Machine learning model based on clinical and imaging features for predicting fungal infections in children with leukemia

Abstract

Background

Due to immunosuppression, mucosal barrier injury, and prolonged neutropenia resulting from both the disease and chemotherapy, along with the frequent use of broad-spectrum antibiotics and glucocorticoids, children with leukemia are at a high risk of invasive fungal disease (IFD). The present study aimed to develop an effective machine learning model to predict fungal infections in children with leukemia.

Methods

A total of 247 pediatric patients diagnosed with leukemia and concurrent infections were evaluated. Five distinct ML classifiers—Random Forest, Logistic Regression, Support Vector Machine (SVM), Naïve Bayes, and K-Nearest Neighbors—were employed to construct predictive classification models. These models were trained using three distinct feature sets: (1) clinical features exclusively, (2) imaging features exclusively, and (3) an integrated feature set comprising both clinical and imaging data. The predictive model was validated prospectively in an independent cohort of 61 patients. Model performance was evaluated through cross-validation techniques to ensure robustness and generalizability. To validate the clinical applicability of the ML models, their diagnostic performance was systematically compared against that of three radiologists with varying experience levels: Reader A (3 years), Reader B (6 years), and Reader C (11 years).

Results

Among the five classifiers evaluated, models using both clinical and imaging features consistently outperformed those relying solely on either clinical or imaging features. Notably, the SVM algorithm exhibited the highest overall predictive performance. Within the SVM algorithm, the validation set achieved the mean area under the curve (AUC) values of 0.825 with clinical features alone, 0.852 with imaging features alone, and 0.947 when both clinical and imaging features were combined. The corresponding mean AUC values for the test set were 0.777, 0.797, and 0.879. Furthermore, a comparative analysis between the classification results of the SVM model and the diagnostic assessments provided by three radiologists demonstrated that the SVM consistently outperformed the radiologists across key performance metrics.

Conclusions

The SVM algorithm demonstrates robust efficacy in predicting fungal infections among pediatric patients diagnosed with leukemia. Within the predictive model, the variables that exhibited the greatest influence included pleural thickening, neutropenia, hormone therapy, CRP level, mediastinal lymphadenopathy, and the presence of pleural effusion.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12887-026-07149-y.

Introduction

Over the years, the overall survival rates for pediatric leukemia have markedly improved, with current overall survival rates approaching approximately 90% for acute lymphoblastic leukemia and 70% for acute myeloid leukemia [1, 2]. Despite these advancements, invasive fungal disease (IFD) continues to represent a significant cause of morbidity and mortality among children undergoing treatment for hematologic malignancies [3]. Notably, the incidence of IFD has shown an upward trend in recent years, a phenomenon largely attributed to enhanced diagnostic methods, intensive chemotherapy regimens, prolonged episodes of neutropenia, and the increased utilization of central venous catheters [4]. Although the therapeutic landscape has evolved with the introduction of novel antifungal agents, IFD-associated mortality remains alarmingly high, with reported fatality rates ranging from 20% to 70% across various clinical studies [3–5].

The current diagnostic framework for fungal infections primarily relies on the galactomannan assay, (1→3)-β-D-glucan (BDG) assay, next-generation sequencing for gene detection, and radiological findings via chest computed tomography (CT). However, the BDG assay may yield negative results during the initial stages of infection. Furthermore, the characteristic chest CT features commonly identified in adult pulmonary mold infections are frequently absent or less pronounced in pediatric patients, particularly those under 5 years of age [6]. Given the high morbidity and mortality rates associated with fungal infections in patients with leukemia, the early identification of patients at high risk is critical to inform timely and effective therapeutic interventions.

Machine learning (ML) methodologies have become increasingly integral to various domains within medicine, particularly in the realms of outcome prediction, diagnostic processes, and analysis of medical imaging data [7, 8]. In the present study, we employed ML techniques to systematically identify clinical risk factors and radiological features that are indicative of fungal infections in pediatric patients diagnosed with leukemia. Furthermore, we constructed and validated a predictive model designed to accurately forecast the likelihood of fungal infections within this vulnerable cohort.

Materials and methods

Patient population

The study was approved by the Ethics Committee of the Children’s Hospital of Soochow University(No.2025CS160). For the retrospective cohort, the requirement for informed consent was waived. For the prospective cohort, Informed consent was obtained from the legal surrogates of the subjects following a detailed description of the purpose of the study. Eligible cases were retrospectively identified from pediatric inpatients admitted to the Department of Hematology at the Children’s Hospital of Soochow University between January 1, 2021, and January 1, 2025. The prospective cohort was identified at the same hospital between February 1, 2025, and February 1, 2026.

Definition

According to the definition of invasive fungal disease (IFD) by the European Organization for Research and Treatment of Cancer/Mycoses Study Group Education and Research Consortium (EORTC/MSGERC) [9], proven IFD requires fungal hyphae identified microscopically, or recovery of fungi from culture of normally sterile specimens. Probable IFD requires the presence of at least 1 host factor, a clinical feature and mycologic evidence and is proposed for immunocompromised patients only. Whereas cases that meet the criteria for a host factor and a clinical feature but for which mycological evidence has not been found are considered possible IFD, and such patients were excluded from the present study. Possible IFD were not included in this study. Excluding children with mixed infections and incomplete data, a total of 97 children were eventually included in the study. The non-fungal infection group was included according to the definition of microbiologically documented infection [10]: identification of a pathogen from blood or other infection sites using standard culture techniques routinely performed at our institution. The control group comprised 74 cases of bacterial condition, 45 cases of viral condition, and 31 cases of noninfectious condition, totaling 150 non-fungal infection cases. Additionally, we enrolled 50 patients between February 2025 and February 2026 from the Children’s Hospital of Soochow University—the prospective test cohort (Fig. 1).

Fig. 1.

Machine learning flowchart of this study

Clinical data were systematically collected and comprised demographic and laboratory parameters, including age, sex, white blood cell (WBC) count, platelet count, absolute neutrophil count (ANC), presence of neutropenia, levels of C-reactive protein (CRP), procalcitonin (PCT) level, and history of hormone therapy. Imaging features were independently evaluated by two experienced radiologists who were blinded to clinical information. The imaging assessment encompassed a comprehensive array of findings: linear or reticular opacities, patchy areas of consolidation, pulmonary nodules, presence of air bronchogram, bronchiectasis, ground-glass opacities (GGOs), cavitary lesions, air crescent signs, mediastinal lymphadenopathy, pleural effusion, and pleural thickening.

Data partitioning

The dataset consisted of 247 pediatric patients, comprising 97 diagnosed with fungal infections and 150 without evidence of infection. To maintain proportional representation of each class, the dataset was stratified into five folds. Three folds were designated for model training, one fold for the validation set, and the remaining fold for the test set. The test set comprised 19 patients with fungal infections and 30 without evidence of infection. The stratified fold partition was randomly generated once and remained fixed for model development and evaluation procedure. In addition, the prospective test cohort of 61 pediatric patients, including 17 diagnosed with fungal infections and 44 without evidence of infection, was used exclusively as an independent test set to evaluate the generalizability of the machine learning model developed from the retrospective cohort.

Model construction and performance assessment

To construct predictive models, five distinct classification algorithms were employed: Random Forest (RF), Logistic Regression (LR), Support Vector Machine (SVM), Naïve Bayes (Bayes), and K-Nearest Neighbor (KNN). These models were trained using the three previously defined feature sets. Hyperparameter optimization for each classification model was conducted through an exhaustive grid search strategy on the validation dataset. The SVM model developed using the integrated feature set achieved the optimal configuration by employing a Radial Basis Function kernel. The most effective parameter settings identified during model tuning were a penalty coefficient (C) of 32 and a kernel coefficient (γ) of 2 × 10^− 9. This configuration, which demonstrated superior performance on the validation set, was subsequently employed for performance evaluation on the test datasets.

The model performance was assessed using a range of standard classification metrics, including accuracy, precision, sensitivity (recall), specificity, F1-score, receiver operating characteristic (ROC) curves, and the area under the ROC curve (AUC-ROC).

Statistical analysis

Continuous variables with a non-normal distribution were expressed as the median (interquartile range, IQR) and tested using the Mann–Whitney U test. Categorical.

variables were expressed as frequencies and percentages and tested using either the Chi-square or Fisher exact probability test. The calibration performance was evaluated using calibration curves and brier score [11]. For radiological features, Kappa analysis was also performed to assess inter-rater agreement. A two-tailed P value < 0.05 was defined as statistically significant. Statistical analysis and graphic production were performed using Python 3.10.12 and SPSS.

Reading protocols and blinding conditions

Three radiologists with varying experience in the interpretation of IFD based on chest CT images and/or clinical history were selected for this study. Observers included: a senior thoracic radiologist (reader A), a junior thoracic radiologist (reader B), and a thoracic radiology fellow in training (reader C). Reader A with 11 years of experience, Reader B with 6 years of experience, and Reader C with 3 years of experience. Each reader was asked to classify a patient into fungal or non-fungal infection based on only radiographic images (CT images). Only CT images from a patient’s initial scan were given to readers. Standardized criteria were utilized to make the IFD diagnosis based on the imaging criteria for IFD proposed by the EORTC/MSGERC consensus guideline [12]. Each reader conducted their assessments independently and were blinded to the diagnosis. No previous studies were provided for comparison.

For the purpose of evaluating diagnostic performance, the radiologists were fully blinded to: (1) the final diagnosis (IFD or non-IFD), (2) all clinical and laboratory data (including host factors, neutropenia status, galactomannan results, antifungal prophylaxis/treatment, and microbiological culture results).

Results

Univariate analysis of fungal infections in leukemia

Univariate analysis of patient data revealed statistically significant differences (P < 0.05) between groups in the following variables: WBC count, platelet count, ANC, presence of neutropenia, CRP levels, history of hormone therapy, and radiological features, including linear or reticular opacities, pulmonary nodules, air bronchogram, bronchiectasis, GGO, cavitation, air crescent sign, mediastinal lymphadenopathy, pleural effusion, and pleural thickening. In contrast, no statistically significant differences (P > 0.05) were observed for age, sex, PCT levels, or the presence of patchy consolidation (Table 1).

Table 1.

Clinical and imaging characteristics of fungal versus non-fungal infections in leukemia patients

Characteristics	Fungal (n = 97)	Non-fungal (n = 150)	P-value
Age (years)	8 (4.5, 11)	6 (4, 10)	0.133
Gender (%)
Male	55 (56.7)	93 (62.0)	0.407
Female	42 (43.3)	57 (38.0)
WBC count (109/L)	2.79 (1.05, 5.59)	3.90 (2.11, 7.28)	0.003
Platelet count (109/L)	67 (33, 147)	120.5 (62.75, 210)	0.000
ANC (109/L)	0.30 (0.05, 1.08)	0.95 (0.36, 1.89)	0.000
CRP (mg/L)	16.20 (3.78, 49.96)	4.37 (0.67, 9.46)	0.000
PCT (ng/mL)	0.16 (0.08, 0.25)	0.19 (0.11, 0.34)	0.057
Neutropenia (%)	54 (55.7)	33 (22.0)	0.000
Hormone therapy (%)	51 (52.6)	50 (33.3)	0.003
Linear or reticular opacities (%)	53 (54.6)	52 (34.7)	0.002
Patchy consolidation (%)	60 (61.9)	79 (52.7)	0.155
Pulmonary nodule (%)	46 (47.4)	25 (16.7)	0.000
Air bronchogram (%)	47 (48.5)	16 (10.7)	0.000
Bronchiectasis (%)	17 (17.5)	5 (3.3)	0.000
GGOs (%)	63 (64.9)	57 (38.0)	0.000
Cavitation (%)	26 (26.8)	6 (4.0)	0.000
Air crescent sign (%)	10 (10.3)	0 (0.0)	0.000*
Mediastinal lymphadenopathy (%)	38 (39.2)	9 (6.0)	0.000
Pleural effusion (%)	34 (35.1)	13 (8.7)	0.000
Pleural thickening (%)	44 (45.4)	11 (7.3)	0.000

WBC White blood cell, ANC Absolute neutrophil count, CRP C-Reactive protein, PCT Procalcitonin, GGOs Ground-glass opacities

Predictive performance of clinical, image, and combined models

A comprehensive summary of performance metrics for the five classifiers was conducted across clinical, imaging, and integrated feature models on both validation and test datasets. The model incorporating combined features consistently demonstrated superior performance compared with models using either clinical or imaging features alone, as detailed in the Supplementary Material (Table S1-5). In the combined feature model, mean AUC values on the validation set were as follows: RF (0.885), SVM (0.947), KNN (0.935), Bayes (0.861), and LR (0.895). On the test set, the corresponding AUC values were as follows: RF (0.860), SVM (0.879), KNN (0.852), Bayes (0.789), and LR (0.824).

Among all evaluated classifiers, the SVM algorithm exhibited superior diagnostic performance. Table 2 summarizes the diagnostic performance of SVM models developed using three distinct feature sets. The combined model peaked at an AUC of 0.947 for predicting IFD in pediatric leukemia patients in the retrospective training set (sensitivity: 90%, specificity: 90%, accuracy: 90%, precision: 85.7%, F1-score: 0.878). Performance remained high in both the retrospective (AUC = 0.879) and prospective (AUC = 0.882) test sets.

Table 2.

Performance of different SVM models

		Clinic model	Image model	Combined model
Validation cohort	AUC (95%CI)	0.825(0.692–0.934)	0.851(0.719–0.950)	0.947(0.873–1.000)
	Sensitivity	0.550(0.333–0.769)	0.700(0.500–0.888)	0.900(0.765–1.000)
	Specificity	0.800(0.647–0.937)	0.866(0.735–0.970)	0.900(0.786–1.000)
	Accuracy	0.700	0.800	0.900
	Precision	0.647	0.777	0.857
	F1-score	0.594	0.736	0.878
Test cohort	AUC (95%CI)	0.777(0.630–0.913)	0.796(0.653–0.917)	0.879(0.745–0.983)
	Sensitivity	0.684(0.466–0.875)	0.631(0.400–0.850)	0.842(0.650–1.000)
	Specificity	0.700(0.535–0.857)	0.800(0.645–0.935)	0.866(0.733–0.969)
	Accuracy	0.693	0.734	0.857
	Precision	0.590	0.666	0.800
	F1-score	0.634	0.648	0.820
Prospective cohort	AUC (95%CI)	0.772(0.625–0.897)	0.839(0.702–0.940)	0.882(0.776–0.963)
	Sensitivity	0.588(0.357–0.833)	0.647(0.400–0.866)	0.706(0.471–0.929)
	Specificity	0.814(0.684–0.919)	0.907(0.805–0.978)	0.884(0.783–0.975)
	Accuracy	0.750	0.833	0.833
	Precision	0.556	0.733	0.706
	F1-score	0.571	0.688	0.705

SVM Support vector machine, AUC Area under the curve, CI Confidence interval

Figure 2a and b presents the ROC curves and decision curve analysis comparing the combined-feature model across the five classifiers. Figure 3 depicts both ROC and precision-recall (PR) curves for the three SVM models evaluated on the test dataset. To further determine the optimal model, we analyzed calibration curves for both the retrospective and prospective test sets. The plots showed that the combined model had the lowest Brier score, reflecting the best calibration (Fig. 4). Figure 5 illustrates the relative importance of features in the combined-feature SVM model. Analysis using Shapley Additive Explanations (SHAP) values (Fig. 6) identified pleural thickening, neutropenia, hormone therapy, CRP, mediastinal lymphadenopathy, and pleural effusion as the top predictive factors.

Fig. 2.

a Receiver operating characteristic curves of five machine-learning classifiers. SVM AUC = 0.879; RF AUC = 0.861; KNN AUC = 0.853; LR AUC = 0.825; Bayes AUC = 0.789. b Decision curve analysis of the five machine-learning classifiers

Fig. 3.

Performance evaluation of the three SVM models: (a) Precision-Recall curves and (b) Receiver operating characteristic curves in the test set

Fig. 4.

Calibration curves and Brier scores of the three models in the retrospective (a) and prospective (b) test sets

Fig. 5.

Ranking of predictive contribution for variables in the combined-feature SVM model in the retrospective (a) and prospective (b) test sets

Fig. 6.

The SHapley additive exPlanations (SHAP) summary plot of the the combined-feature SVM model in the retrospective (a) and prospective (b) test sets. Input variables are ranked in a descending order of feature importance. Red indicates positive contribution; blue indicates negative contribution

Performance comparison: combined model vs. radiologists

This study further conducted a comparative analysis between the classification results generated by the SVM model and the diagnostic assessments performed manually by three radiologists with varying levels of clinical experience: Reader A with 11 years of experience, Reader B with 6 years of experience, and Reader C with 3 years of experience. The SVM model outperformed all radiologists across key evaluation metrics, including accuracy, sensitivity, and F1-score, while also demonstrating robust performance in other indicators. Specifically, the model’s accuracy marginally surpassed that of the most experienced radiologist (Reader C). Furthermore, the model exhibited markedly superior sensitivity compared with all three radiologists, underscoring its enhanced ability to correctly identify positive cases. This finding highlights the model’s considerable potential for clinical application, as illustrated in Table 3; Fig. 7.

Table 3.

Diagnostic results by three radiologists with different levels of experience

	AUC	Sensitivity	Specificity	Accuracy	Precision	F1-score
Reader A	0.623	0.309	0.826	0.623	0.535	0.392
Reader B	0.717	0.474	0.873	0.716	0.707	0.567
Reader C	0.814	0.608	0.946	0.814	0.880	0.719
SVM	0.879	0.842	0.866	0.857	0.800	0.820

SVM Support vector machine, AUC Area under the curve

Fig. 7.

Confusion matrix for pneumonia classification by reader A, reader B, and reader C

Discussion

This study employed five ML classifiers to construct predictive models based on 20 variables. The results demonstrated that the combined clinical-imaging model significantly outperformed single-feature models across all metrics, with the SVM classifier achieving the highest overall accuracy (AUC = 0.947). SHAP analysis further revealed that pleural thickening, neutropenia, hormone therapy, CRP, mediastinal lymphadenopathy, and pleural effusion were the most influential predictors.

SVM is a well-established learning algorithm developed from statistical learning theory. When implemented with a linear kernel, SVM can effectively manage high-dimensional feature spaces while minimizing the risk of overfitting. Owing to its robust generalization capability, it has been widely applied to pattern recognition tasks involving relatively small sample sizes, consistently demonstrating superior performance [13, 14].

In the present study, WBC count, ANC, CRP, and PCT were identified as key clinical biomarkers of systemic inflammation. Although WBC count is the most widely used routine practice, its diagnostic specificity for severe infections is limited due to various pathophysiological and iatrogenic factors [15]. Notably pediatric patients with leukemia and documented fungal infections demonstrated markedly lower WBC counts than those without fungal infection. Further prospective, multicenter studies are required to validate its predictive value for fungal infections. In pediatric patients with acute leukemia, both quantitative depletion and functional impairment of neutrophils markedly increase susceptibility to bacterial, viral, and fungal infections. Chemotherapeutic regimens frequently aggravate these deficits, thereby leading to prolonged neutropenia and further increasing infection risk. Substantial evidence indicates that severe or sustained neutropenia, along with the administration of purine analogs during treatment, represents a significant predisposing factor for fungal infections in patients with acute leukemia [16, 17]. Prior research delineated a critical threshold for the duration of neutropenia, beyond which the risk of infection increases sharply [18]. Furthermore, multiple studies have demonstrated a strong positive correlation between neutropenia duration and invasive fungal infection, indicating that extended periods of neutropenia markedly heighten the probability of breakthrough infections [17, 19].

PCT and CRP have emerged as validated biomarkers for the diagnosis and surveillance of severe infections [20, 21]. CRP remains one of the most widely accessible and clinically used inflammatory markers, exhibiting rapid serum elevation in response to infection or tissue injury [22]. Some studies have reported CRP as a sensitive early marker of fungal infection, reflecting its prompt rise during the initial phase of disease [23]. However, its specificity remains controversial, with conflicting evidence regarding its ability to reliably distinguish fungal infections from other etiologies. Notably, neonatal studies demonstrate significantly higher CRP concentrations in fungal infections than in bacterial infections [24]. In contrast, investigations in adult patient cohorts show no significant difference between fungal and bacterial groups [25]. These discrepancies underscore the heterogeneity of biomarker responses across various patient subgroups and disease etiologies.

Diverse studies across varied patient cohorts and clinical settings have examined the diagnostic and predictive utility of PCT. Several studies demonstrated higher serum PCT concentrations in gram-negative sepsis than in gram-positive or fungal sepsis, indicating a possible role in differentiating bacterial from fungal infections [26]. Nonetheless, the suboptimal sensitivity and specificity reported for PCT challenge its dependability as a standalone marker for this purpose [27]. Emerging data reveal inconsistent patterns in PCT kinetics during fungal infection, with certain investigations reporting persistently elevated concentrations [20], while others describe a progressive decline as disease advances [28]. These limitations diminish the reliability of either CRP or PCT as independent diagnostic indicators, highlighting the necessity for further prospective testing. Notably, two recent studies proposed that elevated CRP with suppressed PCT levels in immunocompromised adults may indicate fungal infection [23, 29]. This observation supports the emerging evidence that combining CRP and PCT measurements may improve diagnostic precision [21, 30]. Our study demonstrated significantly elevated CRP concentrations (P < 0.05) and a near-significant reduction in PCT levels (P = 0.057) among children with leukemia and fungal infections compared with those without fungal disease.

In adults, CT is widely recognized as a valuable modality for predicting early therapeutic response. Segmental or lobar consolidation is considered a typical manifestation of mold-associated pneumonia and is incorporated as a clinical criterion in the European Organization for Research and Treatment of Cancer/Mycoses Study Group Education and Research Consortium (EORTC/MSGERC) definitions for IFD [9]. In pediatric patients, particularly those under 5 years, these adult-defined radiographic hallmarks are frequently absent. Instead, pulmonary mold infections typically present with multiple nodules, nodular halo signs, and mass-like lesions as the predominant imaging abnormalities in this age group [31, 32]. In our cohort of children with leukemia, pleural thickening, mediastinal lymphadenopathy, and pleural effusion emerged as the most significant predictors of invasive fungal infection. Pleural thickening has traditionally been regarded as a manifestation of pulmonary inflammation, presumed to result from the extension of infection from consolidated lung parenchyma to the visceral pleura. The pathobiology of pleural thickening is multifactorial, involving complex interactions among multiple cellular components and inflammatory mediators, modulated by both host immune status and pathogen-specific characteristics. Predisposing host factors for pleural fungal infections include severe inherited immunodeficiencies and low absolute CD4 + T-lymphocyte counts [33, 34]. Notably, children with leukemia exhibit profound immune dysfunction compared with both healthy pediatric cohorts and adults, leading to attenuated antifungal immunity. Pleural thickening is frequently observed in severe infections, and it often reflects an acute exacerbation phase of disease. This finding often correlates with pleural invasion by fungal hyphae and, in a subset of patients, may be further complicated by empyema or the development of a bronchopleural fistula [35]. In pediatric leukemia, impaired lymphocyte maturation results in markedly reduced CD4 + antigen expression, culminating in profound cellular immune dysfunction. Relevant studies indicate a correlation between mediastinal lymphadenopathy and absolute CD4 + T-lymphocyte counts, supporting its use as a reliable surrogate marker for decreased absolute CD4 + T-lymphocyte counts [36, 37]. Lower absolute CD4 + T-lymphocyte counts reflect compromised immune function, thereby predisposing affected children to severe opportunistic infections. In contrast, the association between pleural effusion and fungal infections remains insufficiently characterized, and the underlying pathophysiological mechanisms are yet to be elucidated. Pleural effusions associated with Cryptococcus were first documented in a 1968 review by W. Warr et al. [38]. A recent study indicates that pleural effusion occurs with greater frequency in patients receiving prolonged corticosteroid therapy, immunosuppressive agents, or in those with severe immunodeficiency than in immunocompetent individuals [39]. Nevertheless, the underlying mechanisms driving the development of fungal pleural effusion remain poorly defined and require further investigation.

Additionally, assessment of diagnostic concordance among radiologists with varying levels of expertise demonstrated substantial inter-observer variability in classifying pulmonary fungal infections. Even radiologists with 12 years of professional experience encountered challenges in achieving consistent and accurate identification of these fungal infections. We hypothesize that this diagnostic difficulty arises from the absence of typical adult imaging manifestations in pediatric pulmonary fungal infections, compounded by the interpretive subjectivity inherent to CT-based pneumonia assessment [40]. Our integrated model, which combines clinical and radiological parameters, achieved markedly higher accuracy than radiologists with 3–5 years of experience in predicting fungal infections. It also marginally exceeded the performance of those with 12 years of experience.

Several ML models have been developed within this field. Zhang et al. [41] evaluated a model that clinic-radiological data, CT radiomics, and deep learning for the prediction of invasive pulmonary aspergillosis, with the AUC of 0.881 in the test set. Bhavani et al. [42] constructed logistic regression and gradient boosting machine-based prediction algorithms to identify bacteremia and fungemia among hospitalized patients. However, the majority have focused on adult populations. To our knowledge, this is the first machine learning-based IFD study specifically dedicated to pediatric leukemia patients. According to the EORTC/MSGERC consensus definitions, children are recognized as a distinct population, and radiographic findings of IFD are less specific in children than those reported in adults [9]. This fundamental difference limits the generalizability of adult-derived models to pediatric populations, and thus direct comparisons were not performed. A multicenter external validation will be needed to evaluate the generalizability and robustness of our model across diverse populations, laboratory platforms, and real-world clinical settings.

Limitations

Our study has several limitations. First, a prospective cohort from the same center was used as the test set, which may restrict the external validity of the model when applied to other centers. Second, the potential inclusion of patients with co-infections cannot be entirely ruled out due to prolonged hospital stays, which may affect the results. These considerations highlight the necessity for large-sample, prospective, multicenter studies to further refine and validate the proposed predictive models.

Conclusion

Our SVM-based ML algorithm exhibits robust predictive capability for identifying invasive fungal infections in pediatric patients with leukemia. The key variables contributing most significantly to the model’s performance included pleural thickening, neutropenia, hormone therapy, elevated CRP levels, mediastinal lymphadenopathy, and pleural effusion. Further prospective studies are necessary to validate the clinical utility of this predictive model.

Abbreviations

ML

Machine Learning

RF

Random Forest

LR

Logistic Regression

SVM

Support Vector Machine

Bayes

Naïve Bayes

KNN

K-Nearest Neighbor

IFD

Invasive Fungal Disease

BDG

(1→3)-β-D-Glucan

WBC

White Blood Cell (WBC)

ANC

Absolute Neutrophil Count (ANC)

CRP

C-Reactive Protein (CRP)

PCT

Procalcitonin (PCT)

GGOs

Ground-Glass Opacities

AUC

Area Under The Curve

ROC

Receiver Operating Characteristic (ROC)

SHAP

Shapley Additive Explanations (SHAP)

Authors’ contributions

Wan-liang Guo and Ya-kang Dai designed the study. Peng Ge and Yu He Data acquisition. Peng Ge and Xu-sheng Qian Analysis and data interpretation. Peng Ge, Xu-sheng Qian and Yu He Drafting of the manuscript. Peng Ge and Wan-liang Guo Critical revision.

Funding

This work was supported by Suzhou Science & Technology Projects (SZS2024007), Suzhou Medical Key Discipline (No. SZXK202510), Suzhou Science & Technology Projects (SSD2024021).

Data availability

Availability of data and materials: Data are available upon request from the corresponding author.

Declarations

Ethics approval and consent to participate

The study was approved by the Ethics Committee of the Children’s Hospital of Soochow University（No.2025CS160）. For the retrospective cohort, the requirement for informed consent was waived. For the prospective cohort, Informed consent was obtained from the legal surrogates of the subjects following a detailed description of the purpose of the study.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.