To support the interpretation of the modeling results, we summarized each topic with a name and grouped similar topics together. This summary was developed by the authors through a manual review and discussion of the top words ranked by their probabilities in each topic and the 5 posts with the highest probabilities of each topic. We also compared adult children and spousal caregivers with respect to topic prevalence. The effect for each topic was estimated by regressing the proportion of topics on a binary meta-variable that indicates whether the top post was composed by a spousal caregiver as opposed to an adult-child caregiver. Positive (or negative) effects indicated that the corresponding topics were more likely to be posted by spousal (or adult-child) caregivers.

    Sentiment Analysis

    Sentiment analysis refers to a common natural language processing technique that applies computational methods to determine whether a given post conveys positive, negative, or neutral emotions or tones. In this study, we define the relative days of a post as the number of days since the first post of its author was published. By realigning all the posts along their relative days rather than published dates (eg, October 19, 2022), we analyzed how the sentiment of a caregiver’s posts changes on average as the caregiver interacts with others in the community over time.

    In recognition of the measurement bias that can be introduced by applying off-the-shelf models to this data set, we relied upon 2 existing popular sentiment tools to perform the analysis, which we believed could provide better intuition than the application of a single model. The first tool is the Valence Aware Dictionary for Sentiment Reasoning (VADER) version 3.3.2 [47], which is a rule-based model to evaluate the sentiment of a given text [48]. VADER generates a normalized, weighted compound score that ranges between −1, the most extreme negative sentiment, and +1, the most positive sentiment. The second tool is Linguistic Inquiry and Word Count (LIWC) 2015 [49], which calculates the percentage of each linguistic category by mapping the words in a given text into a predefined word list of the linguistic category [50]. This tool has been widely adopted in online content-based research [51]. In this study, we focused on the negative emotion category in LIWC.

    Classification

    Classification Task

    After performing the topic and sentiment analysis, we built machine learning models to classify caregiver relationships, which can further disclose the difference in temporal posting behaviors between adult-child and spousal caregivers. To do so, we conducted a post-level prediction by predicting the relationship for each post. We then performed a user-level prediction by applying a majority voting model to the prediction labels of all the posts of a caregiver. We represented the post-level prediction as a 3-class classification, where each post was an instance and the label was the corresponding author’s self-reported relationship. The latter was mapped into one of the three relationship categories: (1) adult child, (2) spouse, and (3) other caregiver. This multiclass classification design enabled the trained models to be readily applied to any other ADRD forums.

    Experiment Design

    As the post volume of each online caregiver followed a long-tailed distribution (refer to the Results section), including all the available posts in model training would have biased the fitted model to superactive users. Therefore, to balance between including more training data and mitigating the bias induced by the long-tailed distribution, we proposed a training and evaluation procedure as described in the following subsections.

    Data Preparation

    We removed posts with ≤10 words. This decision was based on a manual review of 100 randomly selected posts, which showed that 83% (83/100) were about greeting or gratitude, website URLs, or incomplete sentences. We applied a stratified shuffle split on user level and selected all the posts for 80% (14,729/18,412) of the users as the training data, denoted by Dtrain, and the remaining posts as test data, denoted by Dtest. In doing so, we ensured that the ratio of the user-level labels remained the same in Dtrain and Dtest and there was no overlap in the authorship across the 2 data sets.

    Models

    We created 4 training subsets by selecting no more than i posts from each caregiver in Dtrain, where i in [5, 10, 15, 20] as selecting an exact i number of posts might result in subsets that were either too small owing to data sparsity or not diverse enough to capture the full range of variability of the data. We used each subset to train a model Mi. We used the Google Pretrained Bidirectional Encoder Representations from Transformers (BERT) fine-tuning model (“bert-base-uncased” in Transformer package v2.8.0) provided by the Hugging Face Transformers as the algorithm. For a fair comparison, we configured each model with an embedding size of 256, batch size of 32, learning rate of 2×10−5, and number of training epochs of 10. It should be noted that in addition to the fine-tuned BERT model, we experimented with other models including logistic regression, random forests, and a bidirectional long short-term memory with attention. However, these models did not outperform the BERT model and hence are not presented in this paper.

    Evaluation

    We grouped the caregivers in Dtest into 5 bins based on their post volumes: [1, 1], (1, 5], (5, 10], (10, 100], and (100, ∞). For instance, the user bin (1, 1) comprises caregivers who have composed merely 1 post, whereas the user bin (100, ∞) contains those who have contributed >100 posts. We calculated and compared the model performance in each user bin. Specifically, in each user bin, we randomly sampled 80% (2946/3683) of caregivers and used all their posts as a test subset to calculate the macro area under the precision-recall curve (AUPRC), which is preferred when the labels are unbalanced. We repeated the process 10 times and reported the mean and SD of AUPRC. We also reported precision, recall, and F1-score for each type of relationship prediction and the overall macro AUPRC by counting the total true positives, false negatives, and false positives.

    Temporal Trend

    To characterize the temporal patterns of the predictability of online posts in determining relationships, we applied the model with the best performance to calculate the predicted probability of the true label for each post. For example, if a post was composed by a spousal caregiver, we estimated the probability that the model predicts the author as a spousal caregiver. In this task, we focused solely on adult children and spousal caregivers and analyzed the predicted probabilities along with the relative days for all the caregivers in the test data set.


    Activity Statistics

    Overview

    Table 1 summarizes the posts, users, and their active days in the 3 ALZConnected forums. In total, we collected 534,205 posts that were published by 18,622 users in the community. Among the 534,205 posts, 56,737 (10.62%) were top posts. Most of the posts were composed in the Caregivers Forum (287,556/534,205, 53.82%) and Spouses Forum (238,068/534,205, 44.56%), whereas only a small proportion (8581/534,205, 1.6%) was posted in the Lost Forum. The users in the Spouses Forum were more active in posting (averaging 50 posts/user) compared with the users in the Caregivers Forum (18 posts/user) and the Lost Forum (12 posts/user), and they also had the largest mean active days of 829 (SD 853) days.

    Table 1. The number of posts, users, and their active days in the 3 ALZConnected forums.

    		OverallForum


    		CaregiversSpousesLost
    Posts, n (%)534,205 (100)287,556 (53.82)238,068 (44.56)8581 (1.61)
    Top posts, n (%)56,737 (100)34,840 (61.41)20,767 (36.6)1130 (1.99)
    Users, n (%)18,622 (100)15,666 (84.13)4726 (25.38)692 (3.72)
    Posts per user, mean (SD)29 (275)18 (158)50 (355)12 (37)
    Active days, mean (SD)689 (791)572 (718)829 (853)751 (708)
    Posts and Relationships

    Figure 1 presents the number of posts (left) and the number of unique caregivers (right) for each type of self-reported relationship in the 3 forums. The overall column counts the total number of posts and caregivers in each forum. Adult-child caregivers (10,997/18,265, 60.21%) and spousal caregivers (4356/18,265, 23.85%) constituted most of the users (15,353/18,265, 84.05%) across the 3 forums. Together, they contributed most of the posts (474,015/508,091, 93.29%) in these 3 forums, with 42.1% (213,851/508,091) of adult children posts and 51.2% (260,164/508,091) of spousal posts. This observation was consistent with a previous study that showed that adult children and spouses comprise the 2 two largest proportions of the informal caregivers for persons living with dementia [41].

    ‎
    Figure 1. The number of posts (left) and the number of users (right) by relationship types in the 3 forums. The overall columns report the total number of posts (left) and unique users (right).

    Although most spousal (or adult children) posts were typically composed in the Spouses (or Caregivers) Forum, we found that a large number of posts were made elsewhere. For example, 45.09% (1964/4356) of the spousal caregivers published 19.98% (54,327/271,887) of the posts in the Caregivers Forum, whereas 6.78% (746/10,997) of the adult-child caregivers generated 9.19% (20,922/227,768) of the posts in the Spouses Forum. In addition, 88.3% (586/664) of the caregivers in the Lost Forum were either adult children or spouses. They composed 96.37% (8125/8431) of the posts in the forum. Other commonly reported relationships were relatives (1173/18,265, 6.42%), grandchildren (825/18,265, 4.51%), friends (322/18,265, 1.76%), siblings (247/18,265, 1.35%), and neighbors (31/18,265, 0.17%), and together they contributed 6.71% (34,076/508,091) of posts in the 3 forums. It should be noted that 1.04% (5308/508,091) of the posts were composed by professional caregivers, which suggested a skewed but diverse ADRD caregiver population in this online community.

    Views and Comments

    Figure 2 (left) compares the views and comments for the topic threads (on a log scale). It was evident that the overwhelming majority of the density for the view of topic threads centered around e7≈1100 (with a range of 43-285,450), with spousal topic threads receiving more views than adult children topic threads (under a 2-sample Kolmogorov-Smirnov test, d=0.24; P<.001). By contrast, the distributions of the comments for both types of caregivers had more modes, and most of the comment numbers were concentrated on 0, 1, and 7 (with a range of 0-2359), which was far less than the views. The 100× difference indicated that the collective knowledge generated in this open community may be consumed by a much larger number of caregivers who either do not have an account or seldom engage in online discussions. A 2-sample Kolmogorov-Smirnov test showed that spousal topic threads received slightly more comments than the adult children topic threads (d=0.10; P<.001).

    ‎
    Figure 2. The distribution of comments and views for each topic thread. The x-axis is log transformed for visualization purposes (left); the scatterplot shows posting volume and active days among the caregivers (right). The density plots of both statistics are illustrated along the top and to the right.
    Posts and Active Days

    Figure 2 (right) depicts the scatterplot of post volume and active days for each user (on a log scale). Our study yielded several noteworthy findings. First, both post volume and active days followed a long-tailed distribution, which revealed that only a small number of caregivers published a large number of posts and were active for a long time. For example, the top 1% (109/10,997) of the adult children (spousal, 43/4356, 1%) caregivers with the largest post volume generated 56.52% (120,863/213,851; 124,474/260,164, 47.84%) of all of the adult-child (spousal) posts. The number of posts ranged from 1 to 18,662, whereas the number of active days ranged from 1 to 3899. Second, except for online caregivers who posted only once, most caregivers were short-term online users who published a limited number of posts during a short period. For instance, 89.58% (8519/9510) of the users published <50 posts, and 87.09% (8283/9510) of the users were active for <2 years. Although not shown in Figure 2 (right), 54.82% (4307/7856) of these short-term users (published <50 posts and active for <2 years) were more likely to participate in online discussions within their own topic threads than respond to other caregivers’ topic threads. For these online users, their primary focus might be on addressing specific caregiving challenges or questions. Third, among long-term users who published >50 posts within an >2-year active period, 89.4% (504/564) posted more on other caregivers’ topic threads than on their own topic threads. These caregivers, who are described as “veteran users” of the online community [52], preferred to interact with other online caregivers intensively. Under the 2-sample Kolmogorov-Smirnov test, spousal caregivers tended to publish more posts (d=0.11; P<.001) and had longer active days (d=0.14; P<.001).

    Caregiving Topics

    The optimal number of topics was determined to be 25. Table 2 presents these topics, along with their most representative words (top words) based on the highest word probabilities in a topic and the expected topic proportion in the entire data that were used to train the STM.

    Table 2. Summary of the 25 topics. The words in each topic are ranked in decreasing order based on their estimated probabilities in the topic.
    Group and topicTop wordsExpected topic proportion
    Daily caring issues

    		#12: Balanced livingtake, home, day, work, will, time, hous, need, come, get0.056

    		#13: Sleepingnight, day, sleep, bed, idn’t, walk, get, last, pain, hour0.039

    		#18: Showeringuse, clean, get, bathroom, shower, put, cloth, room, floor, wear0.032

    		#11: Financial issuespay, money, phone, bill, check, get, insur, idn’t, call, can0.028

    		#1: Transportationcar, door, walk, open, hous, store, idn’t, drive, light, key0.028

    		#14: Travelingdog, get, morn, hope, one, cat, ill, good, today, got0.022

    		#9: Dietseat, food, drink, dinner, meal, feed, cook, water, weight, mouth0.021
    Disease related

    		#20: Loss of expressionsay, tri, thing, get, tell, talk, ask, someth, time, always0.053

    		#16: Disease diagnosisyear, ago, month, now, stage, last, sinc, diagnos, past, week0.053

    		#15: Loss of memorymemori, seem, chang, can, becom, time, often, word, still, even0.043

    		#24: Nursing homecare, home, facil, nurs, place, move, visit, assist, memori, staff0.042

    		#21: Medicationdoctor, med, take, medic, hospit, day, agit, also, help, week0.040

    		#10: Last wishesneed, will, care, medic, make, health, issu, doctor, decis, state0.035

    		#25: Dementiatest, dementia, alzheim, diseas, brain, diagnosi, symptom, neurologist, cognit, result0.030

    		#17: COVID19idn’t, death, patient, program, covid, case, famili, provid, may, state0.026
    Emotion

    		#7: Feelingsfeel, husband, life, never, diseas, like, hard, ever, know, cri0.044

    		#2: Gratitudelove, day, time, friend, wife, happi, miss, watch, one, daughter0.043

    		#3: Blessingsgod, will, love, lord, peac, bless, pray, give, can, jesus0.038
    Relationship

    		#8: Relativesmother, live, father, famili, sister, caregiv, parent, help, dementia, care0.028

    		#6: Relativesmom, dad, shes, brother, sister, visit, also, move, want, help0.020
    Others

    		#4: Common verbsjust, know, dont, want, think, cant, realli, like, can, doesn’t0.082

    		#5: Common questionsthank, help, post, read, caregiv, can, anyon, find, mani, support0.065

    		#19: Common wordsone, good, well, look, like, thing, lot, thought, way, new0.050

    		#22: Common verbssaid, call, told, got, idn’t, ask, went, today, back, took0.050

    		#23: Timingstart, back, get, happen, stop, around, turn, head, just, hand0.032

    Caregiving discussions were mainly focused on 4 primary topic groups: daily caregiving issues, disease related, sentiment and relationship, together with some common verbs in online communications, which we categorized as others. In particular, the disease related topic group had the largest proportion, contributing 32.3% of the discussion. The second largest topic group was about daily caregiving matters, which included #13 (sleeping issues; 3.9%; eg, “My mom didn’t sleep well last night, and now she’s napping during the day. Should I let her nap?”) and #11 (financial issues; 2.8%; eg, “I am now my mom’s social security payee and I had to use my [bank name] to intervene in my mom’s account because she gave the money to a scammer”) and other issues that were commonly reported in the literature [53,54]. The emotion topic group included #7 feelings (4.4%), #2 gratitude (4.3%), and #3 blessings (3.8%), which mainly describe the mixed emotions they experienced. The relationship topic group was embodied by topics #6 (2%) and #8 (2.8%), both of which were about relatives. On the basis of the top words and their related posts, topic #6 was mentioned more in family conflicts, whereas topic #8 was mentioned more in other scenarios.

    Figure 3 compares adult children and spousal caregivers with respect to topic prevalence. There are several findings to highlight. First, although there was a smaller number of spousal caregivers than adult-child caregivers in ALZConnected, spouses discussed a wider range of topics than adult children, including #14 (traveling), #1 (transportation), #25 (dementia), #17 (COVID-19), #18 (showering), #15 (loss of memory), and #16 (disease diagnosis). This aligned well with the literature, which indicated that spousal caregivers report a need to redefine their role and relationship with persons living with dementia [55].

    ‎
    Figure 3. A contrast can be seen in topic prevalence between adult-child and spousal caregivers (with 95% CIs). A positive (negative) value along the x-axis indicates that the topic is more prevalent among spousal (adult-child) caregivers.

    Second, the 2 topics related to #6 and #8 (relatives) and #24 (nursing home) were more frequently discussed by adult children (eg, “my mother will be moving to my sister’s house and she will become my mother’s full time caregiver.... My sister refuses to learn about dementia caregiving. I suggested...move to her house with my mother for a few weeks...she thinks this is unnecessary.”). This suggested that adult children faced more family conflicts when caring for persons living with dementia than spouses [33]. In addition, adult children seemed to discuss the medical topic (#21) more frequently than spouses.

    Third, in the emotion topic group, adult children posted content about “praying” in the forums, such as topic #3 blessings (eg, “And I will ask the Father...give you another Helper...to be with you forever”), whereas spouses expressed more personal feeling during daily care, either gratitude or sadness, such as topic #2 gratitude (eg, “I remember when: We laughed together We cried together.... We loved life together We got through tough times together”) and #7 feelings (eg, “I feel guilty...for not being patient with my DH...guilty about mentioning how hard it is to people who really don’t understand and think I’m exaggerating”).

    Sentiment Analysis

    Figures 4 and 5 illustrate how sentiment changed over time by the VADER compound score and LIWC negative emotion, respectively. To improve readability, we averaged the VADER compound scores or LIWC negative emotion for each of the caregivers during an active day into a single point. The linear interpolation and its 95% CI are shown in Figures 4 and 5.

    ‎
    Figure 4. Sentiment in online posts according to the Valence Aware Dictionary for Sentiment Reasoning (VADER) compound sentiment score for adult-child (left) and spousal (right) caregivers. To improve readability, all of 1 day’s scores are averaged into a single point.

    As shown in Figure 4, which corresponds to the VADER compound score, the average sentiments for both caregiver types were consistently >0.4 and maintained an increasing trend. Compared with spouses, adult children exhibited a higher growth rate in the VADER compound score. The Spearman rank-order correlation between the VADER compound score and relative days for adult children and spouses were 0.42 and 0.17 (P<.001), respectively. This observation is in agreement with a previous finding that spousal caregivers expressed both gratitude and negative feelings relatively more than adult-child caregivers [56]. Meanwhile, according to the LIWC negative emotion, as shown in Figure 5, both caregiver types exhibited a decrease in negative emotion over time. The Spearman rank-order correlation between the LIWC negative emotion and relative days for adult children and spousal caregivers were −0.36 and −0.44 (P<.001), respectively. The improved sentiment over time echoes the finding of a prior study that showed that a caregiver’s engagement in an online community is related to a reduction in depressive symptoms [21].

    ‎
    Figure 5. Sentiment in online posts according to the Linguistic Inquiry and Word Count (LIWC) negative emotion score for adult-child (left) and spousal (right) caregivers. To improve readability, all of 1 day’s scores are averaged into a single point.

    Classification

    Overview

    After filtering out the posts with ≤10 words, there were 508,030 posts composed by 18,412 caregivers. The ratio of user-level labels was 59:24:17 in class of adult child and spouse and other. The 4 training subsets in Dtrain, where i∈[1,5,10,100], were corresponding to 3.57% (14,729/413,101), 9.42% (38,931/413,101), 13.05% (53,924/413,101), and 33.98% (140,379/413,101) percentile of all the caregivers in Dtrain based on their post volume. The 5 bins of the test subsets were corresponding to 40.62% (1496/3683; 1496/94,929, 1.58%), 34.21% (1260/3683; 2520/94,929, 2.66%), 12.76% (470/3683; 3675/94,929, 3.87%), 9.04% (333/3683; 14,442/94,929 15.21%), 3.37% (124/3683; 72,796/94,929, 76.68%) of the caregivers (posts) in Dtest.

    Model Performance

    Table 3 presents the overall model performance, where Mi represents each model training set, where i denotes the maximum number of posts randomly sampled by each user. We want to emphasize several important findings. First, M100 achieved the best AUPRC in post-level prediction 0.737, whereas M10 achieved the best AUPRC in user-level prediction 0.813. Both of them significantly outperformed the second-best model under a 2-tailed t test (P<.001). Second, owing to the imbalanced labels, all the models had lower model performance in classifying other than classifying adult child and spouse. Third, the models trained using a smaller number of user posts (eg, M1 and M5) had better performance in predicting adult child in both tasks, whereas the models trained using a larger number of user posts (eg, M10 and M100) had better performance in predicting Spouse in both tasks. These observations confirm the necessity of examining model performance in different user bins based on the number of their published posts.

    Table 3. Performance of prediction at the post level and user level for the machine learning models. All numbers correspond to percentage values.
    TaskPost-level prediction (%), mean (SD)User-level prediction (%),mean (SD)

    		M1M5M10M100M1M5M10M100
    Adult child

    		Precision61.2 (0.04)63.2 (0.04)a61.6 (0.04)62.3 (0.03)81.4 (0.13)86.4 (0.17)85.7 (0.19)85.4 (0.19)

    		Recall81.8 (0.05)76.7 (0.06)78.7 (0.06)78.5 (0.04)94.3 (0.15)92.7 (0.17)93.4 (0.14)93.0 (0.18)

    		F1-score70.0 (0.04)69.3 (0.04)69.1 (0.04)69.5 (0.03)87.4 (0.09)89.5 (0.12)89.4 (0.12)89.1 (0.12)
    Spouse

    		Precision83.7 (0.06)81.2 (0.03)81.6 (0.05)80.9 (0.05)81.9 (0.40)81.1 (0.33)82.4 (0.38)80.2 (0.52)

    		Recall57.7 (0.06)65.4 (0.60)64.9 (0.06)67.8 (0.07)87.5 (0.41)90.6 (0.34)91.4 (0.34)91.3 (0.33)

    		F1-score68.3 (0.05)72.5 (0.04)72.3 (0.05)73.8 (0.06)85.6 (0.26)85.6 (0.27)86.7 (0.25)85.4 (0.34)
    Other

    		Precision14.8 (0.08)20.0 (0.10)21.6 (0.15)23.4 (0.12)66.6 (0.76)74.9 (0.54)77.6 (0.58)76.3 (0.46)

    		Recall20.9 (0.14)25.0 (0.16)21.7 (0.15)18.8 (0.12)24.8 (0.49)44.6 (0.78)42.6 (0.79)39.5 (0.68)

    		F1-score17.3 (0.10)22.2 (0.13)21.7 (0.15)20.9 (0.12)36.1 (0.60)55.9 (0.70)55.0 (0.75)52.0 (0.66)
    AUPRC71.6 (0.04)72.8 (0.04)72.8 (0.05)73.7 (0.05)75.9 (0.18)80.6 (0.25)81.3 (0.21)80.4 (0.26)

    aThe best performance for each metric is highlighted is italicized.

    Figure 6 illustrates the performance of the AUPRC model for different user bins. As the post volume increased from 1 to 100, all the models had increased AUPRC scores for user-level prediction. However, their AUPRC scores dropped when the post volume of a user exceeded 100. By contrast, there was a consistent decrease in AUPRC scores as the post volume increased for post-level predictions. These opposing observations are primarily owing to the majority voting used for user-level predictions. Although more posts generally improve user-level AUPRC through a majority vote, the dramatic decrease of AUPRC scores in the post-level prediction for the (100,∞) bin negatively affects the user-level prediction. Second, M1 performed worst in most scenarios, suggesting that the sampling of 1 post strategy leads to underfitted models. By contrast, although M100 had the best AUPRC (Table 3 in post-level prediction, it only outperformed other models at a statistically significant level (t test; P<.001) in bin (100,∞) which contained many more posts than other bins. Notably, M10 outperformed M100 in a statistically significant manner (t test; P<.001) in bin (10,100].

    ‎
    Figure 6. User-level (left) and post-level (right) model performance, in terms of area under the precision-recall curve (AUPRC), of different user bins in the test data set. Note that each user bin defines a post volume range. For example, a user bin of (1, 1) includes caregivers who only published 1 post, whereas a user bin of (100, ∞) includes caregivers who published >100 posts.
    Error Analysis

    To understand the challenge of the prediction tasks, we performed an error analysis of the classification outputs from M10, which had the best user-level prediction and less bias toward superactive users when compared with M100.

    Figure 7 depicts the confusion matrix for the entire test for both post-level and user-level predictions. The most difficult task is to distinguish other from adult child. For example, the misclassified ratio between adult child versus other and spouse versus other was (296+57)/(70+6)≈4.6 in user-level prediction and (967+280)/(296+94)≈3.2 in post-level prediction, compared with 59/24≈2.5, the ratio between adult child and spouse in Dtrain. After a manual examination, we observed that many caregivers reporting relative relationships with persons living with dementia were children-in-law of persons living with dementia. Here is one such example that is labeled as other but predicted as adult child: “We moved the MIL to our neighborhood.... She is always mad at FIL.... We would like to know how to ease her mood.” We also observed that most of the 100 randomly selected misclassified posts were short comments expressing gratitude or asking follow-up questions; posts that either mentioned multiple relationships within a long, complex story or did not mention any specific relationship; or posts that asked about balanced living or financial issues. For example, “[NAME], bless you for helping your mom...one person carries a daunting responsibility.” was predicted as adult child but it was actually composed by an other caregiver.

    ‎
    Figure 7. Confusion matrix of the M10 output for post-level (left) and user-level (right) predictions.
    Temporal Patterns

    Figure 8 illustrates the temporal patterns in the prediction probability for the 87,326 posts from 3039 adult children and spousal caregivers in Dtest. To improve readability, each point in the figures corresponds to the mean value of the prediction probabilities for a relative day. A linear interpolation and its 95% CI are also shown in each subfigure. There was a decreasing temporal trend for both caregiver types. For adult child and spouse, the Spearman rank correlation between the probabilities and relative days was 0.40 and −0.09 (P<.001), respectively. Notably, the longer a caregiver was active in the online community, the more challenging it was to classify the author’s relationship for a post. This suggests that the caregivers shifted their discussion focuses over time. Considering that long-term or veteran users were more likely to comment on other caregivers’ topic threads, some of them might become a specialist or a generalist in this ADRD caregiving online community [57]. However, this phenomenon was less evident among spousal caregivers, indicating their relatively more consistent caregiving experience (eg, caring for persons living with dementia in their daily activities [33]), regardless of the dynamic trajectory of informal caregiving for persons living with dementia [58].

    ‎
    Figure 8. Prediction of caregiver relationship with persons living with dementia over time for adult children (left) and spouses (right).

    Principal Findings

    This study reports on the online behaviors of adult-child and spousal caregivers in a large online community. There are several principal findings of merit. First, ADRD caregivers are diverse in terms of their relationships with persons living with dementia as well as in the ways they use social support in the online community. For example, the high volume of views in ALZConnected indicates that many online users are “lurkers” or collective knowledge consumers who mostly read online discussions [59]. However, these online caregivers are often neglected by researchers because of a lack of their online voices. Moreover, a majority of the online caregivers who write posts, known as collective knowledge producers, are short-term users. This brevity in platform use is most likely an artifact of their focus on addressing a specific caregiving issue. In recognition of this observation, in addition to the typical goal in the literature to improve caregivers’ long-term well-being [28,60], research should aim to solve the short-term needs of most online caregivers as well.

    Notably, there were more adult children than spousal caregivers in ALZConnected. This may be partly because the younger generation is more likely to use the internet. However, spouses tended to publish relatively more posts and actively participate longer in ALZConnected than adult children. Spousal topic threads also received more comments than adult children’s topic threads. Adult-child and spousal caregivers shared many common concerns about issues related to living, sleeping, diet, and finance. Spouses were more likely to discuss other caregiving-related issues such as showering, transportation, disease diagnosis, and the COVID-19 pandemic. This may be due to the close physical bond that a spouse has with the care recipient, the emotional bond, or either a combination of both, as suggested in previous research [7].

    Our sentiment analysis showed that both caregiver types experienced an increasing positive sentiment over time in this community, indicating a positive impact on the emotional well-being of caregivers who pursued long-term online support. This is consistent with findings from observational studies that indicate that participating in online communities reduces caregivers’ stress levels [28,30]. However, spousal caregivers had relatively lower sentiment improvement, which, together with their wider range of caregiving discussions (Figures 4 and 5), suggests that they experience more complex challenges than adult-child caregivers.

    The best classifier obtained an AUPRC of 81.3% on the user-level prediction. From an algorithmic perspective, our work contributed to an important biomedical research field in estimating online users’ sociodemographic factors from their posts [61,62] by accurately predicting online caregivers’ relationships with persons living with dementia based on their posts. relationships with persons living with dementia based on their posts. Specifically, these factors can be effectively adjusted in any statistical analysis related to online ADRD caregiving discussion. In addition, given that informal dementia caregivers often share their experiences on any online platform they prefer, our classifiers can be applied to help caregivers locate the posts that are published by the same type of caregivers, which will be likely to meet their specific caregiving needs.

    The observed decreasing trend in the probability curves on the posting timeline for both caregiver types suggests that online caregivers tended to change or expand their discussion scope over time. This is because these long-term online caregivers are more likely to support other caregivers by composing comments. Community organizers should consider fostering these “veteran users” to sustain the community. Again, the trend of a relatively lower decreasing prediction probability for spousal caregivers provided more evidence that they behave differently from adult-child caregivers. Therefore, tailored strategies should be devised to assist these caregivers based on their relationship type. For instance, spousal caregivers could benefit from an emphasis on daily caregiving matters and emotional support, whereas adult-child caregivers might find greater value in discussions concerning nursing homes, medications, symptoms, and family conflicts. The error analysis showed that the model was more likely to make incorrect predictions between adult children and other caregivers, who, based on a manual examination of the posts, were found to be more likely to be the children-in-law of persons living with dementia. This suggests that children-in-law caregivers, although a small proportion, may share similar caregiving experiences or challenges with adult-child caregivers and should receive attention from the research community and society as well [63].

    More broadly, our findings imply that open online communities such as ALZConnected can be leveraged to provide sufficient informational or emotional support to a wide range of informal ADRD caregivers, especially those who live in rural areas where local support services or resources are generally limited [64] and those who hesitate to seek offline support owing to stigma [65]. However, any such translational research is recommended to consider the following issues to ensure fairness and effectiveness of accessing and using online social support: (1) the digital divide caused by either unavailable high-speed internet, low eHealth literacy, or intensive caregiving activities [66]; (2) the misinformation or misconceptions regarding diseases or related treatments [67]; and (3) any potential hatred language that is not uncommon in online environments [68].

    Limitations and Future Work

    Despite the merit of our findings, there are several limitations of our study that can serve as the basis for future research.

    First, our research focused on a single online community, which may have limited the generalizability of our findings. Therefore, it is necessary to validate our findings using other online communities. However, as ALZConnected is the largest ADRD online community in North America, these findings should be informative to guide future research in this area. Second, we were unable to examine the online behaviors of “lurker” caregivers or include the sociodemographic characteristics (eg, age, gender, sexuality, race, and ethnicity) of online users in our analysis because of the lack of availability of such data in general online environments. Thus, we believe that future research can consider incorporating surveys with users to complement the research findings drawn from merely online posts. Third, we only examined the sentiment encoded in the online discussions. There is clearly an opportunity for investigations that integrate observational research into the study to help measure the psychological well-being of online caregivers, which, together with computational methods, examines how exactly online interactions could improve a caregiver’s well-being. Fourth, we did not evaluate the prediction models in other online communities. This is deemed to be outside the scope of this study, as it requires an intensive survey to collect gold standard relationship information. However, given the promising outcomes of this research, we believe that the built models can be easily adapted to predict the kin relationship between caregivers and care recipients based on the posts in other online ADRD caregiving communities, which can be achieved by using either transfer learning [69] or retraining the models using new annotated data.

    Conclusions

    To the best of our knowledge, this is the first study to examine and compare the online behaviors of adult children and spousal ADRD caregivers in a large open online community. Our results show that these 2 caregiver groups exhibited different online behaviors, with spousal caregivers posting more frequently and discussing a wider range of caregiving issues. These observations indicate that spouses experience more complex challenges than adult children. Researchers and community organizers should take into account the heterogeneity of online behaviors, which is likely to be owing to different caregiving experiences, to improve the online experience of different types of caregivers. Further research can be applied to explore how online interactions can address specific caregiving challenges at the topic thread level and benefit both short-term problem-solving and long-term social integration.

    Data Availability

    All the data analyzed in this study are publicly available.

    Authors' Contributions

    ZY conceived the initial idea of this study. CN and QS conducted the experiments, performed data analysis, and validated the computational methods. L Song identified relevant theories. CN, QS, BM, L Song, PC, L Stratton, and ZY actively engaged in the discussion of the results and contributed to the editing of the final manuscript.

    Conflicts of Interest

    None declared.

    1. Chertkow H, Feldman HH, Jacova C, Massoud F. Definitions of dementia and predementia states in Alzheimer's disease and vascular cognitive impairment: consensus from the Canadian Conference on Diagnosis of Dementia. Alzheimers Res Ther. Jul 08, 2013;5(Suppl 1):S2. [FREE Full text] [CrossRef] [Medline]
    2. GBD 2019 Dementia Forecasting Collaborators. Estimation of the global prevalence of dementia in 2019 and forecasted prevalence in 2050: an analysis for the Global Burden of Disease Study 2019. Lancet Public Health. Mar 2022;7(2):e105-e125. [FREE Full text] [CrossRef] [Medline]
    3. 2023 Alzheimer's disease facts and figures. Alzheimers Dement. Apr 2023;19(4):1598-1695. [CrossRef] [Medline]
    4. Manzini CS, do Vale FA. Emotional disorders evidenced by family caregivers of older people with Alzheimer's disease. Dement Neuropsychol. Mar 2020;14(1):56-61. [FREE Full text] [CrossRef] [Medline]
    5. Sajwani Z. Challenges of caregivers of patients with Alzheimer’s disease. Innov Aging. 2020;4(Supplement_1):149. [FREE Full text] [CrossRef]
    6. Brodaty H, Donkin M. Family caregivers of people with dementia. Dialogues Clin Neurosci. 2009;11(2):217-228. [FREE Full text] [CrossRef] [Medline]
    7. Statz TL, Peterson CM, Birkeland RW, McCarron HR, Finlay JM, Rosebush CE, et al. “We moved her too soon”: navigating guilt among adult child and spousal caregivers of persons living with dementia following a move into residential long-term care. Couple Fam Psychol Res Pract. May 13, 2021;11(4):300-314. [CrossRef]
    8. Garcia N, Papazyan A, Choi S, Song Y. Caregiving between spouse and adult child caregivers of older adults with cognitive impairment. Innov Aging. 2021;5(Supplement_1):791-792. [FREE Full text] [CrossRef]
    9. Shi H, Mao C, Tang J, Liang H. Research on the health of and interventions for family caregivers of people with dementia: a bibliometric analysis of research output during 1988-2018. BMC Geriatr. Jan 21, 2020;20(1):20. [FREE Full text] [CrossRef] [Medline]
    10. Kwok T, Au A, Wong B, Ip I, Mak V, Ho F. Effectiveness of online cognitive behavioral therapy on family caregivers of people with dementia. Clin Interv Aging. Apr 2014;9:631-636. [FREE Full text] [CrossRef] [Medline]
    11. Cheng ST, Mak EP, Kwok T, Fung H, Lam LC. Benefit-finding intervention delivered individually to Alzheimer family caregivers: longer-term outcomes of a randomized double-blind controlled trial. J Gerontol B Psychol Sci Soc Sci. Oct 16, 2020;75(9):1884-1893. [CrossRef] [Medline]
    12. Kuzu N, Beşer N, Zencir M, Sahiner T, Nesrin E, Ahmet E, et al. Effects of a comprehensive educational program on quality of life and emotional issues of dementia patient caregivers. Geriatr Nurs. Nov 2005;26(6):378-386. [CrossRef] [Medline]
    13. Sanders S, Ott CH, Kelber ST, Noonan P. The experience of high levels of grief in caregivers of persons with Alzheimer's disease and related dementia. Death Stud. Jul 04, 2008;32(6):495-523. [CrossRef] [Medline]
    14. Woods RT, Orrell M, Bruce E, Edwards RT, Hoare Z, Hounsome B, et al. REMCARE: pragmatic multi-centre randomised trial of reminiscence groups for people with dementia and their family carers: effectiveness and economic analysis. PLoS One. Apr 19, 2016;11(4):e0152843. [FREE Full text] [CrossRef] [Medline]
    15. Cheng ST, Lau RW, Mak EP, Ng NS, Lam LC, Fung HH, et al. A benefit-finding intervention for family caregivers of persons with Alzheimer disease: study protocol of a randomized controlled trial. Trials. Jul 02, 2012;13(1):98. [FREE Full text] [CrossRef] [Medline]
    16. Song L, Chang TY. Do resources of network members help in help seeking? Social capital and health information search. Soc Netw. Oct 2012;34(4):658-669. [FREE Full text] [CrossRef]
    17. Laukka E, Rantakokko P, Suhonen M. Consumer-led health-related online sources and their impact on consumers: an integrative review of the literature. Health Informatics J. Jun 02, 2019;25(2):247-266. [FREE Full text] [CrossRef] [Medline]
    18. Song L. Nan Lin and social support. In: Burt RS, Bian Y, Song L, Lin N, editors. Social Capital, Social Support and Stratification: An Analysis of the Sociology of Nan Lin. Cheltenham, UK. Edward Elgar Publishing; Dec 5, 2019;78-106.
    19. Thoits PA. Mechanisms linking social ties and support to physical and mental health. J Health Soc Behav. Jun 14, 2011;52(2):145-161. [CrossRef] [Medline]
    20. Song L, Zhang Z. Social support. In: McLevey J, Scott J, Carrington PJ, editors. The Sage Handbook of Social Network Analysis. Thousand Oaks, CA. SAGE Publications, Inc; 2023.
    21. Song L, Chen W. Does receiving unsolicited support help or hurt? Receipt of unsolicited job leads and depression. J Health Soc Behav. Jun 14, 2014;55(2):144-160. [CrossRef] [Medline]
    22. Fields NL, Roark EM, Xu L. Leveraging volunteers to support dementia family caregivers: an innovative approach to care and support. In: Gaugler JE, editor. Bridging the Family Care Gap. Cambridge, MA. Academic Press; 2021;387-405.
    23. Cheng ST, Lam LC, Kwok T, Ng NS, Fung AW. The social networks of Hong Kong Chinese family caregivers of Alzheimer's disease: correlates with positive gains and burden. Gerontologist. Dec 31, 2013;53(6):998-1008. [CrossRef] [Medline]
    24. Berry N, Lobban F, Belousov M, Emsley R, Nenadic G, Bucci S. #WhyWeTweetMH: understanding why people use Twitter to discuss mental health problems. J Med Internet Res. Apr 05, 2017;19(4):e107. [FREE Full text] [CrossRef] [Medline]
    25. Pagán-Ortiz ME, Cortés DE, Rudloff N, Weitzman P, Levkoff S. Use of an online community to provide support to caregivers of people with dementia. J Gerontol Soc Work. Jul 29, 2014;57(6-7):694-709. [FREE Full text] [CrossRef] [Medline]
    26. Ni C, Malin B, Song L, Jefferson A, Commiskey P, Yin Z. “Rough day … need a hug”: learning challenges and experiences of the Alzheimer’s disease and related dementia caregivers on Reddit. Proc Int AAAI Conf Web Soc Media. May 31, 2022;16:711-722. [FREE Full text] [CrossRef]
    27. Kim H. Understanding internet use among dementia caregivers: results of secondary data analysis using the US caregiver survey data. Interact J Med Res. Mar 23, 2015;4(1):e1. [FREE Full text] [CrossRef] [Medline]
    28. Wilkerson DA, Brady E, Yi EH, Bateman DR. Friendsourcing peer support for Alzheimer’s caregivers using Facebook social media. J Technol Hum Serv. Apr 05, 2018;36(2-3):105-124. [CrossRef]
    29. Bachmann P, Hruska J. Alzheimer caregiving problems according to ADLs: evidence from Facebook support groups. Int J Environ Res Public Health. May 25, 2022;19(11):6423. [FREE Full text] [CrossRef] [Medline]
    30. Leszko M. The role of online communication among spousal caregivers of individuals with Alzheimer's disease. J Women Aging. Apr 17, 2020;32(4):462-480. [CrossRef] [Medline]
    31. Erdelez S, Tanacković SF, Balog KP. Online behavior of the Alzheimer's disease patient caregivers on Croatian online discussion forum. Proce Assoc Inf Sci Technol. Oct 18, 2019;56(1):78-88. [CrossRef]
    32. Benson JJ, Oliver DP, Washington KT, Rolbiecki AJ, Lombardo CB, Garza JE, et al. Online social support groups for informal caregivers of hospice patients with cancer. Eur J Oncol Nurs. Mar 2020;44:101698. [FREE Full text] [CrossRef] [Medline]
    33. Conde-Sala JL, Garre-Olmo J, Turró-Garriga O, Vilalta-Franch J, López-Pousa S. Differential features of burden between spouse and adult-child caregivers of patients with Alzheimer's disease: an exploratory comparative design. Int J Nurs Stud. Oct 2010;47(10):1262-1273. [CrossRef] [Medline]
    34. Savundranayagam MY. Receiving while giving: the differential roles of receiving help and satisfaction with help on caregiver rewards among spouses and adult-children. Int J Geriatr Psychiatry. Jan 22, 2014;29(1):41-48. [CrossRef] [Medline]
    35. Cheng K, Yin Z. "Please advise": understanding the needs of informal caregivers of people with Alzheimer's disease and related dementia from online communities through a structured topic modeling approach. AMIA Jt Summits Transl Sci Proc. May 23, 2022;2022:149-158. [FREE Full text] [Medline]
    36. Yin Z, Sulieman LM, Malin BA. A systematic literature review of machine learning in online personal health data. J Am Med Inform Assoc. Jun 01, 2019;26(6):561-576. [FREE Full text] [CrossRef] [Medline]
    37. Chancellor S, Baumer EP, De Choudhury M. Who is the "human" in human-centered machine learning: the case of predicting mental health from social media. Proc ACM Hum Comput Interact. Nov 07, 2019;3(CSCW):1-32. [FREE Full text] [CrossRef]
    38. Tsao SF, Chen H, Tisseverasinghe T, Yang Y, Li L, Butt ZA. What social media told us in the time of COVID-19: a scoping review. Lancet Digit Health. Mar 2021;3(3):e175-e194. [FREE Full text] [CrossRef] [Medline]
    39. Lee YJ, Jang H, Campbell G, Carenini G, Thomas T, Donovan H. Identifying language features associated with needs of ovarian cancer patients and caregivers using social media. Cancer Nurs. 2022;45(3):E639-E645. [CrossRef] [Medline]
    40. ALZConnected. URL: https://alzconnected.org/ [accessed 2023-04-14]
    41. Proferes N, Jones N, Gilbert S, Fiesler C, Zimmer M. Studying Reddit: a systematic overview of disciplines, approaches, methods, and ethics. Soc Media Soc. May 26, 2021;7(2) [FREE Full text] [CrossRef]
    42. Roberts ME, Stewart BM, Tingley D. stm: an R package for structural topic models. J Stat Softw. 2019;91(2):1-40. [FREE Full text] [CrossRef]
    43. Chandelier M, Steuckardt A, Mathevet R, Diwersy S, Gimenez O. Content analysis of newspaper coverage of wolf recolonization in France using structural topic modeling. Biol Conserv. Apr 2018;220:254-261. [FREE Full text] [CrossRef]
    44. Lindstedt NC. Structural topic modeling for social scientists: a brief case study with social movement studies literature, 2005–2017. Soc Curr. May 02, 2019;6(4):307-318. [CrossRef]
    45. Kumar V, Srivastava A. Trends in the thematic landscape of corporate social responsibility research: a structural topic modeling approach. J Bus Res. Nov 2022;150:26-37. [FREE Full text] [CrossRef]
    46. Mimno D, Wallach HM, Talley E, Leenders M, McCallum A. Optimizing semantic coherence in topic models. In: Proceedings of the Conference on Empirical Methods in Natural Language Processing. Presented at: EMNLP '11; July 27-31, 2011, 2011; Edinburgh, UK. URL: https://dl.acm.org/doi/pdf/10.5555/2145432.2145462
    47. Hutto C, Gilbert E. VADER: a parsimonious rule-based model for sentiment analysis of social media text. Proc Int AAAI Conf Web Soc Media. May 16, 2014;8(1):216-225. [FREE Full text] [CrossRef]
    48. Elbagir S, Yang J. Twitter sentiment analysis using natural language toolkit and VADER sentiment. In: Proceedings of the International MultiConference of Engineers and Computer Scientists 2019. Presented at: IMECS 2019; March 13-15, 2019, 2019; Hong Kong, China. [CrossRef]
    49. Pennebaker JW, Boyd RL, Jordan K, Blackburn K. The development and psychometric properties of LIWC2015. University of Texas at Austin. 2015. URL: https://repositories.lib.utexas.edu/bitstream/handle/2152/31333/LIWC2015_LanguageManual.pdf [accessed 2023-10-30]
    50. Chung CK, Pennebaker JW. Linguistic Inquiry and Word Count (LIWC): pronounced “luke,” . . . and other useful facts. In: McCarthy P, Boonthum-Denecke C, editors. Applied Natural Language Processing: Identification, Investigation and Resolution. Hershey, PA. IGI Global; 2012.
    51. Saha K, Yousuf A, Boyd RL, Pennebaker JW, De Choudhury M. Social media discussions predict mental health consultations on college campuses. Sci Rep. Jan 07, 2022;12(1):123. [FREE Full text] [CrossRef] [Medline]
    52. McDonald D, Woodward-Kron R. Member roles and identities in online support groups: perspectives from corpus and systemic functional linguistics. Discourse Commun. Jan 25, 2016;10(2):157-175. [CrossRef]
    53. Figueiro MG, Plitnick B, Roohan C, Sahin L, Kalsher M, Rea MS. Effects of a tailored lighting intervention on sleep quality, rest-activity, mood, and behavior in older adults with Alzheimer disease and related dementias: a randomized clinical trial. J Clin Sleep Med. Dec 15, 2019;15(12):1757-1767. [FREE Full text] [CrossRef] [Medline]
    54. Richards KC, Radhakrishnan K, Britt KC, Vanags-Louredo A, Park E, Gooneratne NS, et al. "Please don't forget us": a descriptive qualitative study of caregivers of older adults with Alzheimer's disease and related dementias during the COVID-19 pandemic. Res Gerontol Nurs. Sep 2022;15(5):217-228. [FREE Full text] [CrossRef] [Medline]
    55. Dang S, Looijmans A, Ferraris G, Lamura G, Hagedoorn M. Exploring the needs of spousal, adult child, and adult sibling informal caregivers: a mixed-method systematic review. Front Psychol. Mar 25, 2022;13:832974. [FREE Full text] [CrossRef] [Medline]
    56. Cheung DS, Ho KH, Cheung TF, Lam SC, Tse MM. Anticipatory grief of spousal and adult children caregivers of people with dementia. BMC Palliat Care. Nov 20, 2018;17(1):124. [FREE Full text] [CrossRef] [Medline]
    57. Waller I, Anderson A. Generalists and specialists: using community embeddings to quantify activity diversity in online platforms. In: Proceedings of the World Wide Web Conference. Presented at: WWW '19; May 13-17, 2019, 2019; San Francisco, CA. [CrossRef]
    58. van den Kieboom R, Snaphaan L, Mark R, Bongers I. The trajectory of caregiver burden and risk factors in dementia progression: a systematic review. J Alzheimers Dis. 2020;77(3):1107-1115. [FREE Full text] [CrossRef] [Medline]
    59. Gazit T, Bronstein J, Amichai-Hamburger Y, Aharony N, Bar-Ilan J, Perez O. Active participants and lurkers in online discussion groups: an exploratory analysis of focus group interviews and observation. Inf Res. Jun 2018;23(2) [FREE Full text]
    60. McKechnie V, Barker C, Stott J. The effectiveness of an internet support forum for carers of people with dementia: a pre-post cohort study. J Med Internet Res. Mar 28, 2014;16(2):e68. [FREE Full text] [CrossRef] [Medline]
    61. Shrestha P, Rey-Villamizar N, Sadeque F, Pedersen T, Bethard S, olorio T. Age and gender prediction on health forum data. In: Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC'16). Presented at: LREC'16; May 23-28, 2016, 2016; Portorož, Slovenia. URL: https://aclanthology.org/L16-1541.pdf
    62. Black CM, Meng W, Yao L, Ben Miled Z. Inferring the patient's age from implicit age clues in health forum posts. J Biomed Inform. Jan 2022;125:103976. [FREE Full text] [CrossRef] [Medline]
    63. Pinquart M, Sörensen S. Spouses, adult children, and children-in-law as caregivers of older adults: a meta-analytic comparison. Psychol Aging. Mar 2011;26(1):1-14. [FREE Full text] [CrossRef] [Medline]
    64. Williamson HJ, McCarthy MJ, Garcia YE, Bacon R, Dunn DJ, Baldwin JA. Addressing the needs of rural caregivers of individuals with Alzheimer's disease and related dementias during and beyond coronavirus disease 2019 (COVID-19). Public Policy Aging Rep. Sep 04, 2020;30(4):178-180. [FREE Full text] [CrossRef] [Medline]
    65. Velilla L, Acosta-Baena N, Allen I, Lopera F, Kramer J. Analysis of family stigma and socioeconomic factors impact among caregivers of patients with early- and late-onset Alzheimer's disease and frontotemporal dementia. Sci Rep. Jul 25, 2022;12(1):12663. [FREE Full text] [CrossRef] [Medline]
    66. Arighi A, Fumagalli GG, Carandini T, Pietroboni AM, De Riz MA, Galimberti D, et al. Facing the digital divide into a dementia clinic during COVID-19 pandemic: caregiver age matters. Neurol Sci. Apr 18, 2021;42(4):1247-1251. [FREE Full text] [CrossRef] [Medline]
    67. Cheng TY, Liu L, Woo BK. Analyzing Twitter as a platform for Alzheimer-related dementia awareness: thematic analyses of tweets. JMIR Aging. Dec 10, 2018;1(2):e11542. [FREE Full text] [CrossRef] [Medline]
    68. Castaño-Pulgarín SA, Suárez-Betancur N, Vega LM, López HM. Internet, social media and online hate speech. Systematic review. Aggress Violent Behav. 2021;58:101608. [FREE Full text] [CrossRef]
    69. Torrey L, Shavlik J. Transfer learning. In: Handbook of Research on Machine Learning Applications and Trends: Algorithms, Methods, and Techniques. Hershey, PA. IGI Global; 2010.

    ‎
    ADRD: Alzheimer disease or related dementias
    AUPRC: area under the precision-recall curve
    BERT: Google Pretrained Bidirectional Encoder Representations From Transformers
    LIWC: Linguistic Inquiry and Word Count
    STM: structural topic modeling
    VADER: Valence Aware Dictionary for Sentiment Reasoning

    Edited by T de Azevedo Cardoso; submitted 15.04.23; peer-reviewed by J Guendouzi, Y Chi; comments to author 15.08.23; revised version received 05.09.23; accepted 11.10.23; published 17.11.23.

    Copyright

    ©Congning Ni, Qingyuan Song, Bradley Malin, Lijun Song, Patricia Commiskey, Lauren Stratton, Zhijun Yin. Originally published in the Journal of Medical Internet Research (https://www.jmir.org), 17.11.2023.

    This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in the Journal of Medical Internet Research, is properly cited. The complete bibliographic information, a link to the original publication on https://www.jmir.org/, as well as this copyright and license information must be included.