Stress Management Using Large Language Models: An Integrated Web-
based Intervention Framework
Pradeep Gupta1, Shubham Pandey2 and Moksh Kushwaha3
1
Asst. Prof., Ajay Kumar Garg Engineering College, Ghaziabad-201009, India
2
Student, Ajay Kumar Garg Engineering College, Ghaziabad-201009, India
3
Student, Ajay Kumar Garg Engineering College, Ghaziabad-201009, India
Abstract. The issue of stress has pervaded the psychological landscape especially in the lives of young
adults who are being forced to deal with competitive academic as well as social worlds. Although the tradi-
tional therapeutic methods are still useful, their inaccessibility has inspired the need to find scalable digital
solutions. This paper outlines a full-fledged web-based stress-management system, that integrates structured
daily journalling, ongoing monitoring of stress levels, analytics-based personalisation, and a chat-based sup-
port agent using the large language models (LLMs). A four-week pilot study was conducted on a group of 30
university students (1850) aged 1830 years and all of whom had moderate to high stress levels at baseline
(mean PSS = 24.3). The participants were involved in the intervention process and used journalling (mean =
5.8 entries/week), used chatbot (mean = 2.4 sessions/week), and were provided with coping suggestions tai -
lored individually. The quantitative data depicted that the perceived stress decreased by 14 percent over the
course of week four (mean PSS = 20.9; p = 0.001) and that engagement measures were positively correlated
with chatbot use and perceived reduction of stress (r = 0.42). The qualitative feedback also revealed that
there were high emotional validation, usability, and value of AI-mediated reflection perceptions and that
more diverse personalised content is required. It demonstrates that the effectiveness and use of web-based
stress-management tools can be significantly improved with the help of the interventions that are supported
by LLM. The paper ends with limitations of the system and the recommendation of further work, such as ex-
tended personalisation algorithms, introduction of multimodal data integration, and controlled studies over
the long term to quantify long-term effects.
Keywords: Stress Management; Large Language Models; Web-based Intervention; Mental Health Technol-
ogy; Journalling; Personalised Analytics; Conversational AI; Digital Wellbeing; Human–AI Interaction; Psy-
chological Support Systems
1. Introduction
Stress is a major factor that affects cognition, emotional control and wellbeing in the long run. Adolescents,
especially in universities, usually find themselves in a long period of stress related to the university
requirements, social duties and life transition. Repeated stress has been found to affect the ability to learn,
consolidation of memories as well as flexibility in behaviour [1]. Although there is apparent necessity of timely
support through stress-management resources, the availability of conventional resources, including counselling
services, is not very high due to the financial, logistic, or social restrictions.
Mental-health interventions in the digital format have become available as instruments to aid emotional self-
regulation. Moderate success has been found in the use of web-based journalling, mindfulness apps, and
cognitive behavioural modules in reducing stress and increasing self-awareness [2]. Most of such interventions,
however, are based on one-size-fits-all design which is not responsive and personalised. As a result, user
involvement tends to reduce with time.
New possibilities to develop reflective and adaptive digital interventions are presented by the development of
artificial intelligence and especially large language models (LLM). The LLMs have the ability to respond in a
conversational manner, give an empathetic response, and interpret emotional patterns in user input. Studies
indicate that LLMs can be used in the mental-health setting with a high potential, such as stress detection and
emotional support [3]. The approaches of prompt-engineering have demonstrated significant improvements in
detecting stress cues in the text written by users [4].
However the features are regularly separated by the existing systems: journalling systems exist without chatbots,
LLM chatbots exist without stress tracking, and analytics systems are rarely coupled with reflective
conversation. The proposed study fills in these gaps by creating a single, full-stack platform, combining
organised journalling, customised analytics, and conversation mediated by a LLM in a unified, user-friendly
interface.
The paper provides a critique of the systems effectiveness in perceived stress reduction, patterns of engagement
with the system and emotional resonance, thus providing information on how AI-enabled devices can be
meaningful in augmenting digital stress-management interventions.
2. Related Works
There has been a trend towards the use of digital stress-management systems which are considered to be more
accessible and can aid the process of processing emotions. Structured reflection and cognitive reappraisal have
been demonstrated to enhance the regulation of affects and perceived stress by using web-based cognitive-
behavioural programs and expressive writing tasks [2]. The initial studies conducted by Pennebaker regarding
expressive journalling show the benefits of writing in organising emotions and mental burden [5].
The recent development of LLMs has widened the range of digital mental-health tools. Hua et al. [3] report in a
scoping review that there is an increasing interest in the use of LLMs in empathetic conversation, preliminary
counselling functions and mental-state recognition. Esmi et al. [4] show that the level of stress detection is
enhanced by an order of magnitude when the task-specific, optimised prompts are used with the LLMs- this is a
pointer to the fact that the prompts are sensitive to the language cues of distress.
In spite of this possibility, mental-health tools that operate using LLM have safety issues. Guo et al. [6]
highlight that the responses of LLM can sometimes be incorrect, inflated and ethically vague unless the systems
are put in place to prevent such instances. On the same note, Rousmanere et al. [7] note that AI-generated
emotional support does not seem to be more useful than trained mental-health professionals, but rather
perceived as such
Engagement has also been positive with personalisation in digital wellbeing systems. The machine-learned
stress-prediction frameworks and intelligent devices can offer actionable and timely recommendations based on
user behaviour [8]. Nevertheless, not many platforms combine LLM conversation, adaptive analytics and
structured journalling. This paper places itself at the cross-section of developments through a holistic and
responsive solution to stress-management.
3. Materials and Methods
The system was developed as a full-stack platform integrating frontend, backend, analytics, and LLM layers.
The architecture was designed to ensure seamless transitions between journalling, conversation, and trend-based
recommendations.
3.1 System Architecture
The frontend interface, built using [Link], presents four primary interaction spaces: a journalling dashboard, a
stress-level indicator, an LLM-powered chatbot interface, and a personalisation panel offering suggestions such
as calming stories, breathing exercises, or music therapy. This interface emphasises minimalism, emotional
warmth, and frictionless navigation, as design usability is known to influence psychological engagement in
digital interventions.
The [Link] backend is implemented in Express. It takes care of user authentication (JWT-based), control of
session-state, journalling storage and querying of analytics. Its platform is powered by a PostgreSQL database
containing structured records (stress scores, timestamps, suggestion logs) and document database (MongoDB)
containing semistructured journalling content.
The fundamental component of the system is the LLM Integration Layer upon which the system builds
contextual prompts before the model by combining recent journal entries, history stress scores, and mood
indicators. This model is set to react with reflective inquiring, emotional confirmation, and practical proposals
without use of denominative terms.
The management of this system is an analytics engine that handles trends of stress levels on-the-fly. A 7-day
moving average is calculated by the engine, upward or downward trends are found, recurring high-stress days
are identified, and rule-based strategies are applied to provide personalised recommendations.
3.2 User Workflow
Once the users are logged in, they are instructed to make a journal entry about how they feel during the day, and
then choose a rating on a stress scale between 0 and 10. After it is submitted, data are logged and instantly
displayed at the stress-trend graph. The users can then communicate with the chatbot that processes the most
recent post and presents unique insights or thought-provoking questions. In case the analytics engine identifies
the patterns of increased stress, the suggestions module provides the supportive tasks, which are consistent with
psychological coping styles.
The weekly engagements, stress patterns, and frequency of journalling summaries assist the users with self-
reflections about the progress.
3.3 Pilot Study Design
The pilot study was carried out in a sample of 30 university students aged between 18 and 30 years and who had
a score above 20 on the Perceived Stress Scale (PSS) which depicts moderate and high levels of stress. The first
week was confined to journalling and stress ratings to determine the baseline behaviours. During the next 3
weeks, customers were using the entire platform with chatbot and personalisation modules. At the end of week
four, semi-structured interviews were undertaken to investigate subtle thinking in terms of utility and emotional
appeal and system flaws.
Table 1: Participant Demographics
Variable Value
Total Participants 30
Age Range 18–30 years
Mean Age 21.4 years
Gender Distribution 60% Female, 40% Male
Baseline PSS Mean 24.3
3.4 Measurement Strategy
Quantitative variables were longitudinal measures of PSS, frequency of journalling, the number of chatbot
interactions, and logs of suggestion acceptance. The statistical analyses were based on the use of paired-sample
t-tests to test the difference in PSS scores in week four compared to the baseline, and Pearson correlations to test
the relationships between engagement measures and stress reduction.
To investigate the patterns of recurrent experience, qualitative feedback was subjected to analysis in the form of
an inductive thematic approach.
3.4 Ethics
Participants were informed to give their consent, and the ethical approval was made. Anonymisation of all the
datasets was done and encryption was used to store data. The chatbot had a prominent warning that the assistant
is not a replacement to professional therapy, and those scoring high on the PSS were offered mental-health
resources, according to the recommended standards and guidelines [7].
4. Results and Discussion
The pilot study (4 weeks) produced a rich set of data, which included quantitative trends, engagement
behaviour, and qualitative data, which can be used to evaluate the proposed intervention in a multidimensional
manner. The findings indicate that the combined journalling-analytics-LLM system could be used as a powerful
digital stress reliever, especially in students who are subjected to repeated academic or mood swings.
4.1 Quantitative Outcomes
The participants showed a statistically significant decrease in the level of stress during the study period. The
mean of Baseline Perceived Stress Scale (PSS) scores was 24.3 (SD = 4.7), which is in line with moderate levels
of stress that are normally recorded in university population groups. At week four, the means of PSS scores
dropped to 20.9 (SD = 4.5). Such a reduction of 14% was confirmed with the help of a paired-samples t-test,
which showed that the change was highly significant (t(29) = 4.12, p < 0.001). The degree of improvement is
similar to and in other cases a little higher than the results of previous digital interventions which are only based
on journalling or self-managed modules, which on average indicate decreases in the range of 8-12% [2].
The engagement data will give additional information regarding the platform usage and user behaviour. The
average number of journalling entries per week was 5.8, meaning that the participants are highly adherent,
which is probably explained by the minimalistic design and the lack of complexities of the platform. The
average was 2.4 times a week that the integrated chatbot would be accessed, and in some cases after a
journalling session or stressful event, which could indicate that they considered the chatbot as a continuation of
emotional processing and not a distinct activity.
The analysis of the correlation between the frequency of chatbots interaction and the improvement of PSS
showed that the correlation is moderate and positive (r = 0.42, p = 0.02). This observation confirms the
hypothesis that conversational interaction enhances the emotional processing enabled alone by journalling.
Notably, the rate of suggestion acceptance on the platform was 1.6 per week in terms of suggestion acceptance,
which indicates that not every user was active in receiving suggestions, but those ones who did had a tendency
to regard their suggestions as helpful or soothing.
Metric Mean (SD)
Journal Entries per Week 5.8 (1.2)
Chatbot Sessions per Week 2.4 (0.9)
Suggestions Accepted per Week 1.6 (0.7)
Table 2: Engagement Metrics Summary
4.2 Efficiency and Comparative Evaluation
The three-step integration of reflection (journal) and dialogue (LLM) and action (suggestions) was more
efficient compared to traditional interventions that were based on journalling alone. Traditional journalling
involves long term motivation and lacks instant interpretive feedback. Comparatively, the reflective responses of
the LLM provided real-time emotional reflection, which a number of the participants termed as motivating or
clarifying. This feedback process seemed to enhance compliance as well as the intensity of emotional
processing.
In addition, the analytics engine was helpful in adding efficiency through the automation of pattern recognition.
Users who had a constantly high level of stress were provided with specific interventions, which minimized the
load on cognition of self-monitoring. Other members explicitly pointed out that, the site recognized trends
earlier than I did and that it was psychologically important to outsource emotional tracking. The adaptive system
provided a more personalised time of intervention when compared to more traditional digital tools, including
pre-programmed modules and offered to everyone in a uniform way, similar to early intelligent-tutoring systems
employed in educational psychology.
The platform was also very time efficient. An average user spent 4-6 minutes on a journalling session, and the
conversations with chatbots took 1-3 minutes, which means that valuable emotional interaction is possible
within the micro interventions. This effectiveness is consistent with more recent discoveries that short, frequent,
digital reflections are better than less frequent and longer reflections in predicting better affect regulation [5].
Table 3: Qualitative Themes and Representative Quotes
Theme Description Representative Quote
Emotional Users felt heard, “It felt like someone was
Validation understood, and supported. genuinely listening.”
Platform was simple and “I could use it whenever
Ease of Access
available anytime. I needed, even at night.”
Suggestions became “I wanted more unique
Need for Variety
repetitive over time. or varied suggestions.”
Safety & Users understood AI is not a “I knew it’s AI, but still
Boundaries therapist. helpful in daily stress.”
4.3 Qualitative Themes
Qualitative interviews offered more information about the experience of the users on the platform. One theme
was the emotional validation. A lot of respondents said that the responses of the LLM made them feel listened to
despite the intellectual comprehension that the system was unnatural. This assumption is in line with the
available previous research on the topic of LLM-mediated empathetic dialogue, which can easily give a feeling
of conversational presence [6].
Accessibility was another common theme. The participants also enjoyed having a chance to reflect and be
supported at any hour which is not the case in human counselling as it follows a certain schedule. The fact that
the chatbot is always available led to the lesser emotional bottlenecking, i.e., the users felt the emotions and did
not hold them back.
The third theme was the necessity of content variety. Although suggestions were mostly viewed to be helpful, a
number of the participants believed that the recommendations became redundant towards the end of the study.
This implies that it may be needed to increase the size of the suggestion library, and possibly include machine-
learning models in future prototypes of the adaptive content generation process.
Lastly, the participants mentioned that they were comfortable with the limits of the system. The platform had its
disclaimer, that it is not a replacement of therapy, which was perceived as responsible and sincere. Realistic
expectations were kept in value with the system by users, which portrays an effective ethical communication
design.
4.4 Integrated Interpretation
When combined, the findings suggest that the suggested intervention is effective in integrating the features of
conventional journalling with the quickness and compassion of conversational AI. The statistically significant
stress reduction, and high adherence to it by the users, indicates that this hybrid treatment can be effective in
comparison to a number of current single-component interventions.
The efficiency of the platform is due to three major strengths:
1. Reduced Cognitive Load: Trend monitoring and automated reflection processes help the user to lighten the
load of analysing their history of emotions.
2. Increased Surgery: Engagement Conversational interfaces have a greater emotional commitment than fixed
input fields.
3. Personalised Timing: The analytics engine provides prompt suggestions at the point of need and therefore is
proactive and not reactive in its timing.
Compared to other previous studies on LLM where the main concern was on detection or short term
conversational experiment [3][4], in this study a fully integrated cycle of intervention has been shown with self
reflection, emotional dialogue and guided behaviour action. Such triadic format is closer to such principles of
therapy as cognitive behavioural reflection and emotional regulation methods.
4.5 Limitations and Forward Directions
In spite of the optimistic results, there are a number of constraints. The sample was rather small and
demographically restricted which led to a lack of generalisability. Although self-report measures are valid, they
may lead to biasness. The lack of the control group does not allow isolating the special contribution of the LLM
component. Also, the personalisation engine that is based on the rule-based approach is not as sophisticated as
data-driven adaptive systems.
Future considerations would be to study larger groups of people, use randomised controlled trials, combine
physiological monitoring (e.g., HRV, sleeping patterns), and use personalisation (based on reinforcement
learning) in refining suggestive accuracy and content richness.
5. Conclusion
This paper presented a new integrated web-based stress management architecture, which combines journalling,
conversation through the use of LLM, trend analysis, and customised recommendations. The outcomes of a
four-week pilot are promising to indicate significant changes in perceived stress and high levels of engagement
with the conversational agent. The results demonstrate that AI, in particular, LLMs, has a great potential to
complement digital mental-health tools. Future studies are to develop personalisation algorithms further, include
the use of physiological data, enhance the diversity of content and use controlled trials to evaluate effectiveness
more rigorously. All in all, these systems have potentials as easily available, understanding, scalable everyday
stress control tools.
References
[1] Lupien SJ, McEwen BS, Gunnar MR, Heim C. Effects of stress throughout the lifespan on brain, behaviour
and cognition. Nat Rev Neurosci. 2009;10(6):434–445.
[2] Heber E, Ebert DD, et al. Web-based interventions for stress management: Systematic review. J Med
Internet Res. 2017;19(2):e32.
[3] Hua Y, Na H, Li Z, et al. Large language models in mental health: A scoping review. npj Digit Med.
2025;8:230.
[4] Esmi N, Shahbahrami A, et al. Stress detection via prompt engineering. Acta Psychologica.
2025;260:105462.
[5] Pennebaker JW. Writing about emotional experiences as a therapeutic process. Psychol Sci. 1997;8(3):162–
166.
[6] Guo Z, Lai A, Thygesen JH, et al. LLMs for mental health applications: Systematic review. JMIR Ment
Health. 2024;11:e50110.
[7] Rousmaniere T, Zhang Y, Li X. LLMs as mental-health resources. Practice Innovations. 2025;10(1):34–48.
[8] Wahab O, et al. Stress-management with smart monitoring devices: Review. Procedia Comput Sci.
2025;232:145–156.
