Large Language Model-Assisted Clinicians versus Unassisted Clinicians in Clinical Decision Making: Protocol for a Multi-facility Pragmatic Cluster Randomized Controlled Trial in Nairobi, Kenya

Background In Kenya, primary healthcare is challenged by limited resources, high demand for services, and complex clinical presentations, contributing to risks such as misdiagnosis, delayed treatment, and inappropriate prescribing of treatments. Large language models (LLMs) may offer a scalable means of supporting frontline healthcare workers to deliver safer, more accurate care. However, their effectiveness, safety, and acceptability in real-world, low-resource clinical settings remain largely untested. Methods This multi-facility, parallel-group, pragmatic cluster-randomized trial will evaluate the effectiveness of an LLM-enabled clinical decision support system (CDSS) integrated within an electronic medical record (EMR) in outpatient primary care settings in Nairobi, Kenya. A total of 9,000 patients will be enrolled across 16 primary care facilities. Clinicians will be randomized to an intervention group (EMR with LLM-based CDSS enabled) or a control group (EMR with LLM-based CDSS disabled, i.e., clinician-only). In the intervention arm, clinicians will receive automated, LLM-generated recommendations during consultations. The primary outcome is the proportion of participants experiencing at least one episode of treatment failure, defined as a composite of patient re-presentation to primary care with unresolved symptoms from the initial visit, escalation of care, or related adverse events within 14 days of enrolment. Secondary outcomes include diagnostic accuracy, quality of clinical documentation, alignment of treatment of sentinel diseases with local clinical guidelines, appropriateness of antibiotics and antimalarial prescriptions, and patient satisfaction.  Discussion This study represents the first randomized controlled trial of an LLM-based CDSS in primary care in Africa. Findings will inform the feasibility and safety of integrating generative AI into frontline services in low-resource contexts, with the potential to improve quality of care, reduce healthcare costs, and optimize resource use in Kenya and similar settings. 1-2 Sentence Description This multi-facility, pragmatic cluster-randomized phase III trial will evaluate the effectiveness of a large language model-powered, electronic medical record system-integrated, clinical decision support system in primary care clinics in Kenya.