AI-based Clinical Decision Support for Primary Care: A Real-World Study
1
Ministry of Health, Nairobi, Kenya
2
Penda Health, Nairobi, Kenya
3
Open AI, Carlifornia, United States
Popul. Med. 2026;8(Supplement Supplement 1):A770
ABSTRACT
INTRODUCTION:
Primary care in urban Kenya faces diagnostic and treatment errors driven by high volumes and limited diagnostics. We evaluate an EMR-embedded, large-language-model clinical decision-support (LLM-CDS) tool designed as a safety net that fits routine workflow. Objectives are to quantify effects on error rates, assess usability and adoption, and identify implementation enablers. We hypothesize that LLM-CDS reduces clinically meaningful errors versus usual care, is usable and acceptable to clinicians, and that workflow-aligned design increases adoption of the tool.
METHODS:
In a pragmatic, provider-level, cluster-assigned evaluation, clinicians were block-randomized to LLM-CDS access or usual care. The study took place in high-volume urban primary care clinics with mature EMRs. Eligible in-person clinician visits were included; OTC, lab-only, well-child, chronic-program, telemedicine, and dental visits were excluded. All eligible visits informed endpoints, with a stratified subsample reviewed blindly. De-identified data were analyzed using mixed-effects models, and adoption summarized through surveys/logs under approved ethics.
RESULTS:
Across 39,849 visits (AI 20,589; comparison 18,990) in 15 clinics, access to LLM-CDS reduced physician-rated errors: history taking by 31.8% reduction, investigations by 10.3%, diagnosis by 16.0% and treatment by 12.7%. Projected to annual volume, this equates to 22,102 fewer diagnostic-error visits and 28,880 fewer treatment-error visits. Patient-reported outcomes were similar (not feeling better at day 8: 3.8% AI vs 4.3% comparison). All surveyed clinicians with access reported quality improvement; 75% rated it “substantial.” No patient-safety reports indicated harm attributable to the tool. Effects strengthened post-induction, and red-severity alerts fell from 45% to 35% as adoption improved.
CONCLUSIONS:
Embedding LLM clinical decision support into routine primary care workflows improved patient-safety processes without displacing clinician judgment. Health systems should scale carefully: integrate within EMRs, monitor safety, train providers, and iteratively localize content. Policymakers can adopt this blueprint to reduce preventable harm while strengthening guideline adherence across resource-constrained settings.
Primary care in urban Kenya faces diagnostic and treatment errors driven by high volumes and limited diagnostics. We evaluate an EMR-embedded, large-language-model clinical decision-support (LLM-CDS) tool designed as a safety net that fits routine workflow. Objectives are to quantify effects on error rates, assess usability and adoption, and identify implementation enablers. We hypothesize that LLM-CDS reduces clinically meaningful errors versus usual care, is usable and acceptable to clinicians, and that workflow-aligned design increases adoption of the tool.
METHODS:
In a pragmatic, provider-level, cluster-assigned evaluation, clinicians were block-randomized to LLM-CDS access or usual care. The study took place in high-volume urban primary care clinics with mature EMRs. Eligible in-person clinician visits were included; OTC, lab-only, well-child, chronic-program, telemedicine, and dental visits were excluded. All eligible visits informed endpoints, with a stratified subsample reviewed blindly. De-identified data were analyzed using mixed-effects models, and adoption summarized through surveys/logs under approved ethics.
RESULTS:
Across 39,849 visits (AI 20,589; comparison 18,990) in 15 clinics, access to LLM-CDS reduced physician-rated errors: history taking by 31.8% reduction, investigations by 10.3%, diagnosis by 16.0% and treatment by 12.7%. Projected to annual volume, this equates to 22,102 fewer diagnostic-error visits and 28,880 fewer treatment-error visits. Patient-reported outcomes were similar (not feeling better at day 8: 3.8% AI vs 4.3% comparison). All surveyed clinicians with access reported quality improvement; 75% rated it “substantial.” No patient-safety reports indicated harm attributable to the tool. Effects strengthened post-induction, and red-severity alerts fell from 45% to 35% as adoption improved.
CONCLUSIONS:
Embedding LLM clinical decision support into routine primary care workflows improved patient-safety processes without displacing clinician judgment. Health systems should scale carefully: integrate within EMRs, monitor safety, train providers, and iteratively localize content. Policymakers can adopt this blueprint to reduce preventable harm while strengthening guideline adherence across resource-constrained settings.
| eISSN: | 2654-1459 |
We process personal data collected when visiting the website. The function of obtaining information about users and their behavior is carried out by voluntarily entered information in forms and saving cookies in end devices. Data, including cookies, are used to provide services, improve the user experience and to analyze the traffic in accordance with the Privacy policy. Data are also collected and processed by Google Analytics tool (more).
You can change cookies settings in your browser. Restricted use of cookies in the browser configuration may affect some functionalities of the website.
You can change cookies settings in your browser. Restricted use of cookies in the browser configuration may affect some functionalities of the website.
