Early identification and diagnosis of fournier gangrene: a machine learning approach integrating serological characterization

Abstract

Background

Fournier Gangrene (FG) and Generalized Perianal Abscess (GPA) have similar clinical features. But FG has a high mortality and disability rate and needs to be identified and treated as early as possible. This study utilized machine learning methods to integrate clinical and metabolic features to promote early diagnosis of FG.

Methods

Serological characteristics were screened for patients with FG (n = 20) and GPA (n = 16). The metabolomic changes of FG were described based on untargeted metabolomics. We used machine learning tools to combine demographic data, clinical serology, and metabolomics data to establish disease-specific boundary points.

Results

There were significant differences in the serum metabolic profiles between the FG and GPA groups. 118 different metabolites were detected, mainly fatty acids. Based on machine learning integration of metabolic and clinical features, a differential diagnosis combination of Myo-inositol (MI), Procalcitonin (PCT) and Bistris was established for early identification and diagnosis of FG. The diagnostic performance was evaluated using GBDT, SVM, and LR algorithms, demonstrating robust discriminative ability (AUC: 0.80, 0.82, and 0.95; sensitivity: 0.90, 0.92, and 1.00). In addition, we identified 14 differential metabolic pathways. The activation of Necroptosis may lead to the occurrence of explosive perianal and perineal infections.

Conclusion

Our findings provide a biomarker combination for early diagnosis of FG in clinical applications. On the other hand, it provides important insights into the pathological mechanism differences between FG and GPA.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12879-025-11575-1.

Introduction

Fournier gangrene (FG) is an explosive, necrotic skin and soft tissue invasive infection [1]. This disease is caused by the combined action of aerobic and anaerobic bacteria, often accompanied by septic shock [1–4]. Severe systemic toxic reactions are considered the main cause of death [5]. The high mortality and disability rates have made FG a research hotspot and difficulty in multidisciplinary fields such as anorectal surgery, urology, plastic surgery, critical care medicine, and emergency medicine. The diagnosis of FG is challenging. At present, the diagnosis of FG mainly relies on the experience of clinical doctors, which requires a comprehensive evaluation of clinical manifestations, laboratory tests and imaging examinations. However, due to the widespread deep tissue infection of FG, which may have limited skin changes, there were often case reports that confused FG with Generalized Perianal Abscess (GPA) [6, 7]. Simultaneously completing the series of checks often requires a significant amount of time. The difficulty of accurate and timely diagnosis can lead to delayed intervention and an increase in related mortality and complication rates^[8]. Consequently, early recognition and accurate diagnosis are crucial in enhancing the survival rates and improving the quality of life for patients with FG.

Serum Biomarkers are crucial for clinical decision-making, as they can shorten disease diagnosis time [8]. Therefore, actively seeking serological sensitive indicators for diagnosing FG is of great significance for early identification of the disease and treatment. So far, Wong has developed the Laboratory Risk Indicator for Necrotizing Fasciitis score (LRINEC) commonly used in clinical practice for the diagnosis of necrotizing fasciitis [9]. This scoring system includes traditional laboratory indicators such as CRP, White blood cell count, Hemoglobin, Sodium, Creatinine and Glucose. However, relevant studies have found that the sensitivity of the LRINEC score is poor [10, 11]. In addition, studies have shown that Pentaxin-3 was candidate biomarkers for Necrotizing tis infection. But this report mainly focused on the prognosis of FG, rather than early diagnosis [12]. Therefore, early diagnostic biomarkers for FG urgently need to be developed and explored.

Untargeted metabolomics is a powerful tool for studying metabolic processes. It can identify key biomarkers responsible for metabolic characteristics. In this study, we investigated the effect of serum metabolomics profiles as potential parameters for distinguishing FG and GPA using liquid chromatography tandem mass spectrometry (LC-MS) and gas chromatography mass spectrometry (GC-MS). Moreover, FG is not a disease limited to the perianal and perineal areas. It is usually accompanied by systemic inflammatory reactions. Using a single indicator alone cannot comprehensively and objectively reflect the complex situation of FG patients. Therefore, we utilized machine learning (ML) methods to integrate metabolomics and clinical features, overcoming the limitations of traditional diagnostic models [13–15]. This combination of serum diagnostic markers has provided us with a method for accurately distinguishing FG and GPA from multiple dimensions.

Methods

Patients and sample

This is a retrospective study on FG and GPA patients admitted from July 2022 to September 2023. We recruited these patients at Longhua Hospital affiliated with Shanghai University of Traditional Chinese Medicine. The confirmation of FG patients was based on Fisher’s diagnostic criteria for necrotizing fasciitis [16]. The inclusion criteria for GPA patients was: (1) meeting the diagnostic criteria for perianal abscess; (2) The infection range exceeds half of the perianal area surface and was accompanied by at least one perianal anatomical space infection, supported by ultrasound or MRI evidence. All patients were over 18 years old, and pregnant and lactating patients were excluded. Finally, a total of 20 FG patients and 16 GPA patients were recruited. Information about age, gender, BMI, clinical parameters (heart rate, temperature, respiratory and BP), laboratory findings before surgery (e.g. serum sodium, potassium, and creatinine), and imaging features (e.g. fascial edema, subcutaneous edema, and fluid tracking) of each participant was collected. This clinical trial was conducted in compliance with the principles of the 1975 Declaration of Helsinki and was approved by the Ethics Committees of Ethics Committee of Longhua Hospital Affiliated to Shanghai University of Traditional Chinese Medicine (2022LCSY072). All participants signed the consent forms to allow the extra samples be used for academic purpose.

Serum samples were collected from patients before debridement and drainage surgery. Surgical debridement was performed as follows: The initial incision was made at the center of the visibly necrotic or ulcerated lesion. A hemostat or finger was used to explore and separate the septa within the abscess cavity. Subsequently, multiple auxiliary incisions were made along the lesion and the course of the abscess cavity, preserving skin bridges between the incisions. After clarifying the relationship between necrotic tissue and surrounding healthy tissue, as well as the spatial arrangement of the incisions, the necrotic tissue was completely debrided until fresh granulation tissue was exposed. The serum was obtained according to the standardized protocol: (1) Plasma was collected in 5 ml tubes containing gel (Vakuette ^® Serum gel with activator); (2) Gently invert the tubes five times and placed them vertically for coagulation; (3) After 30 min, centrifuge the sample at 3000xg for 15 min. Visually inspect the serum for any residue, and if present, repeat the centrifugation process; (4) the upper plasma was separated into a centrifuge tube and stored in a − 80 °C refrigerator.

Metabolites extraction and analysis

The extraction procedure for Liquid chromatograph-mass spectrometer (LC-MS) metabolites was performed as follows: After thawing the samples at room temperature, we added L-2-chlorophenylalanine (0.06 mg/mL) dissolved in methanol and rotated for 10 s. Then methanol and acetonitrile mixed solution was added. The mixture was rotated 30 s, ultrasonic extracted in an ice water bath for 10 min, and allowed to sit for 30 min at −20 ℃. Next the samples were centrifuged at 4 ℃ for 10 min (13000 rpm/min). Finally, the samples were filtered using a syringe filter of 0.22 μm pore size and retured to LC glass vials. After that the samples were placed in the autosampler for LC-MS analysis.

The extraction of Chromatography-mass spectrometry (GC-MS) metabolites was conducted according to the following protocol: Except for the final filtering step, the remaining methods are the same as LC metabolite extraction. After the supernatant was dried, we added pyridine solution of methoxamine hydrochloride. The mixture was rotated violently for 2 min and incubated in 37 ° C for 90 min. Then BSTFA and n-hexane were added to the mixture, vortexed for 2 min and reacted at 70 ℃ for 60 min. After leaving the sample at room temperature for 30 min, GC-MS metabolomics analysis was performed.

Machine learning algorithm

In order to reduce feature dimensions and improve the robustness of features, this study used Random Forest (RF), Gradient Boosting Decision Tree (GBDT), and Least Absolute Shrinkage and Selection Operator (LASSO) models to select disease diagnostic factors. Clinical features and metabolic data were used as input variables, while diagnosis was used as output. Unlike previous studies, we constructed a mixed feature dataset. The mixed feature dataset was formed by combining raw clinical variables (e.g. age, PCT) with untransformed metabolic mass spectral peak areas (unit-free). The mass spectral peak areas of metabolites underwent log2 transformation during preprocessing to reduce skewness. Finally, the above three ML methods were used to analyze the clinical feature dataset, metabolic feature dataset, and mixed feature dataset of the participants. Due to FG being a fulminant and mortal disease, it is necessary to avoid missed diagnosis. Therefore, based on AUC > 0.9, we further prioritized a dataset and model with higher sensitivity. Finally, disease diagnostic factors were selected based on the importance ranking generated by the corresponding model of the dataset.

The optimal diagnostic model was constructed using three ML methods: GBDT, Support Vector Machine (SVM) and Logistic Regression (LR). We randomly divided all participants into a discovery queue and a validation queue in a 1:1 ratio, and used 10x cross validation in the discovery queue. Finally, the performance of the model was evaluated using AUC, sensitivity, specificity and clinical practicality. The evaluation results were used to demonstrate the stable contribution of the selected feature combination in disease diagnosis and recognition. All machine learning related operations are completed using Python.

Statistical and data analysis

The research design and data analysis flowchart was shown in Fig. 1. Excluding age and gender, we collected 31 clinical features and 118 metabolic features. Only 1 missing erythrocyte sedimentation rate (ESR) value was identified and imputed using the cohort mean. Traditional descriptive methods were used to describe the clinical features. Categorical variables were described by frequency, while continuous variables were described by mean (SD) or median (interquartile range). For each study characteristic, the Mann-Whitney U test or t-test was used for continuous variables, and chi-square test was used for categorical variables. To analyze metabolic characteristics, principal component analysis (PCA) was used to explore the overall metabolic changes between groups. Supervised Orthogonal Partial Least Squares Discriminant Analysis (OPLS-DA) was used to maximize overall metabolic changes between groups. Firstly, metabolites were screened using Student T-test with P < 0.05 and Variable Important in Projection (VIP) > 1.00. In order to further reduce the false positive rate of differential metabolites, based on the above screening methods, inter group differential metabolites were screened by combining Fold Change (FC) and AUC > 0.85. Subsequently, based on precise molecular weight and ion fragmentation information, preliminary identification of group differential metabolites was carried out using the HMDB database. We conducted pathway analysis using KEGG database (https://www.kegg.jp/) and R 4.2.3. Visualization was implemented with ggplot2. Metabolites were mapped to KEGG pathways based on their KEGG compound IDs. Pathway enrichment significance was evaluated using a hypergeometric distribution test. The resulting P-values reflected the statistical significance of pathway enrichment.

Fig. 1.

Study workflow and design. LC-MS: Liquid chromatograph-mass spectrometer; GC-MS: Chromatography-mass spectrometry; PCA: principal component analysis; RF: Random Forest; GBDT: Gradient Boosting Decision Tree; LASSO: Least Absolute Shrinkage and Selection Operator; SVM: Support Vector Machine; LR: Logistic Regression

Results

Patients population characteristics

The clinical and demographic characteristics of the samples are shown in Table 1 and Supplementary File S1. Age, respiratory rate, Neutrophil-to-Lymphocyte Ratio (NLR), C-Reactive Protein (CRP), Procalcitonin (PCT), D-Dimer, blood glucose increased in FG compared to the GPA group, while Hb, hematocrit, albumin, calcium and sodium decreased. There were no significant differences between the FG and GPA groups in term of gender, BMI, temperature, leucocyte, thrombocyte, creatinine and potassium.

Table 1.

Analysis of patient population characteristics

Variables	FG Group (n = 20)	GPA Group(n = 16)	P-value
Age, mean (SD), y	59.90 (13.55)	41.63 (13.36)	<0.001*
Sex (M/F)	15/5	13/3	0.709
BMI, mean (SD)	23.41 (4.36)	3.22 (0.81)	0.164
Temperature, mean (SD), ℃	37.30 (0.74)	37.04 (0.59)	0.314
Respiration rate, mean (SD), /min	19.55 (1.50)	18.44 (0.89)	0.010*
Laboratory data
WBC, mean (SD), *109/L	13.66 (5.19)	11.45 (4.29)	0.086
CRP, mean (SD), mg/L	82.87 (57.61)	28.28 (33.59)	0.003*
PCT, mean (SD), ng/L	0.81 (1.59)	0.11 (0.24)	<0.001*
NLR, mean (SD)	18.11 (30.23)	6.32 (3.27)	0.017*
Hemoglobin, mean (SD), g/dL	127.55 (25.50)	143.50 (14.86)	0.033*
Hematocrit, mean (SD), %	39.01 (12.39)	42.81 (3.93)	0.007*
Platelet count, mean (SD), *109/L	214.75 (110.40)	252.25 (64.30)	0.237
D-Dimer, mean (SD), mg/L	3.17 (2.23)	0.90 (0.53)	<0.001*
Albumin, mean (SD), g/dL	33.67 (10.09)	41.52 (5.98)	0.002*
Creatinine, mean (SD), mg/dL	81.91 (45.81)	71.05 (14.80)	0.899
Sodium, mean (SD), mmol/L	136.99 (3.03)	139.63 (2.81)	0.011*
Calcium, mean (SD), mmol/L	2.14 (0.23)	2.29 (0.11)	0.023*
Potassium, mean (SD), mmol/L	3.79 (0.71)	4.03 (0.25)	0.062
Glucose, mean (SD), mg/dL	10.72 (5.83)	7.18 (3.94)	0.002*

WBC White blood cell count, PCT Procalcitonin, CRP C-reactive protein, NLR Ratio of neutrophil to lymphocyte counts

*Significant results < 0.05

Metabolome profiling differed between the FG and GPA groups

PCA models were established for the sample data of FG and GPA groups, and there was a trend of separation between the two groups. In addition, there was overlap in some sample points between the two groups, indicating a certain similarity between the FG and GPA groups, and there was a certain difficulty in distinguishing the two. The PCA analysis results were shown in Fig. 2 A-1 and Fig. 2 A-2. In order to further explored the differences in metabolic profiles between the FG group and the GPA group, OPLS-DA analysis was conducted based on the metabolic features designed by PCA analysis. The difference between the FG group and the GPA group was more significant in the OPLS-DA of the latter (Fig. 2 B-1 and Fig. 2 B-2). Based on P<0.05 and variable important in projection (VIP) >1.00 as criteria to screen for group differential metabolites，a total of 118 metabolites were identified (Supplementary File S2). The GC platform found 72 metabolites, and the LC platform found 46 metabolites. 22.88% (27/118) of the differential metabolites were found to be Lipids and lipid-like molecules, 19.49% (23/118) were Organic oxygen compounds, 18.64% (22/118) were Organic acids and derivatives, 17.80% (21/118) were Unclassified, 8.47% (10/118) were Organoheterocyclic compounds, 5.08% (6/118) were Organic nitrogen compounds, 3.39% (4/118) were Phenylpropanoids and polyketides, 1.69% (2/118) were Benzenoids, 1.69% (2/118) were Nucleosides, nucleotides, and analogues, and 0.85% (1/118) was Hydrocarbons.

Fig. 2.

Analysis of differential metabolites between FG and GPA populations. (A-1) GC-MS PCA analysis. (A-2) LC-MS PCA analysis. (B-1) GC-MS OPLS-DA analysis. (B-2) LC-MS OPLS-DA analysis

We performed secondary screening on 118 metabolites based on FC > 1.50 or < 0.67 and AUC > 0.85. A total of 10 metabolites were screened out: Bistris, Galactaric acid, Xylofuranose, L-arabitol, Hexaric acid, Glucose-1-phosphate, Galactose, Phenylacetic acid, Phytosphingosine and Myo-inositol (MI). These 10 metabolites showed an upregulation trend in the serum of the FG group, of which 5 were carbohydrate organic compounds, indicating that the upregulation of glucose metabolism may be involved in the occurrence and development of FG.

Utility of ML in the differential of feature selection

To fully utilize information features, we used three ML methods to screen clinical feature datasets, metabolic feature datasets, and mixed feature datasets. A total of 13 clinical and 10 metabolic features were included for further selection (Supplementary File S3). The diagnostic ability of each group of data for FG was preliminarily evaluated using ROC analysis of area under the curve (AUC) and sensitivity (Fig. 3). We found that when using the GBDT method to run a mixed feature dataset, the AUC was 1.00 and the sensitivity was 100%, which was better than other datasets and machine learning methods. Therefore, based on the importance ranking of 13 clinical features and 10 metabolic features by GBDT, feature factors for establishing diagnostic models are selected.

Fig. 3.

Extracting diagnostic model features from three datasets. A Clinical feature dataset ROC curve. B Metabolic characteristic dataset ROC curve. C Mixed feature dataset ROC curve

A machine learning model for disease diagnosis based on metabolic characteristics and clinical data

We select feature factors based on importance greater than 10%. Ultimately, “PCT, Bistris, and MI” were selected for constructing the diagnostic model of FG. LR, SVM, and GBDT were used to evaluate diagnostic ability of the model. Their AUC values were 0.80, 0.82, and 0.95, with sensitivities were 0.90, 0.92, and 1.00. Fig. 4 summarized the accuracy, sensitivity, specificity and clinical practicality of three ML evaluation models for FG diagnosis.

Fig. 4.

Performances of the diagnostic model. A Normalized radar chart of evaluation indicators for machine learning prediction models. The radar chart normalization was performed using maximum-based linear scaling. B Decision curve analysis (DCA) of machine learning diagnostic model. The DCA assessed the clinical utility of the model by calculating the net benefit at different threshold probabilities. In the diagnosis of FG, a higher net benefit indicated that the model assisted clinicians in more accurately identifying high-risk patients while reducing overtreatment of low-risk individuals. As shown in Fig. 4B, when the threshold probability = 0.4, the GBDT model provides an additional net benefit of 270, 270, and 170 per 1000 patients compared to Treat All, LR, and SVM

Metabolic pathway analysis

The difference in metabolite concentration provides a method for capturing the differential metabolic pathways of different diseases. We used the KEGG database to conduct metabolic pathway enrichment analysis, in order to preliminarily explore the metabolic pathway mechanisms related to the occurrence and development of FG. Fourteen pathways were screened based on P < 0.05, including sphingolipid metabolism, Galactose metabolism, Linoleic acid metabolism, Retrograde endocannabinoid signaling, Glucagon signaling pathway, Necroptosis programmed necrosis, Glyceroptosis metabolism, Tryptophan metabolism, Choline metabolism in cancer (Supplementary File S4). Choline metabolism in cancer, Sphingolipid signaling pathway, Citrate cycle (TCA cycle), Renal cell carcinoma, Autophagy, and Pathogenic Escherichia coli infection. Ranking the above 14 pathways with RichFactor (a metric that assesses the enrichment degree of specific metabolic pathways), it was found that Necroptosis was highly enriched, which may be a key metabolic pathway affecting the occurrence of FG. In addition, 2/14 pathways were related to glucose metabolism, including Galactose metabolism and Glucagon signaling pathway. This indicated that the disruption of these 2 glucose metabolism pathways was closely related to the occurrence of FG. The enrichment analysis results were shown in Fig. 5.

Fig. 5.

The significantly altered pathways in FG and GPA. A Metabolites enrichment analysis. B Metabolism pathway analysis in FG

Discussion

The diagnosis of FG is mainly based on a comprehensive analysis of clinical manifestations, laboratory tests, and imaging examinations. It is difficult for inexperienced doctors to identify FG early. According to literature reports, treatment within 24 h of onset can reduce the mortality rate to 0, while after 6 days of onset, the mortality rate can reach as high as 30–50% [9]. Therefore, delaying treatment significantly increases the mortality and disability rates of FG. Early identification and accurate diagnosis can save the lives of FG patients and improve their quality of life. However, the current research hotspot on FG is mainly on prognostic analysis [17]. Early diagnostic biomarkers for FG are still under development and validation.

Currently, an increasing number of studies indicate that differentially expressed metabolites are closely related to the occurrence and development of inflammatory responses. Meanwhile, the collection of serum metabolites is convenient, non-invasive, and can serve as potential biomarkers for the diagnosis, prognosis and treatment of infectious diseases. Therefore, our study conducted a serum metabolomics study. In this study, it was found that FG and GPA have significant differences in metabolic characterization, with a total of 118 metabolites showing significant differences. The main types of metabolites were lipids and organic acids. To further explore the impact of differentially expressed metabolites, we conducted KEGG pathway enrichment analysis. There were significant changes in 14 metabolic pathways, with the most significant difference observed in Necroptosis. Necroptosis is a programmed inflammatory cell death characterized by cell swelling and membrane rupture. Unlike apoptosis, Necroptosis can directly activate or regulate inflammatory responses by releasing intracellular contents through ruptured plasma membranes [18]. Substantial evidence has associated Necroptosis with inflammation across diverse pathological states, suggesting its potential role in disease pathogenesis [19]. Some studies have revealed that regulating key molecules RIPK1, RIPK3 and MLKL in the Necroptosis pathway affects the progression and severity of inflammatory responses [19, 20]. However, whether Necroptosis actively contributes to FG onset or progression remains speculative and warrants experimental validation through functional studies.

FG is not a disease limited to the perianal and perineal regions, usually accompanied by systemic inflammatory reactions. The comprehensive discrimination of multiple biomarkers can improve the accuracy of diagnosis. Therefore, we constructed a differential diagnostic model for FG and GPA using the clinical characteristics, laboratory tests, and metabolic data of patients. A combination of one laboratory parameter “PCT” and two metabolites “MI and Bistris” was selected using three ML methods. Finally, through 10x cross validation, it was found that the model has good predictive performance.

PCT is secreted by thyroid C cells under normal physiological conditions, and only a small portion of PCT enters the peripheral circulation [21]. When the body experiences severe bacterial and fungal infections, sepsis, and multiple organ failure, various tissues and types of cells in the body can produce PCT and release it into the bloodstream [22, 23]. The level of PCT is influenced by factors such as the size of the infected organ, the degree of inflammation, and the immune response status [24]. In our study, serum PCT levels were significantly elevated in FG patients compared to GPA controls, suggesting distinct inflammatory profiles between these conditions. These observations were consistent with more pronounced systemic inflammation in FG. While PCT dynamics may offer adjunctive diagnostic insights for FG assessment, causative relationships require validation through prospective multicenter studies.

Additionally, our study identified MI as a potential biomarker in FG. In literature reports, MI is closely related to glucose metabolism. MI, as an insulin signaling transducer and regulator of glucose metabolism, plays an important role in promoting normal insulin function, maintaining blood glucose stability, and improving glucose metabolism related diseases [25, 26]. Since diabetes is the most common complication of FG, it may have clinical relevance. In addition, FG is a serious infectious disease, which is prone to excessive metabolism and stress hyperglycemia [27, 28]. Notably, we observed significantly elevated serum MI and fasting glucose levels in FG cohorts, suggesting that increased MI may represent a compensatory response to hyperglycemia [29]. Although MI exhibits insulin-mimetic properties that could modulate glucose homeostasis, this mechanistic association requires validation in controlled interventional studies.

Another metabolic feature in early diagnostic models is Bistris. Lourdes Infants found the presence of Bistris on the active site of the crystal structure of phosphatidylcholine phosphatase (PchP) [30]. PchP is typically present in Pseudomonas aeruginosa. PchP participates in hydrolysis and recognizes different structural thresholds of substrates, thereby disrupting host cell membrane phospholipids [31, 32]. Lourdes Infants’ research indicates that complexes with Bistris provide a molecular basis for the catalytic and regulatory functions of PchP [30]. Studies from different databases have shown that PchP exists in various organisms, including Pseudomonas, Burkholderia, and some plant pathogenic fungi [33]. Coincidentally, the pathogenic bacteria of FG have diversity. FG is the result of the joint action of aerobic bacteria, anaerobic bacteria and fungi. At present, there is no article directly proving the correlation between Bistris and FG. However, in our study, we found that the content of Bistris in serum samples from the FG population was significantly higher than that of GPA. Based on the above-mentioned literature research, we hypothesize that Bistris may be involved in the formation of structural thresholds for FG related pathogenic bacterial toxins or enzymes and catalyze their invasion ability. This idea still needs further validation and research.

This study aimed to describe the metabolic differences between FG and GPA. And based on clinical and metabolic research, a serum diagnostic biomarker combination for early identification and diagnosis of FG was constructed. This combination parameter characterized the unique characteristics of FG from three aspects: inflammation degree, glucose metabolism level, and microbial invasion ability. However, our study has several limitations. Firstly, the sample size was relatively small due to the low incidence of FG. To address this, multicenter studies will be conducted to enroll larger patient cohorts. Secondly, our study was limited to a single cohort and lacked an external validation cohort. A prospective multicenter study will be specifically designed to evaluate the clinical applicability of our model. These planned investigations may systematically evaluate the performance of the model for early diagnosis.

Conclusion

This metabolomics study found different metabolic profiles between FG and the control group GPA. Using a series of ML algorithms, we developed a combination parameter of serum metabolites and clinical features (PCT, MI and Bistris) for early identification and diagnosis of FG. Our research supplemented the exploration results of biomarkers for FG diagnosis, which contributed to the development of precise diagnosis of FG. In addition, the activation of Necroptosis pathway may lead to the occurrence of explosive perianal and perineal infections, providing ideas and possibilities for further research on the mechanisms and regulatory pathways of FG.

Abbreviations

FG

Fournier gangrene

GPA

Generalized perianal abscess

LRINEC

Laboratory Risk Indicator for Necrotizing Fasciitis score

MI

Myo-inositol

PCT

Procalcitonin

CRP

C-Reactive protein

PchP

Phosphatidylcholine phosphatase

PCA

Principal component analysis

OPLS-DA

Orthogonal Partial Least Squares Discriminant Analysis

FC

Fold change

VIP

Variable important in projection

NLR

Neutrophil-to-lymphocyte ratio

AUC

Area under the curve

LC-MS

Liquid chromatography tandem mass spectrometry

GC-MS

Gas chromatography mass spectrometry

ML

Machine learning

RF

Random forest

GBDT

Gradient boosting decision tree

LASSO

Least Absolute Shrinkage and Selection Operator

SVM

Support vector machine

LR

Logistic regression

Authors’ contributions

YYB, LJG, and WC conceived and designed the study. ZJY participated in collecting and processing data. ZJY and XCF conducted deep learning and machine learning development. ZJY and WJW did the LC-MS and GC-MS experiments. ZJY analyzed data and wrote the paper. All authors provided critical feedback on the research, analysis and manuscript.

Funding

This work was supported by Shanghai Hospital Development Center Foundation (SHDC12024135); Shanghai key Clinical Specialty (shslczdzk04301); Development highland construction of TCM-the inheritance and extension plan of traditional Chinese Medicine School in Shanghai(ZY[2021-2023-0209]); The second round of construction project of the national Traditional Chinese Medicine academic school inheritance studio ([2019]62); Shanghai Municipal hospital clinical ability promotion and promotion of traditional Chinese medicine surgery specialist alliance(SHDC22024306); Research-Oriented Traditional Chinese Medicine Hospital Construction (YC-2023-0901); Shanghai University of Traditional Chinese Medicine Science and Technology Development Project (24BZH08).

Data availability

Data is provided within the manuscript.

Declarations

Ethics approval and consent to participate

This study was according to the guidelines of the Declaration of Helsinki and was approved the Hospital Ethics Review Board of the Longhua Hospital Affiliated Shanghai University of Traditional Chinese Medicine (NO.2022LCSY072). All included patients provided oral and written informed consent to allow the extra samples be used for academic purpose. We confirmed that all data was anonymized and maintained as strictly confidential.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.