Over recent years, mental health problems have become increasingly prevalent worldwide [1-3]. However, access to professional therapy remains limited for many individuals due to high treatment costs and a shortage of trained mental health professionals [4-6]. As a result, a growing number of individuals are seeking support for their problems through online mental health communities (OMHCs) [7-9]. These communities provide a more accessible alternative to professional therapy [10,11], allowing individuals to share their distressing experiences and receive peer support to help them cope [11,12]. These problems, typically shared through posts within OMHCs, not only reveal support seekers’ struggles but also reflect the underlying dysfunctional beliefs contributing to their distress. These insights are crucial in shaping more personalized and effective online interventions for support seekers [13,14].

A set of dysfunctional beliefs commonly present in OMHC posts is early maladaptive schemas (EMSs) [15,16]. Central to schema therapy (ST), EMSs are pervasive, self-defeating beliefs formed in childhood and often persist into adulthood. Young et al [15] identified 18 distinct types of EMSs (Table 1) that distort worldviews and are typically associated with mental health problems, such as anxiety [17-21]. Understanding these associations provides a more reliable method for anticipating potential schemas based on a mental health diagnosis, enabling proactive and targeted therapeutic interventions [15,22,23]. Similarly, recognizing the unique features of each schema (eg, “I’m incompetent when it comes to achievement,” corresponding to the failure to achieve schema) helps inform targeted interventions (eg, setting small, realistic goals to foster a sense of accomplishment). By addressing these schemas, these interventions can enhance therapy effectiveness and improve treatment outcomes.

However, the characteristics of these EMSs can vary across demographic groups. For example, Cámara and Calvete [18] identified a strong association between the failure schema and depression in Spanish university students. However, Wright et al [24] did not report such an association in their study of a US cohort. Similarly, Balsamo et al [25] did not observe a relationship between the failure schema and depression among young adults in Italy. Likewise, OMHCs likely present a demographically distinct population of online support seekers reflecting unique EMS characteristics [18,24-26]. Gaining a clearer understanding of these characteristics could provide valuable insights for policy making and the design of more effective digital therapeutic tools for this online support-seeking population. However, how EMSs are reflected in these communities remains largely unexplored, despite their growing importance in mental health support.

This research gap largely stems from the lack of tools for accurately identifying EMSs in the asynchronous, text-based environments of OMHCs. In professional practice, therapists use the Young Schema Questionnaire (YSQ) [27] to assess patients’ EMSs through clinically validated statements of dysfunctional beliefs, which patients rate on a scale from 1 (“completely untrue of me”) to 6 (“describes me perfectly”). Replicating this structured assessment in OMHCs is challenging due to the unstructured nature of online support-seeking interactions [15,28,29]. Overcoming this challenge requires interpreting EMSs directly from the text, which necessitates specialized natural language processing techniques to identify EMSs in OMHC posts. Recent advancements in large language models (LLMs) have demonstrated remarkable performance in complex language tasks, such as text classification, retrieval, and generation, which require nuanced interpretation of unstructured texts [30-32]. These models can capture subtle meanings in textual discourse, such as recognizing that the phrase “dark tunnel” in “I just feel like I’m stuck in a dark tunnel, and there’s no way out” metaphorically conveys despair, isolation, and hopelessness [33]. LLMs have also been used to predict mental health problems, such as depression, from online conversations [34-38] and have recently been applied to predict EMSs in OMHC posts [16].

Using the EMS classifiers from the recent study on EMS prediction [16] and leveraging recent LLMs, we can identify EMS labels within large volumes of unstructured OMHC texts. This enables us to (1) uncover associations between EMSs and mental health problems within OMHCs and (2) examine how features of EMSs are reflected in online support-seeking populations. These insights will guide the development of online interventions powered by artificial intelligence (AI) and grounded in ST, tailored to meet the unique therapeutic needs of online support seekers [15].

We curated posts shared by support seekers in 3 highly frequented OMHCs, namely 7-Cups [39], Beyond-Blue [40], and Patient [41], from 2013 to November 2023. Using a custom Python-based web crawler developed for this study, we collected posts from subcommunities focusing on anxiety, depression, posttraumatic stress disorder (PTSD), personality disorders (PDs), eating disorders (EDs), and substance use disorders (SUDs). These mental health problems are prevalent among adults [42,43] and are often treated with ST [15].

To identify EMSs in OMHC posts, we used the entailment-based prediction model (EPM) from a recent study [16]. EPM predicts EMSs from a post by leveraging textual entailment, a sentence-level inference technique from natural language processing [16]. In textual entailment, a statement entails a hypothesis h if a human reader would infer that h is likely true based on t [44]. EPM computes the entailment between sentences in a post and statements from the YSQ. When a sentence entails a YSQ statement, the latter serves as an “explanation” for assigning the corresponding EMS label. For instance, the sentence “Since then I either try to find excuses not to ride or when I do try to ride, I have to push myself and have massive panic attacks.” entails the YSQ statement “I often feel that I’m going to have an anxiety attack.” This YSQ statement corresponds to the vulnerability to harm or illness schema, and the EPM labels the post with this schema. We used the Text-To-Text Transfer Transformer (Google; T5-LLM [45]) in EPM to compute entailment and assign EMS labels to each post.

The 18 EMSs in Table 1 are each represented as a binary categorical variable, where a value 1 indicates the relevance of a schema to an OMHC post, and 0 indicates nonrelevance. In a similar way, we labeled each post with 6 binary categorical variables, each corresponding to one of the mental health problems examined in this study.

Using these categorical variables, we constructed 2×2 contingency tables (Table S1 in Multimedia Appendix 1) and calculated the expected frequency, E_ij (equation S1 in Multimedia Appendix 1), and the χ² statistic (equation S2 in Multimedia Appendix 1) for each pair of schema and mental health problem [46]. Each χ² value was compared to the critical value from the χ² distribution table with 1 df and a significance level of 0.05. For each χ² value that exceeded the critical value, we rejected the null hypothesis to indicate a statistically significant association between a schema and mental health problem. To quantify the strength of association for each statistically significant pair, we calculated the odds ratio (OR), OR_{ems_mhp} (equation S3 in Multimedia Appendix 1), and its 95% CI, OR_{ems_mhp_95%CI} (equation S4 in Multimedia Appendix 1) [47,48]. Following established guidelines for interpreting ORs [48], we considered an association inconclusive if the 95% CI of the OR included the null value of 1.

We further examined how online support seekers express their schemas by analyzing specific features reflected in their online posts. For this purpose, we used the case conceptualization technique, which is commonly used by therapists to identify key features related to patients’ beliefs and mental health problems, thus facilitating personalized therapeutic interventions. In ST, case conceptualization focuses on several dimensions, including schema triggers (situations that activate the schema), emotions (associated feelings), negative thoughts (irrational beliefs), coping responses (behaviors and reactions), and bodily sensations (physical and physiological experiences) [15].

We adapted this case conceptualization to our dataset using a data-driven AI approach, which we refer to as group-level case conceptualization. First, we prepared the data by using the EPM to identify sentences from OMHC posts that entailed specific YSQ statements, as outlined in the Curating the Dataset and Identifying EMSs section. These sentences reflected the dysfunctional beliefs and respective schemas established in the YSQ. We grouped these sentences according to their schema, resulting in a distribution of sentence counts by EMSs, which ranged from 8494 for social isolation to 98 for subjugation. To derive group-level features for each schema, we randomly sampled 10% of the sentences (in Multimedia Appendix 2, we demonstrate that the similarity scores between features extracted from different groups of randomly sampled sentences consistently average 0.79 (SD 0.09), which indicates substantial overlap in the features generated across groups. This validates the approach of conducting in-depth analysis on a single group of randomly sampled sentences for each schema and instructed GPT-4 (OpenAI; we used the model gpt-4-1106-preview for GPT-4) [32] through a custom prompt (Textbox 1) to extract common features among the sentences and categorize them along the dimensions of case conceptualization.

This study aligns with NUS-IRB Review Not Required guidelines by analyzing nonindividually identifiable data from OMHCs to explore the characteristics of EMSs. Specifically, we used textual content that was already anonymized through usernames and we further decoupled these pseudonyms and other identifying information from the data. In addition, the analysis focused exclusively on examining publicly available data, with no interactions with support seekers. This approach minimized the risks of introducing harmful or unsupportive content while also adhering to ethical guidelines and the responsible use of these online platforms [49,50].

Our dataset comprised 29,329 online posts in which support seekers shared their experiences with various mental health problems. Anxiety was the most frequently mentioned problem, appearing in 51.56% (15,149/20,329) of the posts, followed by depression, which appeared in 33.98% (9966/20,329) of the posts. PTSD, SUDs, PDs, and EDs correspondingly represented 8.75% (2565/20,329), 1.96% (576/20,329), 1.87% (548/20,329), and 1.79% (525/20,329) of the dataset. This distribution aligned with general sociodemographic trends in the prevalence of mental health problems. For example, anxiety is typically more prevalent than depression, while both are more common than PTSD, PDs, and EDs [42,51]. This suggests that the mental health problems discussed in OMHCs may reflect their real-world prevalence, suggesting that policies based on these trends could be effectively applied to online settings.

Regarding post length, the median sentence count was 10 (IQR 5-16). Word count showed similar variability, with a median of 214 words (IQR 124-356). This variability likely reflected the open nature of online support forums, where individuals can express themselves freely, unconstrained by the structured environments of in-person therapy. Unlike therapy sessions, where therapists guide conversations and encourage concise, focused responses, OMHCs allow support seekers to share their experiences through detailed, unfiltered narratives. Consequently, the dataset included both succinct and elaborate descriptions of mental distress.

As illustrated in Table 2, the distribution of EMSs, as predicted by EPM [16], was nonuniform across the dataset. Posts more frequently involved schemas related to social isolation and vulnerability to harm or illness than schemas such as approval seeking and subjugation. We hypothesize that this skewed distribution resulted from a combination of societal, environmental, biological, and psychological factors. For instance, modern societal pressures, pervasive social media use, and global crises, such as the COVID-19 pandemic, may amplify feelings of isolation and vulnerability [52,53]. In contrast, schemas such as approval seeking and subjugation may be less common due to cultural factors, including the emphasis on individualism and self-reliance in many societies [54].

Among all mental health problems, depression had the highest number of statistically significant associations with 12 out of the 18 EMSs (refer to Table S1 in Multimedia Appendix 3 for specific P values, ORs, and 95% CI of these associations). This was followed by PDs and PTSD, which had 9 and 7 associations, respectively. In contrast, anxiety, EDs, and SUDs showed fewer associations, with only 2 to 3 EMSs associated with these problems.

In Table S1 in Multimedia Appendix 3, we present the strength of the association between specific schemas and particular mental health problems using ORs. Consistent with previous works [55,56], we used OR<1.5 to indicate small effect sizes (weak associations), OR<2.5 to indicate moderate effect sizes (moderate associations), OR<4.0 to indicate large effect sizes (strong associations), and OR>4.0 to indicate very large effect sizes (very strong associations). Higher ORs suggest stronger associations, indicating that certain EMSs are relatively more likely to be present in specific mental health problems observed in OMHCs compared to others.

For anxiety, the strongest association was observed with the vulnerability to harm or illness schema, reflecting a very large effect size. In contrast, insufficient self-control and emotional inhibition schemas exhibited weak associations with anxiety.

In the case of depression, schemas with strong associations included social isolation, approval-seeking, and emotional deprivation. Moderate associations were observed for abandonment, defectiveness, dependence, failure to achieve, and negativity. Meanwhile, unrelenting standards, entitlement, punitiveness, and self-sacrifice schemas showed weak associations.

For EDs, moderate associations were observed with the emotional inhibition, defectiveness, and self-sacrifice schemas.

In PDs, the subjugation schema exhibited a very strong association, followed by moderate associations for the abandonment, entitlement, enmeshment, defectiveness, punitiveness, and mistrust schemas. Weak associations were observed for emotional deprivation and social isolation.

PTSD showed a very strong association with the mistrust schema and strong associations with punitiveness and subjugation. The emotional deprivation schema demonstrated a moderate association with PTSD, while weak associations were observed for enmeshment, self-sacrifice, and abandonment.

For SUDs, a moderate association was observed with the failure to achieve schema, while the dependence schema exhibited a weak association.

Features of each schema were identified using GPT-4, as detailed in Multimedia Appendix 4. On average, our group-level case conceptualization approach identified approximately 13 features per schema across the 5 dimensions of case conceptualization. Schema triggers, emotions, and negative thoughts each comprised approximately 3 features, followed by coping responses with about 2 features, and bodily sensations with 1 feature. Most features conveyed negative connotations. For instance, individuals with the abandonment schema often perceived themselves as burdens to others and frequently reported feelings of loneliness, anxiety, and emptiness.

These features also highlighted the unique characteristics of each schema. For instance, the punitiveness schema, which centers on self-blame, involved features, such as “situations where there is a fear of repeating past mistakes or events,” “feelings of guilt or shame over past actions or events,” “thinking that one cannot forgive oneself or others for past wrongdoings,” and “responding with self-harm or destructive behavior.” These features are elaborated on in the Discussion section.

This study focused exclusively on 3 widely accessed OMHCs that use English as the primary mode of communication. To broaden the cultural and linguistic diversity of the sample, future research could explore OMHCs that use other languages and compare the findings of this study with subsequent research to verify whether our results extend to OMHCs in different sociolinguistic and cultural contexts, where schema expression may vary. In addition, while the anonymity of these platforms encourages openness, it limits access to sociodemographic data (eg, age, gender, and socioeconomic status), which might be important in shaping mental health perspectives and schema characteristics. Future studies could explore methods that allow for the ethical anonymous collection of demographic data or use validated techniques to infer demographic characteristics while maintaining user privacy, which is of utmost importance. These suggestions are essential for assessing the generalizability of the EMS characteristics uncovered in this study. Expanding the demographic and cultural scope in future research would help validate and deepen these findings, contributing to a more comprehensive understanding of EMSs in online support-seeking environments.

In this study, we examined the characteristics of EMSs within OMHCs, situating our findings within the existing literature. By analyzing a large dataset of posts from widely accessed OMHCs, annotated with EMS labels using a state-of-the-art EMS classifier, we identified key EMS features and various associations between EMSs and mental health problems. Notably, we found that depression, PDs, and PTSD were associated with a wide range of EMSs, highlighting the complexity of these multifaceted problems. These findings suggest that interventions for these problems could benefit from addressing multiple EMSs simultaneously. In contrast, mental health problems with fewer EMS associations, such as anxiety, may respond more effectively to more targeted interventions focused on specific schemas. In addition, we identified EMSs as more prominently associated with certain mental health problems in OMHCs. For example, anxiety-related posts typically center around vulnerability to harm, while depression-related posts reflect prolonged unmet interpersonal needs. Negative self-perception and self-neglect are prevalent in posts about EDs, while posts about PDs involve themes of subjugation. Posts about PTSD often describe traumatic experiences, and those concerning SUD are marked by themes of personal failure. A key contribution of this study is the novel group-level case conceptualization technique, which uses recent advanced LLMs and prompt engineering techniques to extract EMS features from OMHC posts. This approach provides valuable insights into how EMSs are expressed in online support-seeking environments and has the potential to inform personalized, ST-focused interventions. Overall, this study provides a foundational understanding of EMS characteristics in online support-seeking contexts and their role in shaping mental health experiences. Our findings highlight the potential for tailored, ST-based interventions in online peer support settings, which could enhance therapeutic outcomes and promote the mental well-being of individuals seeking support in these online communities.