CANONIC Foundation

Community Learning Ledgers for Cancer Navigation in Small Island Developing States

TITLE PAGE

Title: Community Learning Ledgers for Cancer Navigation in Small Island Developing States

Short Title: Community Learning Ledgers in Caribbean SIDS

Authors:

Marisa Nimrod, MD, MPH1,2; Allana Roach, PhD, MSB3,4; Britney-Ann Wray, BS, CTBS, CCRP5; M. Amoy Fraser, PhD, CCRP, PMP5; Neville Calleja, MD, PhD, MSc, FFPH6,7; Junaid Kalia, MD8; Anil Bajnath, MD9; Dexter Hadley, MD/PhD4,5,9,10

Affiliations:

1 CARIBBEAN ASSOCIATION OF ONCOLOGY AND HEMATOLOGY (CAOH), Port of Spain, Trinidad and Tobago 2 M.D. MEDICAL CONSULTANCY LTD, Trinidad and Tobago 3 AMERICAN CANCER SOCIETY, Extramural Discovery Science, Atlanta, GA 4 CANONIC FOUNDATION, Orlando, FL 5 COLLEGE OF MEDICINE, UNIVERSITY OF CENTRAL FLORIDA, Orlando, FL 6 FACULTY OF MEDICINE AND SURGERY, UNIVERSITY OF MALTA, Msida, Malta 7 DIRECTORATE FOR HEALTH INFORMATION AND RESEARCH (DHIR), MINISTRY FOR HEALTH, Malta 8 SAVELIFE.AI, Chicago, IL 9 AMERICAN BOARD OF PRECISION MEDICINE (ABOPM), Nashville, TN 10 DEPARTMENT OF PEDIATRICS, STANFORD UNIVERSITY SCHOOL OF MEDICINE, Stanford, CA

Corresponding Author: Dexter Hadley, MD/PhD, CANONIC Foundation, 6900 Lake Nona Blvd, Orlando, FL 32827. Email: dexter@canonic.org

Word Count: 3,000

Figures/Tables: 3 Figures, 3 Tables

References: 34

Funding/Support: CANONIC Foundation.

Conflict of Interest Disclosures: Dr Hadley and Dr Nimrod are co-inventors on US provisional patent applications PROV-007, PROV-008, and PROV-009, assigned to the CANONIC Foundation. Dr Hadley is inventor on PROV-001. Dr Bajnath is founder and president of the American Board of Precision Medicine. Dr Kalia is founder and CEO of SaveLife.AI. No other disclosures were reported.

Data Sharing Statement: The community learning ledger (date, question text, random UUID), raw ChatGPT API responses, and the automated scoring script are available in the governed repository at https://github.com/hadleylab-canonic. Individual-level data beyond the three-field ledger schema does not exist by design.

Role of the Funder/Sponsor: The CANONIC Foundation provided infrastructure support. The funder had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.

Reporting Guideline: STROBE. Checklist provided in Supplement 2.

KEY POINTS

Question. Does a localized, governed AI cancer navigation platform outperform out-of-the-box AI agents on geographic navigation specificity for Caribbean patients?

Findings. In head-to-head comparison on 28 cancer screening queries across ten Caribbean jurisdictions, a localized platform (CaribChat) cited verified Caribbean facilities in 100% of responses versus 46.4% for ChatGPT (GPT-4o) and provided actionable navigation with specific facility names, addresses, or telephone numbers in 100% versus 35.7% (P<.001 for both). Over 30 days, the community’s engagement matured from screening exploration to active treatment navigation, with substantive query rates rising from 73.7% to 92.7%.

Meaning. Out-of-the-box AI agents fail the majority of Caribbean cancer navigation queries. Localized governance, comprising verified facility registries, culturally governed healing tradition maps, and institutional evidence layers, significantly outperforms general-purpose AI for health navigation in resource-constrained settings, establishing a replicable model for Small Island Developing States worldwide.

ABSTRACT

Importance. Cancer is the second leading cause of death in the Caribbean, where breast cancer mortality ranges from 14 to 30 per 100,000 women and public mammography access is as low as 0.19 per 10,000 population. General-purpose AI agents like ChatGPT are increasingly used for health navigation, but no study has tested whether they can serve populations whose hospitals, referral pathways, healing traditions, and regulatory environments differ fundamentally from the US training context.

Objective. To test whether a localized, governed AI platform significantly outperforms an out-of-the-box AI agent (ChatGPT) on cancer navigation specificity for Caribbean patients, to characterize the thematic evolution of community engagement over time, and to demonstrate that clinically relevant navigation intelligence can accumulate across jurisdictions without data protection legislation and without collecting any personally identifiable information.

Design, Setting, and Participants. Retrospective observational study of community learning ledger data from CaribChat.ai, deployed across ten Caribbean jurisdictions (Trinidad and Tobago, Jamaica, Barbados, Bahamas, Guyana, Saint Lucia, Dominica, Antigua and Barbuda, St. Vincent and the Grenadines, and Grenada) from March 2 through March 31, 2026, with head-to-head comparative analysis against ChatGPT (GPT-4o, OpenAI) conducted April 2, 2026. Community members, clinicians, and cancer society personnel accessed the platform without registration. Filed for exempt determination under 45 CFR 46.104(d)(4)(ii).

Main Outcomes and Measures. Primary: thematic classification and temporal evolution of community learning sessions. Secondary: head-to-head comparative navigation specificity (CaribChat vs ChatGPT on 28 screening queries), healing tradition evidence-tagging completeness, and herb-drug interaction flagging accuracy.

Results. The ledger accumulated 136 sessions over 30 days (4.5/day). Of 114 substantive sessions (83.8%), the thematic landscape evolved across three phases: early exploration dominated by screening and access queries (weeks 1-2), deepening engagement with epidemiology and healing traditions (week 3), and active treatment navigation dominated by survivorship and support (weeks 4-5). Survivorship overtook screening as the largest domain (34 [29.8%] vs 28 [24.6%]), and treatment queries doubled from 11.4% to 15.8%. In head-to-head comparison, CaribChat cited verified Caribbean facilities in 28/28 (100%) screening responses versus 13/28 (46.4%) for ChatGPT (P<.001) and provided actionable navigation in 28/28 (100%) versus 10/28 (35.7%) (P<.001). Indigenous healing tradition queries constituted 11.4% of substantive sessions, all receiving structured evidence-tagged responses with herb-drug interaction flagging. The Organisation of Eastern Caribbean States endorsed the platform across 15 member states.

Conclusions and Relevance. Out-of-the-box AI agents fail the majority of Caribbean cancer screening queries. A localized, governed platform achieved 100% facility citation and 100% actionable navigation where ChatGPT achieved 46.4% and 35.7%, demonstrating that geographic governance is not an incremental improvement but a structural requirement for health AI outside the US training context. The community learning ledger architecture, which accumulates navigation intelligence that matures from exploration to active care without collecting any personally identifiable information, provides a replicable model for Small Island Developing States worldwide.

INTRODUCTION

Cancer is the second leading cause of death across the Caribbean Community (CARICOM) member states.1,8 Breast cancer mortality rates range from 14 to 30 per 100,000 women, with public mammography access as low as 0.19 per 10,000 population in some jurisdictions.7 The emergence of large language models in healthcare has produced AI navigation tools that assume the world looks like the United States: comprehensive cancer centers within driving distance, insurance coverage for guideline-concordant care, and patients who interact through portals in English.11,12 None of these assumptions hold across SIDS. Yet no study has directly compared whether US-default AI systems can serve Caribbean cancer navigation needs.

Caribbean populations maintain healing traditions including bush medicine, faith-based healing, dietary traditions rooted in indigenous foodways, community caregiving networks, and spiritual healing modalities including Obeah.13,14 When a patient asks “Is soursop good for breast cancer?”, the clinically accurate answer requires a structured evidence-tagging schema that no US-based health AI platform provides.15,18 The absence of unified data protection legislation across several Caribbean jurisdictions creates a governance gap for health AI research.17 Conventional compliance frameworks (HIPAA, GDPR) offer no guidance, and research ethics governance must be constructed from institutional authority rather than statutory mandate.10

This paper introduces the community learning ledger as a structural solution. The ledger is an append-only record of anonymized questions, compiled into navigation intelligence that is itself the primary content layer of the platform. The “.ai” in CaribChat.ai refers to this community intelligence, not to artificial intelligence in the conventional sense. We report the thematic evolution of the ledger across ten Caribbean jurisdictions, the head-to-head comparison against ChatGPT on geographic navigation specificity, and the governance architecture that makes both possible.

METHODS

Study Design and Oversight

Retrospective observational study of community learning ledger data from CaribChat.ai across ten Caribbean jurisdictions, with prospective head-to-head comparative analysis against ChatGPT. Arm A of the CANONIC Community Learning Study. Filed for exempt determination under 45 CFR 46.104(d)(4)(ii).9 Institutional governance comprises seven roles (eMethods in Supplement 1).

Platform Architecture

CaribChat.ai requires no registration, authentication, or demographic information. Each session receives a random UUID with no linkage table. The intelligence layer comprises 10 evidence sources including CAOH guidelines, CARPHA surveillance data, NCCN Resource Stratification Framework,2 PAHO/WHO epidemiology,3 and a per-country screening infrastructure registry with 15 verified facilities across 10 jurisdictions (eTable 1 in Supplement 1).

Community Learning Ledger

The ledger records three fields per interaction: date, question text, and random session identifier. No IP addresses, geolocation, device fingerprints, cookies, or PII are collected. Anonymization occurs at capture, not through post-hoc de-identification (eFigure 1 in Supplement 1).

Thematic Classification and Temporal Analysis

All 136 sessions were independently classified using a codebook with 6 thematic domains, 9 cancer types, and 5 jurisdiction categories (eTable 2 in Supplement 1). To assess thematic evolution, the accumulation period was divided into three phases: early (March 2-12, weeks 1-2), middle (March 13-21, week 3), and late (March 22-31, weeks 4-5). Domain proportions were compared across phases. Five healing traditions were mapped with a four-level evidence schema (eTable 3 in Supplement 1).13,14

Comparative Navigation Analysis

All 28 screening and access queries were submitted verbatim to ChatGPT (GPT-4o, OpenAI; temperature=0, no system prompt, no geographic context) via the OpenAI API on April 2, 2026. Each response pair was scored on three binary outcomes: (1) cited a verified Caribbean facility from the screening infrastructure registry, (2) cited a US-specific resource, and (3) provided actionable navigation (specific facility name with address, telephone, or referral pathway). Scoring was performed programmatically using a reproducible script (score_comparison.py) with keyword matching against the facility registry and independently validated by two reviewers (MN, AR). Raw API responses are archived in the governed repository.

Statistical Analysis

Descriptive statistics with exact binomial 95% CIs (Clopper-Pearson). Temporal trends by Fisher exact test. Paired comparison by McNemar test. Python 3.11, scipy 1.12, pandas 2.2.

RESULTS

Community Learning Accumulation and Thematic Evolution

The ledger accumulated 136 sessions over 30 days (March 2-31, 2026), yielding 4.5 sessions/day (31.7/week; 95% CI, 26.5-37.0). Of these, 114 (83.8%) were clinically substantive. The non-substantive rate dropped from 26.3% in weeks 1-3 to 7.3% in weeks 4-5, indicating word-of-mouth acquisition replacing cold discovery (Figure 1).

The thematic landscape evolved across three distinct phases (Figure 2; Table 1). In the early phase (weeks 1-2; 53 substantive sessions), screening and access dominated (34.0%), with healing tradition queries comprising 17.0% of sessions as community members tested the platform’s cultural competency (“What is bush medicine?”, “Is obeah good for cancer?”, “Is castor oil good for lumps”). In the middle phase (week 3; 29 substantive sessions), epidemiology and treatment queries grew as clinicians and public health professionals engaged (“Which CARICOM countries have implemented national cancer control programs”, “What essential medicines for cancer are not available in the Caribbean?”). In the late phase (weeks 4-5; 32 substantive sessions), survivorship and support dominated (37.5%), with patients in active treatment seeking guidance on chemotherapy side effects, radiation timing, lymphedema, Zometa, Herceptin, mastectomy reconstruction, and metastatic diagnosis. Treatment queries doubled across the full period from 11.4% to 15.8%.

This three-phase trajectory, from “what can you tell me?” through “what does the data say?” to “I am in treatment and need help,” demonstrates community maturation rather than novelty-driven decay. The second half of the accumulation period showed higher volume (78 vs 58 sessions), more active days (13 vs 9), and deeper clinical engagement than the first.

One navigation arc illustrated the full depth: a community member reported a delayed biopsy at Mt. Hope, expressed fear of mastectomy (“I worried that they going to have to cut my breast off and it will look ugly”), and was directed to TT Cancer Society navigators with a specific telephone number. Another community member spontaneously contributed HIV testing sites and emergency hotlines from the Trinidad Ministry of Health, demonstrating community ownership of the platform’s intelligence.

Queries spanned 9 cancer types (Table 2) and 10 jurisdictions, with sub-national granularity referencing specific towns (Toco, Gasparillo), hospitals (Mt. Hope, St. James Medical), and services (PET scan, HPV self-testing). Two new jurisdictions (St. Vincent, Antigua) appeared in weeks 4-5, both OECS member states, suggesting endorsement-driven geographic expansion.

Comparative Navigation Analysis

When the same 28 screening queries were submitted verbatim to ChatGPT (GPT-4o), the responses differed dramatically on all three dimensions (Table 3). CaribChat cited verified Caribbean facilities in 28/28 (100%; 95% CI, 87.7%-100%) responses; ChatGPT cited Caribbean facilities in 13/28 (46.4%; 95% CI, 27.5%-66.1%) (P<.001, McNemar test). CaribChat provided actionable navigation in 28/28 (100%) responses; ChatGPT in 10/28 (35.7%; 95% CI, 18.6%-55.9%) (P<.001). ChatGPT cited US-specific resources in 3/28 (10.7%) responses; CaribChat cited zero.

The gap was most pronounced for hyper-local queries. For “Where can I get screened for cancer in Toco?”, CaribChat responded with the TT Cancer Society mobile mammography units serving eastern Trinidad, Sangre Grande Hospital for referral, and the Cancer Society telephone (+1-868-226-1221). ChatGPT mentioned “local health centers” and the “Ministry of Health” generically, with no telephone, no mobile unit schedule, and no Toco-specific navigation. For “What about Gasparillo?”, CaribChat cited San Fernando General Hospital; ChatGPT did not recognize the query as cancer-related. For “What’s their number?”, CaribChat maintained conversational context and provided the facility telephone; ChatGPT responded “I’m sorry, but I can’t assist with that request.”

Secondary Outcomes

All healing tradition queries received four-level evidence-tagged responses (13/13, 100%; 95% CI, 75.3%-100%). Five bush medicine queries with pharmacological activity triggered herb-drug interaction flagging (5/5, 100%; 95% CI, 47.8%-100%), including soursop (CYP3A4 interaction with taxanes), castor oil (GI absorption effects on oral chemotherapy), ganja (cannabinoid-drug interactions), curcumin (anticoagulant potentiation), and apricot seeds (amygdalin/cyanide toxicity).

DISCUSSION

The central finding of this study is that localized governance significantly outperforms out-of-the-box AI agents for health navigation outside the US training context. ChatGPT, the world’s most widely used AI chatbot, cited a verified Caribbean facility in fewer than half of screening responses and provided actionable navigation in barely a third. This is not a marginal gap that prompt engineering can close; it is a structural failure of the data layer. No amount of instruction tuning will teach a model the telephone number of the Trinidad and Tobago Cancer Society, the mobile mammography schedule for eastern Trinidad, or the fact that there are no full oncology services at Mt. Hope Hospital. These facts must be governed, verified, and coupled to the model at the architecture level, which is precisely what the community learning ledger provides.

Second, the community’s engagement evolved rather than decayed. The three-phase trajectory, from screening exploration through epidemiological inquiry to active treatment navigation, demonstrates that community learning compounds over time. By weeks 4-5, patients in active treatment were asking about Zometa, Herceptin, radiation timing, lymphedema, and metastatic diagnosis, queries that no US health AI utilization study has reported from Caribbean populations. The non-substantive rate dropping from 26.3% to 7.3% confirms that word of mouth is replacing cold discovery as the primary acquisition channel.

The 11.4% prevalence of healing tradition queries has no equivalent in published US health AI utilization data. The four-level evidence schema provides a third path between dismissal and endorsement. When the platform acknowledges that soursop acetogenins demonstrate antiproliferative activity in vitro15,18 while noting the absence of clinical trial evidence and flagging CYP3A4 interactions with taxanes, it treats the question with the same clinical seriousness as a chemotherapy query.

The structural anonymization architecture eliminates re-identification risk by collecting only three fields per interaction with no linkage table.16 This provides stronger privacy guarantees than any regulatory framework, because there is no identifiable data to protect regardless of legislative environment.9,17 The same architecture that governs without law in the Caribbean also satisfies GDPR data minimization and EU AI Act auditability requirements.30,31

This study has limitations. The retrospective design precludes causal inference about navigation outcomes. Structural anonymization prevents demographic characterization and distinction between clinician and patient users. The 30-day period is relatively short. Geographic concentration in Trinidad and Tobago (49.1%) reflects initial dissemination channels, though St. Vincent and Antigua queries in weeks 4-5 suggest OECS-driven diversification. The ChatGPT comparison used a single model version (GPT-4o) at a single time point; performance may vary with model updates. Inter-rater reliability will be calculated for the full dataset after the June 2026 data freeze.

Conclusions

Localized, governed AI significantly outperforms out-of-the-box agents for cancer navigation in Small Island Developing States. Over 30 days across ten Caribbean jurisdictions, the platform achieved 100% facility citation and 100% actionable navigation where ChatGPT achieved 46.4% and 35.7%, while the community matured from exploration to active care navigation without collecting any personally identifiable information. The implication extends beyond the Caribbean: every population whose hospitals, referral pathways, and healing traditions differ from the US training context is currently underserved by general-purpose AI. The woman in Sangre Grande who typed “I have a lump in my breast” did not need an AI that knows everything about cancer. She needed one that knows her hospital, her grandmother’s remedies, and the phone number for the navigators who can help her.

REFERENCES

  1. Caribbean Public Health Agency (CARPHA). Cancer Incidence in the Caribbean, Volume I. Port of Spain, Trinidad and Tobago: CARPHA; 2018.
  2. National Comprehensive Cancer Network (NCCN). NCCN Framework for Resource Stratification of NCCN Guidelines. Plymouth Meeting, PA: NCCN; 2024.
  3. Pan American Health Organization (PAHO). Cancer in the Americas: Country Profiles 2023. Washington, DC: PAHO/WHO; 2023.
  4. International Atomic Energy Agency (IAEA). Technical cooperation in breast cancer screening and diagnosis in the Caribbean. Vienna: IAEA; 2022.
  5. World Health Organization. WHO Guideline on Health Policy and System Support to Optimize Community Health Worker Programmes. Geneva: WHO; 2018.
  6. Organisation of Eastern Caribbean States (OECS). Endorsement of CaribChat Community Health Navigation Platform. Castries, Saint Lucia: OECS Secretariat; 2026.
  7. Springer S, McFarlane S, Guthrie K, et al. Mammography utilization in the Caribbean: a situation analysis of six countries. Cancer Causes Control. 2017;28(11):1319-1326.
  8. Cawich SO, Harti G, Griffith J, et al. Cancer in the Caribbean: current status and a framework for cancer control in the region. Ecancermedicalscience. 2023;17:1638.
  9. Protection of Human Subjects, 45 CFR §46.104(d)(4)(ii) (2018).
  10. Hadley D. IRB without a law: governed research ethics in zero-legislation jurisdictions. Published March 18, 2026. Accessed March 31, 2026. https://hadleylab.org/blogs/irb-without-a-law
  11. Singhal K, Azizi S, Tu T, et al. Large language models encode clinical knowledge. Nature. 2023;620(7972):172-180.
  12. Thirunavukarasu AJ, Ting DSJ, Elangovan K, Gutierrez L, Tan TF, Ting DSW. Large language models in medicine. Nat Med. 2023;29(8):1930-1940.
  13. Clement YN, Morton-Gittens J, Basdeo L, et al. Perceived efficacy of herbal remedies by users accessing primary healthcare in Trinidad. BMC Complement Altern Med. 2007;7:4.
  14. Lans C. Ethnomedicines used in Trinidad and Tobago for reproductive problems. J Ethnobiol Ethnomed. 2007;3:13.
  15. Lowe HI, Toyang NJ, Watson CT, Ayeah KN, Bryant J. Unlocking the anticancer potential of soursop (Annona muricata): a systematic review. Molecules. 2017;22(8):1398.
  16. Benitez K, Malin B. Evaluating re-identification risks with respect to the HIPAA privacy rule. J Am Med Inform Assoc. 2010;17(2):169-177.
  17. CARICOM Secretariat. CARICOM Model Harmonisation Legislation on Data Protection. Georgetown, Guyana: CARICOM; 2020.
  18. Moghadamtousi SZ, Fadaeinasab M, Nikzad S, Mohan G, Ali HM, Kadir HA. Annona muricata (Annonaceae): a review of its traditional uses, isolated acetogenins and biological activities. Int J Mol Sci. 2015;16(7):15625-15658.
  19. Balboni TA, Paulk ME, Balboni MJ, et al. Provision of spiritual care to patients with advanced cancer: associations with medical care and quality of life near death. J Clin Oncol. 2010;28(3):445-452.
  20. Martei YM, Pace LE, Brock JE, Shulman LN. Breast cancer in low- and middle-income countries: why we need pathology capability to solve this challenge. Clin Lab Med. 2018;38(1):161-173.
  21. Hadley D, Nimrod M, inventors; CANONIC Foundation, assignee. Community learning ledger for small island developing states. US provisional patent application PROV-007. Filed 2026.
  22. Hadley D, Nimrod M, inventors; CANONIC Foundation, assignee. Governed AI navigation with cultural competency attestation. US provisional patent application PROV-008. Filed 2026.
  23. Hadley D, Nimrod M, inventors; CANONIC Foundation, assignee. Federated health governance across zero-legislation jurisdictions. US provisional patent application PROV-009. Filed 2026.
  24. Hadley D, inventor; CANONIC Foundation, assignee. MAGIC governance language with explainable AI. US provisional patent application PROV-001. Filed February 25, 2026.
  25. CARICOM. Communique of the Caribbean Commission on Health and Development. Georgetown, Guyana: CARICOM Secretariat; 2023.
  26. Roach A, Scantlebury-Manning T, Gittens-St Hilaire M, et al. Cancer incidence in Barbados, West Indies. Cancer Causes Control. 2019;30:1279-1293.
  27. Roach A. Research ethics in resource-limited settings: establishing institutional review boards in communities without prior research infrastructure. Bioethics. 2018;32(8):554-563.
  28. Hadley D. The $255 billion wound: American healthcare wastes $255 billion a year on governance it cannot prove. Published February 28, 2026. Accessed March 31, 2026. https://hadleylab.org/papers/the-255-billion-dollar-wound
  29. Luciani S, Alcantar-Lopez M, Goss P, et al. Breast cancer in the Americas: an update on recent epidemiological, clinical, and programmatic developments. Lancet Reg Health Am. 2024;31:100694.
  30. European Parliament and Council. Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (AI Act). Official Journal of the European Union. 2024.
  31. European Parliament and Council. Regulation (EU) 2025/327 on the European Health Data Space (EHDS). Official Journal of the European Union. 2025.
  32. Bossert T, Bowser D, Amenyah J. Is decentralization good for logistics systems? Evidence on essential medicine availability and stockouts from Ghana. Health Policy Plan. 2007;22(2):73-82.
  33. Lofters A, Slater M, Angl EN, Leung FH. Brain drain and not enough gain: Canadian-based medical trainees from the Caribbean. Can Fam Physician. 2014;60(3):277-281.
  34. Aguilera B, Carracedo S, Saenz C. Systemic assessment of research ethics systems across Latin America and the Caribbean. Lancet Glob Health. 2022;10(8):e1204-e1208.

FIGURE LEGENDS

Figure 1. Session accumulation by date (N=136) with thematic phase overlay. Early phase (weeks 1-2): screening-dominated exploration. Middle phase (week 3): epidemiology and treatment queries emerge. Late phase (weeks 4-5): survivorship and active treatment navigation dominate. Non-substantive rate drops from 26.3% to 7.3%.

Figure 2. Thematic domain evolution across three phases (N=114 substantive sessions). Survivorship and support (29.8%) overtook screening and access (24.6%) as the dominant domain, while treatment queries doubled from 11.4% to 15.8%.

Figure 3. Comparative navigation specificity: CaribChat vs ChatGPT (GPT-4o) on 28 screening queries. CaribChat achieved 100% on both Caribbean facility citation and actionable navigation; ChatGPT achieved 46.4% and 35.7% respectively.

TABLES

Table 1. Thematic Domain Distribution of 114 Substantive Sessions

Domain No. (%) Early (wk 1-2) Late (wk 4-5) Representative Queries
Survivorship and Support 34 (29.8) 20.8% 37.5% “I just got diagnosed with metastatic breast cancer”; “How does one treat lymphedema”; “How do I deal with depression”
Screening and Access 28 (24.6) 34.0% 18.8% “Where can I get screened in Port of Spain?”; “What cancer services exist in St. Vincent”; “Where can we find a PET scan?”
Epidemiology and Surveillance 18 (15.8) 11.3% 15.6% “How many cancer registries are there in the Caribbean”; “Incidence of cervical cancer in Jamaica”
Treatment and Guidelines 18 (15.8) 11.3% 21.9% “What are the treatment options in Trinidad for multiple myeloma”; “What is the survival rate for triple positive breast cancer”
Healing Traditions 13 (11.4) 17.0% 6.3% “What is bush medicine?”; “Is soursop good for breast cancer?”; “Are apricot seeds good for cancer?”
Platform Discovery 3 (2.6) 5.7% 0.0% “Who is Marisa Nimrod?”; “I heard there is a conference coming up”

Table 2. Cancer Type Distribution (N=63 Cancer-Specific Queries)

Cancer Type No. (%)
Breast 22 (34.9)
General oncology 21 (33.3)
Prostate 7 (11.1)
Hematologic (myeloma, leukemia, lymphoma) 6 (9.5)
Colorectal 3 (4.8)
Cervical 2 (3.2)
Esophageal 1 (1.6)
Glioblastoma 1 (1.6)
Kidney (transplant-related) 1 (1.6)

Table 3. Comparative Navigation Specificity: CaribChat vs ChatGPT (N=28 Screening Queries)

Metric CaribChat ChatGPT (GPT-4o) P Value
Cited verified Caribbean facility, No. (%; 95% CI) 28 (100; 87.7-100) 13 (46.4; 27.5-66.1) <.001
Cited US-specific resource, No. (%) 0 (0) 3 (10.7) .25
Provided actionable navigation, No. (%; 95% CI) 28 (100; 87.7-100) 10 (35.7; 18.6-55.9) <.001

Queries submitted verbatim via OpenAI API on April 2, 2026. Model: GPT-4o, default settings, temperature=0, no system prompt, no geographic context added. P values by McNemar test. Scoring script (score_comparison.py), raw API responses (chatgpt_responses.json), and structured scores (comparison_scores.json) available in governed repository for full reproducibility.

AUTHOR CONTRIBUTIONS

Concept and design: Hadley, Nimrod. Acquisition, analysis, or interpretation of data: Nimrod, Roach, Wray, Fraser, Hadley. Drafting of the manuscript: Hadley. Critical revision of the manuscript for important intellectual content: All authors. Statistical analysis: Hadley. Administrative, technical, or material support: Wray, Fraser, Kalia, Bajnath. Supervision: Hadley, Nimrod, Calleja. Clinical content validation: Nimrod, Bajnath. Research ethics oversight: Roach, Calleja. European regulatory expertise: Calleja.

eSUPPLEMENT (Supplement 1)

The following supplemental materials are available in Supplement 1:

Supplement 2: STROBE Checklist

COMMUNITY-LEARNING | SUBMISSION | JAMA NETWORK OPEN

CANONIC ∩ FOUNDATION