There is a crucial need for a reliable prediction model for infection to guide risk stratification and precise preventive treatment on patients with hematological conditions, in which infection is a major cause of morbidity and mortality. In the absence of such a prediction model, many patients receive prophylactic antibiotics roughly based on the overall infection risk evaluated by disease and therapy [4,5]. In this prospective multicenter study, we characterized the different microecological landscapes between patients without neutropenia prechemotherapy and patients with neutropenia postchemotherapy with hematological diseases by performing mNGS. Using mNGS data and machine learning, we derived a prediction model for infection that can accurately distinguish patients with hematological conditions at low risk versus high risk. The prediction model for infection is particularly important because it will be crucial for precise preventive treatment and reducing the occurrence of infection, as well as reducing the abuse of antibiotics and decreasing antimicrobial resistance.

Patients with hematological conditions frequently develop treatment-induced neutropenia. Under this critical immunosuppressed state, most become vulnerable to life-threatening infections. In our study, following the disruption of the granulocyte barrier, cohort B demonstrated obvious microbial alterations compared with cohort A, characterized by an increased proportion of Acinetobacter and Pseudomonas. Notably, certain Acinetobacter including Acinetobacter sp. MYb10, Acinetobacter sp. TAC-1, and Pseudomonas such as Pseudomonas brenneri, Brevundimonas vesicularis, Pseudomonas otitidis, became predominant in cohort B. Among them, Pseudomonas otitidis is a novel species of Pseudomonas bacteria that has been associated with otic infections, pneumonia, and epididymo-orchitis [25,26]. Compared with cohort A, Burkholderia contaminans was also increased in cohort B, which is capable of airway colonization and may subsequently lead to infections [27,28]. Concurrently, the important microorganisms used to distinguish between patients without neutropenia prechemotherapy and patients with neutropenia postchemotherapy, such as Acinetobacter johnsonii, Caulobacter flavus, Neisseria elongata, and Corynebacterium propinquum showed obvious elevation in patients with neutropenia, all representing clinically significant pathogens. Specifically, Acinetobacter johnsonii has been recognized as a multidrug-resistant pathogen capable of causing rare bloodstream infections in immunocompetent patients [29,30,31]. Caulobacter spp. has been associated with bloodstream infections and postoperative meningitis [32,33]. While Neisseria elongata is typically a commensal in the nasopharyngeal tract, it has increasingly been identified as a cause of serious human infections, including endocarditis and infected lung bullae [34,35]. Additionally, Corynebacteria have been implicated in pulmonary infections [36]. These findings suggest the disruption of microbial ecological balance following immune barrier compromise. Damaged gastrointestinal mucosal barrier and skin barrier frequently occurs in patients with hematological conditions postchemotherapy, which leads to impaired immunity, uncontrolled opportunistic pathogens invasion, and proliferation, clinically manifested as prolonged fever and severe infections. Our early identification of these microbial shifts enables timely, cost-effective interventions to prevent microbial expansion and proliferation, offering a promising strategy for mitigating subsequent severe infections.

To reduce the incidence of infections in patients with hematological conditions, prophylactic antibiotics are clinically administered. The Infectious Diseases Society of America (IDSA) recommends antibiotic prophylaxis for patients with cancer with an absolute neutrophil count (ANC) < 100 cells/mm3 when the anticipated duration exceeds 7 days [4]. To further narrow the population for prophylactic antibiotic therapy, NCCN guidelines proposed an overall infection risk stratification roughly evaluated by disease and therapy [5]. However, these methods have limitations: the former cannot objectively calculate the duration of neutropenia, while the latter fails to adequately consider factors which may contribute to increased infection rates, such as patient sex, age, performance status, comorbidities (e.g., diabetes, autoimmune diseases), and prior infection history. These issues make it difficult for clinicians to accurately identify the patient population requiring prophylactic treatment. The immediate negative consequences include suboptimal preventive efficacy and inappropriate antibiotic use, while long-term consequence manifests as increased bacterial resistance rates [7,8]. Here, the study provides a paradigm: our microbial-based infection prediction model does not need to consider the duration and degree of neutropenia. It can identify high risk infected populations based on the microecological landscape in patients with different sex, ages, performance status, disease types, comorbidities, prior infection history and treatment measures, thereby enabling precise prevention.

In our study, we first established microbial profiling as the primary factor for predicting the high risk infected population, successfully constructing an infection prediction model with an excellent performance score of 0.942. The model demonstrated high accuracy, correctly identifying 116 of 117 subsequently infected patients as high-risk (99.1% sensitivity) and 56 of 77 infection-free patients as low-risk (72.7% specificity). Furthermore, we conducted a detailed analysis of the correlation between clinical factors and the model’s risk score, identifying advanced age, elevated CRP, and PCT levels, and the presence of underlying comorbidities as significant high-risk factors for infection. These findings are highly consistent with clinical observations. In our prediction model, Cutibacterium avidum, Acinetobacter schindleri, Cupriavidus metallidurans, Staphylococcus haemolyticus, Pseudomonas putida, Brevibacterium luteolum, and Burkholderia stabilis were listed as the top predictive microorganisms for infection. Most of these are opportunistic pathogens—microorganisms often underestimated—yet capable of causing clinical infections in immunocompromised patients. Cutibacterium avidum, though primarily recognized as a skin commensal, can also act as an opportunistic pathogen following bacterial seeding, leading to both superficial and deep/invasive infections [37,38]. Similarly, Acinetobacter schindleri is a frequently misidentified opportunistic pathogen [39]. In rare cases, even typically environmental bacteria can cause severe infections, as demonstrated by the first case of nosocomial septicemia caused by Cupriavidus metallidurans [40]. Among coagulase-negative staphylococci, Staphylococcus haemolyticus stands out as the most prevalent and antibiotic-resistant species associated with bacteremia [41]. It is capable of causing severe infections, including meningitis, endocarditis, prosthetic joint infections, septicemia, peritonitis, and otitis, particularly in immunocompromised patients [42]. Pseudomonas putida, another opportunistic pathogen, is primarily linked to nosocomial infections [43]. While it rarely causes disease in healthy individuals, it poses a significant risk to those with weakened immune systems [44]. Notably, one report describes the first fatal case of Pseudomonas putida bacteremia secondary to skin and soft tissue infections [45]. Brevibacterium luteolum, though an uncommon opportunistic pathogen, has been implicated in infections ranging from cutaneous manifestations to bacteremia, with one documented case of B. luteolum bacteremia occurring in an immunocompromised host [46]. Finally, Burkholderia stabilis, a clinically significant gram-negative bacterium, has been associated with multiple nosocomial outbreaks, often due to contamination of medical devices and supplies [47,48]. These opportunistic pathogens may be the ones causing fever of unknown origin (FUO). Taken together, our results suggest that the microorganism-based model may have potential for predicting infection in patients with hematological conditions, enable precise prevention to dually reduce infection rates and prophylaxis-related resistance.

Our study has several notable strengths, including the innovative use of plasma cfDNA metagenomics for infection prediction, development of an infection prediction model combining microbial mNGS data and clinical metrics, detailed clinical phenotyping, and a large prospective cohort of patients with hematological diseases. Besides, hematologic patients with neutropenia are highly susceptible to infections caused by environmental, skin surface, and gut colonizing microorganisms, which have no effect on healthy individuals. mNGS offers a broader pathogen detection spectrum and higher sensitivity, making it particularly suitable for pathogen identification in the neutropenia population. However, it also has some limitations. Firstly, more convincing results would be achieved if plasma mNGS testing were conducted in the same patient both at the non-neutropenic phase prechemotherapy and neutropenic phase postchemotherapy. Secondly, data were not obtained from the healthy population as a baseline microbial community; therefore, it was hard to characterize the microbiome of patients and to generate an optimal threshold for pathogen identification. Thirdly, future studies in a larger cohort are needed to further validate these findings, strengthen the utility of this model, and assess the impact on clinical outcomes. These limitations are the direction for future improvement. Once these challenges are overcome, implementation of the prediction model has the potential to significantly help improve the management and control of hematological diseases and could potentially be applied to other immunocompromised patient populations.

The lack of precise risk stratification for patients with neutropenia drives noselective prophylactic antibiotic use, increasing costs, hospital stays, and healthcare burdens. Our model enables targeted prophylaxis for high-risk patients, reducing unnecessary antibiotic use and downstream costs associated with adverse events and antimicrobial resistance (AMR)—a major threat in China. By lowering selective pressure, this strategy may delay AMR emergence. Long-term regional or national implementation could substantially reduce antibiotic consumption, resistance rates, and treatment costs. Although plasma mNGS has higher upfront costs than conventional methods, its ability to simultaneously detect pathogens and predict infection risk offers cost-effectiveness potential as sequencing costs decline, though prospective health economic evaluations are needed.