知镜 Cognispect · 医学文献与研究工作台

From microscopy to antimicrobial decisions: a clinically grounded roadmap for critical care infectious diseases.

Qu HL, Li JN, Gao Y, Xu XM, Zhang XB, Yang SD.
Frontiers in artificial intelligence · 2026-04-13 · PMID 42052210

文献摘要

In the intensive care unit (ICU), antibiotics often begin under extreme uncertainty. Fever, leukocytosis, hypotension, and organ dysfunction may signal bacterial infection, but the same findings are common with aspiration, post-operative inflammation, drug reactions, or sterile systemic inflammation. Cultures take time and their yield falls after antibiotics. Rapid molecular tests and metagenomics can add actionable information, but they also raise the burden of interpreting complex results. Microscopy is one of the few inputs that can shift management within minutes to hours: Gram-stain patterns from positive blood-culture bottles, respiratory specimens, cerebrospinal fluid, and wound material can reshape initial coverage and support early de-escalation when negative. Tissue and cytology help distinguish invasion from key mimics. The gap is consistency-reads vary across observers, workflows differ, and results do not always translate into reliable bedside actions. This review focuses on infectious-disease artificial intelligence (AI) as ICU bedside decision support, rather than as a survey of models. Using ICU sepsis as the primary use case-and neurocritical care as a challenging setting where sedation, brain injury, and noninfectious inflammation often mimic infection-we separate evidence into pathogen signals and host-response signals. We then map both streams to six decisions over the first 72 hours: start now versus pause, choose initial spectrum, reassess and narrow, escalate diagnostics and source control, act on high-risk resistance or invasive pathogens, and stop safely. We summarize where AI is most credible today (Gram-stain assistance, culture-plate triage, urine-culture screening, infection-focused digital pathology, host-response classifiers, and selected metagenomics) and what makes outputs actionable: calibrated probabilities, explicit confidence with safe deferral when uncertain, validation across hospitals and instruments, and endpoints tied to stewardship and safety (time to appropriate therapy, antibiotic days, de-escalation within 72 hours, missed bacteremia). Evidence was updated through February 28, 2026.

在知镜精读原文来源DOI 原文

知镜提供医学文献检索、阅读笔记、人体图谱、统计绘图与研究复现工具。