EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Assessing Socioeconomic Bias in AI-Powered Maternal Health Messaging: Evidence from Jacaranda Health’s PROMPTS Platform in Kenya

Anneka Wickramanayake, Chemeli Priscah Cheruiyot, Jay Patel, Katherine Jennings, Sharon Ochieng, Catherine Gakii, Lyvia Lusiji, Kelvin Ndambuki

Oxford Open Digital Health · 13.11.2026

DOI: 10.1093/oodh/oqag021 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Oxford Open Digital Health, online erschienen am 13.11.2026. Verfasst von Anneka Wickramanayake und 7 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Background

In many low- and middle-income countries (LMICs), mothers lack timely, accurate health information, which can delay lifesaving care. Jacaranda Health’s PROMPTS platform reaches 3 million+ mothers in Kenya via SMS, using artificial intelligence (AI) to triage questions, provide rapid responses, and connect urgent cases to human agents. While AI can expand access, if not carefully designed it may reinforce inequities. This study examines whether PROMPTS’s AI models perform differently across socioeconomic groups.

Methods

We applied a four-phase bias assessment framework. First, we surveyed 4569 PROMPTS users across 22 counties and assessed socioeconomic vulnerability using a validated typing tool. Second, we examined whether training datasets reflected users across vulnerability strata. Third, we evaluated AI performance in classifying question topics. Finally, we audited AI-generated outputs for medical accuracy, grammar, and thoughtfulness, stratified by vulnerability.

Results

PROMPTS engages women across vulnerability levels. The most vulnerable mothers were less likely to ask questions, suggesting underrepresentation in training data. The AI model correctly classified 90% of questions, with no significant differences by vulnerability, though confidence was lower for messages from more vulnerable mothers. AI-generated responses were equally accurate across groups, but answers to more vulnerable users were sometimes rated as less thoughtful.

Conclusion

PROMPTS demonstrates a fair application of AI for maternal health, with strong performance across diverse groups. However, underrepresentation of vulnerable mothers and subtle differences in response quality highlight areas for improvement. Engaging vulnerable voices, embedding safeguards, and routine monitoring will be critical to ensure AI advances health equity in LMIC contexts.

Zitieren

APA

Wickramanayake, A., Cheruiyot, C. P., Patel, J., Jennings, K., Ochieng, S., Gakii, C., Lusiji, L., & Ndambuki, K. (2026). Assessing Socioeconomic Bias in AI-Powered Maternal Health Messaging: Evidence from Jacaranda Health’s PROMPTS Platform in Kenya. Oxford Open Digital Health. https://doi.org/10.1093/oodh/oqag021

Vancouver

Wickramanayake A, Cheruiyot CP, Patel J, Jennings K, Ochieng S, Gakii C, et al. Assessing Socioeconomic Bias in AI-Powered Maternal Health Messaging: Evidence from Jacaranda Health’s PROMPTS Platform in Kenya. Oxford Open Digital Health. 2026. doi:10.1093/oodh/oqag021

BibTeX
@article{Wickramanayake2026,
  author = {Wickramanayake, Anneka and Cheruiyot, Chemeli Priscah and Patel, Jay and Jennings, Katherine and Ochieng, Sharon and Gakii, Catherine and Lusiji, Lyvia and Ndambuki, Kelvin},
  title = {Assessing Socioeconomic Bias in AI-Powered Maternal Health Messaging: Evidence from Jacaranda Health’s PROMPTS Platform in Kenya},
  journal = {Oxford Open Digital Health},
  year = {2026},
  doi = {10.1093/oodh/oqag021}
}
Kostenlos starten →