Clinical Decision Support with Large Language Models: A Theoretical Framework for Safe Integration into Resource-Constrained Health Systems

Abstract

Large language models (LLMs) can now summarise records, suggest differential diagnoses, support triage and converse with patients, and the health systems with the fewest clinicians face the strongest pressure to adopt them. Yet the evidence on their clinical use comes mostly from high-income settings, and existing guidance treats LLM-based clinical decision support as one undifferentiated technology rather than a set of tasks with different risks. This study aimed to specify where in the clinical workflow an LLM may safely sit in resource-constrained health systems, the failure modes characteristic of each position, and the safeguards each position requires. A theoretical, integrative theory-building design was used. The study population comprised peer-reviewed journal articles published from 2023 to 2026 on LLMs and tool-using agentic systems, clinical decision support, patient safety, governance, implementation and the health workforce in low-resource settings. Purposive, relevance-ranked sampling across eight search strings in PubMed/MEDLINE yielded 265 unique records; 213 were screened against predefined criteria, and 42 were retained, together with five landmark articles added by digital object identifier, giving a final sample of 47 articles. Every included article was verified against its Crossref publisher deposit, and none had been retracted. Data were extracted into a structured matrix and analysed by deductive and inductive thematic synthesis. Only 13 of the 47 articles were set in Africa, and only two reported prospective use of an LLM by African clinicians. The synthesis produced three outputs. The first is a taxonomy of four workflow positions of rising risk: documentation support, triage and referral support, differential diagnosis and management planning, and direct patient communication. The second is a five-domain safeguard architecture whose intensity scales with the risk of the position. The third is six testable propositions and seven design principles for bounded clinical autonomy. The central argument is that risk depends less on model capability than on interception distance, defined as the number of independent human checks between a model's output and the patient. The study recommends that health ministries, regulators and developers choose the workflow position before choosing the model, ground outputs in nationally adapted guidelines, default to referral under uncertainty, and grant greater autonomy only on the evidence of post-deployment monitoring.