As health systems have moved away from institutional care, community mental health services have developed internationally, and outreach in people’s living environments has become more prevalent [4]. In Japan, psychiatric home-visit nursing is a representative outreach service within community mental health. Provided under the national health insurance scheme, nurses, occupational therapists, and psychiatric social workers deliver support in service users’ homes. The scope of care is broad, ranging from psychological support to assistance with activities of daily living, support for employment and social participation, daytime leisure activities, and family support. Qualitative evidence from staff and users indicates that psychiatric home-visit nursing, marked by acceptance, equal partnership, expansion of clients’ worlds, future-oriented care, and support for hobbies and empowerment, aligns with personal recovery [5,6].

To enhance the quality of recovery-oriented support in mental health, international research has sought to clarify the concept of recovery and develop its assessment. Service users’ narratives have played a central role, with knowledge accumulating through qualitative analyses by experts. Indeed, the very definition of recovery has emerged from narrative research that synthesizes the lived experiences of people with mental disorders, positioning narrative as indispensable for understanding and practicing recovery [2,7]. The CHIME framework arises from an integration of qualitative studies grounded in the accounts of service users [3]. These frameworks rest on a substantial body of interview-based and first-person research. For example, Davidson and colleagues, drawing on in-depth interviews and qualitative analyses with people living with schizophrenia and other serious mental illnesses, have shown that recovery in-volves reconstructing a valued sense of self, renegotiating relationships with mental health services, and reclaiming everyday roles and citizenship in the community [8,9]. Similarly, classic first-person and narrative accounts of recovery, such as those by Anthony and Deegan, highlight themes of hope, meaning, and active self-determination in building a life beyond illness [2,10]. In this sense, qualitative research has provided a foundational empirical and conceptual basis for existing recovery frameworks. At the same time, qualitative analysis has limitations. While it excels at grasping rich context and generating new concepts, generalization is difficult and the volume of analyzable data is constrained [11]. As a complementary quantitative approach, Recovery Assessment Scale (RAS) [12] has been widely used. However, the RAS depends on self-evaluation at a single point in time and may insufficiently reflect aspects of social participation and narrative [13].

A key challenge in mental health, including recovery research, is the difficulty of quantitatively analyzing subjective narratives. In response to this challenge, NLP has attracted attention. NLP refers to computational methods that enable machines to understand and analyze natural languages. Across studies of schizophrenia and mood disorders, reduced speech output, impoverished vocabulary, and semantic fragmentation are associated with poorer social functioning and outcomes, and measures such as discourse coherence and the frequency of emotion words correlate significantly with clinician-rated symptom scales [14,15,16]. Although diagnosis and assessment have traditionally relied on questionnaires and clinician observation, growing evidence that patients’ speech and conversations can yield clinical insights suggests that NLP can provide new objective indicators that complement these subjective approaches.

This study focuses on BERTopic [17], an AI-based method built upon Bidirectional Encoder Representations from Transformers (BERT) [18]. Unlike simple word counts, BERT captures context-dependent meaning. Applying BERT to topic extraction yields BERTopic, which can infer the themes of texts while taking context into account. Using BERTopic, a study analyzing mental-health helpline conversations reported more interpretable topic structures and superior performance compared with non-contextual baselines [19]. Another study found that narratives by individuals with severe mental illness included more food-related topics with positive sentiment than those of healthy controls [20]. Together, these findings suggest that BERTopic is a promising approach in the mental health domain.

With respect to quantifying recovery, studies have analyzed social media posts using topic modeling. Sik et al. [21] identified intimacy, social criticism, and despair as novel recovery-related topics. However, social media posts are fragmentary and insufficient for capturing the recovery process as a whole. Applying NLP to life history narratives told by service users may objectively quantify subjective experiences and reveal aspects of recovery not fully captured by existing scales or qualitative frameworks. In particular, analyzing the structure of narratives spanning past to future is expected to offer a new perspective that complements prior research.

The objectives of this exploratory study were to quantify the subjective life history narratives of psychiatric home-visit nursing users using NLP and to clarify how linguistic features relate to recovery-related indicators.

This study is significant in that it treats service users’ life history narratives themselves as quantitative data. This approach goes beyond methods that rely only on questionnaires or conventional qualitative analysis and makes it possible to view recovery from several complementary perspectives. By examining how linguistic features are related to measures of outcome domains relevant to recovery, including mental health status, global functioning, social functioning, and subjective recovery, we can explore how different aspects of recovery are reflected in the ways people talk about their lives. Furthermore, because life history narratives naturally include memories of the past, awareness of the present, and expectations for the future, they can provide, even in a cross-sectional study, an overall picture of the recovery process. By using NLP, this study seeks to capture service users’ experiences in a way that does not depend on clinicians’ subjective judgment, and aims to contribute to new developments in recovery research and in recovery-oriented clinical practice and evaluation, particularly in community-based services such as psychiatric home-visit nursing.

Data were obtained from Kusunoki Mental Hospital, Specific Medical Corporation Kusunokikai, a psychiatric hospital in Japan. This hospital is a psychiatric specialty institution that provides inpatient and outpatient services, including outpatient consultation, psychiatric home-visit nursing, ambulatory psychotherapy, and psychiatric rehabilitation. Users of the hospital’s psychiatric home-visit nursing service, which offers care comparable to typical services in Japan and is delivered at home by nurses and occupational therapist, were targeted in this study. All users had a psychiatric diagnosis and were outpatients of the hospital. The inclusion criteria were: age ≥ 18 years; a diagnosis classified under Mental, Behavioural or Neurodevelopmental disorders in the International Classification of Diseases 11th Revision (ICD-11) [22]; current use of psychiatric home-visit nursing services; and informed consent provided by the participant and their primary physician. The exclusion criteria were: difficulty understanding an interview conducted in Japanese; and prominent symptoms such as delusions that would preclude appropriate understanding of the study and provision of consent. Recruitment took place from August 2024 to March 2025. This study was approved in 2024 by the Research Ethics Committee of the Kusunoki Mental Hospital, Specific Medical Corporation Kusunokikai, Japan (approval No. 2024-0001). All participants received oral and written explanations regarding the study purpose and methods and provided informed consent.

Data analysis comprised text preprocessing, use of a sentiment dictionary, topic modeling, and statistical analyses examining relationships between recovery-related measures and interview data.

The mean age of the participants was 58.6 years (standard deviation [SD] = 14.5); 10 (37.0%) were men and 17 (63.0%) were women. Diagnoses were schizophrenia (6A2) in 10 (37.0%), bipolar disorder (6A6) in 6 (22.2%), depression (6A7) in 6 (22.2%), and other (e.g., neurodevelopmental or substance use disorders) in 5 (18.5%). Mean questionnaire scores were 79.1 for RAS (SD = 18.0), 31.1 for GHQ-30 (SD = 16.7), and 110.4 for SFS (SD = 23.5). More details can be seen in Table 2.

The mean interview duration was 58.0 min (SD = 12.6). Participants produced an average of 228.6 utterances (SD = 61.4) and 5353.0 words (SD = 2254.7). BERTopic identified 12 topics. The total number of utterances assigned to topics was 3824, accounting for 62.0% of all utterances; hence, 38% were unassigned. Although the proportion of unassigned items depends on parameter settings, prior research reports approximately 20%–50% or higher [19,38,39,40]. For conversations or short texts, 10–20 topics are commonly used to ensure interpretability [19,41,42]. The indices in the present study fall within these ranges.

The 12 topics were: Backchannels (aizuchi) (non–self-initiated responses to the interviewer; P/N ratio = 0.05), Thinking (utterances with unclear purpose; 9.87), Physical condition (primarily health complaints; 0.24), Eating (interest in food and routines; 5.13), School (recollections of learning and experiences; 9.34), Sleep (routines and bodily sensations; 0.81), Family (family history and parent–child conflict; 0.99), Conversation (awareness of speaking and relations with others; 2.44), Work (experiences of employment; 3.07), Hobbies (leisure including creative activity and learning; 4.11), Housework (routines to maintain daily life and perceived burden; 2.21), and Hospital (engagement with healthcare and its impact; 0.21). Further details can be seen in Supplementary Materials S1, characteristic words and exemplar sentences in Table 3, and word counts and sentiment tendencies in Table 4.

Topic 5 (School) showed significant positive correlations with RAS (r = 0.39, p = 0.05) and SFS (r = 0.57, p < 0.01). Topic 7 (Family) was negatively correlated with RAS (r = –0.46, p = 0.02). Regarding GAF, total word count (r = 0.44, p = 0.02) and topic 12 (Hospital) (r = 0.42, p = 0.03) were positively correlated, whereas Negative Emotion in J-LIWC (r = –0.49, p = 0.01) and topic 1 (Backchannels) (r = –0.41, p = 0.03) were negatively correlated. In addition, Positive Emotion in J-LIWC showed a significant negative correlation with GHQ-30 (r = –0.44, p = 0.02). The results of the correlation analyses among demographics, J-LIWC category proportions, BERTopic topic proportions, and recovery-related measures are summarized in Table 5.

In interpreting these findings, it is important to note that personal recovery, mental health status, global functioning, and social functioning were treated as related but distinct constructs in this study. Mental health status was assessed by the GHQ-30, global functioning by the GAF, social functioning by the SFS, and personal recovery by the RAS. Accordingly, the associations involving emotion words, backchannels, and narrative length primarily reflect current mental health status and overall functioning, whereas the topic-level findings regarding School and Family are more directly related to subjective recovery and social functioning.

These findings indicate that the narratives of users of psychiatric home-visit nursing contain signals related to mental health status and functioning, consistent with previous NLP research showing that linguistic features of patients’ speech and text are associated with psychopathology and functional outcomes [14,15,16,20]. In this study, the use of emotion words, the amount of speech, and talk about medical care were systematically linked to psychological distress and global functioning, suggesting that similar linguistic indicators can provide clinically meaningful information even when applied to life history interviews conducted in everyday settings.

At the same time, the focus of this study was not symptom or risk detection per se, but the understanding of personal recovery [1,2,3]. The associations between topics such as School and Family and subjective recovery suggest that how individuals narrate their life histories—what experiences they view as resources and which relationships they experience as difficult—is closely related to their sense of recovery, in line with narrative accounts of recovery as an ongoing process of meaning-making and identity reconstruction [2,3,44]. This implies that NLP can be positioned not only as a tool for early problem detection, but also as a means of visualizing recovery processes and supporting recovery-oriented practice in community-based mental health care. Future work applying similar approaches to more diverse service contexts and longitudinal data may further clarify the role of language in personal recovery.

This study has several limitations.

First, the sample size was small (n = 27) and restricted to users of psychiatric home-visit nursing from a single hospital in Japan, which constrains the generalizability of the findings. Participants were mainly middle-aged and older long-term service users with severe mental illness, so the results may not directly extend to younger people, those with different diagnostic profiles, or users of other services and inpatient care. Even so, the study demonstrates the feasibility of applying NLP to life history narratives of core community mental health service users and examining their associations with recovery-related indicators. The present findings should therefore be interpreted as exploratory and context-specific, and larger multi-site studies with more diverse cohorts are needed to test reproducibility.

Second, the cross-sectional design precludes causal conclusions about whether narrative features predict changes in recovery and functioning or merely reflect current status. Although life history narratives naturally contained past, present, and future-oriented content and thus allowed us to capture part of the broader recovery process, narratives were obtained at only one time point. Longitudinal studies that track changes in both narratives and recovery-related indicators over time are needed to examine how narrative features and recovery trajectories co-vary, and which kinds of linguistic change precede or accompany advances in recovery.

Third, the NLP approach itself has limitations. While BERTopic and J-LIWC yielded interpretable topic structures and sentiment categories, 38% of utterances remained unassigned. These segments may contain low-frequency, highly individualized, context-bound narratives in which personal aspects of recovery are embedded, and excluding them may have attenuated or obscured some associations. Future work should qualitatively inspect unassigned utterances and consider modeling their recovery-related meanings using alternative topic models, adjusted parameters, or hybrid qualitative–quantitative frameworks.

Fourth, topic labels and their interpretation were based on the clinical judgments of the research team. Although this approach is more objective and reproducible than traditional qualitative analyses, some subjectivity remains, and no independent validation by external experts or service users was conducted. Nonetheless, the extracted topics corresponded well to clinically familiar themes (e.g., school, family, work, hospital). Future studies should enhance objectivity and transparency by involving multiple independent raters, including people with lived experience.

Fifth, data collection procedures may have influenced both narratives and questionnaire responses. Interviews were conducted immediately after routine home-visit nursing sessions, so day-to-day condition, preceding interaction, and the home recording environment may have affected speech quantity and content, although established therapeutic relationships likely facilitated detailed and personal narratives. In addition, participants completed three relatively lengthy self-administered questionnaires (RAS, SFS, GHQ-30) in addition to the 45–60-min interview, which may have caused respondent burden and fatigue, potentially reducing concentration and increasing “good-enough” responding. Questionnaires were distributed after the interview, so the reflective process during the interview may also have influenced responses, for example by temporarily heightening subjective distress or, conversely, strengthening perceived recovery. Future research should reduce assessment burden (e.g., through shorter scales or distribution across visits) and systematically examine the impact of the order and timing of interviews and questionnaires.

Sixth, the corpus consisted of Japanese life history interviews, and caution is needed when generalizing to other languages and cultures. Frequent subject omission in Japanese complicates comparison with English-language indicators such as first-person singular pronoun frequency, and differences in J-LIWC coverage and conversational practices (e.g., frequent backchannels) may affect how narrative quantity and responsiveness appear.

Even so, this study represents an early attempt to apply BERTopic to life history narratives in psychiatric home-visit nursing and to examine their associations with recovery-related indicators, providing a basis for applications in other languages and cultures and for comparisons with other NLP methods. Future work should apply similar approaches in different linguistic and cultural settings and formally evaluate cross-linguistic and cross-cultural validity to strengthen the generalizability and clinical utility of this approach.