Perspective | Open Access | Volume 9 (4): Article 197 | Published: 18 Nov 2026
Menu, Tables and Figures
Table 1: Proposed AI use and assessment across stages of advanced field epidemiology training
| Training stage | Permitted AI assistance | Independent competence | Assessment evidence |
|---|---|---|---|
| Foundation Months 1–3 | Tutoring after an initial attempt. Language editing of trainee-developed text when consistent with the learning objective. | Problem framing, case definitions with person, place and time, line lists, epidemic curves by onset date, attack rates with stated denominators and an initial analysis plan. | Written and oral assessment of core methods. Observation of a line-list exercise and discussion of decisions. Detection of a planted denominator error. |
| Supervised field practice Months 4–12 | Code drafting after plan approval and demonstration of relevant calculations. Literature scoping, hypothesis comparison, draft critique and questionnaire formatting. | Initial hypotheses, control selection, measures of association that fit the study design and matching, interpretation, recommendations and investigation review criteria. Explanation and verification of analytic choices. | Oral assessment of major deliverables. Reproduction of a key estimate, such as an attack rate or odds ratio, using conventional tools. Review of assistance and verification. |
| Advanced practice Year 2 | Wider support for tasks the trainee can perform and evaluate. Supervised use of more automated workflows within agency rules. | Investigation leadership, adaptation to new evidence, communication with decision-makers and justification of consequential decisions. | Observed field performance. Simulation without AI. Critique of an AI-generated recommendation to close an investigation after negative tests. Review of decisions and verification across the portfolio. |
Abbreviation: AI, artificial intelligence.
Note: Timing is illustrative for a two-year advanced program. Progression depends on demonstrated competence in each task. All uses require institutional approval, data protection and proportionate disclosure. Independent assessment excludes generative AI but may permit specified conventional tools and references. Language assistance follows the learning objective. These proposals require evaluation and adaptation to other training tiers.
Eric Nzirakaindi Ikoona1,&, Lucy Namulemo2, Godfrey Lumbuye Kayita3, Ronald Kaluya4, Rebecca Sakwa5, Foday Sahr1
1National Public Health Agency, Freetown, Sierra Leone, 2Foothills Community-Based Interventions, Monticello, Kentucky, United States of America, 3African Field Epidemiology Network, Kampala, Uganda, 4Uganda Counselling and Support Services, Kampala, Uganda, 5Makerere University–Johns Hopkins University Research Collaboration, Kampala, Uganda
&Corresponding author: Eric Nzirakaindi Ikoona, Office of the Executive Director, National Public Health Agency, 42A Main Motor Road, Wilberforce, Freetown, Sierra Leone, Email: ikoonae@yahoo.com, ORCID: https://orcid.org/0000-0003-3402-1961
Received: 18 Oct 2026, Accepted: 16 Nov 2026, Published: 18 Nov 2026
Domain: Field Epidemiology, Artificial Intelligence
Keywords: Field epidemiology training programs; artificial intelligence; competency-based education; assessment; outbreak investigation
©Eric Nzirakaindi Ikoona et al. Journal of Interventional Epidemiology and Public Health (ISSN: 2664-2824). This is an Open Access article distributed under the terms of the Creative Commons Attribution International 4.0 License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Cite this article: Eric Nzirakaindi Ikoona et al., Strengthening competence assessment for artificial intelligence use in field epidemiology training. Journal of Interventional Epidemiology and Public Health. 2026; 9(4):197. https://doi.org/10.37432/jieph-d-26-00437
Generative artificial intelligence (AI) now drafts code, text and analyses in field epidemiology training programs. A survey of fellows in Canada, the United States and Europe found that two thirds used AI in their work, mostly for analytic code, while one in five had received AI training. Evaluations in China and Switzerland found that AI tools gave uneven answers to field epidemiology examination questions, cited guidance that did not exist and misreported a measure of association. In our view, assessment should therefore examine the trainee’s reasoning as well as the quality of the submitted product. Drawing on our experience of mentoring and examining trainees during outbreak responses, we propose five connected arrangements. Permissions to use AI should expand as trainees demonstrate each skill. Major deliverables should include proportionate records of substantive assistance. Oral assessment and observed performance should test whether trainees can reproduce key estimates, defend denominators and control selection, and interpret negative laboratory findings. Outbreak workplans should state in advance when inconclusive findings require review or escalation. National public health agencies and training networks, including those in Africa, should evaluate these arrangements together. Evaluation should examine independent reasoning, error detection, mentor workload, cost and equity. The intended outcome is a graduate who can verify assisted work, explain consequential decisions and take responsibility for public health recommendations.
Field epidemiology training programs (FETPs) prepare epidemiologists through supervised public health service [1]. Outbreak reports, surveillance evaluations, analyses and presentations provide evidence of core competencies [2]. Each product records decisions about case definitions, denominators, comparison groups and control measures.
Generative artificial intelligence (AI), including large language models that produce text and code, can now contribute to each product. We write as epidemiologists who mentor and examine FETP trainees and who have worked in national responses to Ebola, Lassa fever, coronavirus disease 2019 (COVID-19), mpox and measles. In our view, programs should permit AI use and assess the reasoning behind every product that earns credit.
The United States Centers for Disease Control and Prevention (CDC) issued draft FETP guidance in August 2026 [3]. The guidance names oral assessment, observed simulation and documentation of assistance as safeguards [3]. The World Health Organization (WHO) addresses governance of AI in health [4]. The United Nations Educational, Scientific and Cultural Organization (UNESCO) asks institutions to protect human agency and validate tools before learners use them [5]. The African Union’s 2024 AI strategy gives priority to skills, ethics and data governance [6].
An earlier viewpoint set out professional responsibilities for health workers who use AI [7]. We apply them to FETP assessment where mentoring and technical resources are limited.
Evidence from training and public health practice
In a 2024 survey of fellows in Canadian, United States and European FETPs, 69 of 105 respondents (66%) used AI in their work [8]. Of the 69 users, 63 (91%) used AI to troubleshoot analytic code. Only 21 of 105 respondents (20%) had received AI training [8].
In China, six large language models answered 20 multiple-choice questions from the 2025 Zhejiang FETP entrance examination [9]. The best model answered 15 of 20 (75%) correctly and two ChatGPT versions each answered 8 of 20 (40%). The 35 junior epidemiologists answered a median of 12 of 20 questions (60%) correctly. Several models cited guidelines that did not exist [9].
In a Swiss randomized trial, 30 students and staff analyzed simulated survey data with ChatGPT, either alone or alongside statistical software [10]. An unplanned analysis found that only 1 of the 15 participants who used ChatGPT alone (6.7%) completed the analysis without serious mistakes. In one error, ChatGPT reported an unconverted model estimate as an odds ratio, which compares the odds of an outcome between groups [10]. Both groups used ChatGPT, so the trial shows verification failures rather than loss of skill.
Between June and October 2023, epidemiologists at the Korea Disease Control and Prevention Agency screened 99,945 open-source articles through WHO’s Epidemic Intelligence from Open Sources system and verified each signal [11]. Eight events were detected a mean of 20 days before they appeared on WHO’s Event Information Site or in Disease Outbreak News [11]. The report concerns screening of news and other public sources for outbreak signals, and its lesson lies in verification.
We conclude that trainees use AI most for analytic code, so mentors should check that code against the data. Examination performance varied widely between tools, and ChatGPT reported a wrong measure of association. The findings point to one competence: checking a result against the data, the study design and the field context.
Protecting independent competence
Studies outside field epidemiology show why assisted and independent performance need separate assessment. In a randomised study of nearly 1,000 mathematics students, unrestricted GPT-4 access improved practice performance, but later examination scores without AI were about 17% lower than control scores [12]. A tutoring version largely avoided this reduction [12].
In four Polish centers, detection of precancerous growths during colon examinations without AI fell from 28.4% (226 of 795) to 22.4% (145 of 648) after AI was introduced [13]. The study observed routine practice, yet the decline justifies checking independent performance. Among 50 physicians, access to a large language model did not significantly improve diagnostic reasoning, although the model alone scored higher [14].
Translating guidance into educational practice
Making professional accountability assessable
In our view, the steps of an outbreak investigation show where assisted work most needs checking. A case definition needs explicit criteria of person, place and time. An epidemic curve should plot onset dates where available, because report dates add surveillance delays. An attack rate divides new cases by the population at risk, and AI-drafted code can wrongly include people outside that population. Before accepting an AI-generated explanation for a rise in notifications, trainees should check duplicate records, reporting completeness, changes in the case definition and the population covered.
Trainees should also choose and interpret measures of association that fit the study design and sampling strategy, and account for matching in the analysis. In a case-control study, controls should represent the population that produced the cases, and exposure should refer to the relevant period. Logistic regression is a model for yes-or-no outcomes. Trainees should check whether exposure estimates already appear as odds ratios or need conversion, the step ChatGPT omitted in the Swiss trial [10]. Trainees should trace the analyzed records to the line list, the table of individual cases, and distinguish missing values from negative responses. Assessment should also examine whether other factors explain group differences and distinguish prediction from cause and effect [7].
In our experience as examiners, asking a trainee to recompute one estimate from the line list and defend its denominator provides direct evidence of independent reasoning. Wider permission to use AI should follow demonstrated competence in each task.
Teaching decisions after inconclusive testing
CDC’s fictional outbreak scenario illustrates premature closure after targeted testing fails to explain a cluster [3]. In a supervised exercise, trainees can critique an AI-generated recommendation to close an investigation after negative tests. They should explain what remains unknown and propose proportionate next steps.
Premature closure involves ending consideration of alternative explanations too early [15]. Trainees should question the completeness of an initial list of possible causes, including one developed with AI assistance. We propose recording review and escalation criteria in the outbreak workplan before testing begins.
The right response depends on the epidemiologic pattern, specimen quality, timing of collection and available tests. Trainees should justify their choice, including when continued surveillance with planned review is proportionate.
The 2015 Kampala typhoid outbreak offers a published teaching case. Uganda’s Public Health Fellowship Program compared 33 case-patients with 78 controls chosen to resemble them as a group (frequency matching) [16]. Water sold in small plastic bags was drunk by 7 of 31 case-patients (22.6%) and 2 of 78 controls (2.6%), with an odds ratio of 8.90 after accounting for matching [16]. Water and juice samples showed fecal contamination, and authorities closed unsafe water sources [16]. Trainees can explain what the environmental results add, how matching shaped the analysis and why action proceeded under uncertainty.
Combining portfolios with direct assessment
Miller’s framework distinguishes knowledge, its application, demonstration of a skill and performance in practice [17]. A polished portfolio cannot show the trainee’s independent contribution, so programs should add oral assessment and observation against criteria known in advance.
Entrustable professional activities, tasks that supervisors entrust after observing competent performance, offer a model for AI permissions [18]. Before using AI to draft analytic code, a trainee should explain the analysis plan and calculate key measures with conventional tools.
A published protocol for a 24-month Advanced FETP in Sierra Leone shows how to organize such assessment [19]. Assessment combines scoring guides, contribution statements, direct observation, oral defense and independent double rating of high-stakes products [19]. Evaluating these arrangements with the proposed AI checks could add African evidence to AI-specific FETP studies, which so far concern North America, Europe and China [8,9].
Supporting equitable scientific communication
CDC guidance recognizes grammar and structure review as possible AI uses [3]. Programs should state this permission for trainees writing in an additional language. Trainees remain responsible for the question, analysis, interpretation and recommendations.
Working in English as an additional language cost extra time and effort [20]. AI-text detectors often misclassify writing by non-native English speakers [21], so detector scores should not serve as sufficient evidence of misconduct. Examiners can assess epidemiologic reasoning in a mutually understood language where feasible and assess professional communication separately.
Preparing mentors and institutions
Host agencies carry the governance responsibilities that WHO and the African Union describe [4,6], and trained staff depend on such organizational conditions [22]. Agencies should specify approved tools, permitted data categories and routes for resolving uncertainty about data use. Where approved services are unavailable, artificial teaching datasets and secure conventional software can support learning.
Epidemiology mentors can assess design and interpretation, laboratory colleagues can advise on specimens, and data specialists can review complex code [7]. Programs should prepare mentors through shared scoring exercises, record their workload and arrange cover when they deploy to outbreaks.
Five arrangements for program implementation
We propose five arrangements for programs to adapt and evaluate. Table 1 links possible AI uses with evidence of independent competence across training stages.
First, programs should link AI permissions to demonstrated competence. Month ranges in Table 1 guide scheduling, while observed performance determines progression for each task. Early exercises can require an independent attempt before comparison with AI output.
Second, major deliverables should include a proportionate record of substantive AI assistance naming the tool, date, purpose, outputs used and checks performed. Routine language editing may need only a short declaration. Mentors should examine verification, since an absence of rejected output does not show that checking occurred.
Third, programs should pair major deliverables with oral assessment and observed independent performance. Trainees should explain the question, case definition, denominators, measure of association and recommendations. They should then reproduce a key estimate, such as an attack rate or odds ratio, with conventional tools. Programs can pilot one supervised simulation without AI each year alongside authentic field investigations [19].
Fourth, outbreak workplans should include review and escalation criteria agreed by the trainee and mentor. The criteria identify findings that prompt reassessment, who reviews them and feasible next steps. Negative or inconclusive tests provide one trigger. Explicit links between public health intelligence and emergency operations decisions serve the same purpose [23].
Fifth, programs should evaluate these arrangements before wider adoption. National public health agencies, CDC, the Training Programs in Epidemiology and Public Health Interventions Network, the African Field Epidemiology Network and other partners could share instruments across regions. Evaluation should compare independent reasoning, error detection, retention of core skills and assisted work quality. The evaluation should also measure mentor time, cost, access and language background [7]. Evaluation should remain separate from trainee progression decisions [19].
Limitations
Our argument draws on published literature and our experience as mentors and examiners. The AI-specific FETP studies cited here concern North America, Europe and China, and other evidence comes from schools and clinical practice. The proposed stages, records and assessment schedule await validation and adaptation to training tier, institutional rules, resources and language.
FETPs can integrate AI while requiring graduates to demonstrate independent epidemiological reasoning. Fellows already use AI widely for analytic code, and AI tools can misreport measures of association. Linking permissions to observed competence, documenting assistance, reproducing key estimates and agreeing review criteria before testing can protect independent skill. Agencies and training networks should evaluate these arrangements together. The intended outcome is an epidemiologist who verifies assisted work and takes responsibility for decisions affecting communities.
What is already known about the topic
What this study adds
The authors did not receive any specific funding for this work.
Abbreviations
AI: artificial intelligence
CDC: Centers for Disease Control and Prevention
COVID-19: coronavirus disease 2019
FETPs: Field epidemiology training programs
UNESCO: United Nations Educational, Scientific and Cultural Organization
WHO: World Health Organization
Conceptualization: Eric Nzirakaindi Ikoona, Lucy Namulemo,
Methodology: Eric Nzirakaindi Ikoona, Lucy Namulemo
Data Curation: Eric Nzirakaindi Ikoona, Lucy Namulemo, Godfrey Lumbuye Kayita
Formal Analysis: Eric Nzirakaindi Ikoona, Lucy Namulemo, Godfrey Lumbuye Kayita
Investigation: Eric Nzirakaindi Ikoona, Lucy Namulemo, Godfrey Lumbuye Kayita
Project Administration: Eric Nzirakaindi Ikoona, Lucy Namulemo
Resources: Eric Nzirakaindi Ikoona, Lucy Namulemo, Godfrey Lumbuye Kayita, Ronald Kaluya, Rebecca Sakwa, Foday Sahr
Validation: Eric Nzirakaindi Ikoona, Lucy Namulemo, Godfrey Lumbuye Kayita, Ronald Kaluya, Rebecca Sakwa, Foday Sahr
Visualization: Eric Nzirakaindi Ikoona, Lucy Namulemo, Godfrey Lumbuye Kayita, Ronald Kaluya, Rebecca Sakwa, Foday Sahr
Supervision: Eric Nzirakaindi Ikoona, Lucy Namulemo, Godfrey Lumbuye Kayita, Ronald Kaluya, Rebecca Sakwa, Foday Sahr
Writing – Original Draft: Eric Nzirakaindi Ikoona, Lucy Namulemo, Godfrey Lumbuye Kayita, Ronald Kaluya, Rebecca Sakwa, Foday Sahr
Writing – Review & Editing: Eric Nzirakaindi Ikoona, Lucy Namulemo, Godfrey Lumbuye Kayita, Ronald Kaluya, Rebecca Sakwa, Foday Sahr
Table 1: Proposed AI use and assessment across stages of advanced field epidemiology training
| Training stage | Permitted AI assistance | Independent competence | Assessment evidence |
|---|---|---|---|
| Foundation Months 1–3 |
Tutoring after an initial attempt. Language editing of trainee-developed text when consistent with the learning objective. | Problem framing, case definitions with person, place and time, line lists, epidemic curves by onset date, attack rates with stated denominators and an initial analysis plan. | Written and oral assessment of core methods. Observation of a line-list exercise and discussion of decisions. Detection of a planted denominator error. |
| Supervised field practice Months 4–12 |
Code drafting after plan approval and demonstration of relevant calculations. Literature scoping, hypothesis comparison, draft critique and questionnaire formatting. | Initial hypotheses, control selection, measures of association that fit the study design and matching, interpretation, recommendations and investigation review criteria. Explanation and verification of analytic choices. | Oral assessment of major deliverables. Reproduction of a key estimate, such as an attack rate or odds ratio, using conventional tools. Review of assistance and verification. |
| Advanced practice Year 2 |
Wider support for tasks the trainee can perform and evaluate. Supervised use of more automated workflows within agency rules. | Investigation leadership, adaptation to new evidence, communication with decision-makers and justification of consequential decisions. | Observed field performance. Simulation without AI. Critique of an AI-generated recommendation to close an investigation after negative tests. Review of decisions and verification across the portfolio. |
Abbreviation: AI, artificial intelligence.
Note: Timing is illustrative for a two-year advanced program. Progression depends on demonstrated competence in each task. All uses require institutional approval, data protection and proportionate disclosure. Independent assessment excludes generative AI but may permit specified conventional tools and references. Language assistance follows the learning objective. These proposals require evaluation and adaptation to other training tiers.