Liu BQ, Sun ZF, Ning CH, Xiao J, Wu D, Lin CY, Hong XY, Guo R, Chen L, Cao XT, Shen DC, Huang GW. Machine learning-driven pathogen cluster analysis identifies high-risk subtypes of infected pancreatic necrosis in a multi-center cohort. World J Gastroenterol 2026; 32(27): 118717 [DOI: 10.3748/wjg.118717]
Corresponding Author of This Article
Geng-Wen Huang, MD, PhD, Full Professor, Division of Pancreatic Surgery, Department of General Surgery, Xiangya Hospital, Central South University, No. 97 Xiangya Road, Kaifu District, Changsha 410008, Hunan Province, China. huanggengwen@csu.edu.cn
Research Domain of This Article
Gastroenterology & Hepatology
Article-Type of This Article
research-article
Open-Access Policy of This Article
This article is an open-access article which was selected by an in-house editor and fully peer-reviewed by external reviewers. It is distributed in accordance with the Creative Commons Attribution Non Commercial (CC BY-NC 4.0) license, which permits others to distribute, remix, adapt, build upon this work non-commercially, and license their derivative works on different terms, provided the original work is properly cited and the use is non-commercial. See: http://creativecommons.org/licenses/by-nc/4.0/
Baishideng Publishing Group Inc, 7041 Koll Center Parkway, Suite 160, Pleasanton, CA 94566, USA
Share the Article
Liu BQ, Sun ZF, Ning CH, Xiao J, Wu D, Lin CY, Hong XY, Guo R, Chen L, Cao XT, Shen DC, Huang GW. Machine learning-driven pathogen cluster analysis identifies high-risk subtypes of infected pancreatic necrosis in a multi-center cohort. World J Gastroenterol 2026; 32(27): 118717 [DOI: 10.3748/wjg.118717]
Bai-Qi Liu, Ze-Fang Sun, Cai-Hong Ning, Chia-Yen Lin, Xiao-Yue Hong, Rong Guo, Lu Chen, Xin-Tong Cao, Ding-Cheng Shen, Geng-Wen Huang, Division of Pancreatic Surgery, Department of General Surgery, Xiangya Hospital, Central South University, Changsha 410008, Hunan Province, China
Jie Xiao, Di Wu, Department of Emergency, The Third Xiangya Hospital, Central South University, Changsha 410008, Hunan Province, China
Author contributions: Liu BQ designed the study and drafted the manuscript; Liu BQ, Sun ZF, Ning CH, Xiao J, Wu D and Shen DC extracted and collected; Liu BQ analyzed the collected data; Sun ZF, Ning CH, Lin CY, Hong XY, Guo R, Cao XT, Shen DC, Chen L and Huang GW reviewed the results and revised the manuscript; Huang GW supervised the study; and the corresponding author attests that all listed authors meet authorship criteria and that no others meeting the criteria have been omitted.
AI contribution statement: ChatGPT was used exclusively for language polishing and translation. All scientific content was drafted and revised by the authors. All figures were created by the authors from original data.
Supported by National Natural Science Foundation of China, No. 82570772 and No. 82403227; and China Postdoctoral Science Foundation, No. 2024M763715.
Institutional review board statement: This study was reviewed and approved by the Ethics Committee of Xiangya Hospital (No. 201012067) and the Third Xiangya Hospital (No. 21019).
Informed consent statement: Written informed consent was obtained from all participants or their legal representatives for the use of their clinical data for research purposes.
Conflict-of-interest statement: All the authors report no relevant conflicts of interest for this article.
Data sharing statement: The data supporting the findings of this study are available upon reasonable request from the corresponding author.
Corresponding author: Geng-Wen Huang, MD, PhD, Full Professor, Division of Pancreatic Surgery, Department of General Surgery, Xiangya Hospital, Central South University, No. 97 Xiangya Road, Kaifu District, Changsha 410008, Hunan Province, China. huanggengwen@csu.edu.cn
Received: January 12, 2026 Revised: February 2, 2026 Accepted: March 30, 2026 Published online: July 21, 2026 Processing time: 185 Days and 18.7 Hours
Abstract
BACKGROUND
Infected pancreatic necrosis (IPN) presents with highly variable clinical trajectories that are significantly influenced by the underlying microbial profile. Although established prognostic models, such as the Acute Physiology and Chronic Health Evaluation II and Bedside Index of Severity in Acute Pancreatitis scoring systems, evaluate host physiological severity, they do not adequately account for the impact of specific pathogen compositions on patient survival.
AIM
To utilize machine learning-driven pathogen cluster analysis to stratify patients with IPN into distinct clusters for precise risk prediction.
METHODS
We conducted a hierarchical clustering analysis on microbiological data from 396 IPN patients, using the Jaccard distance to identify distinct pathogen-driven clusters. Clinical outcomes were compared across these clusters. External cohort validation was conducted to assess the effect of clusters.
RESULTS
Four distinct pathogen clusters (α–δ) were identified, including an Enterococcus faecium/Enterobacter cloacae–predominant cluster (α), an Escherichia coli–dominant cluster (β), a multidrug-resistant organism-enriched cluster (γ), and an Acinetobacter baumannii–Candida glabrata co-infection cluster (δ). Cluster α, characterized by Enterococcus faecium and Enterobacter cloacae, demonstrated moderate severity and 20.5% in-hospital mortality. Cluster β, dominated by Escherichia coli, showed the lowest mortality (10.3%) and clinical severity. Cluster γ, enriched with multidrug-resistant organisms (e.g., Klebsiella pneumoniae and Pseudomonas aeruginosa), had a high mortality (26.3%) and more severe clinical manifestations. Cluster δ, marked by Acinetobacter baumannii and Candida glabrata, had the highest mortality (31.5%). External cohort validation confirmed the robustness of these four subtypes, supporting their clinical relevance.
CONCLUSION
The study highlights the role of pathogen compositions in IPN, revealing how specific microbial profiles influence prognosis. It provides a novel approach for pathogen-based risk stratification in IPN.
Core Tip: Infected pancreatic necrosis (IPN) is clinically heterogeneous, however current risk stratification relies largely on physiological scores and overlooks pathogen patterns. Using unsupervised machine learning, we identified four reproducible pathogen-based IPN subtypes with markedly different mortality risks in a multicenter cohort. High-risk subtypes were driven by co-occurrence of multidrug-resistant organisms and fungi, whereas Escherichia coli–dominant polymicrobial infections were associated with favorable outcomes. This study highlights pathogen clustering as a clinically actionable approach for prognostic stratification and personalized infection management in IPN.
Citation: Liu BQ, Sun ZF, Ning CH, Xiao J, Wu D, Lin CY, Hong XY, Guo R, Chen L, Cao XT, Shen DC, Huang GW. Machine learning-driven pathogen cluster analysis identifies high-risk subtypes of infected pancreatic necrosis in a multi-center cohort. World J Gastroenterol 2026; 32(27): 118717
Infected pancreatic necrosis (IPN) represents a severe and potentially life-threatening complication of acute pancreatitis (AP), with a mortality rate as high as 20%-30%[1]. Despite advancements in critical care and surgical or endoscopic step-up approach, the responses to treatment show great variability among patients[2,3]. Common scoring systems, such as Ranson, Bedside Index of Severity in Acute Pancreatitis (BISAP) and Acute Physiology and Chronic Health Evaluation II (APACHE II), provide valuable prediction to evaluate the severity and mortality of AP[2,4]. However, these scoring systems primarily reflect patient physiology but fail to consider the influence of pathogen characteristics on prognosis.
Microbial infections play critical roles in the progress of IPN, with polymicrobial infection, multidrug-resistant organisms (MDROs) and fungal infection being closely associated with the patient outcomes[5,6]. The microbial spectrum of various bacteria and fungi exhibits significant heterogeneity, which influences the anti-infection strategies[7]. The diversity of these microorganisms and their interactions complicate infection control and individualized treatment. Recent research has demonstrated that the composition of microbes can activate systemic inflammatory responses via pathogen-associated molecular patterns, subsequently affecting the advancement of multiple organ dysfunction[8]. The co-occurrence of fungal infections and Gram-negative bacteria can significantly elevate the secretion of interleukin-6 and tumor necrosis factor-alpha, thereby accelerating the onset of sepsis[9,10]. However, most previous studies have primarily focused on the role of individual pathogens or a limited number of pathogens in IPN, often neglecting the interactions among pathogens and the effects of various pathogen combinations on patient outcomes. Therefore, understanding the interactions and assemblages of pathogenic microorganisms is of paramount importance for evaluating disease severity and predicting clinical outcomes in IPN.
In recent years, machine learning (ML) has been increasingly used in the identification of disease subtypes and risk stratification. Unsupervised learning algorithms are capable of uncovering hidden patterns within complex and heterogeneous datasets without reliance on pre-labeled data, such as clustering algorithms[11]. Unsupervised clustering offers significant advantages in the classification of infectious diseases, as it can effectively identify subgroups with similar microbiological characteristics[12,13]. However, there is a lack of systematic utilization of clustering models to stratify patients with IPN.
The primary objective of this study is to apply unsupervised ML algorithms to categorize patients with IPN into distinct subtypes based on the establishment of microbial clusters, thereby achieving precise risk stratification. By integrating clinical data with microbial results, we employed hierarchical clustering to elucidate the subtype characteristics of various microbial profiles and validated their roles in prognosis.
MATERIALS AND METHODS
Study cohort
Two cohorts of IPN were included in this study to develop and validate the clustering model for IPN patients. The derivation cohort consisted of 396 IPN patients consecutively enrolled from Xiangya Hospital, Central South University, China between January 2011 and June 2024, while the external validation cohort consisted of 101 IPN patients admitted to the Third Xiangya Hospital, Central South University, China between January 2018 and June 2023. Patients with a history of chronic pancreatitis, with chronic organ dysfunction, during pregnancy and patients with incomplete data were excluded. This study was reviewed and approved by the Ethics Committee of Xiangya Hospital (No. 201012067) and the Third Xiangya Hospital (No. 21019). This study was registered on www.researchregistry.com (unique identifying number: Researchregistry9293, https://www.researchregistry.com). The study was reported according to the STROCSS guidelines (Strengthening the reporting of cohort, cross-sectional and case-control studies in surgery)[14]. Written informed consent was obtained from all participants or their legal representatives for the participation in the study.
Data collection and definition
All collected clinical data were extracted from the dedicated electronic database by trained health professionals. The primary endpoint was in-hospital mortality. Overall survival (OS) was defined as the interval from the date of diagnosis to date of death. Patient data included demographic characteristics (age, gender, comorbidity, smoking or drinking and etiology), illness severity, scoring systems (APACHE II score and BISAP score), intensive care unit (ICU) stay, organ failure (number and type of failing organ systems), therapeutic procedures [step-up or step-down surgical approach, the number of surgery, invasive ventilation, central venous catheter (CVC) indwelling and continuous renal replacement therapy (CRRT)], variables associated with infection (Peripancreatic pathogens, peripancreatic polymicrobial infection, peripancreatic fungal infection, peripancreatic MDRO infection and bloodstream infection), complications and outcomes (gastrointestinal fistula, hemorrhage, pancreatic fistula, in-hospital mortality and OS). The definition of other variables was detailed in the Supplementary material.
Microbiological sampling
Microbiological data were obtained at the time of initial clinical intervention, including percutaneous catheter drainage, minimally access retroperitoneal pancreatic necrosectomy, or surgical necrosectomy. Pathogen identification was based exclusively on positive cultures obtained from either peripancreatic drain fluid or pancreatic necrotic tissue. All samples were processed in the clinical microbiology department using standardized aerobic and anaerobic cultures. Identification and antimicrobial susceptibility testing were conducted according to Clinical and Laboratory Standards Institute guidelines.
Clustering model
Cluster analysis, a foundational unsupervised ML technique, groups samples based on inherent similarities without the necessity of prior class labels. In this study, we employed hierarchical clustering utilizing the Jaccard distance and the Ward clustering method to delineate pathogen-driven clusters in IPN. The Jaccard distance was selected to quantify co-occurrence patterns, emphasizing common pathogens while downweighing rare taxa. This approach provided a significant advantage in microbiome studies, where rare species introduce considerable noise. The number of clusters was determined using complementary internal criteria computed within the same distance space, including the elbow method in within-cluster Jaccard dispersion, the mean silhouette curve, and inspection of the silhouette distribution at the selected k. Cluster stability was quantified by subsampling consensus to construct a consensus matrix and compute the proportion of ambiguous clustering, defined as the fraction of co-assignments in an intermediate range, where lower values indicate greater stability. Subsampling involved 80% resamples, repeated 500 times with a fixed random seed. A dendrogram illustrates the hierarchical structure, while t-distributed stochastic neighbor embedding (t-SNE) is utilized for visualization.
Statistical analysis
We subsequently evaluated the impact of unsupervised clustering based on peripancreatic pathogens in patients with IPN. We compared baseline characteristics and clinical outcomes across groups. Summary statistics were presented as total frequencies and percentages for categorical variables, and as medians with interquartile ranges (IQR) or means with SD for continuous variables, as appropriate. Differences among groups were assessed using the Student’s t-test for continuous variables, the Kruskal-Wallis test for continuous variables that did not follow a normal distribution, and the χ² test for categorical data. A two-sided P value of less than 0.05 was deemed statistically significant. All figures were generated, and corresponding statistical analyses were performed using R version 4.0.2.
RESULTS
Patients in the derivation cohort
A total of 396 patients were included in the analysis, of whom 294 (74.2%) were male. The average age of the cohort was 46.9 ± 12.8 years. The distribution of etiology was 49.7% of hypertriglyceridemia, 30.6% of biliary, 4.8% of alcoholic, and 14.9% of other causes. According to the determinant-based classification, critical AP (CAP) was seen in 53.5% of patients. A total of 276 patients were admitted to the ICU with a median (range) length of ICU stay of 6.0 (0.0-18.0) days. Multiple organ failure was present in 34.1% of patients, while 21.0% had single organ failure. 84.8% of patients underwent a step-up approach, while 15.2% underwent a step-down approach. The median (range) number of surgical interventions per patient was 3.0 (2.0-5.0). The proportion of mechanical ventilation, CRRT and CVC retention were 41.7%, 23.5% and 42.2%, respectively. Vasopressors were administered to 33.1% of patients. The types of pancreatic infections included 60.9% of polymicrobial infection, 53.5% of MDRO infection and 25.0% of fungal infection. 27.8% of patients developed bloodstream infection. The complications included hemorrhage in 87 (22.1%), gastrointestinal fistulas in 62 (15.7%) and pancreatic fistulas in 155 (39.5%). The overall mortality rate in the study cohort was 22.2% (Supplementary Table 1).
Clustering in the derivation cohort
In the derivation cohort, hierarchical clustering analysis identified a four-class model as the optimal fit, comprising cluster α [Enterococcus faecium (E. faecium)/Enterobacter cloacae (E. cloacae–predominant)], β [Escherichia coli(E. coli)–dominant], γ (MDRO-enriched), and δ [Acinetobacter baumannii (A. baumannii)–Candida glabrata (C. glabrata) co-infection] (Supplementary Figure 1). The distinction among these four classes was clearly demonstrated in the t-SNE plot (Figure 1). Of 112 patients (28.3%) were classified into cluster α, 78 patients (19.7%) into cluster β, 152 patients (38.4%) into cluster γ, and 54 patients (13.6%) into cluster δ. The distribution of microorganisms within the four-class model was detailed in Supplementary Figure 2 and Supplementary Table 2. Cluster α displayed a significant predominance of Candida albicans(C. albicans) and E. cloacae. Notably, cluster α was associated with a higher isolation rate of E. faecium compared to clusters β and δ (P < 0.001). Cluster β was characterized by a significantly elevated isolation rate of E. coli (P < 0.001). In cluster γ, Klebsiella pneumoniae (K. pneumoniae) was significantly predominant (P < 0.001), and this cluster also showed higher isolation rates of Pseudomonas aeruginosa (P. aeruginosa) and E. faecium than clusters β and δ (P < 0.001). Cluster δ was linked to a higher isolation rate of A. baumannii and C. glabrata (P < 0.001) as indicated in Supplementary Table 2.
Figure 1 t-distributed stochastic neighbor embedding plot of the derivation cohort.
The main purpose of t-distributed stochastic neighbor embedding (t-SNE) plot is the visualization of high-dimensional data. In t-SNE plot, circles represent individual patients. cluster α (Enterococcus faecium/Enterobacter cloacae–predominant) in blue, cluster β (Escherichia coli–dominant) in yellow, cluster γ (multidrug-resistant organism-enriched) in green and cluster δ (Acinetobacter baumannii–Candida glabrata co-infection) in orange. t-SNE: T-distributed stochastic neighbor embedding.
Clinical characteristics and outcomes of each cluster in the derivation cohort
Cluster α (E. faecium/E. cloacae–predominant) served as the reference group, exhibiting moderate clinical severity. This cluster displayed transitional characteristics: ICU admission rates (65.2%) and median duration of ICU stay (5.0 days, IQR = 0.0-17.0) were intermediate between cluster β’s minimal critical care requirements and the prolonged intensive management observed in clusters γ and δ. The in-hospital mortality was 20.5% (Table 1).
Table 1 Baseline of clusters in the derivation cohort, n (%).
Cluster β (E. coli–dominant) exhibited a generally more favorable clinical course, characterized by the lowest severity scores, interventions, and mortality. Patients within this cluster demonstrated the lowest frequency of ICU admission (55.1%) and the shortest median duration of ICU stay (2.5 days, IQR = 0.0-13.0). This group required less intensive support, including vasopressor use (15.4%) and CRRT (9.0%). The in-hospital mortality rate was only 10.3% (Table 1).
Cluster γ (MDRO-enriched) was characterized by the highest prevalence of MDRO infection (73.7%). This group exhibited severe clinical manifestations. The in-hospital mortality rate reached 26.3%, accompanied by a high incidence of multi-organ failure (61.8%). Additionally, the median duration of ICU stay was significantly longer at 10.0 days (IQR = 3.0-21.5 days), and more interventions were required, with a median of 4.0 (IQR = 3.0-6.0). Furthermore, the risk of pancreatic fistula was the highest among the groups, reaching 42% (Table 1).
Cluster δ (A. baumannii–C. glabrata co-infection) represented the cohort with the most critical illness and poorest prognosis. These patients were predominantly classified as CAP (66.7%) and exhibited the highest APACHE II and SOFA scores. The cluster suffered the highest rates of multi-organ failure (44.4%), which received more mechanical ventilation (57.4%) and CRRT (40.7%) therapies. There was a high proportion of MDRO infection (61.1%) in this cluster. Overall, in-hospital mortality was 31.5%, with the highest mortality rate in the cohort (Table 1).
Kaplan-Meier analysis showed an association between clusters and 90-day mortality (P = 0.008) (Figure 2). Cluster δ showed the worst survival outcome whereas cluster β had the best prognosis. Pairwise comparisons revealed differences in in-hospital mortality among all four groups (Table 2). In adjusted Cox models, cluster β showed a lower hazard of death (HR = 0.34, 95%CI: 0.12–0.97), whereas cluster δ showed a higher hazard (HR = 2.35, 95%CI: 1.14–4.86), relative to the reference cluster (Table 3).
Figure 2 Kaplan-Meier survival analysis of the derivation cohort.
All show significant differences in 90-day mortality by clusters. E. faecium: Enterococcus faecium; E. cloacae: Enterobacter cloacae; E. coli: Escherichia coli; MDRO: Multidrug-resistant organism; A. baumannii: Acinetobacter baumannii; C. glabrata: Candida glabrata.
Table 2 Statistical test results of in-hospital mortality among the clusters in the derivation cohort.
One hundred one patients with IPN were included in the external validation cohort. In the cohort, we also obtained four clusters (cluster 1, 2, 3 and 4). A distinct differentiation among the four-class model was observed in the t-SNE plot for the external validation cohort (Supplementary Figure 3). Supplementary Figure 4 and Supplementary Table 3 illustrated the distribution of pathogens and the characteristics of patients within each cluster of the external validation cohort, respectively. The predominant pathogens in each category were consistent with those observed in the derivation cohort. Clusters 1 and δ shared similarities, primarily characterized by A. baumannii and C. glabrata. Cluster 2 was predominantly associated with MDRO infection, including K. pneumoniae, similar to cluster γ. Cluster 3 exhibited the highest proportion of E. faecium, whereas cluster 4 was primarily composed of E. coli, corresponding to clusters β and α in the derivation cohort, respectively. Notably, clusters 1 and 2 were associated with the worst prognosis, while cluster 4 was linked with the best prognosis (Supplementary Figure 5).
DISCUSSION
To our knowledge, this is the first study employing a ML-based approach to focus on stratifying IPN from the perspective of microbial cluster analysis. Four distinct subtypes, characterized with microbial composition patterns, exhibited marked differences in terms of disease severity, organ dysfunction and death risk. These findings underscore the heterogeneity of pathogenic infection in IPN and provide new insights for precision medicine in clinical judgement.
In this study, we identified IPN subtypes characterized by distinct combinations of pathogens through cluster analysis, revealing the significant role of MDROs in patient prognosis. Notably, the γ subtype, enriched with MDROs such as K. pneumoniae, P. aeruginosa and E. faecium, exhibited high prevalence of multidrug resistance in severe cases. Research has demonstrated that MDRO infection was becoming increasingly prevalent among patients with IPN, which correlated with a significant rise in mortality[15,16]. Specifically, infection caused by Gram-negative bacteria, such as Klebsiella pneumonia and P. aeruginosa, has been identified as primary contributors to the poor prognosis[17,18]. The frequent occurrence of MDROs may be partially explained by the high and inappropriate use of antibiotics and invasive approach. Furthermore, the clustering results suggest that these multidrug-resistant (MDR) bacteria may be in a co-infection state, where the interactions among these pathogens enhance their resistance and complicate treatment strategies. K. pneumoniae and P. aeruginosa, both categorized as the ESKAPE pathogens, are known to cause hospital-acquired infections, exhibiting extensive resistance patterns, especially in intensive care settings where the use of broad-spectrum antibiotics and invasive procedure are prevalent[19]. These bacteria are capable of forming biofilms, which provide a protective environment that enhances their resistance to antimicrobial agents and fosters genetic exchange through horizontal transfer, particularly plasmids that carry resistance genes[19]. In a coinfection scenario, both K. pneumoniae and Enterococcus faecalis can coexist in a patient’s system, enabling the transfer of genetic material that may enhance multidrug resistance, ultimately complicating treatment[20]. For instance, it has been observed that co-infection may facilitate the exchange of extended-spectrum beta-lactamase (ESBL) genes among Enterobacteriales, including Klebsiella and Enterococcus, which can lead to heightened levels of resistance against commonly used antibiotics[21]. Our results indicated that patients in the γ subtype required more surgical interventions, reflecting the complexity of treatment. However, frequent surgical interventions, the placement of long-term drainage tubes and the regular replacement and irrigation of these tubes, may further increase the incidence of MDRO infection[22]. In summary, the interplay between MDRO infection, surgical interventions, and antibiotic usage manifests a complex clinical scenario. This scenario necessitates diligent management strategies to combat the rise of resistance, address infections comprehensively, and optimize patient outcomes through controlled antibiotic use and preventive measures.
Cluster δ was primarily characterized by infections caused by A. baumannii and C. glabrata, reflecting the significant trend of co-occurrence of bacteria and fungi in hospital-acquired infections. A. baumannii is widely recognized for its MDR strains, which complicate treatment options. As the data suggested, prior antibiotic use was the most significant risk factor for the acquisition of MDR A. baumannii[23]. Candida fungal infection in pancreatic necrosis represents a serious complication associated with SAP[24]. A retrospective analysis found fungal infection in 52 of 113 patients, with C. albicans accounting for 46% of infections in that cohort[25]. Previous studies indicated that pancreatic Candida infection alone was not linked to prognosis[6], but our findings showed that co-infection with A. baumannii and Candida species led to a worse prognosis, with multi-organ failure as the main clinical manifestation. The enhanced virulence and severity of this co-infection might result from bacterial-fungal interactions. Research has demonstrated that Candida colonization promoted A. baumannii pneumonia in animal models, with rats colonized by Candida showing higher pneumonia incidence, increased bacterial load, and more severe inflammation than controls. The colonized group exhibited lower concentrations of interleukin-17 and higher levels of interferon-gamma compared to the non-colonized control group. This interaction modulated the host’s inflammatory response, favoring A. baumannii survival[26]. Furthermore, the presence of A. baumannii has been observed to facilitate the growth and persistence of Candida[27]. The presence of A. baumannii and Candida in Cluster δ points towards a potential shared ecological niche where both organisms can thrive. This co-occurrence is likely facilitated by conditions common in ICU settings, such as prolonged hospitalization, antibiotic usage and the presence of sophisticated medical devices like ventilators and catheters[26], which has also been confirmed by our study. Recognition of such mixed bacterial–fungal clusters may also support early antifungal vigilance and timely escalation of drainage or debridement in patients with rapidly progressive organ dysfunction, while at the same time highlighting the importance of avoiding unnecessary invasive procedures or prolonged catheter placement that could further predispose to nosocomial co-infections.
E. coli was identified as a prominent microbial agent within cluster β. Compared to other clusters, this group was associated with a lower incidence of multiple organ failure, fewer surgical interventions and lower mortality. These findings align with our previous research[28]. Importantly, the presence of polymicrobial infections was more common among patients infected with E. coli (77.5%) compared to those without this infection (65%)[28]. This suggests that E. coli may contribute to beneficial co-dependencies or maintain a balanced ecological role, potentially coexisting with other pathogens without negatively impacting overall patient outcomes. Studies have demonstrated that the polymicrobial biofilm composed of C. albicans and E. coli exhibited dynamic interspecies interactions[29]. This phenomenon may be attributed to the presence of lipopolysaccharides from E. coli, which appear to be a critical component of the biofilm[30]. Furthermore, research indicated that E. coli released soluble factors analogous to EDTA, which sequestered essential metals from the surrounding environment, thereby suppressing the pathogenic characteristics of C. albicans[31]. These findings necessitate further investigation into the mechanisms behind this observed protective effect and the overall role of microbiota in modulating disease outcomes in pancreatic conditions, particularly those involving necrotizing complications. From a clinical perspective, the identification of a β cluster-dominant microbial profile may suggest a relatively favorable prognosis. For such patients, priority should be given to achieving adequate source control, with effective drainage of infected collections, while antimicrobial de-escalation may be considered once clinical condition is improved. Understanding these intricate relationships might pave the way for more targeted therapeutic strategies, including microbiome modulation and selective antibiotic use tailored to specific bacterial profiles present in patients.
Our subtyping method diverges from traditional risk stratification methods, such as the APACHE II or BISAP scores, by directly correlating pathogen characteristics with clinical features. While conventional scoring systems are instrumental in predicting outcomes and stratifying patient risk in intensive care settings, they predominantly rely on aggregated clinical and physiological data, neglecting the specific pathogens involved in the infection. This study illustrated that an exclusive focus on clinical parameters might lead to an underestimation of the mortality risk associated with MDRO-related subtypes. For instance, among patients with comparable APACHE II score, the mortality in cluster γ was higher. This observation underscores the significant influence of microbial characteristics on prognosis. However, prior research has largely concentrated on the impact of individual pathogens or a limited number of pathogens in IPN, often overlooking the interactions among various pathogens and the effects of different pathogen combinations on patient outcomes. Accordingly, pathogen-based subtyping may offer complementary prognostic information to traditional scoring systems by incorporating microbial characteristics into risk stratification.
ML clustering analysis plays a significant role in disease prognosis by enabling the identification and classification of distinct patient subgroups based on complex patterns within clinical, genomic or behavioral data. A study involving 75762 adult patients admitted to the emergency department employed unsupervised ML algorithms to analyze vital signs recorded within the first six hours of admission[32]. The analysis revealed four distinct patient clusters, each characterized by unique physiological patterns and corresponding clinical outcomes. Furthermore, by clustering patients based on immune indicators, the researchers identified different immune phenotypes within the sepsis population, which were associated with varying survival rates[33]. However, there are few studies that utilize clustering analysis of pathogen characteristics. Zhang et al[12] conducted an analysis based on blood culture pathogens to elucidate the heterogeneity of bloodstream infections. They demonstrated that data-driven blood culture classification could identify unique physiological characteristics and prognosis. Unlike the study conducted by Zhang et al[12], this study employed hierarchical clustering rather than K-means clustering. Hierarchical clustering has several advantages over K-means clustering. One of the notable strengths of hierarchical clustering is its ability to accommodate varying group structures and relationships, particularly in the context of dealing with large datasets of pathogen types[34]. Moreover, hierarchical clustering performs well in visual representation through dendrograms, which illustrate the relationship among various entities at different similarity levels[35]. Thus, when it comes to maximizing the utility of diverse and imbalanced datasets typical in pathogen analysis, hierarchical clustering provides a robust alternative by not only handling various data distributions effectively but also by offering comprehensive insights that can facilitate a deeper understanding of pathogen dynamics.
However, there were several limitations which should be acknowledged in this study. First, the model was created through a single-center, prospective cohort spanning an extended duration, which raises questions regarding its applicability to other centers and populations worldwide. The lengthy study duration could lead to variations in clinical management practices and data collection methods, influenced by changing guidelines and treatment strategies for IPN, thereby possibly resulting in diverse effects on clinical outcomes and restricting its generalizability. Second, the documentation regarding antibiotic use, including type, duration, and whether therapy was prophylactic or therapeutic, was incomplete because of the complexity of antimicrobial management in IPN. Given the strong association between antibiotic exposure and the emergence of MDRO and fungal infections, this limitation prevents us from fully disentangling the effects of antimicrobial selection pressure from those of pathogen composition itself. Third, this analysis focused on common pathogens for clinical clarity, which may overlook rare but highly virulent organisms. Future studies should employ alternative distance metrics or weighted preprocessing to better account for these rare pathogens, while integrating NGS data, such as relative abundance and virulence genes, to refine clinical stratification. Fourth, the mechanistic explanations proposed for the identified clusters are hypothesis-generating rather than causal. This study was designed as a clinical, data-driven stratification analysis to identify prognostically relevant pathogen constellations, rather than to directly interrogate microbial interactions. The proposed bacterial–fungal synergy in the δ cluster and the potential modulatory or antagonistic effects in the β cluster are supported by prior experimental literature but were not experimentally validated in this cohort. Consequently, whether these interactions causally drive the observed outcome differences requires future confirmation using in vitro co-culture systems and in vivo infection models. In addition, the exploration of the relationship between pathogen-defined subtypes and host inflammatory responses was limited by the absence of certain laboratory variables and biomarkers, which restricted our ability to elucidate immune–microbial interactions across clusters. Future studies would focus on constructing multicenter large-scale databases and combining precision medicine techniques and advanced artificial intelligence to develop and validate high-quality models.
CONCLUSION
This study demonstrates that unsupervised ML effectively stratifies IPN into prognostically distinct subtypes based on pathogen profiles (Figure 3). The approach reveals that specific pathogen combinations significantly dictate mortality, organ failure, and surgical complexity. Critically, microbial community interactions substantially influence clinical trajectories. Future research must prioritize deciphering microbial ecological networks and developing targeted strategies to disrupt high-risk consortia, advancing toward microbiome-guided precision therapy for IPN.
Banks PA, Bollen TL, Dervenis C, Gooszen HG, Johnson CD, Sarr MG, Tsiotos GG, Vege SS; Acute Pancreatitis Classification Working Group. Classification of acute pancreatitis--2012: revision of the Atlanta classification and definitions by international consensus.Gut. 2013;62:102-111.
[RCA] [PubMed] [DOI] [Full Text][Cited by in Crossref: 5667][Cited by in RCA: 4849][Article Influence: 373.0][Reference Citation Analysis (13)]
Lu JD, Cao F, Ding YX, Wu YD, Guo YL, Li F. Timing, distribution, and microbiology of infectious complications after necrotizing pancreatitis.World J Gastroenterol. 2019;25:5162-5173.
[PubMed] [DOI] [Full Text]
Bereanu AS, Bereanu R, Mohor C, Vintilă BI, Codru IR, Olteanu C, Sava M. Prevalence of Infections and Antimicrobial Resistance of ESKAPE Group Bacteria Isolated from Patients Admitted to the Intensive Care Unit of a County Emergency Hospital in Romania.Antibiotics (Basel). 2024;13:400.
[RCA] [PubMed] [DOI] [Full Text][Cited by in RCA: 39][Reference Citation Analysis (0)]
Neidhöfer C, Sib E, Neuenhoff M, Schwengers O, Dummin T, Buechler C, Klein N, Balks J, Axtmann K, Schwab K, Holderried TAW, Feldmann G, Brossart P, Engelhart S, Mutters NT, Bierbaum G, Parčina M. Hospital sanitary facilities on wards with high antibiotic exposure play an important role in maintaining a reservoir of resistant pathogens, even over many years.Antimicrob Resist Infect Control. 2023;12:33.
[RCA] [PubMed] [DOI] [Full Text][Cited by in Crossref: 8][Cited by in RCA: 16][Article Influence: 5.3][Reference Citation Analysis (0)]
Quan J, Hu H, Zhang H, Meng Y, Liao W, Zhou J, Han X, Shi Q, Zhao D, Wang Q, Jiang Y, Yu Y. Investigating Possible Interspecies Communication of Plasmids Associated with Transfer of Third-Generation Cephalosporin, Quinolone, and Colistin Resistance Between Simultaneously Isolated Escherichia Coli and Klebsiella Pneumoniae.Microbiol Spectr. 2023;11:e0355422.
[RCA] [PubMed] [DOI] [Full Text][Cited by in RCA: 6][Reference Citation Analysis (0)]
Creativity or innovation: Grade B, Grade B, Grade B
Scientific significance: Grade A, Grade B, Grade B
P-Reviewer: Jiang J, Associate Professor, China; Peng XW, PhD, Principal Investigator, Research Assistant Professor, China S-Editor: Li L L-Editor: A P-Editor: Lei YY
