Talaromycosis is classified as an invasive fungal infection, predominantly affecting individuals with weakened immune systems, particularly those with HIV. It is widespread in Asia's tropical and subtropical regions, notably in Southeast Asian nations like Vietnam, Thailand, and southern China [1]. In Southeast Asia, it accounts for between 6.4% and 11% of hospital admissions related to HIV in Vietnam and 3.3% in Thailand. In southern China, the prevalence is higher, with Guangxi and Guangdong reporting rates of 16.1% and 17.3%, respectively, highlighting its significant burden in these areas [1–6]. Despite the implementation of antifungal therapy, delayed diagnosis contributes to a high in-hospital mortality rate among patients with talaromycosis, ranging from 16.7–30% [7–9]. However, the efficacy of these interventions relies on the timely identification of high-risk individuals.

The typical clinical features of Talaromyces marneffei (TM) infection commonly include fever, weight loss, lower HB levels, generalized fatigue, skin lesions, peripheral or abdominal lymphadenopathy (POAL), and hepatosplenomegaly [10]. Studies have demonstrated that skin lesions of TM infection often present as central umbilicated lesions, which can serve as critical diagnostic indicators for early detection. Approximately one-third of patients with TM infection exhibit skin lesions [11]. Microbiological culture is currently considered the definitive standard for diagnosing talaromycosis. However, this method is time-consuming, and the isolation or identification of pathogens from clinical specimens often takes more than 10 days. Such delays may result in the postponement of antifungal therapy initiation, potentially adversely affecting patient outcomes [12–14]. Despite the development of several highly specific serological diagnostic techniques, false-negative outcomes continue to pose a challenge for the prompt identification of TM infection [13].

Early initiation of antifungal therapy in talaromycosis can significantly reduce mortality and mitigate the severity of the disease [7, 15–17]. However, the efficacy of targeted interventions depends on the timely identification of high-risk individuals. Although recent studies have used dynamic nomograms to evaluate the potential for TM infection in hospitalized individuals with HIV, relying on skin lesions as a risk factor may reduce diagnostic ACC for those without skin lesions [18]. There is a critical demand for quantitative, accessible, and efficient approaches to evaluate the potential risk of talaromycosis in HIV-infected patients lacking skin lesions.

In recent years, ML techniques utilizing EMR have gained significant traction and acceptance among healthcare professionals. Unlike conventional statistical approaches, ML algorithms impose fewer limitations on data requirements and excel at modeling intricate datasets [19], driving their growing application in the medical field. However, the inherent complexity of ML models, especially their “opaque system” nature, often limits transparency in understanding their decision-making processes [20]. To overcome this limitation, model interpretation tools are essential for elucidating the underlying mechanisms of ML models. This study utilized the Shapley Additive Explanations (SHAP) methodology to develop an interpretable model. This tool enhances the interpretability and clinical utility of ML-based predictions by clearly visualizing each variable's contribution to the model's outcomes. Additionally, this study developed an accessible web-based application for implementing the predictive model, enabling clinicians to utilize the model for program prediction without necessitating the installation of Python or the acquisition of programming skills.

The diagnostic model's outcome variable determined the presence of talaromycosis in patients. Talaromycosis was diagnosed when TM was isolated from clinical specimens using standard culture techniques [12] or identified in biopsy tissue histopathology.

The sample size was determined using the events per variable (EPV) metric [21, 22], a widely recognized approach in statistical analyses. In southern China, the incidence of HIV co-infection with Talaromyces marneffei is 0.16. Considering our objective to include twelve predictor variables and establish the EPV at 10, the required sample size was calculated using the following formula:

The patient data were retrospectively extracted from hospital admission records covering the period from August 3, 2010, to January 17, 2019. Missing data in the clinical records were managed through mean substitution. Initially, our model was developed through a systematic review, meta-analysis, and expert consensus, incorporating a comprehensive array of variables. These variables encompassed the presence of oral candidiasis, tuberculosis, bacterial pneumonia, pneumocystis pneumonia, cryptococcosis, cytomegalovirus disease, herpes simplex virus disease, and lymphoma. Additionally, demographic and clinical factors such as sex, age, body mass index (BMI), nationality, occupation, marital status, and injection drug use were included. The model also considered clinical symptoms and laboratory findings, including fever, cough, poor appetite, hepatomegaly, splenomegaly, hemoglobin (Hb) levels, white blood cell count (WBC), platelet count (PLT), absolute lymphocyte count (ALC), C-reactive protein (CRP), AST, ALT, AST/ALT ratio, albumin (ALB), blood urea nitrogen (BUN), creatinine (Cr), (1–3)-β-D glucan (G) levels, CD4⁺ T-cell count(CD4), and the presence of hepatitis C and hepatitis B.

Furthermore, potential variables were subjected to an additional screening process based on dimensionality reduction criteria to evaluate their appropriateness for inclusion in the model [23]. The Random Forest (RF) algorithm is an ensemble method that combines several decision trees. It generates models by applying random sampling to the training dataset and determining optimal split points. Individual decision trees in RF are constructed based on feature metrics derived from dataset attributes [24, 25], facilitating a robust assessment of each feature's significance [26]. Lasso regression performs variable selection and controls model complexity through regularization, effectively preventing overfitting. The parameter λ controls regularization strength, leading to a simplified model with fewer variables. The optimal value of λ(λ₋min) is identified using 10-fold cross-validation, where the point of minimum error is used to select the most relevant predictive variables [27, 28]. Unlike traditional feature selection methods, Boruta uses a wrapper-based method to select features. It aims to determine the feature set that demonstrates the strongest association with the dependent variable, prioritizing comprehensive relevance over the creation of a minimal, model-specific subset [29]. Through the iterative elimination of low-correlation features, this method significantly reduces noise and enhances the consistency of classification outcomes [30]. Meanwhile, as a highly potent ensemble learning technique, Extreme Gradient Boosting (XGBoost) is established based on classification trees. By integrating multiple weak learners into a robust ensemble model through sequential boosting steps, it constructs a tree-based classifier, providing a reliable approach for precise classification tasks [31]. Finally, Mutual Information (MI) effectively captures non-linear relationships, complementing RF for robust feature importance, Lasso for sparsity, Boruta for comprehensive relevance, and XGBoost for high-precision ensemble learning in medical data analysis [32].

Specifically, we utilized Lasso regression, RF, Boruta, XGBoost, and MI to analyze the 36 potential risk factors. Variables with non-zero coefficients were ranked according to their contribution to the outcome, and common variables were identified by intersecting the results from all five methods.

Interpretability refers to the clarification of how ML models generate outcomes. The inherent opacity of machine learning models often hinders their clinical application, prompting significant research to improve interpretability [33, 34]. This study presents an intuitive model explanation framework employing the SHAP (Shapley Additive exPlanations) BreakDown package, which quantifies the extent of individual predictors' contributions to model predictions. Furthermore, we developed an intuitive model interpreter that works independently of any specific model and uses input predictor variables to interpret results. The Gradio package facilitated the incorporation of predictive variables and optimal modeling into an interactive web application.

This multicenter retrospective cohort study utilized electronic medical record data from endemic and non-endemic regions in China to develop a machine learning model that accurately predicts talaromycosis risk in HIV-infected patients without skin lesions, demonstrating strong discriminative capacity and clinical applicability. The model's performance was evaluated at an independent medical center. This new model offers personalized risk prediction via a comprehensive multi-step process that includes variable selection, model development, validation, analysis, understanding, and online deployment, surpassing previously published models in terms of accuracy and usability. Featuring a user-friendly interface and practical application, it provides significant clinical value to aid clinicians in swift decision-making. This can facilitate targeted interventions, such as whether to initiate antifungal treatment. These measures can substantially decrease patient mortality and economic burden.

It is commonly recognized that ML models provide enhanced predictive capabilities compared to conventional models in the medical field [40]. This benefit facilitates the development of a more robust model from complex datasets [41]. This study underscores the proficiency of machine learning in managing heterogeneous multidimensional data through the integration of twelve predictive variables. For example, the elevation of AST and the reduction of PLT show high heterogeneity. Additionally, we employed visualization techniques to clarify the forecast model and create functional web applications. This step is crucial for addressing concerns related to the opacity of black-box models and for rebuilding trust in machine learning within the medical sector [42]. To enhance reproducibility, the model development process incorporated cross-checking, self-validation, and outside confirmation, all conducted with predetermined random seeds. This is a crucial step to ensure the clinical significance of the model through its interpretability [43].

In recent years, studies on HIV-infected patients hospitalized with TM infection have gradually increased. This infection is more common in Southeast Asia, especially among immunocompromised patients. However, while many studies focus on the prognostic risks of these patients, there are relatively few studies predicting the risk of infection. Xu L et al [18]. Employed multivariable logistic regression to create a forecasting model using clinical indicators for ten key factors from 1,564 cases in four tertiary hospitals in Southwest China to evaluate talaromycosis risk in hospitalized HIV-positive patients. The model attained AUC values of 0.883 in training and 0.889 in validation, highlighting the significant role of skin lesions as a key variable in disease prediction within diagnostic models. However, for patients with HIV-infected TM infection who do not have skin lesions, the use of this model may have certain limitations.

We created and externally validated a talaromycosis prediction model for HIV-infected patients lacking skin lesions, using 12 clinical variables from the electronic medical records of four hospitals. To address collinearity in the model development data, we used RF, Lasso, XGBoost, Boruta, and MI to select the 36 influential variables, employing L1 regularization. The study aimed to create a practical clinical application by selecting a limited number of readily accessible patient markers. Taking practicality into account, if a single marker can fulfill the prediction objective, there is no need to incorporate additional variables. More significantly, these predictive variables serve as baseline markers. Additionally, drawing on the findings from previous literature, these 36 variables are frequently used indicators in similar research studies. The study retrospectively analyzed 9 years of patient data. Moreover, these 36 variables represent the most extensive and readily accessible data within our electronic medical records system. We employed eight machine learning algorithms to individually fit each of the 12 predictive variables, aiming to achieve optimal fitting results. The SVM algorithm generated a model with a high ACC (0.714) and an AUC of 0.809. This model performed exceptionally well on both internal and external validation datasets.

Xu L et al [18]. Reported that the AST, PLT, age, POAL, and CD4⁺ T-cell count were independent factors influencing the diagnosis of TM infection and subsequently formulated a clinically applicable predictive model based on these findings. Our study has arrived at similar conclusions, with these five variables standing out as key contributors to the diagnosis of TM infection. This finding is further corroborated by the network application we developed using Gradio.

Studies indicate that TM infections are more prevalent and deadly among individuals infected with HIV in Southeast Asia [1]. Upon diagnosis, immediate and active treatment measures must be taken. High model ACC is essential to reduce false negatives (missed diagnoses) and false positives (misdiagnoses). The SVM model demonstrated robust diagnostic performance during the validation process. During the training phase, it achieved an ACC of 0.714, an AUC of 0.809, a specificity of 0.795, and a sensitivity of 0.631. Notably, the model's performance metrics, including ACC, AUC, specificity, sensitivity, and F1-score, showed improvement in both the testing and external validation datasets compared to the training phase. For instance, in the external validation cohort, the SVM model attained an AUC of 0.921, an ACC of 0.853, and an F1-score of 0.819. These enhancements underscore the model's strong generalization capability and reliability. The model's high accuracy, specificity, and sensitivity in detecting TM infection render it highly promising for clinical application in the early diagnosis and risk stratification of TM infection in HIV-positive patients without skin lesions. Its consistent performance across multiple validation stages further emphasizes its potential utility in real-world clinical settings.

Our study has several limitations that we must acknowledge. Firstly, the study was carried out in China, with the selection of participants mainly based on local populations. Consequently, applying these results to global populations may introduce potential biases, a frequent issue with predictive models [44–46]. The method can be readily adapted for application in other countries, and we foresee future opportunities for its implementation [47]. Consistent with other studies, HIV-infected individuals with talaromycosis often exhibit severe immunosuppression, as indicated by lower CD4⁺ T-cell counts [6, 9, 14]. This study targeted individuals with CD4⁺ T-cell counts under 200 cells/µL, who are more susceptible to opportunistic infections and health issues. Utilizing a multi-stage validation approach, which encompassed 10-fold cross-validation and independent external validation, we achieved a substantial enhancement in the predictive accuracy of the model for assessing the risk of Talaromyces marneffei infection in HIV-infected individuals lacking skin lesions. The calibration curve exhibited a slope approaching 1.0, signifying an almost linear correlation between predicted and observed probabilities, which is considered optimal. The model's capability for risk stratification is consistent with the clinical objectives outlined in the World Health Organization's 2023 guidelines on priority fungal pathogens [48]. This alignment provides evidence-based support for decision-making regarding the early administration of antifungal treatment to high-risk patients in endemic regions. Thirdly, the development model's 12 predictive variables are available upon admission. In brief, clinicians need to be alert to possible co-infections with talaromycosis in patients with lower CD4⁺ T-cell counts, especially in areas where such infections are common, even in the absence of skin lesions. However, in real-world clinical settings, the diagnosis of TM infection is influenced by a variety of factors, such as disease severity, physician experience, and the choice of diagnostic tools. Fortunately, in recent years, with the growing attention to rare infectious diseases, awareness of TM infection diagnosis has also gradually increased. By optimizing these factors, the accuracy and timeliness of the diagnosis can be improved, thus enhancing patient prognosis. In addition, when applied to populations in non-endemic areas, our model may have a higher misdiagnosis rate. In recent years, we have expanded data collection from additional hospitals and aim to enhance the model using the latest data from non-endemic regions and various centers. Simultaneously, we are undertaking prospective multicenter studies (ChiCTR1900021195).

To conclude, we have created a clinically beneficial and user-friendly network application. While prospective validation remains essential, predicting the risk of major complications in HIV-associated talaromycosis patients without skin lesions is now possible. This advancement facilitates proactive and targeted clinical decision-making, enabling proactive and targeted clinical decision-making.