Cognify: A Modular Privacy-Conscious AI-Driven Mobile App for

Mental Health Based on Cognitive Distortion Detection

Mariam Dawoud

, Mohamad Rasmy

and Alia El Bolock

Department of Computer Science and Engineering, The American University in Cairo, Cairo, Egypt

Computer Science Department, Faculty of Computer and Information Sciences, Ain Shams University, Cairo, Egypt

Keywords:

Mental Health, Mobile App, Learning Model, Security, Cognitive Distortions.

Abstract:

Cognitive distortions—irrational thought patterns contributing to emotional distress—are central to cognitive

behavioral therapy (CBT), but early detection often depends on clinical assessments, limiting opportunities

for timely self-reﬂection. Cognify is a cross-platform mobile application designed to help users detect and

understand these distortions by analyzing daily journal entries using a ﬁne-tuned NLP model that classiﬁes

entries into 14 cognitive distortion types. The app offers real-time feedback, weekly summaries highlighting

recurring patterns, and an intuitive interface that promotes ongoing engagement. This paper presents Cognify’s

system architecture, AI model integration, and results from a pilot study, which demonstrated improved user

awareness of cognitive patterns, high user satisfaction, and increased journaling consistency over time. The

app’s modular design also allows for optional integration of privacy-preserving features, ensuring ﬂexibility

to address evolving user needs. By combining AI-driven distortion detection with an adaptable journaling

experience, Cognify offers a practical and engaging tool for enhancing cognitive awareness and supporting

personal growth.

1 INTRODUCTION

Mental health applications have gained signiﬁcant

traction in recent years, providing accessible and ef-

fective support for users managing many psycholog-

ical disorders and maintaining their emotional well-

being. Due to the increased awareness on mental

health, it is important to be able to detect early symp-

toms of mental disorders called cognitive distortions,

which are thoughts that cause inaccurate perceptions

of reality due to exaggerations or irrationality (Beck,

2022). These distortions can be analyzed on the

basis of 15 characteristics(including neutral) to help

guide mental health professionals (psychiatrists, psy-

chologists, therapists, and others) to susceptible dis-

orders. Most existing applications either lack a ro-

bust AI-based cognitive distortion detecting mecha-

nism or fail to implement a privacy-preserving mech-

anism, causing drawbacks in participation due to con-

cerns on trust and integrity. We propose a compre-

hensive approach by addressing challenges through

a user-friendly cross-platform mobile application that

can accurately detect cognitive distortions through an

AI robust model using RoBERTa, as well as maintain

privacy of sensitive data through a pluggable frame-

work of integrated techniques. The user will be log-

ging in their journal entry each day, and each entry

will be classiﬁed with the relevant cognitive distor-

tion, then the data will be stored safely on the database

for retrieval when needed. We aim to ensure users are

comfortable using the application and providing their

journal data while trusting the application to provide

meaningful results based on insights from their data.

At the core of the application is a robust AI-

driven cognitive distortion detection module, which

analyzes users’ daily journal entries using a ﬁne-tuned

RoBERTa model. This AI model automatically clas-

siﬁes entries into one or more of 15 recognized cogni-

tive distortion categories, providing users with clear,

actionable feedback to help them recognize problem-

atic thought patterns over time. The application is

designed with modularity in mind, ensuring that the

AI model, user interface, and data handling compo-

nents are decoupled to allow for easy updates and

customization. This modular design also leaves room

for incorporating privacy-preserving technologies if

needed, particularly to address user concerns when

handling sensitive mental health data. By providing

a user-friendly journaling interface combined with in-

telligent feedback, the application empowers users to

Dawoud, M., Rasmy, M. and El Bolock, A.

Cognify: A Modular Privacy-Conscious AI-Driven Mobile App for Mental Health Based on Cognitive Distortion Detection.

DOI: 10.5220/0013571400003964

In Proceedings of the 20th International Conference on Software Technologies (ICSOFT 2025), pages 361-368

ISBN: 978-989-758-757-3; ISSN: 2184-2833

361

become more aware of their thinking patterns while

offering valuable insights for both personal reﬂection

and professional therapeutic support.

The paper presents the following contributions:

• develop a full-stack cross-platform journaling

mobile application to securely detect cognitive

distortions.

• construct a machine learning model for detecting

cognitive distortions and their types.

• implement a privacy-preserving framework inte-

grating ECC and Blockchain latest technologies.

• conduct a pilot study to ensure the usability of the

app from the user’s perspective.

By combining a powerful AI analysis engine with

an intuitive and adaptable mobile interface, the appli-

cation offers a comprehensive, ﬂexible tool for cog-

nitive distortion awareness and self-reﬂection, adapt-

able to both individual use and potential integration

with therapeutic interventions.

The rest of the paper is designed as followed:

Section 2 discusses existing literature in the ﬁelds of

mental health mobile apps, AI detection models, and

privacy-preserving algorithms. Section 3 discusses

the system architecture and how the application ﬂow

works. Section 4 discusses the methodology in more

detail, focusing on the technical implementations and

evaluation metrics. Section 5 discusses results and

what they infer in the ﬁeld, followed by Section 6 to

suggest future work based on the provided results.

2 RELATED WORK

This section discusses existing literature for mobile

mental health applications speciﬁcally to survey avail-

able methodologies and identify gaps in research. We

also look at AI detection models and brieﬂy discuss

the possibility of privacy preserving techniques.

2.1 Mobile Mental Health Applications

Awareness of mental health disorders has increased

signiﬁcantly in recent years, driving the develop-

ment of mobile applications designed to support users

through features like mood tracking, guided exercises,

and assessments. A narrative review categorized

mental health apps into distinct types such as mood

trackers, mindfulness-based apps, self-care apps, and

treatment apps. This categorization highlighted the

need for mental health apps in widespread disorders

such as anxiety, depression, and suicidal ideations,

and the need for incorporation with psychotherapy

for enhanced outcomes and interventions. Blending

these applications with Cognitive Behavioral Therapy

(CBT) techniques can be very efﬁcient, but challenges

such as low user engagement, inconsistent evaluation

methodologies, and privacy concerns are emphasized.

Moreover, the lack of standardized assessments and

fragmentation issues of these apps limit the research

in deducing a reliable conclusion (Diano et al., 2022).

Mindful Meadows is a mobile mental health app

that utilizes self-assessments, mood tracking, music,

yoga, a chatbots and therapist booking all in one

framework to offer comprehensive support. However,

the system raises some concerns as it lacks clear clin-

ical validations and weak privacy protections (Gaik-

wad et al., 2024). Another app called Mindset pro-

vides similar comprehensive support through journal-

ing, mood tracking, guided breathing, and anony-

mous peer support. Results and reviews prove its

convenience in use, although no evaluation has been

done on the long-term health beneﬁts (Samuel and

Shirley, 2023). MindWell Solace is a chatbot-driven

application targeting student mental health that uses

a classiﬁer to detect disorders and inform counselors

if needed. The model achieves a good accuracy for

some symptoms but is depth-limited due to the na-

ture of yes/no questions and symptom self-reporting

(Bhave et al., 2024). Finally, an app developed dur-

ing the COVID-19 pandemic offered remote counsel-

ing with mental health professionals, integrating their

records and journal entries. While it proved effective

for remote access, it is relatively limited in its scala-

bility (Krisnanik et al., 2020).

2.2 Cognitive Distortion Detection

Models

In the technical scope, detecting cognitive distortions

has been done through gaming; ARCod is an aug-

mented reality (AR) interactive game to assess cogni-

tive distortions according to its 5 different levels (Tas-

nim and Eishita, 2022). Detection was also imple-

mented using a machine learning-based text analysis

of online blogs and journal entries, similar to what is

implemented in this app (Shickel et al., 2020a).

The detection of cognitive distortions has been ap-

proached as a text classiﬁcation problem. A major

challenge in this task is obtaining a sufﬁciently large

labeled dataset to train deep learning models effec-

tively. The literature contains extensive efforts to uti-

lize deep learning for cognitive distortion detection.

Several works have focused on collecting and anno-

tating datasets for this purpose, as well as training ma-

chine learning classiﬁcation models on these datasets.

Most of these datasets are in English, including those

ICSOFT 2025 - 20th International Conference on Software Technologies

362

compiled by (Shickel et al., 2020b; Elsharawi and

El Bolock, 2024; Lim et al., 2024; Shreevastava and

Foltz, 2021). Additionally, a few datasets have been

developed in Chinese, such as those by (Wang et al.,

2023; Qi et al., 2023; Na, 2024).

Our work speciﬁcally builds upon the founda-

tional dataset introduced by (Mostafa et al., 2021),

which was the ﬁrst publicly available cognitive dis-

tortion detection dataset. This English-language

dataset comprises 2,409 text entries categorized into

two types of cognitive distortions (overgeneraliza-

tion and should statements) along with non-distorted

texts. The authors evaluated several machine learn-

ing and deep learning models employing pre-trained

embeddings, ultimately identifying a tuned Long

Short-Term Memory (LSTM) network with 300-

dimensional GloVe embeddings as the most effective

model.

Subsequently, (Elsharawi and El Bolock, 2024)

expanded this dataset, creating the largest open-

source collection available to date. This extended

dataset contains 34,370 sentences annotated across

14 cognitive distortion categories, along with neutral

(non-distorted) examples. Their highest-performing

model was a Convolutional Neural Network (CNN)

employing pre-trained BERT embeddings. This

dataset was sourced from existing social media emo-

tion datasets, annotated by one of the authors with a

psychology background, and subsequently veriﬁed by

a certiﬁed psychologist.

Given the limited availability of high-quality an-

notated datasets, recent studies have explored ar-

tiﬁcial dataset augmentation techniques. Notably,

(Rasmy et al., 2024) proposed four data augmentation

methods speciﬁcally to expand the dataset introduced

by (Elsharawi and El Bolock, 2024), thereby enhanc-

ing training set size and improving model perfor-

mance. They demonstrated an improvement of up to

5.9% in F1-score using a ﬁne-tuned RoBERTa model

trained on the augmented dataset, compared to the

previously identiﬁed CNN-BERT model evaluated on

the original non-augmented dataset. In our frame-

work, we integrate this best-performing ﬁne-tuned

RoBERTa model trained on the augmented dataset.

2.3 Privacy Preservation in Mobile

Apps

Privacy in mobile health apps remains a major con-

cern due to excessive data collection, unclear poli-

cies, and security vulnerabilities that undermine user

trust. Lightweight, transparent privacy frameworks

are needed for sensitive health data. Privacy tech-

niques in mental health mobile applications vary

widely, from basic encryption methods like AES and

RSA to more advanced approaches such as Homo-

morphic Encryption, differential privacy, and secure

cloud storage frameworks (Vichare et al., 2017; Zhou

et al., 2018; Inakollu et al., 2024; Latif et al., 2020).

While these methods offer varying levels of data pro-

tection, many suffer from high computational costs,

performance delays, reliance on external auditors, or

reduced data accuracy, making them difﬁcult to im-

plement seamlessly in real-world mobile apps (Su-

guna and Shalinie, 2017; Vijay Sai et al., 2024;

Whaiduzzaman et al., 2020). Although these tech-

niques are not the core of our proposed framework, it

is important to highlight them due to the highly sensi-

tive nature of data collected in mental health applica-

tions, where strong privacy protections are critical to

maintaining user trust and encouraging sustained app

engagement (O’Loughlin et al., 2019; Parker et al.,

2019).

3 COGNIFY OVERVIEW &

FEATURES

We present the system features highlighting the use

cases available on the app, then walk through the de-

velopment process from reading in the user’s journal

entry to its storage in the database.

3.1 System Features

Cognify was designed to deliver real-time cognitive

distortion feedback through a journaling experience

at the frontend for an intuitive user interaction to a

sophisticated backend processing and privacy main-

tenance. The app is a feature-rich, cross-platform

mobile application designed to enhance self-reﬂection

and mental well-being through AI-driven cognitive

distortion detection. The system provides users with

an user-friendly journaling experience while leverag-

ing advanced NLP techniques to analyze thought pat-

terns in real time. By automatically classifying jour-

nal entries into 15 cognitive distortion categories, the

app offers immediate feedback to help users better

understand their thought processes. Additionally, it

generates structured weekly insights, allowing users

to track recurring patterns over time and gain deeper

self-awareness. Seamless data storage and retrieval

and possible secure data handling, through encryption

and account deactivation, provide accessibility and

protection of user sensitive data. By combining AI-

powered analysis with an adaptable and user-centric

interface, Cognify aims to bridge the gap between

self-guided mental health tools and professional ther-

Cognify: A Modular Privacy-Conscious AI-Driven Mobile App for Mental Health Based on Cognitive Distortion Detection

363

apeutic interventions. In the section below, we take

a look at the development of the application to fulﬁll

these features seamlessly.

3.2 Data Storing

The single string of data is then immediately en-

crypted locally; we generate an encryption key that

is unique per user and derived using a combination

of device-stored keys protected via the Flutter Secure

Storage module. This ensures that if the database

is ever compromised, the entries remain unreadable

without the user’s local key. Upon encryption, the

data is stored in a hybrid model of a blockchain on

Firebase; this is described further in section 4.6.2.

3.3 Generating Analytics

Once the entry is saved safely on the database, we

can view a scrollable view of all previous entries, and

opening any entry card reveals the collected data: the

title, date and time, mood and intensity, journal en-

try, and the classiﬁcation result. This provides users

an opportunity to view and reﬂect back on previous

entries and emotions.

When the user views their proﬁle, they are met

with a couple of analytics and summaries. An aver-

age mood display calculates the user’s average mood

throughout the week, updating each day and restarting

at the beginning of the new week. Moreover, a sum-

mary display of the week’s entries is presented, with

only the title visible under each day of the week. Fig-

ure 1b shows the interface developed for the proﬁle

page described.

3.4 Privacy & Account Deactivation

At any given time, the user can decide to deactivate

or delete their account. Taking this action triggers a

full data export process, permanently removing data

from the database and exporting it in a ZIP archive

ﬁle locally on the user’s mobile device.

3.5 Application Prototyping

In the following two sections, we delve deeper into

the model used and the privacy optional layer utilized.

3.5.1 Cognitive Distortion Detection Algorithm

In this study, we utilize the dataset provided by

(Elsharawi and El Bolock, 2024), as it is the largest

publicly available dataset in English for cognitive

distortion detection. We enhance this dataset with

(a) Entry Editing Screen (b) Proﬁle Page

Figure 1: Screenshots from the Cognify App.

augmentations provided by (Rasmy et al., 2024).

The best-performing model identiﬁed by the authors

for this dataset, which we adopt, is the ﬁne-tuned

RoBERTa model.

RoBERTa (Liu et al., 2019) is pre-trained on a

large and diverse textual corpus, using a masked lan-

guage modeling objective to predict missing tokens

based on their surrounding context. This approach en-

ables the model to develop a nuanced understanding

of linguistic structures and meaning. It has demon-

strated strong performance across a wide range of

NLP tasks, particularly in text classiﬁcation, which

makes it a well-suited choice for identifying subtle

linguistic patterns in cognitive distortions.

The authors in (Rasmy et al., 2024) experimented

with four data augmentation techniques: synonym re-

placement (SR), random insertion (RI), word embed-

ding substitution, and back-translation. Their results

indicated that the highest performance was achieved

using a combination of SR and RI. Consequently, we

train our model using datasets augmented with these

two techniques.

We follow the preprocessing, data splitting, and

hyperparameter recommendations provided by the

authors. The preprocessing steps include lowercas-

ing text, removing punctuation, unrecognized sym-

bols, and tags, and eliminating duplicate entries. The

dataset is split into 70% training, 10% validation, and

20% testing. For ﬁne-tuning, we employ the best-

identiﬁed learning rate of 4e-5 and use early stopping

to monitor validation loss instead of a ﬁxed number

of epochs. Additionally, we optimize the batch size to

32, as it yields the best results.

ICSOFT 2025 - 20th International Conference on Software Technologies

364

3.5.2 Privacy Preserving Algorithm

The development process of Cognify was guided by a

modular framework design, ensuring that each core

component—journal management, cognitive distor-

tion detection, and data handling—can operate in-

dependently while still integrating seamlessly. This

modularity allows for the optional inclusion of a

privacy-preserving layer, which can be enabled or

omitted depending on the sensitivity of the data be-

ing processed. For less sensitive data or during initial

development phases, the app can function without en-

cryption, maintaining ﬂexibility for different deploy-

ment needs.

To validate the modular design and test the sys-

tem’s ability to handle encrypted data ﬂows, we in-

tegrated a pluggable privacy-preserving layer using

AES-based encryption. This test implementation en-

sures that the application can fetch, store, and pro-

cess encrypted data without disrupting the journaling

or cognitive distortion detection workﬂows. This pri-

vacy module remains open for further enhancement,

allowing future contributors to replace or upgrade the

encryption mechanism based on evolving privacy re-

quirements or regulatory standards.

4 SYSTEM ARCHITECTURE

The system architecture of Cognify is designed to

deliver a cross-platform mobile application that sup-

ports journaling, cognitive distortion detection, and

optional privacy-preserving features. The architec-

ture follows a modular client-server model, consist-

ing of a mobile frontend, a cloud-hosted backend,

and supporting services for storage, classiﬁcation,

and encryption. The mobile client, built using Flut-

ter, ensures a uniﬁed experience across Android and

iOS, simplifying maintenance and ensuring consis-

tent functionality. Journal entries are processed by a

cognitive distortion classiﬁer deployed as a serverless

function, activated only when new entries are submit-

ted. The system is designed with modularity at its

core, allowing each component—including data han-

dling, AI processing, and privacy layers—to operate

independently, enabling future upgrades or replace-

ments without disrupting the overall system. Impor-

tantly, the privacy-preserving component, including

encryption and secure storage, is implemented as a

ﬂexible, optional add-on that can be customized or re-

placed based on evolving privacy requirements. This

architecture balances computational efﬁciency, plat-

form ﬂexibility, and extensibility to accommodate fu-

ture enhancements, including integration with thera-

Figure 2: Cognify Architecture.

pists’ platforms or evolving security needs.

Figure 2 presents our proposed architecture for the

Cognify app.

4.1 Development Process

The system is divided into the following key compo-

nents:

• Mobile Client: responsible for the user interac-

tion, journaling input, and encrypting the data.

• Processing: responsible for detecting the cogni-

tive distortion characteristic; a model ﬁne tuned

on cognitive distortion datasets.

• Data Storage: handles encrypted data storage

and retrieval.

• Analytics Generation: computing weekly sum-

maries, average mood, and using entries to gener-

ate trends.

• Deactivation & Data Export: enables users to

securely download their data before account dele-

tion.

Figure 3 presents the basic processing layer of the

Cognify system. Once the user is logged in, they can

create their daily journal entry, where a simple inter-

face is provided for a distraction-free entry editing, as

seen in Figure 1a. The user titles the entry, records

their mood on a scale from 1 to 5 (1 being not well

and 5 being very well) and how intense are they feel-

ing that emotion on a similar scale of 1 to 5. They

then proceed to enter their thoughts in a form of jour-

naling, which can be one or more sentences.

Once saved, this entry is fed into the trained BERT

model to classify the entry into one of the 15 prede-

ﬁned cognitive distortions, including but not limited

to catastrophizing, black-and-white thinking, and oth-

ers. This classiﬁcation result is concatenated to the

original raw entry string to form a single string ready

to be stored. More details on the construction of the

model can be found in section 3.5.1.

5 RESULTS & DISCUSSION

In this section, we provide evaluations of the Cognify

app based on user reviews and model accuracy. We

conduct a pilot study in the ﬁrst section to evaluate

Cognify: A Modular Privacy-Conscious AI-Driven Mobile App for Mental Health Based on Cognitive Distortion Detection

365

Figure 3: Cognify Processing Flowchart.

usability and trust of users for the app, then we discuss

the result accuracy of the model compared to previous

similar models.

5.1 Pilot Study of Cognify App

A pilot study was conducted with ﬁve participants to

evaluate the app’s usability, effectiveness, and user

engagement, and edits were applied accordingly to

ensure the app aligns with user expectations and ad-

dresses common concerns in mental health apps.

The participants for the pilot study were recruited

through convenience sampling from a university set-

ting, friends, and family. They were between the ages

of 20 and 30 and represented diverse educational ma-

jors and backgrounds, including engineering, social

sciences, and health-related ﬁelds. This variety en-

sured a basic level of heterogeneity in user perspec-

tives, aiding the evaluation of the usability of the app

across non-specialist populations.

The two-hour session includes a brief introduction

to the mobile app and the concept of cognitive distor-

tions, app download and exploration, followed by a

feedback session featuring the System Usability Scale

(SUS) to assess usability and user experience. Partic-

ipants ﬁnally engaged in a discussion to share their

comfort levels, impressions of the app, and alignment

with their understanding of cognitive distortions. We

discuss the analysis and results below according to us-

ability, willingness to use as a concept, and trust in the

app.

The usability analysis includes evaluating the re-

Figure 4: Cognify User Feedback Summary.

sponses from the SUS questions and calculating an

overall score to determine usability, along with iden-

tifying concerns from the the open-ended responses.

Calculating the average SUS value for the app yielded

a score of 89.38 out of 100, which is remarkable; a

SUS score above 80 indicates high usability, mean-

ing the users found the app intuitive and easy to

use. Overall, the participants had a smooth experi-

ence with minimal usability issues.

Regarding the willingness to use an app that de-

tects and classiﬁes cognitive distortions, we analyzed

responses about whether users found the app useful

for managing negative thinking, assessed their readi-

ness to continue using the app and recommend it to

their friends, and obtained feedback on their overall

comprehension. 4 out of 5 participants (80%) said

they would be open to using the app, with one par-

ticipant saying maybe. Similarly, 4 of 5 believed the

application could improve mental health, and would

recommend the app to friends and family. The ﬁfth

participant voiced the need for improvements in clar-

ifying the topic further and increasing engagement in

the app for higher effectiveness.

Finally, few questions were asked on users’ trust

in using the app to emphasize the need for a privacy

integration in any mental health application. This

included examining responses regarding trust, par-

ticularly data privacy and AI analysis to determine

whether users feel comfortable sharing their journal

entries and having an AI model detect their cognitive

distortions if any. 4 of 5 participants stated that they

would trust the app if it applies a strong privacy mech-

anism and if there is a clear readable privacy policy.

All participants expressed their concerns about AI an-

alyzing their journal entry data, indicating a need for

transparency in the way data is handled and a need

for more control and clarity on the whereabouts of

the datasets collected. Figure 4 highlights results of

the most prominent questions.

In conclusion, the pilot study shows a promising

future for Cognify as a cognitive distortion journaling

ICSOFT 2025 - 20th International Conference on Software Technologies

366

detection app. User feedback informed modiﬁcations

will include higher transparency on how AI handles

journal data. The clear usability of the app, along with

the public interest in mental health recognition, paves

way for further enhancements and extensions to ful-

ﬁll further objectives such as therapist integration and

privacy modules.

5.2 Cognitive Distortions Detection

Model

To evaluate the model’s performance, we use a dataset

comprising 74,055 training samples—24,685 from

the original dataset, 24,685 from SR augmentation,

and 24,685 from RI augmentation. The validation

and test sets contain 3,526 and 7,054 samples, re-

spectively. We assess the model using standard text

classiﬁcation metrics, including precision, recall, F1-

score, and accuracy. The ﬁne-tuned RoBERTa model

achieves scores of 63.05% for precision, 68.32% for

recall, 64.27% for F1-score, and 68.32% for accuracy.

When evaluating such an automatic model for

classifying cognitive distortions, it is crucial to con-

sider the inherent risks of misclassiﬁcation due to

overlapping characteristics among various categories.

For instance, the sentence ”I failed this interview, I’ll

probably fail all interviews I get” simultaneously rep-

resents overgeneralization, magniﬁcation, and catas-

trophizing (Mostafa et al., 2021). Such overlaps

present signiﬁcant challenges for models typically

trained to select a single deﬁnitive category, poten-

tially overlooking other relevant distortions.

The complexity and inherent ambiguity of natural

language further exacerbate this issue, contributing

directly to low inter-annotator agreement rates. Anno-

tation of cognitive distortions, which the model relies

on during training, is inherently subjective, as anno-

tators often struggle to choose a single dominant cat-

egory, frequently inadvertently prioritizing secondary

distortions (Pico et al., 2025). This subjectivity sig-

niﬁcantly complicates the creation of consistent and

reliable training datasets.

Although experienced psychologists may detect a

broader range of distortions in their patients, some

types may still go undetected; automated cognitive

distortion detection remains an invaluable tool within

Cognitive Behavioral Therapy (CBT). The primary

goal of CBT is to help patients self-identify their

distorted thought patterns, thereby promoting self-

awareness and therapeutic progress. Moreover, these

automated systems support therapists by highlighting

cognitive distortions that may not be explicitly evi-

dent during sessions, extending the therapists’ obser-

vational capabilities.

6 CONCLUSION & FUTURE

WORK

Cognify is the ﬁrst mobile application to automati-

cally detect and classify cognitive distortions neces-

sary for the diagnosis of many psychological disor-

ders while maintaining user privacy. Combining real-

time feedback with privatized journaling, we are able

to bridge the gap between self-help tools and thera-

pist guided Cognitive Behavioral Therapy (CBT). Our

solution provides a scalable proactive tool to support

mental health with its privacy ensuring design and fo-

cus on accurate cognitive detection, making it an im-

portant contribution in the personalized digital mental

health care sector in research.

Extensions on this research can include a chatbot

system to converse with the user instead of having

them write an entry directly; this can further elimi-

nate user bias and produce more concrete data. Given

the inherent ambiguity and subjectivity in cognitive

distortion annotations, future research could focus on

improving dataset quality and investigating the efﬁ-

cacy of multi-label classiﬁcation frameworks to better

capture overlapping distortions. Moreover, the pri-

vacy preserving framework can be further enhanced

to be later applied to any mobile application hold-

ing sensitive data. Further user feedback can be ob-

tained by expanding the pool size of participants over

a longer period of time.

REFERENCES

Beck, J. S. (2022). Cognitive behavior therapy: Basics and

beyond.

Bhave, U., Narendra, M. M., Bhadresh, D. J., Suhas, J. A.,

and Ajit, R. R. (2024). Mindwell solace: Your mental

health companion. In 2024 4th Asian Conference on

Innovation in Technology (ASIANCON), pages 1–4.

Diano, F., Ponticorvo, M., and Sica, L. S. (2022). Men-

tal health mobile apps to empower psychotherapy: A

narrative review. In 2022 IEEE International Con-

ference on Metrology for Extended Reality, Artiﬁcial

Intelligence and Neural Engineering (MetroXRAINE),

pages 306–311.

Elsharawi, N. and El Bolock, A. (2024). C-journal: a jour-

naling application for detecting and classifying cogni-

tive distortions using deep-learning based on a crowd-

sourced dataset. In Proceedings of the 2024 Joint

International Conference on LREC-COLING, pages

3224–3234.

Gaikwad, A., Nimbolkar, G., Keswani, R., Kolhe, V., and

Navale, G. (2024). Mindful meadows: A mental

health app. In 2024 8th International Conference on

Computing, Communication, Control and Automation

(ICCUBEA), pages 1–7.

Cognify: A Modular Privacy-Conscious AI-Driven Mobile App for Mental Health Based on Cognitive Distortion Detection

367

Inakollu, A., Kranthi, S., and A, J. (2024). A novel approach

to data security in cloud storage using erasure coding

and re-encryption. In 2024 8th International Confer-

ence on I-SMAC (IoT in Social, Mobile, Analytics and

Cloud) (I-SMAC), pages 972–976.

Krisnanik, E., Isnainiyah, I. N., and Resdiansyah, A. Z. A.

(2020). The development of mobile-based applica-

tion for mental health counseling during the covid-19

pandemic. In 2020 International Conference on In-

formatics, Multimedia, Cyber and Information System

(ICIMCIS), pages 324–328.

Latif, S., Hao, Y., Zhang, H., Bassily, R., and Rountev, A.

(2020). Introducing differential privacy mechanisms

for mobile app analytics of dynamic content. In 2020

IEEE International Conference on Software Mainte-

nance and Evolution (ICSME), pages 267–277.

Lim, S., Kim, Y., Choi, C.-H., Sohn, J.-y., and Kim, B.-H.

(2024). ERD: A framework for improving LLM rea-

soning for cognitive distortion classiﬁcation. In Pro-

ceedings of the 6th Clinical Natural Language Pro-

cessing Workshop, pages 292–300, Mexico City, Mex-

ico. Association for Computational Linguistics.

Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D.,

Levy, O., Lewis, M., Zettlemoyer, L., and Stoyanov,

V. (2019). Roberta: A robustly optimized bert pre-

training approach. arXiv preprint arXiv:1907.11692.

Mostafa, M., El Bolock, A., and Abdennadher, S. (2021).

Automatic detection and classiﬁcation of cognitive

distortions in journaling text. In WEBIST, pages 444–

452.

Na, H. (2024). CBT-LLM: A Chinese large language

model for cognitive behavioral therapy-based mental

health question answering. In Proceedings of the 2024

Joint International Conference on Computational Lin-

guistics, Language Resources and Evaluation (LREC-

COLING 2024), pages 2930–2940, Torino, Italia.

ELRA and ICCL.

O’Loughlin, K., Neary, M., Adkins, E. C., and Schueller,

S. M. (2019). Reviewing the data security and pri-

vacy policies of mobile apps for depression. Internet

Interventions, 15:110–115.

Parker, L., Halter, V., Karliychuk, T., and Grundy, Q.

(2019). How private is your mental health app data?

an empirical study of mental health app privacy poli-

cies and practices. International Journal of Law and

Psychiatry, 64:198–204.

Pico, A., Taverner, J., Vivancos, E., and Garcia-Fornes, A.

(2025). Comparative analysis of the efﬁcacy in the

classiﬁcation of cognitive distortions using llms. In

Proceedings of the 17th International Conference on

Agents and Artiﬁcial Intelligence - Volume 1: EAA,

pages 957–965. INSTICC, SciTePress.

Qi, H., Zhao, Q., Li, J., Song, C., Zhai, W., Luo, D., Liu,

S., Yu, Y. J., Wang, F., Zou, H., et al. (2023). Super-

vised learning and large language model benchmarks

on mental health datasets: Cognitive distortions and

suicidal risks in chinese social media. arXiv preprint

arXiv:2309.03564.

Rasmy, M., Sabty, C., Sakr, N., and El Bolock, A. (2024).

Enhanced cognitive distortions detection and classiﬁ-

cation through data augmentation techniques. In Pa-

ciﬁc Rim International Conference on Artiﬁcial Intel-

ligence, pages 134–145. Springer.

Samuel, M. and Shirley, C. (2023). Mindset, an android-

based mental wellbeing support mobile application. In

2023 3rd International Conference on Pervasive Com-

puting and Social Networking (ICPCSN), pages 989–

996.

Shickel, B., Siegel, S., Heesacker, M., Benton, S., and

Rashidi, P. (2020a). Automatic detection and clas-

siﬁcation of cognitive distortions in mental health

text. In 2020 IEEE 20th International Conference

on Bioinformatics and Bioengineering (BIBE), pages

275–280.

Shickel, B., Siegel, S., Heesacker, M., Benton, S., and

Rashidi, P. (2020b). Automatic detection and clas-

siﬁcation of cognitive distortions in mental health

text. In 2020 IEEE 20th International Conference

on Bioinformatics and Bioengineering (BIBE), pages

275–280. IEEE.

Shreevastava, S. and Foltz, P. (2021). Detecting cognitive

distortions from patient-therapist interactions. In Pro-

ceedings of the Seventh Workshop on Computational

Linguistics and Clinical Psychology: Improving Ac-

cess, pages 151–158.

Suguna, M. and Shalinie, S. M. (2017). Privacy preserv-

ing data auditing protocol for secure storage in mobile

cloud computing. In 2017 International Conference

on Wireless Communications, Signal Processing and

Networking (WiSPNET), pages 2725–2729.

Tasnim, R. A. and Eishita, F. Z. (2022). Arcod: A serious

gaming approach to measure cognitive distortions. In

2022 IEEE 10th International Conference on Serious

Games and Applications for Health(SeGAH), pages

1–8.

Vichare, A., Jose, T., Tiwari, J., and Yadav, U. (2017). Data

security using authenticated encryption and decryp-

tion algorithm for android phones. In 2017 Interna-

tional Conference on Computing, Communication and

Automation (ICCCA), pages 789–794.

Vijay Sai, R., Geetha B, G., Yogeshwaran, S., Vignesh,

A., and Santhosh, D. (2024). Implementation mod-

ular encryption to safeguard health data in mobile

cloud environments. In 2024 3rd International Con-

ference on Applied Artiﬁcial Intelligence and Comput-

ing (ICAAIC), pages 1352–1357.

Wang, B., Deng, P., Zhao, Y., and Qin, B. (2023). C2d2

dataset: A resource for the cognitive distortion analy-

sis and its impact on mental health. In Findings of the

Association for Computational Linguistics: EMNLP

2023, pages 10149–10160.

Whaiduzzaman, M., Hossain, M. R., Shovon, A. R., Roy,

S., Laszka, A., Buyya, R., and Barros, A. (2020). A

privacy-preserving mobile and fog computing frame-

work to trace and prevent covid-19 community trans-

mission. IEEE Journal of Biomedical and Health In-

formatics, 24(12):3564–3575.

Zhou, T., Cai, Z., Xiao, B., Wang, L., Xu, M., and Chen, Y.

(2018). Location privacy-preserving data recovery for

mobile crowdsensing. volume 2, pages 1–23.

ICSOFT 2025 - 20th International Conference on Software Technologies

368