目的

血液肿瘤患者因长期免疫抑制及粒细胞缺乏，成为铜绿假单胞菌（PA）等革兰氏阴性菌感染的高危人群。PA感染相关死亡率和对碳青霉烯类药物的耐药率（CRPA）显著高于肠杆菌科，CRPA的持续定植与反复感染特性，加之抗生素选择压力下基因组的快速进化，导致治疗窗口期缩短、经验性抗生素选择困难，成为威胁患者生存的突出问题。为此，本研究旨在：（1）基于多中心临床数据与机器学习构建PA多药耐药预测模型，优化经验性抗生素决策；评估短程抗生素（7-11天）及联合治疗在血液肿瘤患者中的安全性与疗效，优化治疗策略；（2）探索CRPA在免疫抑制宿主中的定植-感染转化规律；（3）探索CRPA在头孢他啶-阿维巴坦（CZA）压力下的耐药进化机制和生物学行为改变。

方法 

1. 临床队列研究：整合中国三家血液病中心（2014-2024年）的血液肿瘤患者合并PA血流感染的临床和微生物数据（924例），回顾性分析耐药特征及预后危险因素，利用机器学习构建多种耐药预测模型（碳青霉烯类耐药，哌拉西林他唑巴坦耐药，头孢菌素类耐药，喹诺酮类耐药，多重耐药菌），通过多模型预测结果和指南共识指导用药体系构建数据驱动的临床决策系统，在测试集中评估模型对临床实践的影响，结局指标为不恰当经验性抗生素使用率和经验性碳青霉烯使用率；开展短疗程（7-11天）与长疗程（12-21天）治疗疗效对比；开展单药与联合抗生素在CRPA治疗中的疗效对比。 

2. 分子机制研究：对本中心血液肿瘤患者的339株CRPA分离株进行全基因组测序和生物膜能力测定，解析耐药、毒力基因特征；设计病例对照研究（Case组，47例定植感染患者（140株分离株） vs  Control组，43例持续定植患者（106株分离株）），进行生物膜测定、全基因组测序、转录组测序，利用比较基因组学分析，结合贝叶斯广义线性混合模型（GLMM）和逻辑回归模型（LR）筛选感染相关基因，利用SNP分析明确进化规律，利用比较转录组学分析明确感染特定基因表达。 

3. 耐药演化研究：利用表型基因型关联分析筛选CZA耐药菌株高频变异位点；纵向追踪11例CZA治疗后诱导耐药患者的50株CRPA（血流、黏膜及肛周分离株），通过全基因组SNP分析、转录组分析探索体内CZA耐药进化规律，利用体外传代耐药诱导实验验证，揭示耐药分子机制和生物学行为改变。 

结果

1. PA耐药预测模型和抗生素策略优化： 来自三个血液中心的924例PA BSI患者被纳入研究， CRPA感染占 20.2% (n=187)，不恰当经验性治疗（IET48h，OR=11.991，95%CI 6.419-22.398，P<0.001）是独立预后不良因素，多种耐药预测模型驱动的决策支持系统可显著降低不恰当经验治疗比例（测试集1，实际11.86%，模型推荐3.95%，p=0.0105；测试集2，实际20.81%，模型推荐7.11%，p=0.0002），并减少经验性碳青霉烯类使用率（测试集1，实际使用vs模型一线推荐vs模型二线推荐，68.9% vs 1.7% vs 41.3%, p<0.0001；测试集2，实际使用vs模型一线推荐vs模型二线推荐，40.1% vs 0.5% vs 43.2%, p<0.0001）。在抗生素管理规范研究中，短疗程（7-11天）与长疗程疗效无差异（30天死亡率5.9% vs 4.6%，p>0.05），但住院时间缩短4天（p=0.011）；在CRPA治疗中，单药治疗和联合治疗具有相似的临床结局和耐药菌再定植率（p>0.05），而IET48h是30d死亡率（p<0.001）和获得性CRPA定植（p=0.013）的独立危险因素。

2. 耐药菌定植感染机制：血液肿瘤CRPA分离株基因组呈现多样性，耐药以外排泵为核心机制，碳青霉烯酶（IMP-1/DIM-1）仅占3%以下；毒力基因中鞭毛蛋白、分泌系统和生物膜相关基因高度保守，高毒力菌株exoS+/exoU+的检出率为2.9%。GLMM模型显示宿主因素（合并症、接受免疫抑制剂治疗、淋巴细胞减少、单核细胞增高、肺炎、既往暴露有喹诺酮类药物）及菌株因素（低生物膜能力）为感染独立危险因素。比较基因组显示分解代谢酶相关基因（藻酸盐裂解、戊糖和糖醛酸相互转化通路、维生素B6代谢通路）和毒力基因（铁载体）是感染独立基因，比较转录组分析提示提示感染组高表达基因富集在硫代谢和CAMP通路，而群体感应通路、吩嗪合成、细菌趋化、O-Antigen核苷糖生物合成基因下调。

3. CZA耐药演化：血液肿瘤CRPA分离株的CZA耐药率逐年攀升，CZA耐药株生物膜合成能力增加（2.261 vs 1.388, p=0.0005），体内外研究显示耐药菌双组分系统基因高频变异，代谢活性降低，激活生物膜合成和外排泵，导致CZA治疗失败，同时耐药菌株运动能力下降，中性粒细胞清除效率下降，促使免疫逃逸。

结论

本研究构建的PA多药耐药预测模型显著提升经验性治疗准确性，短程抗生素治疗在部分血液肿瘤患者中安全可行，联合治疗在降低CRPA血流感染死亡率方面较单药并无获益。通过基因组与转录组分析，发现血液肿瘤患者中CRPA耐药基因以外排泵为主，代谢基因是CRPA感染特定基因。CRPA通过生物膜-免疫逃逸模式主导CZA耐药，并达到宿主内长期持留的目的。研究结果为血液肿瘤患者耐药菌感染的精准防控提供了数据驱动的决策支持与分子干预靶点，对改善临床预后及延缓耐药进化具有重要意义。

Objective

Patients with hematologic malignancies are at high risk of infections caused by Gram-negative pathogens such as Pseudomonas aeruginosa (PA) due to prolonged immunosuppression and neutropenia. PA infections exhibit higher mortality and carbapenem-resistance rates (CRPA) compared to Enterobacteriaceae. Persistent colonization and recurrent infections by CRPA, complicated with rapid genomic evolution under antibiotic selection pressure, lead to narrowed therapeutic windows and challenges in empirical antibiotic selection, posing critical threats to patient survival. This study aims to: (1) Develop a machine learning-based multi-drug-resistance prediction model for PA using multicenter clinical data to optimize empirical antibiotic decision-making; evaluate the safety and efficacy of short-course (7-11 days) and combination therapies in hematologic malignancy patients; (2) Investigate the colonization-infection transition dynamics of CRPA in immunocompromised hosts; (3) Elucidate resistance evolution mechanisms and phenotypic adaptations of CRPA under ceftazidime-avibactam (CZA) pressure.

Methods

Clinical Cohort Study: Retrospective analysis of clinical and microbiological data from 924 hematologic malignancy patients with PA bloodstream infections (BSI) across three Chinese centers (2014–2024). Machine learning models were developed to predict resistance to carbapenems, piperacillin-tazobactam, cephalosporins, fluoroquinolones, and multidrug resistance. A data-driven clinical decision support system (CDSS) was constructed using model predictions and guideline recommendations. Outcomes included inappropriate empirical therapy (IET) rates and carbapenem usage. Short-course (7-11 days) vs. long-course (12-21 days) therapies and monotherapy vs. combination therapy for CRPA were compared.

Molecular Mechanism Investigation: Conducted whole-genome sequencing and biofilm quantification on 339 CRPA isolates. A case-control design (47 colonization-infection cases [140 isolates] vs. 43 persistent colonization controls [106 isolates]) was employed for comparative genomics, transcriptomics, biofilm analysis, and Bayesian generalized linear mixed models (GLMM) to identify infection-associated genes and evolutionary patterns.

Resistance Evolution Analysis: Phenotypic-genotypic correlation analysis identified high-frequency CZA resistance variants. Longitudinal tracking of 50 CRPA isolates (bloodstream, mucosal, and perianal) from 11 patients with CZA-induced resistance was conducted. Whole-genome SNP analysis, transcriptomics, and in vitro resistance induction experiments were performed to elucidate resistance mechanisms and phenotypic adaptations.

Results

Resistance Prediction Model and Antibiotic Strategy Optimization: Among 924 PA BSI cases, CRPA infections accounted for 20.0% (n=185). IET within 48 hours (OR=11.991, 95% CI 6.419–22.398, P<0.001) was an independent prognostic risk factor. The CDSS significantly reduced IET rates (Test Set 1: actual 11.86% vs. model-recommended 3.95%, P=0.0105; Test Set 2: 20.81% vs. 7.11%, P=0.0002) and empirical carbapenem use (Test Set 1: 68.9% vs. 1.7% vs. 41.8%, P<0.0001; Test Set 2: 40.1% vs. 0.5% vs. 43.2%, P<0.0001). Short-course therapy showed comparable 30-day mortality to long-course therapy (5.9% vs. 4.6%, P>0.05) but reduced hospitalization by 4 days (P=0.011). For CRPA treatment, monotherapy and combination therapy exhibited similar outcomes (P>0.05), while IET independently predicted 30-day mortality (P=0.001) and CRPA colonization (P=0.041).

Colonization-Infection Mechanisms: The genome of CRPA isolates from hematologic patients showed diversity, drug resistance extracellular pump was the core mechanism, and carbapenemase (IMP-1/DIM-1) accounted for less than 3%. In virulence genes, flagellin, secretory system and biofilm-related genes were highly conserved, and the detection rate of exoS+/exoU+ was 2.9%. GLMM identified host factors (comorbidities, immunosuppressive therapy, lymphopenia, monocytosis, pneumonia, prior fluoroquinolone exposure) and strain factors (low biofilm capacity) as infection risks. Comparative genomics revealed infection-associated genes in catabolism (alginate lyase, pentose/glucuronate interconversion, vitamin B6 metabolism) and virulence (siderophores). Transcriptomics highlighted upregulated sulfur metabolism/CAMP pathways and downregulated quorum sensing, phenazine synthesis, chemotaxis, and O-antigen biosynthesis in infection strains.

CZA Resistance Evolution: CRPA resistance to CZA increased annually, with resistant strains exhibiting enhanced biofilm formation (2.261 vs 1.388, p=0.0005). In vivo and in vitro studies identified frequent mutations in two-component systems, reduced metabolic activity, activated biofilm and efflux pump system, and impaired motility. These adaptations facilitated CZA treatment failure. At the same time, the locomotor ability and neutrophil clearance efficiency of drug-resistant strains decreased, which promoted immune escape and host persistence.

Conclusions

The PA multidrug resistance prediction model significantly improved empirical therapy accuracy. Short-course antibiotic therapy is safe for part of hematologic malignancy patients, while combination therapy offers no mortality advantage over monotherapy for CRPA BSI. Genomic-transcriptomic analyses revealed efflux pump-driven resistance and metabolic gene signatures in CRPA infections. CRPA employs biofilm-mediated immune evasion to achieve CZA resistance and host persistence. This study provides data-driven decision support and molecular targets for precision management of resistant infections in hematologic malignancies, with critical implications for clinical outcomes and resistance mitigation.