Digitising Disappearing Knowledge: Artificial Intelligence Applications in the Documentation and Conservation of West African Ethnomedicinal Plant Traditions

Eniola Morufat Azeez¹*, Babatunde Semiu Adeleke², Zainab Abisola Azeez²

1 Department of Biology, Emmanuel Alayande University of Education, Oyo

2 Department of Pure and Applied Biology, Ladoke Akintola University of Technology, Ogbomoso

*Correspondence: eniazeez1@gmail.com;

DOI: https://doi.org/10.71431/IJRPAS.2026.5608

INTRODUCTION

Each community in most West African towns is blessed with a healer who knows things that no database currently records. Such healer is aware of which combinations of Vernonia amygdalina leaves and the bark of Khaya senegalensis is used for malaria in her particular district/town. She knows the season in which to harvest, the parts to avoid, the preparation method that reduces bitterness without destroying efficacy and the contraindications learned from generations of observation, an information that cannot be found in databases such as PubMed and others.

West Africa’s Ethnomedicinal plant knowledge is extensive and a systematic review of ethnobotanical literature on Nigerian plants from 1964 to 2024 identified 963 plant species across 144 familes being used for medicinal, spiritual, cosmetic and other purposes, with the top-cited species including Carica papaya (42 citations), Vernonia amygdalina (42) and MAngifera indica (41) (Rafiu et al., 2025a). However, the same review documented persistent problems such as insufficient documentation of indigenous documentation, limited collaboration between researchers and non-adherence to best-practice survey methods. The implication being that the knowledge exists but the infrastructure to preserve the knowledge systematically is lacking.

This lack matters beyond cultural loss because traditional medicine has been the trusted source of healthcare for African populations for centuries with 70-80% of the continent’s population relying on herbal therapies as their primary form of care (WHO, 2022; Ikhoyameh et al., 2024). These figures reflect real clinical dependence. In rural communities where hospital access is limited, expensive or unreliable, the traditional healer is the primary care provider. The plants she uses and the knowledge of how to use them constitute a functioning healthcare system.

At the same time, there is an untapped pharmacological potential that conventional drug discovery has barely touched. Of the estimated 400,000 vascular plants on earth, less than 10% have been systematically evaluated for their therapeutic potential and the proportion of African medicinal plants is very low. Natural products have a strong historical record with examples seen in artemisinin derived from Artemisia annua following ethnopharmcological leads, vinblastine and vincristine used in cancer therapy derived from Catharanthus roseus, a plant used in Carribbean folk medicine. The challenge thus lies in how to bridge the distance between what traditional healers know and what pharmaceutical science can validate.

Artificial intelligence (AI) has started offering credible solutions to bridging this gap. Recent researches have shown that deep learning can identify medicinal plant species from leaf images with accuracy exceeding 97-99% under controlled conditions (Mulugeta et al., 2024). AI tools are being applied to mine existing ethnobotanical literature, structure informal knowledge and predict bioactive properties of compounds before any laboratory work is done (Suryawanshi et al., 2026). Muthuraj and Chandrasekaran, (2026) argue that natural language processing can now systematically mine ethnopharmacological records at a scale no human research team could manage, accelerating the identification of biologically active plant compounds.

Therefore, the missing link is a careful application of these tools specifically to the West African ethnomedicinal context, with honest attention to what they can and cannot do, and with the communities whose knowledge is at stake placed at the center of analysis; a gap this paper attempts to fill.  

The Ethnomedicinal Landscape of West Africa: What is at Stake

Scale and Cultural Depth

The Yoruba, Igbo and Hausa medical traditions in Nigeria alone encode centuries of empirical plant pharmacology. Lawal et al., (2022) surveyed 100 participants across five areas of Lagos State and identified 183 medicinal plant species from 61 families used for health conditions ranging from malaria and respiratory disease to skin disorders, insomnia and convulsions. The most cited plants in that survey including Carica papaya, Morinda lucida, and Vernonia amygdalina are used for fevers and malaria, which had the highest consensus factor (ICF=0.95), an indication of strong agreement across informants about their efficacy for those conditions. Such consensus is itself a form of clinical signal worth investigating pharmacologically.

In Ibadan, Oyo State, an ethnobotanical survey of plants used in treatment of sexually transmitted diseases reported 52 plant species from 34 families used by traditional healers (Omilani, 2021).  In Ede, Osun State, a survey of 100 participants documented 87 plant species used for cough and respiratory conditions with approximately 32% of documented plants reported as cough remedies for the first time in the literature (Lawal et al., 2020). Each of these surveys expands the documented record but they also illustrate the problem of scale. Three surveys of three communities, each producing tens of species and use reports, each done with slightly different methodologies and each requiring searate analytical effort to compare or integrate.

Across Benin and other West African countries, the problem increases. Ahoyo et al., (2024), surveyed 590 traditional medicine practitioners across Benin’s phytodistricts and discovered that the least available tree species are the most frequently requested by healer, a pattern indicating that overharvesting is pushing key medicinal plants towards local extinction. Species including Azadirachta indica, Khaya senegalensis and Pterocarus erinaceus were found to be both highly medicinally and increasingly scarce in their natural habitats.

Why the Knowledge is Disappearing

Several intersecting pressure drive traditional loss. Urbanization draws younger generations away from rural communities where practical botanical knowledge is transmitted. Formal education systems in West Africa rarely incorporate or validate indigenous botanical knowledge. The economic incentives for young people to pursue traditional healing as a livelihood are weak compared to salaried employment. Many traditional healers are reluctant to share detailed formulations out of concern for intellectual property or religious obligations of secrecy, and this reluctance is rotational given historical cases of bioprospecting that benefitted pharmaceutical companies without compensating source communities.

There is also a straightforward demographic reality. Most practicing traditional healers in West Africa are older adults. When they die, the specific, localized knowledge they hold dies with them unless it has been recorded. In his bibliometric analysis of Nigerian ethnobotanical research, Rafiu et al., (2025) identified ‘insufficient documentation of indigenous practices’ and ‘limited collaboration’ as the primary structural gaps in the existing body of research, problems that have persisted across all three phases of Nigerian ethnobotanical research history (the fundamental period, 1964-1989; the expansion phase, 1990-2009 and the modernization phase, 2010-2024). Six decades of research, and documentation remains insufficient.

AI TOOLS APPLICABLE TO ETHNOBOTANICAL DOCUMENTATION AND CONSERVATION

Computer Vision and Plant Identification

 The most immediate barrier in ethnobotanical fieldwork is accurate plant identification. Semi-structured interviews with traditional healers often involve collected plant specimens or descriptions of plants. This makes errors in species identification propagate throughout the research record. A plant misidentified in a survey becomes a misidentified entry in a database, a misidentified compound in a pharmacological study and a potential clinical safety problem if the wrong species is collected for medicinal use.

Artificial intelligence has however proffered solution to this through deep learning-based identification. Deep learning-based identification has reached a level of accuracy that makes it practical for field use. Mulugeta et al., (2024) systematically reviewed 31 deep learning studies on medicinal plantclssification published between 2018 and 2022, finding that convolutional neural networks (CNNs) were used in 64.5% of studies, transfer learning was applied in 83.8% and leaf images especially leaf shape were used in 74% of approaches. The consistent finding across this literature is that deep learning substantially outperforms classical morphometric approaches, which plateud around 66% accuracy while modern CNNs regularly achieve above 94% (Mulugeta et al., 2024).

Dey et al., (2023) compared seven advanced deep learning architectures including VGH16, VGG19, DenseNet201, ReNet50V2, Xception, InceptionResNetV2 and InceptionV3 for automated identification of 30 medicinal species across 20 plant families. DenseNet201 performed best, achieving 99.64% accuracy on public image data and 97% on field images with complex backgrounds. This is relevant for ethnobotanical applications where photographs are taken in variable lighting with cluttered backgrounds and often only of partial plant specimens, particularly the conditions that make accurate identification difficult for non-specialist researcher.

The qualification that matters here is dataset composition. Mulugeta et al., (2024) documented that 67.7% of deep learning studies used private datasets and that there is a clear absence of globally available, publicly accessible datasets for medicinal plants indigenous to specific countries and regions. Most high accuracy models have been trained on South Asian and Southeast Asian species. West African medicinal plants remain underrepresented in these training corpora. DenseNet201 achieving 97% accuracy on South Asian medicinal species does not automatically translate to comparable accuracy on Vernonia amygdalina or Newbouldia laevis. This however is not a reason to dismiss the technology, rather it is areason to invest in building West African plant image datasets, which is a fundable research programme.

Natural Language Processing for Mining Ethnobotanical Records

West Africa’s ethnomedicinal knowledge exists in several forms; in the minds of traditional healers, in unpublished fieldwork notes, in published surveys scattered across dozens of journals, in grey literature such as government health reports and herbarium recoeds, and increasingly in digital texts. The challenge is extracting structured, comparable information from this dispersed and heterogeneous corpus.

Natural language processing (NLP) tools, including large language models trained on biomedical text are now capable of this extaraction. Suryawanshi et al., (2026) found that AI algorithms can analyse large ethnobotanical records to produce predictions about biological agents, evaluate herbal mixtures and improve quality standards in traditional medicine practices. Muthuraj and Chandrasekaran, (2026) document the role of NLP in mining ethnopharmacological knowledge especially in platforms that enable scalable processing of natural product data.

A direct proof of concept is the data mining study by Bibak et al., (2024) which used machine learning to analyse ethnopharmacological documents and found that ML-powered research improved the effectiveness and efficiency of expert researchers by factors of 3.1 and 5.14 respectively, retrieving 420 relevant ethnopharmacological documents in seven hours compared to an estimated 36 hours of human effort. For West African ethnobotany, where published records are distributed across the Journal of Ethnobiology and Ethnomedicine, Plants (Basel), Journal of ethnopharmacology and many other outlets, systematic NLP-based literature synthesis could generate the integrated database that currently does not exist.

The practical shape of this would involve training or fine-tuning a biomedical NLP model on the existing West African ethnobotanical literature, then using it to extract structured plant-condition-preparation-dosage relationships from both published and unpublished sources. The output would be a queryable database of ethnobotanical knowledge in the region, tagged by species, condition, plant part, preparation method, geographic origin and ethnic group.

Machine Learning for Bioactivity Prediction

Once a plant species has been identified and its traditional uses documented, the next challenge is determining which of its chemical constituents might be pharmacologically active. Traditional bioassay-guided fractionation is slow and expensive; isolating, purifying and testing individual compounds from a plant extract can take years and significant laboratory resources. Machine learning-based bioactivity prediction allows this to happen computationally, prioritizing compounds for laboratory investigation based on structural features associated with known pharmacological activity.

 Muthuraj and Chandrasekaran, (2026) reviewed the current state of AI-driven natural product discovery, documenting ML-driven dereplication, chemiformatics for structural characterization and genome mining for identifying biosynthetic gene clusters that might produce bioactive compounds. Wang et al., (2025) achieved an AUC of 0.9873 using a deep learning framework for predicting ingredient-target interactions in traditional medicine, validated through molecula docking studies on Scutellaria baicalensis.

For West African ethnomedicine, this pipeline would work thus: a documented plant with specific traditional use is identified using AI-assisted species authentication in the field, its phytochemical profile is characterized by GC-MS or LC-MS, the resulting compound library is fed into a machine learning bioactivity prediction model and compounds predicted to have activity against the relevant target are prioritized for laboratory testing. The key insight is that traditional use data themselves function as a training signal. An example is if V. amygdalina has been used by multiple ethnic grouos across West Africa for malaria, and if informant consensus is high, this will be genuinely informative about where to look for pharmacological activity.

Citizen Science Platforms and Community Data Collection

The pl@ntNet platform surpassed 100 million identifications in early 2026, contributed by millions of users across more than 200 countries (Pl@ntNet, 2026). This demonstrates that citizen science approaches to plant identification and data collection can scale in ways that no team of professional researchers can match. For West African ethnobotany, the principle is transferable in such a way that community members, students and traditional healers themselves can collect and submit plant photographs that feed into identification and documentation workflows.

The important modification is that citizen science must be designed differently fro general biodiversity observation. Identifying that a specimen of A. indica exists at a particular GPS coordinate is useful but limited. Recording that it is used by a specific healer for a specific condition, prepared in a specific way and administered at a specific dose is what makes the data ethnobotanically meaningful. Platform designs must accommodate this richer data structure and it must do so in ways accessible to participants with limited digital literacy or intermittent internet connectivity.

THE DOCUMENTATION PROBLEM PRACTICE: SELECTED WEST AFRICAN SPECIES

Several species illustrate concretely why AI-assisted documentation matters for both conservation and drug discovery. V. amygdalina is used across virtually all West African ethnic groups. It appears as one of the top cited plants in multiple Nigerian ethnobotanical surveys (Lawal et al., 2022; Rafiu et al., 2025). Its uses spans malaria, diabetes, liver disease, helminthiasis and skin conditions. Laboratory studies have confirmed antiplasmodial, antidiabetic and hepatoprotective activities in various extracts. Yet there is no authoritative, geographically resolved database of how preparation methods vary across ethnic groups, which plant parts are used for which conditions in which regions, or how patterns of use have changed over time. AI-assisted text mining of the existing survey literature could generate this map within weeks.

Azadiractha indica is also documented in 19 of Nigeria’s six geopolitical regions by Rafiu et al., (2025), making it one of 19 plants with pan-Nigerian documented use. Ahoyo et al., (2024) found it among the species most at risk of local extinction in Benin due to overharvesting driven by high demand. Its leaf essential oil has been studied for molecular-level insecticidal activity against Anopheles gambiae odorant-binding proteins, with nerolidol identified as a lead compound with exceptional binding affinity at both DEET and secondary icaridin binding sites. An AI-assisted field identification tool calibrated for Azadiractha indica and its common morphological lookalikes in the Sudano-Guinean zone would allow community members to confirm species identity, authenticate harvest quality in the supply chain and flag harvesting pressure in conservation-mapped areas.

Newbouldia laevis and Cassia serberiana appear in Ahoyo et al., (2024) as one of the least available yet most requested species I Benin’s medicinal plant markets, a pattern signaling unsustainable harvesting. Neither of these species is presently well-represented in AI plant identification training datasets. Documenting them digitally with multi-organ images linked to GPS coordinates and traditional use records would simultaneously support conservation monitoring and future AI model development.

CHALLENGES: BEING HONEST ABOUT THE DIFFICULTIES

The Dataset Problem is Structural, Not Technical

The accuracy figures cited above, 97-99% for leading plant identification models are real but context-dependent. They apply to species wee represented in training data. West African medicinal plants are not well represented. Building adequate datasets therefore requires sustained fieldwork, botanical expertise to verify identifications and institutional infrastructure to host and maintain the data over time. None of this is impossible but none of it is also free and the history of research funding in African scientific institutions suggests it will not happen without specific, sustained investment.

Connectivity and Hardware Constraints

Most AI-assisted plant identification currently require cloud connectivity. Rural communities in West Africa, particularly those where traditional medicine is most actively practiced, often have poor mobile data coverage. Offline capable apps with compressed local models exist but cover fewer specoes and have lower accuracy. Designing effective tools for these environments is a technical challenge that he developers of pl@ntNet, iNturalist and similar platforms have acknowledged but not fully solved.

Intellectual Property and Community Rights

This is the most politically charged dimension of AI-assisted ethnobotanical documentation. A dataset of traditional medicinal knowledge, once digitized and uploaded to a global platform can be accessed and commercially exploited by pharmaceutical companies without any obligation to compensate the communities whose knowledge generated it. This is not hypothetical, rather it is the documented history of bioprospecting. Protocols such as the Nagoya Protocol on Access and Benefit sharing under the Convention on Biological Diversity exist precisely to address this but implementation in West African countries is uneven. Any AI-assisted documentation programme that does not have a legally enforceable community benefit-sharing arrangement built into its design is repeating the bioprospecting pattern with new technology.

Interpretability and the Limits of Computational Prediction

Machine learning bioactivity predictions are probabilistic. A model predicting that a compound from Morinda lucida  is likely to inhibit Plasmodium falciparum is generating a testable hypothesis not a clinical finding. The risk is that computational predictions circulate without the experimental validation needed to confirm or refute them. Transparent reporting of uncertainty and clear communication of what predictions do and do not mean is essential if AI tools are to support rather than distort the ethnopharmacological research process.

A FRAMEWORK FOR RESPONSIBLE AI-ASSISTED ETHNOMEDICINAL DOCUMENTATION

Several principles should govern any serious programme in this area. Community partnerships from the start and not after. Traditional healers and community members should be involved in defining what knowledge is documented, in what form, with what access restrictions and with what benefit arrangements before data collection begins. The Nagoya Protocol provides a legal framework. However, the practical challenge with it is implementing it in contexts where formal contracts may be unfamiliar and where communities have legitimate reasons to be cautious.

Regional dataset building should be a priority. West African Universities, herbaria and research institutes are the logical homes for regionally specific medicinal plant image datasets. A coordinated initiative, perhaps modelled on GBIF’s distributed data-contribution model could produce the training data that current AI tools lack. Universities situated within the Sudano-Guinean savanna zone have access to vegetation that is underrepresented in global herbarium collections and are well placed to lead this effort.

Integration across the documentation-to-validation pipeline. AI-assisted documentation is only valuable if the documented knowledge can be followed through to laboratory investigation. A programme that simply creates a larger database of traditional uses without pharmacological follow-up, does not serve either the communities or the scientific goals of ethnobotany. The pipeline must connect field documentation, AI-assisted identification and literature mining, computational bioactivity prediction, laboratory validation and benefit sharing.

Offline-first tool design. Applications designed for these environments should prioritise offline functionality, minimal data requirements and interfaces usable low-cost devices without assuming high digital literacy.

CONCLUSION

West African ethnomedicinal knowledge is a pharmacological resource of genuine importance and it is being lost faster than it is being documented. The pressures are real as there is an increase in ageing healer populations, urbanization, habitat degradation and fragmented research. The tools that AI now offers, computer vision for species authentication, NLP for literature mining, machine learning for bioactivity prediction and citizen science platforms for scalable data collection, are capable in principle of addressing each of these problems.

These conditions for this to work are not primarily technical. The accuracy of plant identification models can be improved by building better training datasets. The reach of citizen science platforms can be extended by investing in offline functionality. The pharmacological follow-through can be accelerated by connecting documentation programmes with computational phytochemical screening. All of this is achievable.

What is difficult is ensuring that the communities whose knowledge is at the centre of this effort are genuine partners rather than data sources. Without that, AI-assisted ethnobotanical documentation risks being another chapter in a long history of knowledge extraction from African communities that produces scientific papers and pharmaceutical patents without meaningful local benefit.

If this is done right, it could be something different. A technology-assisted project of knowledge preservation, pharmacological discovery and community recognition that serves both African public health and global science. The plants are there. The knowledge is still, for now, accessible. The tools exist. What remains is the institutional will and the ethical seriousness to use them responsibly.

ACKNOWLEDGEMENTS

The authors acknowledge the authority of the Emmanuel Alayande University of Education, Oyo for providing the framework for embarking on a research of this nature.

CONFLICT OF INTEREST

The authors declare that there are no conflicting interests during the course of the study which could jeopardize the findings expressed in the presentation

REFERENCES

1.        African Traditional Medicine Day 2022 | WHO | Regional Office for Africa. (2026, April 5). https://www.afro.who.int/regional-director/speeches-messages/african-traditional-medicine-day-2022

2.        Ahoyo, C. C., Houéhanou, T. D., Yaoitcha, A. S., Akpi, B. P., Natta, A., & Houinato, M. R. B. (2024). How do plant demographic and ecological traits combined with social dynamics and human traits affect woody plant selection for medicinal uses in Benin (West Africa)? Journal of Ethnobiology and Ethnomedicine, 20(1), 15. https://doi.org/10.1186/s13002-024-00655-2

3.        Bibak, H., Heydari, F., & Sadat-Hosseini, M. (2024). Exploring the power of data mining for uncovering traditional medicinal plant knowledge: A case study in Shahrbabak, Iran. PLOS ONE, 19(6), e0303229. https://doi.org/10.1371/journal.pone.0303229

4.        Dey, B., Ferdous, J., Ahmed, R., & Hossain, J. (2024). Assessing deep convolutional neural network models and their comparative performance for automated medicinal plant identification from leaf images. Heliyon, 10(1), e23655. https://doi.org/10.1016/j.heliyon.2023.e23655

5.        Ikhoyameh, M., Okete, W. E., Ogboye, R. M., Owoyemi, O. K., & Gbadebo, O. S. (2024). Integrating traditional medicine into the African healthcare system post-Traditional Medicine Global Summit: Challenges and recommendations. The Pan African Medical Journal, 47. https://doi.org/10.11604/pamj.2024.47.146.43011

6.        Lawal, I. O., Olufade, I. I., Rafiu, B. O., & Aremu, A. O. (2020). Ethnobotanical Survey of Plants Used for Treating Cough Associated with Respiratory Conditions in Ede South Local Government Area of Osun State, Nigeria. Plants, 9(5), 647. https://doi.org/10.3390/plants9050647

7.        Lawal, I. O., Rafiu, B. O., Ale, J. E., Majebi, O. E., & Aremu, A. O. (2022). Ethnobotanical Survey of Local Flora Used for Medicinal Purposes among Indigenous People in Five Areas in Lagos State, Nigeria. Plants, 11(5), 633. https://doi.org/10.3390/plants11050633

8.        Mulugeta, A. K., Sharma, D. P., & Mesfin, A. H. (2024). Deep learning for medicinal plant species classification and recognition: A systematic review. Frontiers in Plant Science, 14, 1286088. https://doi.org/10.3389/fpls.2023.1286088

9.        Muthuraj, R., & Chandrasekaran, J. (2026). Nature meets machine: The AI renaissance in natural product drug discovery. Natural Products and Bioprospecting, 16(1), 37. https://doi.org/10.1007/s13659-025-00589-6

10.    Omilani, A. (2021). ETHNOBOTANICAL SURVEY OF THE MEDICINAL PLANTS USED IN THE TREATMENT OF SEXUALLY TRANSMITTED DISEASES IN IBADAN,OYO STATE, NIGERIA. ScienceOpen Preprints. https://doi.org/10.14293/S2199-1006.1.SOR-.PPH0HJ1.v1

11.    Pl@ntNet (2026) The Pl@ntNet API reaches 100 million identifications. Available at: https://plantnet.org/en/2026/02/02/the-plntnet-api-reaches-100-million-identifications/ (Accessed: 6 April 2026).

12.    Rafiu, B. O., Omotayo, A. O., Lawal, I. O., & Aremu, A. O. (2025). Ethnobotanical uses of plants in Nigeria: An analysis of current research trends and patterns. Journal of Ethnobiology and Ethnomedicine, 21(1), 57. https://doi.org/10.1186/s13002-025-00788-y

13.    Suryawanshi, R. M., Chaudhari, P. D., & Gilhotra, R. M. (2026). AI in ethnopharmacology, the pharmaceutical industry, and its applications. Annales Pharmaceutiques Françaises, 84(2), 161–174. https://doi.org/10.1016/j.pharma.2025.11.011

14.    Wang, X., Li, R., Zou, Y., & Wang, J. (2025). MMFi-DPBML: Multi-molecular fingerprint feature fusion for predicting ingredient-target interactions in traditional Chinese medicine. Journal of Ethnopharmacology, 353, 120451. https://doi.org/10.1016/j.jep.2025.120451