Large Language Model-Based Clinical Decision Support for Antibiotic Selection and Dose Recommendation in Hospitalized Patients With Pneumonia: Multicenter Retrospective Study.

Background: Pneumonia is a common infectious disease, and antibiotic treatment in hospitalized patients must balance efficacy, safety, and resistance risk. However, antibiotic selection and dose adjustment still rely heavily on clinician experience. Although large language models (LLMs) are promising for clinical reasoning, their direct use for antibiotic selection and dose recommendation is limited by hallucinations and weak adherence to clinical constraints. Objective: This study aimed to develop and externally validate a constrained LLM-based clinical decision support pipeline for antibiotic selection and dose recommendation in hospitalized patients with pneumonia. Methods: We conducted a multicenter retrospective study using electronic health record narratives, antibiotic orders, and laboratory indicators of hepatic and renal function from 331 hospitalized patients with pneumonia from 2 hospitals in China. The development cohort included 233 patients, and the external validation cohort included 98 patients. The pipeline integrated dual-branch retrieval (similar-case vector retrieval plus guideline-based knowledge graph retrieval), clinician-defined rule constraints, and hybrid-context reasoning. DeepSeek-V3, GLM-4.6, and GPT-4o were evaluated using F1-score and Jaccard accuracy. Results: On the internal test set, the full pipeline using DeepSeek-V3 achieved the best performance, with an F1-score of 0.8110 (95% CI 0.7371-0.8762) and Jaccard accuracy of 0.7624 (95% CI 0.6810-0.8386) for antibiotic selection and an F1-score of 0.7538 (95% CI 0.6671-0.8329) and Jaccard accuracy of 0.7076 (95% CI 0.6145-0.7938) for joint antibiotic selection plus dosing recommendation. On the external validation set, performance remained high, with an F1-score of 0.8605 (95% CI 0.7891-0.9252) and Jaccard accuracy of 0.8571 (95% CI 0.7857-0.9184) for antibiotic selection, and an F1-score of 0.8503 (95% CI 0.7789-0.9150) and Jaccard accuracy of 0.8469 (95% CI 0.7755-0.9133) for antibiotic selection plus dosing recommendation. The system also provided traceable evidence and rule trigger information to support clinician review. Conclusions: A constrained, retrieval-augmented LLM pipeline improved the consistency and interpretability of antibiotic selection and dose recommendation for hospitalized patients with pneumonia and provided preliminary evidence of cross-site generalizability.