This study builds on a previous exploratory observational study by analyzing the existing narrative data to explore whether LLM-based analyses can transform these narratives of daily life into interpretable insights of QoL in people living with dementia.

The dataset used in this study was originally collected as part of the “Green Care Farms—an innovative care environment for adults living with dementia” project [36]. Similar to the current study, the Green Care Farms project was conducted within the Living Lab in Ageing and Long-Term Care in South Limburg, the Netherlands [37]. The Green Care Farms project aimed to monitor the daily lives of people living with dementia residing in Dutch care farms and traditional nursing homes, examining how the care farm environment influences their well-being.

Data were collected using the MEDLO tool, a tablet-based observation instrument specifically developed to capture detailed, real-time descriptions of the daily lives of people living with dementia in long-term care settings [14,15].

Data were collected over a period of 2 months per Green Care Farm, and the period of the day of the observation was selected randomly. Three researchers and 2 research assistants carried out observations at the 4 green care farms. All observers were trained by an experienced user of the MEDLO tool. This training consisted of studying the manual, practicing the training, and discussing the data. Within this training, both researchers observed the same people living with dementia and discussed their findings. The researchers conducted observations independently when they reached a complete agreement.

People living with dementia were observed for one morning (7 AM-11:30 AM), one afternoon (11:30 AM-4 PM), and one evening (4 PM-8:30 PM), with a 30-minute break each day. Every 20 minutes, a maximum of 11 people living with dementia were observed for 1 minute in a random sequence. This led to 12 observations per individual living with dementia per observation day and 36 momentary assessments per individual living with dementia in total. The people living with dementia were observed for 1 minute. By randomly selecting time points within the days of a person with dementia, insight into their daily lives was gained [15]. All observed people living with dementia had a formal diagnosis of dementia, confirmed by medical staff.

During each 1-minute observation, observers recorded both quantitative scores and narrative descriptions on four domains of daily living: (1) activity performance, (2) physical environment, (3) social interaction, and (4) emotional well-being. The narrative descriptions were written as free-text entries in natural language, without predefined categories or templates, allowing observers to capture contextual details and subtle nuances often overlooked in standardized formats [13,14]. For example (translated from Dutch): “She is sitting at the dining table next to the kitchen unit, staring at a newspaper in front of her.” These narrative descriptions served as the primary dataset for the current study’s NLP analysis.

Data were derived from the Green Care Farms study. This study was reviewed and approved by the Medical Ethics Committee of Zuyderland, the Netherlands (METCZ20210097). An amendment to the study protocol received a positive decision from the same ethics committee on February 20, 2024. The present study used anonymized observational data collected within this approved study protocol for secondary analysis. All procedures were conducted in accordance with the ethical standards of the institutional and/or national research committee and with the Declaration of Helsinki and its later amendments [38]. The data used in the present study had been anonymized by the researchers from the original study before being transferred to the current research team.

Preprocessing is a critical step in text mining, as it prepares raw text for analysis by removing noisy or unreliable data that may interfere with accurate interpretation [19,39]. In this context, “noise” refers to inconsistencies in the data that do not carry meaningful information for analysis [40]. For the current study, preprocessing and analysis were performed using Python version 3.11.8, a free software package for general-purpose programming language [41]. The code for all the analyses can be found on GitHub [42].

While traditional NLP pipelines often involve extensive text cleaning, such as tokenization, normalization, stop-word removal, stemming, and handling punctuation and numbers, minimal preprocessing was conducted in this study [43]. The rationale for this approach was to preserve the natural language, contextual detail, and subtle nuances in the narrative data [44]. These nuances, including emotional tone and subtle behavioral cues, are essential for understanding the QoL of people living with dementia.

Although this study applied minimal preprocessing, several preprocessing steps were carried out in collaboration with researchers from the original Green Care Farm project, leveraging their familiarity with the data to ensure high data quality. (1) All narrative data from the MEDLO tool were extracted from the Excel file into a plain text file and loaded into Python; (2) all words were lowercased to enable consistent matching of identical terms regardless of case and to facilitate abbreviation expansion; (3) commonly used abbreviations, introduced during data collection to expedite note-taking, were replaced with their full-length equivalents (eg, lr: living room, bk: big kitchen); (4) double spaces and stray blank spaces were removed, and numbers were written out in words to improve textual consistency; and (5) the final preprocessed text file was converted into a PDF format, which was then used as input for LLM-based analyses, allowing the model to process the narrative data in its original form. The conversion to PDF was a pragmatic workflow decision rather than an analytical preprocessing step. The secure analysis pipeline had originally been developed to process PDF input files, and therefore, the final preprocessed text file was converted to PDF to ensure compatibility with this existing pipeline.

This study analyzed the narrative data using an LLM approach, using GPT-4.0-mini deployed through the Azure OpenAI Service (Microsoft) [45]. The analysis was conducted in Python 3.11.8 using the OpenAI Python package (openai==1.59.6), which was used to send requests to the Azure OpenAI chat-completion end point [46,47]. Supporting packages included PyPDF2 for extracting text from the PDF input file and pandas for structuring the extracted text during analysis [48,49]. The extracted narrative text was combined into a single Dutch text corpus and submitted to the deployed model through the local Azure deployment named 4OMINI using API version 2023-09-01-preview. The request specified a maximum generated output length of 4096 tokens. Temperature, seed, and structured response format were not explicitly set in the request. Therefore, outputs were generated without explicitly controlling randomness or enforcing a structured response format [50-52]. The analyses were performed via the model between November 2024 and May 2025. The model was not trained or fine-tuned on these data but was used only for inference. All preprocessing, prompt preparation, and output storage were conducted within the local secure research environment of Maastricht University. This secure and scalable blueprint of the model enables organizations to use LLMs, such as GPT-4.0-mini, while safeguarding data privacy.

Expert evaluation was used to assess whether selected model outputs were meaningful and contextually plausible in relation to long-term care practice. For sentiment analysis, 2 researchers (LF and KR, both with 5 years of research experience in long-term care) from the original Green Care Farm project independently reviewed the top positive and negative N-grams per QoL domain [62]. They assessed whether the assigned sentiment direction and intensity were plausible, given the wording of the N-gram and its interpretation within the context of observational dementia care. Disagreements between reviewers were recorded and discussed, and the final interpretation was based on consensus.

For topic modeling, the model-generated clusters were reviewed by 4 members of the research team (DS, HC, SA, and MCS), with expertise in dementia care and QoL research, long-term care research, and health care AI methods. The topic clusters were assessed using 2 qualitative criteria: coherence and distinctiveness. Coherence referred to the internal consistency of each topic cluster, that is, whether the terms grouped under a given label were meaningfully related [58]. For example, a cluster containing the terms music, dancing, laughter, and clapping was assessed to be coherent and appropriately labeled as leisure activities. In contrast, if an unrelated term, such as agitation, appeared within that same cluster, it indicated a lack of coherence, which resulted in further fine-tuning of the prompt. Distinctiveness was assessed by examining the boundaries between topic clusters to avoid redundancy or overlap. If an overlap between topic clusters was identified by the research team, the model was iteratively prompted to generate more distinct clusters. This process of prompt engineering continued until the research team reached consensus that the topic clusters were sufficiently distinct, with minimal thematic overlap.

In total, 5652 observations were conducted on 151 people living with dementia during the Green Care Farms study [62]. Narrative descriptions were missing in 30 (0.53%) observations. The narrative descriptions totaled 50,106 words. An average of 10.5 (SD 5.8) words was noted per narrative. Table 1 presents the background characteristics of the 151 individuals living with dementia who were observed.

The N-gram frequency analysis revealed language patterns related to QoL. Frequently observed terms included laughing, smiling, satisfied, and nice, suggesting that people living with dementia were often described as being in a state of positive emotional well-being. Negative affective expressions, such as becoming angry, crying, and feeling distressed, appeared far less frequently than positive affective expressions.

Ten of the 15 N-grams within the QoL domain behavior competence reflected low-activity behavior. Observers frequently noted people living with dementia sitting at the table, reading the newspaper, or looking outside. Although terms that encompass “activity” such as helps with cooking or participates in gymnastics were present, they were less common. This suggests a pattern of predominantly passive engagement in daily life, despite generally positive emotional expressions.

The daily lives of people living with dementia were most often observed in 2 core settings: the private room and the living room. Additionally, a few observations of people living with dementia were made in outdoor spaces, such as a terrace, the garden, or in the sun.

Similar to other QoL domains, N-grams within the perceived QoL domain were mainly positive affective expressions, such as satisfied, nice, beautiful, and happy. The partial overlap in N-grams from the domain of psychological well-being was observed (ie, satisfied and happy). The 15 most frequent N-grams identified within each QoL domain are presented in Figure 1.

Figure 1 shows the 15 most frequent N-grams identified in the observational field notes for each of Lawton QoL domains. Frequencies indicate the number of times each N-gram occurred in the contextualized text segments included in the N-gram analysis. The panels show that psychological well-being was mainly represented by affective expressions such as “lachen,” “glimlacht,” and “blij;” behavioral competence by passive activity phrases such as “zit aan tafel” and “kijkt naar buiten;” objective environment by indoor settings such as “eigen kamer” and “woonkamer;” and perceived QoL by evaluative or experiential expressions such as “tevreden,” “leuk,” and “mooi.”

Although the model was instructed to extend ambiguous N-grams by one word to improve clarity, some N-grams still lacked semantic clarity and contextual relevance. For instance, certain N-grams live or direct reaction offered limited insight.

Figure 2 presents the 5 most positively and negatively scored N-grams within each QoL domain, showing a wide range from strongly negative to strongly positive sentiment. The analysis of the N-grams identifies distinct sentiment patterns across QoL domains. High positive sentiment-scoring N-grams (around +0.8 to+0.9) show engagement and enjoyment of people living with dementia, such as experiencing fun during games, being cheerful with visitors, and going walking with a group. On the contrary, strongly negative sentiment-scoring N-grams (around −0.8) appear in expressions such as crying, appearing confused, and feeling anxious, which predominantly capture emotional distress in people living with dementia.

Within each QoL domain, the sentiment polarity of the N-grams is clear. Perceived QoL ranges from female client seems satisfied to feels bad. Psychological well-being ranges from happy with visitors to crying. Behavioral competence ranges from goes walking with a group to complaining.

While these examples reflect a mix of positive and negative-scoring N-grams, the overall sentiment distribution in the narrative data was skewed. Across all QoL domains, negative sentiment outweighed positive sentiment. Positive sentiment made up between 17.9% and 24.9% of the total weighted sentiment, depending on the QoL domain. This suggests that the average sentiment per QoL domain was driven downward by the higher intensity of negative sentiment-scoring N-grams. Even when positive sentiment-scoring N-grams appeared more frequently, their lower sentiment scores were insufficient to offset the impact of fewer but more intensely negative sentiment-scoring N-grams. Table 2 presents the weighted positive and negative sentiment scores for each QoL domain.

Figure 2 shows sentiment scores for the top 5 most positive and negative N-grams categorized by Lawton QoL domains. Bars to the right of 0 represent positively scored N-grams, whereas bars to the left of 0 represent negatively scored N-grams. Colors indicate the QoL domain to which each N-gram was assigned: behavioral competence, psychological well-being, objective environment, or perceived QoL. These sentiment scores reflect the polarity of documented textual expressions and should not be interpreted as direct measures of participants’ QoL.

Topic modeling identified 8 main topics in the narrative data. Each topic was defined by a set of representative keywords and was linked to one or more QoL domains, as shown in Table 3.

The model inferred thematic correlations based on the co-occurrence of keywords across different QoL domains. This showed that narrative data involving social interaction and daily activities (ie, eating and walking) frequently appeared alongside positive emotional expressions (ie, smiling and feeling relaxed). On the contrary, negative emotional states (ie, restlessness or sadness) were often reported in contexts of social isolation or reduced stimulation, such as when people living with dementia remained in bed or stayed alone in their rooms.

Perceived QoL was linked to the highest number of topics (6/8). The other QoL domains appeared in 4 topics. Many topics were assigned to multiple QoL domains, indicating that a single thematic cluster often reflects multiple QoL domains. For example, the topic eating and drinking was linked to behavioral competence, objective environment, and perceived QoL, reflecting the multidimensional nature of QoL.

The study aimed to explore whether an LLM could be used to structure existing MEDLO narrative data into interpretable QoL-relevant patterns in people living with dementia residing in nursing homes in the Netherlands. Specifically, it examined whether N-gram analysis, sentiment analysis, and LLM-based topic modeling could identify recurring language patterns, emotional tone, and semantic clusters that could be mapped to the Lawton QoL domains.

A first implication from this study concerns the need to incorporate context into observational data when reporting and analyzing affect in people living with dementia. The findings from this study indicated that while people living with dementia were predominantly observed during passive activities and in a limited range of indoor settings, they were often observed expressing positive emotions during the day. In addition, the sentiment analysis showed that although negative expressions were documented less frequently, they received more extreme negative sentiment scores. This finding is consistent with the literature, suggesting that people living with dementia are frequently documented as comfortable and content in their daily lives; however, their well-being remains sensitive to moments when physical, emotional, or social needs are not fully met [3,63-65]. The reported high-intensity negative expressions may reflect moments when these needs are not adequately addressed [66,67]. However, measuring emotions in people living with dementia is challenging because their emotional expression is often characterized by neutral affect [63]. Consequently, in this study, observations reflecting neutral affect may have contributed little to the weighted sentiment scores, while less common negative expressions could have stood out against this neutral baseline and received relatively high-intensity scores, thereby influencing the weighted sentiment results.

Hence, there is a need to develop documentation guidelines to encourage observers to standardize their data collection methods and practices. While traditional NLP methods, such as term frequency-inverse document frequency, rely on word occurrences, LLMs interpret meaning from the surrounding context [23]. Consequently, the descriptive detail of the narrative data directly affects the outputs. This dependency on context-rich narrative data means that variation in how observers phrase their descriptions can lead to systematic differences in model outputs, even when experiences are similar [67-69]. For example, what to record, how detailed to be, and which emotions or behaviors to emphasize are each influenced by time constraints and professional judgment. Therefore, developing documentation guidelines could encourage observers to standardize their documentation. An example of such documentation guidelines is to include narrative information about the social context, setting, and activity, alongside emotional expressions, in care descriptions, which could provide richer input for LLM-based models while remaining feasible within care routines [70-72]. This also aligns with broader debates in health care–based LLM use, where the quality and style of input text have been shown to shape model performance in long-term care, potentially improving care for people living with dementia [70-72].

A second implication of this study concerns the ways in which LLM-generated conceptualizations of QoL can provide insight into how different aspects of QoL can co-occur and interact in dementia care and in what circumstances they do so [62,66,73-75]. The Kitwood personhood framework, along with later work on person-centered care, emphasizes that QoL derives from opportunities for purposeful activity, social reciprocity, and recognition of individuality rather than from momentary affective expressions alone [67,76,77]. From this perspective, positive documented expressions during passive activities in low-variation indoor settings may be interpreted as reflecting momentary positive affect, while leaving unanswered questions about the extent of meaningful engagement [78,79]. In our analysis, the LLM-generated outputs linked affective terms with descriptions of activity, agency, and social interaction. These textual patterns may provide contextual information that could support further exploration of the distinction between momentary positive affect and meaningful engagement. Furthermore, the topic modeling in this study provides support for the idea that all of Lawton QoL domains are interwoven [3,6,80]. For example, the description of enjoying helping with cooking may simultaneously reflect behavioral competence, psychological well-being, and perceived QoL. In the future, LLM-generated clusters may help describe how emotional, physical, social, and environmental aspects co-occur within observational narratives, while existing questionnaires currently isolate experiences into predefined categories [81-83].

In sum, this study aligns with broader developments in health care, where increasing attention is being given to the analysis of narrative data [24,84-87]. The current findings demonstrate that even brief narrative care data can be structured into recurring textual patterns. This creates opportunities to integrate near–real-time feedback loops into tools such as the MEDLO. Embedding LLM-based analysis within such tools may enable the immediate analysis of narrative descriptions, transforming routine documentation into dynamic QoL feedback of people living with dementia. For example, the automatic detection of subtle changes in emotional tone or activity diversity could help care teams recognize early signs of reduced well-being. Such insights may enable staff to identify emerging needs earlier and adjust activities, environments, or interactions in a more proactive and person-centered manner [67,88,89]. Importantly, these integrated feedback loops should complement, not replace, the expertise of care professionals by offering timely, data-informed insights to support daily practice [90,91]. Evidence from long-term care indicates that these technologies must be developed in close collaboration with domain experts, including care professionals, researchers, and technology developers, to ensure integration into daily workflows, meet usability requirements, and uphold ethical standards [92-95].

A key strength of this study was the unique nature of this real-world dementia-specific dataset. The manual analysis of this dataset would be very time-consuming. However, analyzing these data with an LLM enabled the systematic exploration of thousands of brief descriptions. Another strength lies in the decision to use minimal preprocessing, preserving the authenticity and contextual richness of the observational language. This approach enabled the LLM to capture the nuances of natural expression, thereby maintaining practical relevance in real-world care contexts. Transparency and reproducibility were also prioritized: all coding materials and prompts were made openly available, supporting replication and further adaptation of the analytical workflow in other settings. Finally, the black-box nature of LLMs limits insight into how specific outputs were generated. To mitigate this, 2 experts manually reviewed the most positive and most negative-scoring N-grams to assess whether the assigned sentiment scores aligned with their contextual meaning.

Several limitations should also be acknowledged. First, individual narrative descriptions were often brief and fragmented with an average length of 10.5 (SD 5.80) words. This limited the semantic richness available for LLM-based analysis and may have affected the quality of all outputs. However, analyzing the entire corpus at scale partly offset this limitation, allowing recurring patterns to emerge, despite the brevity of individual notes. Second, the data were collected and analyzed in Dutch nursing home settings, which may limit the generalizability and transferability of the findings to other long-term care contexts and languages. However, analyzing the narratives in their original language avoided potential loss of meaning that could have occurred through translation before analysis. Third, although the final prompts were developed through an iterative refinement process, each final prompt was executed only once for the final analysis. LLM-based outputs may vary across repeated executions and model versions, even when prompts and input data remain unchanged [50,96]. We documented the model, deployment, API version, software package version, analysis period, prompts, and code; however, exact computational reproducibility cannot be guaranteed. Fourth, although the expert panel evaluated the sentiment scores of the top-scoring N-grams, the analysis relied on a general-language lexicon rather than a dementia-specific lexicon. Consequently, some context-dependent terms may have been mis-scored and, if they did not appear among the reviewed N-grams, remained underutilized. Finally, the expert review assessed whether selected outputs were coherent, plausible, and meaningful in relation to long-term care practice but did not perform formal validation or allow for the calculation of agreement metrics. Because QoL in dementia is subjective and lacks a single objective gold standard for validation, the findings should be interpreted as exploratory QoL-relevant patterns rather than validated classifications or measurements of QoL [66].

Future research should build on these findings in several directions. First, larger and more comprehensive datasets, including greater and richer volumes of observational narrative descriptions, are needed to strengthen the robustness and generalizability of the analysis. Second, a dementia-specific lexicon should be developed and validated to ensure that terms are weighted appropriately in this context. Third, prospective validation studies are required to examine whether LLM-based themes and sentiment patterns correspond to established QoL instrument scores. Fourth, research should explore the real-time integration of LLM-based analysis into observation tools such as MEDLO, evaluating usability, interpretability, and workflow alignment through co-design with care professionals.

These findings should be interpreted as exploratory. Although the current analysis suggests that LLM-based analyses may help structure observational narratives into QoL-relevant patterns, the approach is not yet validated against QoL measures. Before such outputs can inform person-centered care, further research is needed to examine whether LLM-generated outputs correspond to established QoL measures if validated and integrated into care workflows.

LLM-based analyses predominantly identified passive activities in little variation in indoor settings; however, people living with dementia were often described as having a positive affect. These findings suggest that LLM-based analyses can structure observational narratives into interpretable patterns of activity, emotional expression, and social interaction. Further validation is required before such outputs can support multidisciplinary reflection and person-centered dementia care.