Dialogues of delivery: a multilingual question-answer dataset for maternal healthcare in East African languages

Abstract

Objective

There is a critical scarcity of domain-specific, clinically grounded Natural Language Processing (NLP) resources for African languages. In Western Uganda, linguistic diversity creates a barrier to maternal healthcare, as mothers lack access to health information in their native languages. The objective of this dataset is to provide a high-quality, in-language medical corpus to enable the development and fine-tuning of Large Language Models (LLMs) and conversational AI tools for maternal health in resource-constrained settings.

Data description

The “Dialogues of Delivery” dataset is a multilingual, parallel corpus comprising 3,694 question-and-answer pairs presented in four languages: English, Luganda, Runyankore, and Swahili (14,800 total entries). Using facility-based convenience sampling at two health facilities in Western Uganda, primary data was collected via structured, open-ended questionnaires from 150 participants (expectant/postpartum mothers and maternal healthcare providers). The dataset underwent a rigorous forward-backward translation protocol by certified linguists and human-in-the-loop clinical validation by independent medical professionals. The dataset captures core maternal health domains, providing a culturally and clinically validated foundation for Afrocentric AI development.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13104-026-07899-z.

Objective

Maternal health outcomes in East Africa remain a pressing concern, with high mortality rates persisting in the region [1, 2]. In rural and peri-urban areas, mothers face numerous barriers to care, among which language and communication gaps are especially critical [3–5]. In linguistically diverse countries like Uganda, which has over 40 languages [6], clinical encounters frequently occur in nonnative languages. This linguistic mismatch compromises diagnostic accuracy, shared decision-making, and patient trust [7–9], ultimately elevating the risk of adverse outcomes and delayed emergency care [10, 11].

While existing mHealth platforms have improved healthcare access [12–14], they often lack interactivity and exclude non-literate populations [15]. Large Language Models (LLMs) offer promising solutions for accessible, conversational patient education. However, their effectiveness in low-resource environments is fundamentally constrained by a lack of high-quality, domain-specific training datasets in local languages [16]. Existing models remain overwhelmingly English-centric. Although multilingual African NLP initiatives are growing [17–19], there remains a severe scarcity of clinically grounded, in-language corpora specifically focused on maternal health [20].

To address this critical gap, we introduce the Dialogues of Delivery dataset: a multilingual, clinically validated question-and-answer corpus in English, Luganda, Runyankore, and Swahili. The primary objective of this dataset is to support the fine-tuning of LLMs for prenatal and postnatal care in East African languages. By providing a culturally tailored and medically accurate foundation, this work aims to enable the development of conversational AI tools that are linguistically appropriate. This resource bridges critical data gaps in low-resource settings, contributing to an equitable digital health ecosystem that empowers mothers and frontline healthcare workers with accessible information.

Data description

The dataset captures maternal health dialogues to facilitate conversational AI development in East African languages. Data was gathered at Mbarara Regional Referral Hospital and Kabwohe Health Centre IV in Western Uganda. Participants were recruited via facility-based convenience sampling, yielding 150 individuals: 65 expectant mothers, 45 postpartum mothers, and 40 healthcare workers (clinicians, midwives, nurses, and Village Health Team members). Primary data collection was conducted over a two-week period by six trained research assistants. To capture rich dialogues, collection utilized multiple formats: semi-structured private interviews (n = 90, 60%) averaging 35 min, focus group discussions (n = 8 groups) lasting approximately 90 min, and strictly anonymous questionnaires (n = 52, 34.7%). Interactive sessions were audio-recorded with consent, transcribed verbatim, and supplemented by observational field notes.

We simulated authentic clinical interactions by having healthcare workers answer questions typically posed by mothers, and vice versa. The questions were systematically categorized into prenatal and postnatal themes. To ensure balanced representation, each thematic category was explicitly structured to contain a minimum of 250 question-answer pairs. The raw responses were digitized, and all Protected Health Information (PHI) was removed to ensure anonymity. A dedicated linguist generated a parallel dataset in English, Luganda, Runyankore, and Swahili using a forward-backward translation protocol (initial translation, independent back-translation, and consensus reconciliation). Translation quality was validated via semantic similarity metrics, achieving average BLEU scores of 0.78 (English-Luganda), 0.81 (English-Runyankore), and 0.83 (English-Swahili).

Subsequently, three independent medical professionals (an obstetrician, a midwife, and a public health specialist) performed human-in-the-loop validation to guarantee clinical integrity and cultural relevance. This audit resulted in the revision of 127 question-answer pairs (3.4%).

The finalized corpus contains 3,694 unique Q/A pairs, totaling 14,800 parallel entries across the four languages. The average question length is 12.4 words, and the average answer length is 38.7 words. The dataset file is approximately 8.2 MB and utilizes UTF-8 encoding to accurately preserve special characters. Thematically, the data highlights the intersection of medical advice and communal life, with the largest segments focusing on community and cultural considerations (31.4%), medical history and lifestyle (24.5%), and symptoms and concerns (24.4%). The high prevalence of cultural inquiries aligns with evidence that mothers navigate complex tensions between traditional practices and biomedical recommendations [21], often driven by deep-rooted cultural trust [22]. Furthermore, the smaller subset of mental health inquiries (6.5%) underscores the clinical reality of stigma, limited mental health literacy, and structural barriers surrounding perinatal depression in the region [23]. An overview of the data file is provided in Table 1.

Table 1.

Overview of data files/data set

Label	Name of data file/data set	File types (extension)	Data repository and identifier (DOI or accession number)
Data file 1	Maternal_Multlingual_Dataset_Uganda_2026	.xls	10.7910/DVN/JS3ILT

Limitations

While this dataset provides a culturally resonant baseline for low-resource languages, it has some limitations. Primarily, the data exhibits limited generalizability, as it is confined to a specific linguistic and geographical scope. Consequently, it may not fully capture the dialectal variations, diverse challenges, or cultural nuances present in other regions of East Africa. Future work should expand upon this high-quality baseline to include broader regional dialects and audio-based modalities. Finally, any conversational AI trained on this dataset must be positioned strictly as a supplementary information-dissemination tool. It is intended to complement, rather than supplant, consultation with qualified healthcare professionals.

Author contributions

R.K. contributed to project administration, conceptualization, supervision, and writing of the original draft. F.K. contributed to project administration, methodology, data curation, investigation, and resources. R.Kuyeso. contributed to software development, methodology, and investigation. E.M. contributed to investigation, and resources. D.S.B. contributed to data curation, investigation, and validation. R.M. and W.W. contributed to formal analysis and visualization. F.T. contributed to project administration and resources. J.N. && B.A. contributed domain-specific medical expertise, including the design and validation of data collection instruments and survey frameworks, A.N. && P.N. were responsible for data collection, clinical validation, and comprehensive review of the dataset, with particular involvement in midwifery-related components, C.N. contributed to the translation of the dataset and its linguistic validation to ensure adherence to standardized and contextually appropriate language, L.A.C. contributed to conceptualization, methodology, and supervision. All authors reviewed and approved the final manuscript.

Funding

This work was supported by the Science for Africa (SFA) Foundation through the Grand Challenges Africa program under Grant Reference Number SFA-R15-145.

Data availability

The primary dataset developed in this study has been deposited in the Havard Dataverse Repository and is accessible at the following link: https://doi.org/10.7910/DVN/JS3ILT.

Declarations

Ethics approval and consent to participate

Institutional approval was obtained from Mbarara University of Science and Technology Research Ethics Committee (MUST-REC) (Reference Number: MUST-2024-1779) and national clearance from the Uganda National Council for Science and Technology (UNCST) (Reference Number: HS5739ES). All procedures involving human participants were conducted in accordance with the principles of the Declaration of Helsinki. Written informed consent was obtained from all participants prior to data collection.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.