Background

F1000Research

2046-1402

F1000 Research Limited

London, UK

10.12688/f1000research.168964.2

Software Tool Article

Articles

MOODMIND: A Pilot Feasibility Study of Artificial Intelligence for Major Depressive Disorder Screening in Tuberculosis Patients

[version 2; peer review: 1 approved, 2 approved with reservations]

Wijayanti

Erlina

Conceptualization Data Curation Funding Acquisition Methodology Supervision Writing – Original Draft Preparation Writing – Review & Editing https://orcid.org/0000-0001-8081-9902 a 1 Abror

Ammar

Software Writing – Original Draft Preparation 2 Rachmawati

Ummi Azizah

Conceptualization Methodology Validation https://orcid.org/0000-0002-4610-3350 2 Agustina

Citra Fitri

Conceptualization Methodology Validation 3 Umniyati

Helwiah

Investigation Project Administration 4 Munti

Diana Batara

Investigation Resources 1 Rahmat

Exir Najib

Investigation Resources 5 Diansyah

Athoillah Ahkam

Investigation Resources 5 1Family Medicine Primary Care Study Program, Faculty of Medicine, Yarsi University, Central Jakarta, Jakarta, Indonesia 2Faculty of Information Technology, Yarsi University, Jakarta, Indonesia 3Department of Psychiatry, Faculty of Medicine, Yarsi University, Jakarta, Indonesia 4Faculty of Dentistry, YARSI University, Jakarta, Indonesia 5Faculty of Medicine, Yarsi University, Jakarta, Indonesia

a erlina.wijayanti@yarsi.ac.id

No competing interests were disclosed.

19 3 2026

2025

1079

6 3 2026

2026

This is an open access article distributed under the terms of the Creative Commons Attribution Licence, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Background

Major Depressive Disorder (MDD) can occur in patients with tuberculosis. The purpose of this research was to develop an early detection system for MDD and conduct an accuracy test.

Methods

The MOODMIND application uses Natural Language Processing (NLP) with sentiment analysis techniques. MOODMIND offers both speech and text options and is available in Indonesian/English. The screening results were compared with physician clinical interview. Single blinding was used so that doctor was unaware of the application test.

Results

The app asks open- and closed-ended questions for MDD identification based on the DSM-5. The test results were divided into non-depressive (none or at-risk) and suspected depression groups. Among the 21 subjects, MOODMIND showed 67% (95% CI: 9.4–99.2%) sensitivity and 100% (95% CI: 81.5–100%) specificity.

Conclusions

MOODMIND demonstrated accuracy results in pilot study but required advanced research with more sample and diverse settings. Ease is advantageous because the steps are simple, but it can be improved by adding words related to depression in the lexicon adjustment for increasing diagnostic performance.

Artificial intelligence depression detection tuberculosis Natural Language Processing 

YARSI University

1365/REK/PN.00/VII/2024

This research was supported by a grant from YARSI University (number 1365/REK/PN.00/VII/2024).

The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Revised Amendments from Version 1

We have carefully revised the manuscript in response to the reviewers’ feedback and made several important improvements. First, the title has been adjusted to Pilot Feasibility Study. This change ensures that the study is positioned appropriately as an early-stage evaluation rather than a definitive diagnostic validation. Second, 95% confidence intervals have been added to the sensitivity and specificity results. We also clarified that structured diagnostic interviews such as SCID or MINI were not used as the reference standard. Instead, physician clinical interviews were employed, and this is acknowledged as a methodological limitation. In addition, expert validation was conducted prior to pilot testing, involving primary care physicians and IT specialists. The rationale for the selected risk thresholds has also been clarified. We further added a statement acknowledging that suicidality screening is not yet integrated into the current version of MOODMIND, recognizing this as an important area for future development to ensure patient safety. Finally, claims regarding performance and applicability have been made more cautious, emphasizing that these findings are preliminary and require validation in larger and more diverse populations.

1. Introduction

Tuberculosis (TB) is a chronic infectious disease that requires at least 6 months of therapy. Psychiatric conditions are important because patients with TB can experience social stigma, worries about their illness, or difficulties during treatment. Depression has a strong effect on negative outcomes. ¹ Individuals who undergo treatment with second- and third-line medications are at a greater risk of stigma and depression. ²

Depression also affects the immune system by lowering CD3, CD4, C8, and lymphocyte. ³ Low serum anti-inflammatory cytokine levels are observed in patients with Major Depressive Disorder (MDD)-TB. Recognition of MDD in patients with TB will be more appropriate for diagnosis, treatment, and prognosis. ⁴

Afaq et al. (2023) reported that 35% of TB patients diagnosed with depression in varying levels. ⁵ Researchs in Indonesia stated that the proportion of depression in TB patients is 5.38% ⁶ and in MDR TB (Multidrug Resistance Tuberculosis) of 68.3% consists of mild, moderate, and severe. ⁷ The integration of mental health services in the management of TB patients still faces obstacles, namely the lack of patient knowledge about depression. ⁸ Patients feel unnecessary or even reluctant to have a depression screening because they are worried about receiving a double stigma. The obstacle experienced by officers is limited time in service. ^{5,
9}

The prevalence of major depression is 322 million worldwide ¹⁰ and some patients do not seek help. Major depression has the potential to lead to suicide. Questionnaires and screening tools have been developed, but most use closed-ended questions, such as the Mental Health Screening Tool for Depressive Disorders (MHS:D). ¹¹

Zotova et al. (2024), researched on the use of Patient Health Questionnaire-9 (PHQ-9) and stated that respondents’ understanding of the PHQ-9 question is sometimes incorrect, one of which is due to different cultures. ¹² The Beck Depression Inventory-Second Edition (BDI-II) uses longer questions. ⁸ PHQ-9 and BDI-II have been validated in a wide range of populations, but limited patient involvement and understanding of the questions. Obstacles can occur in individuals with low literacy, or in diseases (e.g. TB) that are susceptible to stigma. A dialogue-based digital approach can be more interactive and adapted to the context of the disease. The conversation-based screening method is expected to be more convenient and accepted by users in multicultural background.

Natural Language Processing (NLP) is an artificial intelligence capable of analyzing and interpreting words. ¹³ NLP can be used remotely for the real-time detection of depression. Studies have built systems with NLP to analyze the signs of depression based on comments on social media, such as mental health. The researchers compared mental health with the PHQ-9 to determine the accuracy of the system. ¹⁴

The NLP techniques used include sentiment analysis, linguistic markers, word embedding, convolutional neural networks, recurrent neural networks, and large language models. Sentiment analysis examines the tone of emotions in a text, referring to depression if a negative language is identified. ¹⁵

Based on the above description, a web-based application was built to screen for MDD using sentiment analysis. The software provides an alternative with open-ended questions on the two key symptoms for the diagnosis of major depression in both Indonesian and English. Through early detection, it is hoped that depression can be treated immediately and that this will increase the chances of successful treatment.

2. Methods

A. MOODMIND development

The project is part of an effort to examine tuberculosis patients holistically by developing AI-based tools for detecting MDD.

1. Ethical considerations

The ethics committee of YARSI University reviewed the ethical clearance number 114/KEP-UY/EA.20/III/2025.

2. Implementation

MDD is diagnosed if it meets the criteria of five or more symptoms (there is at least one symptom point a or b) for at least two weeks. ¹⁶ Figures 1 and 2 illustrate the concept of MOODMIND, respectively.

Figure 1. MOODMIND application concept for Major Depressive Disorder (MDD) screening. Figure 2. Conceptual framework for MOODMIND application development. 3. Operation

The software can be accessed via the following link: https://moodmind-two.vercel.app/.

3.1 Technologies

MOODMIND used Next.js for Frontend Framework, Tailwind CSS, and Web Speech API for Speech Recognition. The Programming Languages are TypeScript and JavaScript.

3.2 Main components

VoiceChat.tsx manages the voice input, transcripts, and conversation flow control. UseSpeech.ts for customizing hooks to control speech recognition status. The scripts provide questions and response scripts.

3.3 Depression detection methodology

The detection approach was based on several text-based indicators derived from voice transcription, namely, language patterns and depression-related keywords.

3.4 User experience flow

Users open the web-based application and answer system questions using voice or text. The system processes the transcription using sentiment analysis. The results of the analysis are displayed in visual and narrative forms.

3.5 Adaptation for tuberculosis

MOODMIND was adapted with a custom sentiment dictionary, focusing on common terms in Bahasa Indonesia that were reported by patients with TB when experiencing emotional distress.

3.6 Implementation details in sentiment analysis integration

As part of its natural language processing features, this system is equipped with a sentiment analysis module to evaluate the emotions contained in voice recognition transcripts. Sentiment analysis aimed to identify the emotional orientation (positive, negative, or neutral) of a statement, which, in this context, was used to detect indications of mood and enthusiasm in patients. Sentiment analysis was performed using the sentiment library, an open-source JavaScript library that supports lexicon-based analysis.

Lexicon adjustments for Indonesian

By default, a sentiment library supports the English language. To support Indonesians, a special dictionary (lexicon), consisting of a list of words and their sentiment scores, was identified.

This list of words was based on commonly used terminology to express negative emotional states, and was obtained through discussions between research members ( Figure 3).

Figure 3. Special dictionary related to depression in Indonesian.

Sentiment analysis process

After the user provides voice input, which is then transcribed into text, the system performs sentiment analysis of the text. The following functions were used to perform the analysis ( Figure 4). The getSentiment function accepts three parameters: the transcribed text, the sentiment dictionary, and the language code (“id” for Indonesian or “en” for English). If the selected language was Indonesian, the library was registered using a specially compiled dictionary.

Figure 4. Sentiment analysis process in MOODMIND.

Analysis results

The result object returned by the analysis () function contains several attributes that provide an overview of the emotional content of the analyzed text, including the score of text sentiment (positive, negative, or neutral), comparative (the normalized score value relative to the number of tokens), tokens resulting from text segmentation, words identified as having sentiment meaning, and positive/negative words recognized in the text.

By integrating this sentiment analysis, the system automatically detects emotional indicators and provides additional data for depression-screening processes. If negative sentiments related to feelings or interests are found in the last two weeks, then it is followed by closed questions.

Classification Rules

Given the pilot nature of this study, the threshold is not statistically optimized but is intended to detect MDD with clinical logic. The threshold values used for risk categorization are derived from clinical references (DSM-5).

Expert Validation

Prior to the pilot implementation, MOODMIND was reviewed through expert validation. A psychiatrist and two primary care physicians evaluated the clinical relevance of the question flow, mapping the symptoms to the DSM-5 criteria, and categorization of screening results. An information technology expert assesses technical implementations, including speech-to-text processing and lexicon integration.

B. Accuracy test

Quantitative research was carried out with a cross-sectional design and aimed at testing the accuracy of MOODMIND. The research population was drug-sensitive TB patients accompanied by YARSI TB Care cadres. The inclusion criteria were patients aged 17-65 years, had undergone TB treatment for more than 1 month, and were willing to be the subject of the study. Exclusion criteria include patients who could not be contacted and had incomplete data.

Informed Consent was carried out in writing using an electronic questionnaire. Parents or guardians would be asked for written consent (using an electronic questionnaire) for patients who are 17 years old. The samples were taken by purposive sampling in the May-July 2025.

Data collection was obtained by interview, comparing the results of detection with MOODMIND physician clinical interview. The standard reference for MDD in this study was a clinical interview by a physician using diagnostic criteria based on the DSM-5. ¹⁶ The doctor asked 9 questions systematically consisting of MDD symptoms (2 core symptoms and 7 additional symptoms), then classified as MDD if there were at least 5 symptoms (at least 1 core symptom accompanied by additional symptoms). The interview was conducted for 10-15 minutes (A list of questions is available in the Data Availability section link).

Structured interviews such as SCID (Structured Clinical Interview for DSM Disorders) or MINI (Mini International Neuropsychiatric Interview) were not used, this was a methodological limitation. Univariate analysis using Microsoft Excel to calculate sensitivity, specificity, positive predictive value, and negative predictive value. Single blinding was done to the doctor so that she did not know the results of detection with MOODMIND.

4. Results 4.1 Expert validation and usability feedback

Expert feedback results in improvements in audio clarity and transcription synchronization. Improvements were made to the naturalness of the sound. After revision, the system was considered stable and easy to use for pilot testing.

4.2 Use cases

MOODMIND users can select the languages (English and Indonesian) ( Figure 5a). Users can choose either the written or voice mode of conversation ( Figure 5b). Users’ answers were categorized into 3, namely not depressed (score/symptom = 0), at risk of depression (score/symptom = 1-4), and suspected depression (score/symptom ≥ 5) ( Figure 5c). The word “Suspected depression” was used because the diagnosis by the doctor must be carried out and the patient should receive the necessary consultation. The role of a doctor/officer cannot be replaced by AI because of empathy and direct interaction with a human being. MOODMIND does not currently include a suicidality referral flow. In clinical implementation, it is necessary to ensure patient safety. Figure 6 shows the MOODMIND System Processing Pipeline.

Figure 5. a. Front page of MOODMIND. b. Conversation flow in MOODMIND. c. Result of test in MOODMIND. Figure 6. MOODMIND system processing pipeline. 4.3 Accuracy test

We conducted tests on 21 patients with TB in Central Jakarta between May and July 2025 ( Figure 7). The average age of patients was 41.4 years with an age range of 19-64 years. The patient was guided by the researcher when using MOODMIND, whereas the doctor was blinded and did not know the results of the software detection.

Figure 7. The flowchart of patients recruitment.

Table 1 shows a comparison of MOODMIND detection with the physician clinical interview, while Table 2 shows the accuracy level of the software. The sensitivity was 67% (95% CI: 9.4–99.2%), and specificity was 100% (95% CI: 81.5–100%), reflecting the limited precision due to small sample size. The small number of samples was considered in interpreting the results of sensitivity and specificity.

Table 1. MOODMIND screening and Physician Clinical Interview results.

AI MOODMIND	Autoanamnesis		Total
AI MOODMIND	Negative	Positive	Total
Negative	18	1	19
Positive	0	2	2
Total	18	3	21

Table 2. Analysis of MOODMIND screening results on Physician Clinical Interview.

Test	Percentage
Sensitivity	67%
Specificity	100%
Positive predictive value	100%
Negative predictive value	95%

5. Discussion

The MOODMIND application was equipped with sentiment analysis by searching for keywords and analyzing sentiments in Indonesian. The Lexicon technique is used to make a list of words and score sentiments for each word. ¹⁷ Other research has identified the keywords depression, symbols, and expressions through social media. ^{18,
19} Existing depression detection systems/applications such as “Mental Care” which asked 21 questions to respondents, ²⁰ Multi-Gated LeakyReLU processed depressive language using CNN, ²¹ while another study analyzed expressions that did not directly use specific words. ²²

The results of this pilot study obtained a sensitivity of 67%. MOODMIND can be used as an initials screening tool in a variety of settings, not only in healthcare but also in the community. However, for negative cases with high risk, it is recommended to continue undergoing further clinical assessments. High-risk MDD-TB patients include MDR TB, ²³ have comorbidities, ²⁴ and get stigmatized. ²⁵

Artificial intelligence usually requires the ability of the user. ²⁶ However, MOODMIND is very easy to operate, which can reduce issues related to human resources. The main requirements are a device and an internet connection. This tool is an inspiration for the development of similar types in other countries according to the local language, minimizing the gap between the detected cases and the actual number of cases. The variation of words related to depression still adjusts to the current condition, so it must be continuously updated to increase sensitivity from time to time. PHQ-9 is a questionnaire that has been tested to have high validity. However, conversation-based MOODMIND with open-ended questions can offer advantages compared to standard questionnaires that are underutilized.

More sample research is needed to determine the accuracy of MOODMIND in a real-world setting. The absence of structured diagnostic interviews (e.g., SCID or MINI) limits diagnostics to reference standards. In addition, bridging the results of screening to electronic medical records can be a useful alternative for monitoring the mental health of patients with chronic diseases such as tuberculosis. Advanced versions should incorporate suicide risk screening and referral mechanisms.

6. Conclusion

MOODMIND, an artificial intelligence based on Natural Language Processing, can be used as an MDD detection tool. The diagnostic performance in this pilot study still required further exploration accompanied by research with a larger sample. This tool supports mental health monitoring but does not replace the role of doctors. This could also be an idea for AI development in some countries to detect MDD as early as possible.

Software availability

Source code available from: https://github.com/incrementalstudios/mood-mind

Archived software available from: https://doi.org/10.5281/zenodo.16793110 ²⁷

License: MIT License

Data availability

The dataset as the basis for the accuracy test findings can be accessed at the link: https://doi.org/10.5281/zenodo.17114938. ²⁸ We also include the approval sheets and interview guides in the link.

Data are available under the terms of the Creative Commons Zero v1.0 Universal

Acknowledgements

We thank the YARSI Foundation for supporting this study.

References 1

Ruiz-Grosso

Cachay

De La Flor

: Association between tuberculosis and depression on negative outcomes of tuberculosis treatment: A systematic review and meta-analysis. PLoS One. 2020;15(1):e0227472–e0227413. 31923280

10.1371/journal.pone.0227472

PMC6953784

Sweetland

Kritski

Oquendo

: Addressing the tuberculosis–depression syndemic to end the tuberculosis epidemic. Int. J. Tuberc. Lung Dis. 2017;21(8):852–861. 28786792

10.5588/ijtld.16.0584

PMC5759333

Liu

Bai

Ren

: Association between depression or anxiety symptoms and immune-inflammatory characteristics in in-patients with tuberculosis: A cross-sectional study. Front. Psych. 2022;13.

Alvarez-Sekely

Lopez-Bago

Báez-Saldaña

: Major Depressive Disorder and Pulmonary Tuberculosis Comorbidity Exacerbates Proinflammatory Immune Response—A Preliminary Study. Pathogens. 2023;12(3). 36986283

10.3390/pathogens12030361

PMC10059645

Afaq

Ayub

Faisal

: Depression care integration in tuberculosis services: A feasibility assessment in Pakistan. Heal. Expect. 2024;27(1):e13985. 39102704

10.1111/hex.13985

PMC10849063

Regina

Nasution

Syukri

: Depression-Related Factors In Pulmonary Tuberculosis Patients: Secondary Analysis Of The Indonesian Health Survey 2023. Period Epidemiol. J. 2026;14(1):62–71. 10.20473/jbe.V14I12026.62-71

Susanto

Widysanto

Cipta

Tanara

Wirawan

Kosim

: Anxiety and depression level of patients with multidrug-resistant tuberculosis (MDR-TB) in two hospitals in Banten province, Indonesia. Dialogues Heal[Internet].2023;2(January):100115. 10.1016/j.dialog.2023.100115

Todowede

Afaq

Adhikary

: Barriers and facilitators to integrating depression care in tuberculosis services in South Asia: a multi-country qualitative study. BMC Health Serv. Res. 2023;23(1):1–13. 10.1186/s12913-023-09783-z

Reyes Rodríguez

Cuellar-Pompa

Rodríguez Novo

: Validated Tools for Assessing Anxiety and Depression in Nurses: A Systematic Review. Int. J. Environ. Res. Public Health. 2025;22(11):1–17. 10.3390/ijerph22111714

World Health Organization: Depression and Other Common Mental Disorders Global Health Estimates. 2017.

Park

Yoon

Cho

: Final validation of the mental health screening tool for depressive disorders: A brief online and offline screening tool for major depressive disorder. Front. Psychol. 2022;13(October):1–12.

Zotova

Watnick

Ajeh

: Understanding depression and the PHQ-9 items among people living with HIV: A multiple methods qualitative study in Yaoundé, Cameroon. SSM - Ment. Heal. 2024;6:100353. 10.1016/j.ssmmh.2024.100353

Tyagi

Bhushan

: Demystifying the Role of Natural Language Processing (NLP) in Smart City Applications: Background, Motivation, Recent Advances, and Future Research Directions. Wirel. Pers. Commun. 2023;130:857–908. 37168438

10.1007/s11277-023-10312-8

PMC10019426

Salas-Zárate

Alor-Hernández

Paredes-Valverde

: Mental-Health: An NLP-Based System for Detecting Depression Levels through User Comments on Twitter (X). Mathematics. 2024;12(13):1–30.

Teferra

Rueda

Pang

: Screening for Depression Using Natural Language Processing: Literature Review. Interact. J. Med. Res. 2024;13:e55067. 39496145

10.2196/55067

PMC11574504

American Psychiatric Association: Diagnostic and Statistical Manual of Mental Disorders (DSM-5). Encyclopedia of Applied Psychology, Three-Volume Set. 2013;1:160–168. Reference Source

Herdiansyah

Roestam

Kuhon

: Their post tell the truth: Detecting social media users mental health issues with sentiment analysis. Procedia Comput. Sci. 2023;216:691–697. 10.1016/j.procs.2022.12.185

Cha

Kim

Park

: A lexicon-based approach to examine depression detection in social media: the case of Twitter and university community. Humanit. Soc. Sci. Commun. 2022;9(1):1–10.

Huang

: Automatic construction of a depression-domain lexicon based on microblogs: Text mining study. JMIR Med. Informatics. 2020;8(6):1–17.

Gustiadi

Lazuardi

Kedokteran

: PENGEMBANGAN APLIKASI SKRINING KESEHATAN MENTAL. J. Inf. Kesehat Indones. 2024;10(2):67–77.

Rao

Zhang

: MGL-CNN: A Hierarchical Posts Representations Model for Identifying Depressed Individuals in Online Forums. IEEE Access. 2020;8:32395–32403. 10.1109/ACCESS.2020.2973737

Chiong

Budhi

Dhakal

: A textual-based featuring approach for depression detection using machine learning classifiers and social media texts. Comput. Biol. Med. 2021;135:104499. 34174760

10.1016/j.compbiomed.2021.104499

Shrestha

Joshi

Bhattarai

: Prevalence and risk factors of depression in patients with drug-resistant tuberculosis in Nepal: A cross-sectional study. J. Clin. Tuberc. Other Mycobact. Dis [Internet]. 2020;21:100200. 10.1016/j.jctube.2020.100200

Azam

Fibriana

Indrawati

: Prevalence and Determinant of Depression among Multi-Drug Resistance Tuberculosis: Study in Dr. Kariadi General Hospital. J. Respirologi Indones. 2020;40(2):88–96. 10.36497/jri.v40i2.106

Shen

Zong

Liu

: Risk Factors for Depression in Tuberculosis Patients: A Meta-Analysis. Neuropsychiatr. Dis. Treat. 2022;18(March):847–866. 35431546

10.2147/NDT.S347579

PMC9012238

Zafar

Fakhare Alam

Vivas

: The Role of Artificial Intelligence in Identifying Depression and Anxiety: A Comprehensive Literature Review. Cureus. 2024;16(3).

Abror

Wijayanti

Rachmawati

: Moodmind. Zenodo. 2025. 10.5281/zenodo.16793110

Wijayanti

: Data Availability MOODMIND. Zenodo. 2025. 10.5281/zenodo.17114938

10.5256/f1000research.197138.r468964

Reviewer response for version 2

Mansoor

Masab

1 Referee https://orcid.org/0009-0007-4501-7016 1Edward Via College of Osteopathic Medicine, Blacksburg, Virginia, USA

Competing interests: No competing interests were disclosed.

6 4 2026

2026

This is an open access peer review report distributed under the terms of the Creative Commons Attribution Licence, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

recommendation

approve

Thank you for your revisions. I endorse this paper for Indexing.

Are the conclusions about the tool and its performance adequately supported by the findings presented in the article?

Is the rationale for developing the new software tool clearly explained?

Partly

Is the description of the software tool technically sound?

Are sufficient details of the code, methods and analysis (if applicable) provided to allow replication of the software development and its use by others?

Is sufficient information provided to allow interpretation of the expected output datasets and any results generated using the tool?

Partly

Reviewer Expertise:

Machine Learning, artificial intelligence, health informatics, Large language models, Machine vision

I confirm that I have read this submission and believe that I have an appropriate level of expertise to confirm that it is of an acceptable scientific standard.

Wijayanti

Erlina

Family Medicine Primary Care, YARSI University, Jakarta, Indonesia

Competing interests: No competing interests were disclosed.

6 4 2026

I sincerely appreciate your very valuable feedback. Thank you very much.

10.5256/f1000research.186213.r447567

Reviewer response for version 1

Al-Hindy

Hayder

1 Referee https://orcid.org/0000-0001-6232-8501 1College of Pharmacology, University of Babylon (Ringgold ID: 125654), Babylon Governorate, Babylon Governorate, Iraq

Competing interests: No competing interests were disclosed.

29 1 2026

2026

recommendation

approve-with-reservations

F1000Research Software Tool Peer Review: Full Report

Article Summary

This software tool article presents MOODMIND, a web-based application for screening Major Depressive Disorder (MDD) in tuberculosis (TB) patients using Natural Language Processing (NLP), specifically lexicon-based sentiment analysis on voice or text inputs in Indonesian and English. The tool poses DSM-5-derived open- and closed-ended questions, processes responses via a custom Indonesian sentiment dictionary and the JavaScript 'sentiment' library, and categorizes users as not depressed (score 0), at-risk (1-4), or suspected depression (≥5). A pilot accuracy test on 21 drug-sensitive TB patients (mean age 41.4 years) compared MOODMIND outputs to blinded physician anamnesis, yielding 67% sensitivity, 100% specificity, 100% PPV, and 95% NPV. Source code and data are openly available via GitHub and Zenodo. The manuscript highlights the tool's ease-of-use and potential for early detection in resource-limited settings but has undergone prior open peer reviews (one "approved with reservations," one "not approved").

Questionnaire Responses with Detailed Explanations

1. Is the rationale for developing the new software tool clearly explained? Partly

The introduction establishes TB-MDD comorbidity risks, stigma, immune effects, and limitations of closed-ended questionnaires, motivating NLP for natural responses. However, it lacks quantification of MDD prevalence in Indonesian TB cohorts, explicit shortcomings of validated tools (e.g., PHQ-9's rigidity in low-literacy groups), and evidence of screening gaps this addresses.

To address: Add epidemiological data (e.g., local MDD-TB rates), a table comparing MOODMIND to PHQ-9/GAD-7 in TB contexts, and barriers like digital literacy or stigma. Not essential for soundness but strengthens justification.

2. Is the description of the software tool technically sound? Partly

The architecture (Next.js, Web Speech API, custom lexicon) is outlined, with TB-specific adaptations and sentiment scoring described. However, lexicon-based analysis is outdated (vs. BERT/transformers); speech recognition lacks dialect validation; scoring logic (sentiment-to-DSM-5 symptom mapping) is opaque; neutral sentiment handling undefined.

To address: Provide algorithmic flowchart (input → transcription → lexicon scoring → thresholds → categorization); validate lexicon against clinical corpora with inter-rater stats; test Web Speech API accuracy on Indonesian TB dialects; specify all thresholds empirically. Essential: Scoring/threshold details for technical validity.

3. Are sufficient details of the code, methods, and analysis (if applicable) provided to allow replication of the software development and its use by others? No

GitHub link aids access, but is missing: full get Sentiment code; lexicon weights; dependency versions/commit hash; browser/system requirements; physician interview guide; sample size/power calculation; STARD flowchart; handling of incompletes/technical failures. Purposive sampling vague (no exclusions for psychosis/substances).

To address (all essential for soundness):

Supplementary file with complete lexicon (words + weights), full code snippets, version specs.

Detailed methods: physician protocol (SCID/MINI preferred over "anamnesis"), exclusions, recruitment flowchart.

Stats: 95% CIs, F1/AUC in tables; acknowledge n=21 pilot.

These gaps prevent replication; must be fixed.

4. Is sufficient information provided to allow interpretation of the expected output datasets and any results generated using the tool? Partly

Figures 5a-c illustrate user flow/outputs clearly, with appropriate "suspected depression" disclaimer. However, no clinical action guidance (e.g., "at-risk" referrals); underlying scores/symptoms not shown; no rescreening protocols; false-negative risks (33% with 67% sensitivity) unquantified; suicidality absent.

To address: Add output interpretation guide (e.g., "at-risk → prompt PHQ-9"); symptom-level feedback; suicidality question with crisis links. Include workflow integration (e.g., TB clinic timing). Essential: Suicidality/safety protocols to prevent harm.

5. Are the conclusions about the tool and its performance adequately supported by the findings presented in the article? No

"Adequate accuracy" overstates n=21 pilot (only 3 MDD cases; specificity CI likely 82-100%); no PHQ-9 benchmark; ignores false-negatives' TB adherence risks. Generalizability limited (supervised use, single-center).

To address (all essential):

Revise to "preliminary pilot accuracy requiring validation."

Add CIs (e.g., sensitivity 9-99%); compare to PHQ-9 (88% sensitivity).

Explicit limitations section: small n, overfitting risk, no external cohort.

Tone down: "Proof-of-concept for larger studies." Overclaims undermine credibility; must be corrected.

Strengths

Innovative open-ended NLP for TB-MDD screening; bilingual/open-source; ethical (blinding, consent); addresses real comorbidity need in low-resource Indonesia.

Overall Concerns and Minor Issues

Small sample yields unreliable metrics; outdated lexicon risks bias; safety gaps (no suicidality); typos ("Accuration," inconsistent "autoanamnesis"); vague English lexicon testing. Prior reviews echo these—manuscript needs tracked revisions.

Points That Must Be Addressed for Scientific Soundness

Technical replication: Full lexicon/code/thresholds (supplementary).

Validation rigor: CIs/F1; structured gold-standard (SCID); recruitment flowchart.

Safety: Suicidality screening/referrals.

Tempered claims: "Pilot" framing; PHQ-9 comparisons; limitations section.

Methods clarity: Sampling exclusions, speech validation, scoring algorithm.

Without these, the article risks misleading clinical adoption.

Final Recommendation

Approved with reservations. Promising pilot but requires major revisions above for soundness. Re-review post-changes recommended.

Are the conclusions about the tool and its performance adequately supported by the findings presented in the article?

Is the rationale for developing the new software tool clearly explained?

Partly

Is the description of the software tool technically sound?

Partly

Are sufficient details of the code, methods and analysis (if applicable) provided to allow replication of the software development and its use by others?

Is sufficient information provided to allow interpretation of the expected output datasets and any results generated using the tool?

Partly

Reviewer Expertise:

General and cardiac physiology, genetics, Pediatrics,

I confirm that I have read this submission and believe that I have an appropriate level of expertise to confirm that it is of an acceptable scientific standard, however I have significant reservations, as outlined above.

10.5256/f1000research.186213.r435547

Reviewer response for version 1

Espinola

Caroline W.

1 Referee https://orcid.org/0000-0002-5479-8914 1McMaster University, Hamilton, Canada

Competing interests: No competing interests were disclosed.

13 1 2026

2026

recommendation

approve-with-reservations

This manuscript describes an AI tool that uses natural language processing from questions derived from the DSM-5 for a screening of depression in a convenience sample of 21 patients with tuberculosis. They report an sensitivity of 67% and specificity of 100%.

The Abstract is generic and unclear. Needs a more detailed description of the background, rationale, population studied. A clear description of the methods, including model used, is also missing. No information on the sample is provided. In the Methods, the reader would appreciate a more detailed description of the MOODMIND tool. How was NLP obtained? Which sentiment analysis techniques were used? Results are poorly reported, e.g., F1 score, AUC are not reported. Conclusion does not mention TB (the condition being studied) and makes strong assumption without justifying it. What is "Ease" in the conclusion?

Introduction: I would add more information on the standard treatment for TB. I would also recommend adding more details, such as magnitude of risk of MDD associated with TB compared with the general population or risk of depression in people with TB receiving second and third-line therapies. Additionally, what is the impact of these changes in immune system associated with MDD on TB treatment outcomes? In the fifth paragraph, CNN and RNN are ML architetures that are used in NLP datasets, not NLP techniques by themselves. I suggest revising.

Methods: More information would be appreciated on the MOODMIND model and what it entails. What data was used for input - just the open-ended questions or also the other 7 closed-ended questions? Considering a diagnosis of depression requires at least 5 criteria, this could have a significant impact on the content generated, and consequently, the prediction ability of the NLP model.

- 3.4 User experience flow: In the voice converted to text, what was the transcription software used -your own or a third-party one? In case of using your own, was it validated prior to this research? Where was data stored (this could be a privacy concern). Did you run any analysis on potential mistakes in the transcription software?

3.5 Adaptation for tuberculosis: MOODMIND was adapted with a custom sentiment dictionary, focusing on common terms in Bahasa Indonesia that were reported by patients with TB when experiencing emotional distress. - How are these terms different between TB and the general population. A non-Indonesian reader would appreciate more information. What about its use in English? What percentage of the sample responded in English? Was there any difference in accuracy between the English speaking and Indonesian speaking samples?

- Figure 4: Code of the function getSentiment is not provided. Unsure of what it does. I suggest either removing this figure or adding the code for the function

Results:

- 4.1 Use cases: First sentence should be moved to Methods. " The role of a doctor/officer cannot be replaced by AI because of empathy and direct interaction with a human being." - Not only because of these. The clinical interview is the gold standard for the diagnosis of depression. Suggest revising.

- 4.2 Accuration test: Since you used mean age, was the SD of age of the sample? How many patients were approached but either did not consent of were excluded. A detailed description of the eligibility criteria is needed.

Results also need to be provided in plain text, not only in tables

Discussion: The discussion is confusing and does not summarize the goals and results of the study. Furthermore, I understand the greatest contribution of this tool would be as a screening tool for MDD. A sensitivity of 67% is modest for screening tool, which should be able to detect true positives.This consideration should be included in the discussion. The discussion should contemplate how this tool could help detect TB patients at a higher risk for depression, which would warrant a detailed clinical assessment. Are these patients automatically flagged for the healthcare team or do they need to seek help on their own?

Are the conclusions about the tool and its performance adequately supported by the findings presented in the article?

Partly

Is the rationale for developing the new software tool clearly explained?

Is the description of the software tool technically sound?

Partly

Are sufficient details of the code, methods and analysis (if applicable) provided to allow replication of the software development and its use by others?

Is sufficient information provided to allow interpretation of the expected output datasets and any results generated using the tool?

Partly

Reviewer Expertise:

Major depressive disorder, digital mental health, brain stimulation

References 1

: Addressing the tuberculosis–depression syndemic to end the tuberculosis epidemic. The International Journal of Tuberculosis and Lung Disease .2017;21(8) : 10.5588/ijtld.16.0584 852-861

10.5588/ijtld.16.0584

10.5256/f1000research.186213.r435555

Reviewer response for version 1

Mansoor

Masab

1 Referee https://orcid.org/0009-0007-4501-7016 1Edward Via College of Osteopathic Medicine, Blacksburg, Virginia, USA

Competing interests: No competing interests were disclosed.

30 12 2025

2025

recommendation

reject

PEER REVIEW REPORT

SUMMARY

This manuscript presents MOODMIND, a web-based Natural Language Processing (NLP) application designed to screen for Major Depressive Disorder (MDD) in tuberculosis patients. The tool employs lexicon-based sentiment analysis on both voice and text inputs in Indonesian and English, utilizing DSM-5 criteria. A pilot validation study with 21 TB patients demonstrated 67% sensitivity and 100% specificity compared to physician assessment.

While the authors address a clinically relevant problem and present an accessible tool, significant methodological and technical limitations require substantial revision before indexing.

DETAILED RESPONSES TO REVIEW QUESTIONS

1. Is the rationale for developing the new software tool clearly explained?

Answer: PARTLY

Strengths:

The co-morbidity between TB and MDD is well-established in the introduction

The need for accessible screening tools is apparent

The rationale for using open-ended questions is mentioned

Weaknesses:

Lacks justification for why existing validated tools (PHQ-9, GAD-7, BDI-II) are inadequate for this population

Does not explain why sentiment analysis is superior to structured questionnaires

Insufficient epidemiological data on MDD prevalence specifically in Indonesian TB patients

No discussion of barriers to current screening practices that this tool addresses

Missing cost-effectiveness or accessibility arguments compared to existing digital mental health tools

Recommendations:

Provide systematic comparison of existing MDD screening tools and their limitations in TB populations

Clarify specific advantages of MOODMIND over validated instruments

Include data on mental health service gaps in the target population

2. Is the description of the software tool technically sound?

Answer: NO

Critical Technical Deficiencies:

A. Sentiment Analysis Methodology:

Oversimplified approach: Lexicon-based sentiment analysis is outdated; modern approaches utilize transformer-based models (BERT, GPT) or at minimum, more sophisticated machine learning classifiers

No description of sentiment score thresholds for categorization into non-depressed/at-risk/suspected depression

Lack of validation for the custom Indonesian lexicon (Figure 3 shows word list but no validation process)

No explanation of how sentiment scores correlate with DSM-5 diagnostic criteria for MDD

B. Algorithm Description:

Figure 4 shows code snippet but insufficient detail for replication:

What constitutes a "negative sentiment" threshold?

How are comparative scores normalized?

What happens with neutral sentiments?

Missing information on decision tree logic for transitioning from open to closed questions

No description of how the algorithm weighs different DSM-5 symptoms

C. Speech Recognition:

Uses Web Speech API but no validation of transcription accuracy for Indonesian language

No discussion of handling dialectal variations or accented speech

Error handling procedures for misrecognition not addressed

D. Scoring System:

Critical gap: The manuscript states results are categorized as score 0, 1-4, or ≥5, but never explains:

How these numeric scores are generated from sentiment analysis

What each point represents (symptom count? Severity weighting?)

How DSM-5's requirement of "5 or more symptoms with at least 1 core symptom" maps to the scoring

Recommendations:

Provide detailed algorithmic flowchart from input to classification

Specify all threshold values and their empirical basis

Consider upgrading to modern NLP architectures

Validate Indonesian lexicon against clinical datasets

Provide comprehensive technical documentation in supplementary materials

3. Are sufficient details provided for replication?

Answer: NO

Missing Critical Information:

A. Lexicon Development:

Figure 3 shows word list but:

No systematic methodology for term selection

"Obtained through discussions between research members" is insufficient

No validation against clinical depression corpora

Sentiment weights/scores for each term not provided

No inter-rater agreement statistics for lexicon development

B. Software Implementation:

While GitHub repository is referenced:

Version control information missing (which commit/release tested?)

Dependency versions not specified

Deployment environment specifications absent

Browser compatibility not documented

C. Validation Protocol:

Insufficient detail on physician assessment:

What specific questions did physicians ask?

How long were clinical interviews?

What documentation was completed?

Single physician or multiple (inter-rater reliability)?

Sampling procedure: "Purposive sampling" requires more specificity

Inclusion/exclusion criteria need expansion (e.g., cognitive impairment, substance use, psychotic disorders)

D. Statistical Analysis:

No sample size calculation or power analysis

Confidence intervals not provided for sensitivity/specificity

Missing information on handling:

Incomplete responses

Technical failures

Participant withdrawals

Recommendations:

Publish complete lexicon with sentiment weights as supplementary data

Provide structured interview guide used by physicians

Include detailed statistical analysis plan

Add flowchart of participant recruitment and assessment

Specify all software versions and system requirements

4. Is sufficient information provided to interpret expected outputs?

Answer: PARTLY

Strengths:

Figure 5c shows example output interface

Three-category classification (not depressed/at-risk/suspected) is clear

Appropriate disclaimer about need for professional diagnosis

Weaknesses:

A. Output Interpretation:

No guidance on clinical action for each category:

What should clinicians do with "at-risk" patients?

Referral pathways not discussed

Urgency assessment absent

Score explanation missing: Users see categorical result but not underlying score or contributing factors

No feedback on specific symptoms identified

B. Clinical Utility:

Unclear integration with clinical workflow:

When in TB treatment should screening occur?

How often should rescreening happen?

Documentation recommendations?

False negative implications not discussed (with 67% sensitivity, 33% of MDD cases missed)

C. Limitations Communication:

Tool appropriately states it doesn't replace clinical diagnosis, but:

Doesn't explain to users why (especially for low-literacy populations)

Risk of delayed care if patients with negative screens don't seek help

Suicidality assessment completely absent

Recommendations:

Provide detailed clinical implementation guidelines

Include symptom-level feedback in output

Add explicit suicidality screening and immediate referral protocols

Create user education materials explaining tool limitations

Develop clinician interpretation guide with case examples

5. Are conclusions adequately supported by findings?

Answer: NO

Major Concerns:

A. Overstated Claims:

"Adequate accuracy" (67% sensitivity) is debatable for screening tool:

Missing 1 in 3 cases is problematic for MDD screening

Compare to PHQ-9: sensitivity ~88%, specificity ~88% for MDD diagnosis

No justification for why 67% is acceptable

"100% specificity" is misleading:

Based on zero false positives in sample of 18 true negatives

With 95% confidence interval, true specificity could be as low as ~82%

Overfitting likely with such small sample

B. Statistical Limitations:

Sample size (n=21) severely underpowered:

Only 3 MDD-positive cases

Cannot reliably estimate diagnostic accuracy

No subgroup analyses possible

Results not generalizable

Selection bias: Patients "accompanied by YARSI TB Care cadres" may not represent broader TB population

Verification bias: Single physician assessment without structured interview or validated scales

C. Comparative Evidence:

No comparison with validated screening tools in same population

No benchmark against PHQ-9, which is:

Free, brief (9 items)

Extensively validated in medical populations

Available in Indonesian

D. Generalizability:

Single-center study in Central Jakarta

TB patients only - MDD presentation may differ in other chronic diseases

Researcher-supervised administration - real-world performance likely lower

Indonesian language validation insufficient (no dialectal testing)

E. Missing Discussions:

False negative consequences: Untreated depression worsens TB outcomes, adherence

Screening frequency: Optimal timing during 6-month TB treatment unclear

Cost-effectiveness: Not analyzed

Implementation barriers: Internet access, device availability, digital literacy

Recommendations:

Revise conclusions to acknowledge substantial limitations

Clearly state this is a proof-of-concept pilot requiring extensive validation

Compare performance to PHQ-9 in future studies

Conduct multi-center validation with ≥300 participants

Include external validation cohort

Perform head-to-head comparison with validated instruments

ADDITIONAL MAJOR CONCERNS

Methodological Issues:

Gold Standard Inadequate:

Physician "autoanamnesis" is not validated

Should use Structured Clinical Interview for DSM-5 (SCID) or MINI

Consider including validated self-report measures as convergent validity

Blinding Incomplete:

Researchers present during MOODMIND administration could influence responses

Should be independently administered without researcher presence

Missing Data on:

TB disease characteristics (drug-sensitive only stated, but severity, treatment phase?)

Psychiatric history (first episode vs. recurrent MDD?)

Current psychotropic medications

Comorbid psychiatric conditions

Sociodemographic factors

Ethical Concerns:

Vulnerable Population:

TB patients with MDD are doubly stigmatized

Data security measures not described

Privacy protections for voice recordings unclear

Suicidality Risk:

No suicide risk assessment in the tool

Critical safety gap: MDD screening without suicide screening is dangerous

No crisis referral pathway described

Technical Concerns:

Open-Ended Question Analysis:

How are diverse, unstructured responses converted to binary symptom presence/absence?

Inter-rater reliability for human coding not established

Automated coding validation absent

Closed Question Integration:

Manuscript states both open and closed questions used

Logic for triggering closed questions not explained

Scoring methodology for combined responses unclear

SPECIFIC CORRECTIONS REQUIRED

Methods Section:

Line describing sample: Change "21 patients" to "21 patients (pilot feasibility study)"

Add: Structured clinical interview protocol used by physician

Add: Sample size justification or acknowledge as convenience sample

Add: Inter-rater reliability assessment (if multiple raters) or acknowledge single-rater limitation

Results Section:

Add: 95% confidence intervals for all diagnostic accuracy metrics

Add: Participant flowchart (STARD guidelines)

Add: Description of any technical failures or incomplete assessments

Modify Table 2: Include confidence intervals

Discussion Section:

Add: Direct comparison of 67% sensitivity to literature values for validated tools

Add: Clinical implications of 33% false negative rate

Add: Limitations section discussing:

Small sample size and wide confidence intervals

Lack of external validation

Single-center, single-assessor design

Absence of comparison to validated instruments

Add: Implementation research needs before clinical deployment

Conclusion Section:

Modify: Change "adequate accuracy" to "preliminary accuracy estimates requiring validation in larger studies"

Add: Explicit statement: "This tool requires extensive validation before clinical implementation"

Add: Specific next steps for validation research

MINOR ISSUES

Writing Quality:

Generally clear but some grammatical errors

"Autoanamnesis" - unusual term; clarify or use "clinical interview"

Inconsistent terminology (MDD vs. depression vs. major depression)

Figures:

Figure 1: Simplistic - could be removed or enhanced with algorithm specifics

Figure 2: Helpful conceptual framework

Figure 3: Shows code but insufficient explanation

Figure 4: Code snippet needs more context

Figure 5: Good interface examples but needs annotation

References:

Appropriate selection

Missing key references on digital mental health tools

Should cite PHQ-9 validation studies in TB populations (if available)

VERDICT AND RECOMMENDATIONS

Overall Assessment: MAJOR REVISIONS REQUIRED

Must Be Addressed for Scientific Soundness:

Acknowledge severe limitations of 21-patient pilot study throughout manuscript

Revise conclusions to reflect preliminary nature of findings

Add confidence intervals to all diagnostic accuracy estimates

Provide complete technical documentation sufficient for replication

Include suicide risk assessment in tool or acknowledge dangerous omission

Explain scoring algorithm in detail

Validate lexicon using established methodology

Compare to validated screening tools (add as limitation if not done)

Strongly Recommended:

Conduct adequately powered validation study (n≥300) before claiming clinical utility

Perform external validation in different TB treatment settings

Include structured clinical interview as gold standard

Assess inter-rater reliability

Upgrade NLP methodology to modern standards

Publish complete source code and lexicon with version control

Suggested Title Revision:

"MOODMIND: A Pilot Feasibility Study of Artificial Intelligence for Major Depressive Disorder Screening in Tuberculosis Patients"

Alternative Publication Path:

Given the early-stage development and small sample, authors might consider:

Repositioning as a "Software Tool Note" rather than validation study

Focus on technical description with clear acknowledgment that clinical validation is pending

Present diagnostic accuracy data as preliminary feasibility only

CONCLUSION

While MOODMIND addresses an important clinical need and demonstrates creative application of NLP to mental health screening, the manuscript requires substantial revision to meet scientific standards for a clinical validation study. The combination of outdated NLP methodology, inadequate technical description, severely underpowered validation study, and absence of critical safety features (suicide screening) precludes recommendation for approval in current form.

The authors should be commended for open-source development and bilingual implementation, but must conduct rigorous validation research before clinical deployment recommendations can be supported.

Recommendation: MAJOR REVISIONS REQUIRED before this manuscript can be considered for Indexing.

Are the conclusions about the tool and its performance adequately supported by the findings presented in the article?

Is the rationale for developing the new software tool clearly explained?

Partly

Is the description of the software tool technically sound?

Are sufficient details of the code, methods and analysis (if applicable) provided to allow replication of the software development and its use by others?

Is sufficient information provided to allow interpretation of the expected output datasets and any results generated using the tool?

Partly

Reviewer Expertise:

Machine Learning, artificial intelligence, health informatics, Large language models, Machine vision

I confirm that I have read this submission and believe that I have an appropriate level of expertise to state that I do not consider it to be of an acceptable scientific standard, for reasons outlined above.