Multicentre derivation and validation of a prognostic scoring system for mortality assessment in HIV‐infected patients with talaromycosis

Abstract Background Although the widespread use of modern antiretroviral therapy (ART) has reduced the incidence of talaromycosis in people living with HIV, mortality remains as high as 20% in this population, even after appropriate antifungal treatment. Objectives The objective of our study was to develop a risk assessment system for HIV‐infected patients with comorbid talaromycosis, in order to provide these patients with appropriate, effective and potentially life‐saving interventions at an early stage of their illness. Patients/Methods This was a multicentre, retrospective cohort study conducted in China. We built a predictive model based on data from 11 hospitals, and a validated model using the data of 1 hospital located in an endemic area. Results Forward stepwise multivariate statistical calculations indicated that age, aspartate aminotransferase/alanine transaminase ratio and albumin levels, and BUN levels were valid, independent predictors of the risk of death in HIV‐infected patients with talaromycosis. Our developed and validated risk scoring system is effective for the identification of HIV‐infected patients with talaromycosis at high risk of death at hospital admission (p < .001; AUC = 0.860). In our study, our risk prediction model provided functional and robust discrimination in the validation cohort (p < .001; AUC = 0.793). Conclusion The prognostic scoring system for mortality assessment developed in the present study is an easy‐to‐use clinical tool designed to accurately assist clinicians in identifying high‐risk patients with talaromycosis.

Conclusion: The prognostic scoring system for mortality assessment developed in the present study is an easy-to-use clinical tool designed to accurately assist clinicians in identifying high-risk patients with talaromycosis.

K E Y W O R D S
AIDS, HIV, mortality, risk scoring system, Talaromycosis

| INTRODUC TI ON
Talaromyces marneffei (T. marneffei), formerly known as Penicillium marneffei (P. marneffei), is a dimorphic fungus that causes life-threatening infection in immunocompromised patients, especially in endemic areas such as South-East Asia, Southern China and India. 1 As one of the AIDS-defining diseases, talaromycosis represents a significant burden on patients with AIDS, and mortality and morbidity rates are second only to the mortality and morbidity wreaked by tuberculosis and cryptococcosis in Thailand. 2 Although the widespread use of modern antiretroviral therapy has reduced the incidence of talaromycosis in people living with HIV, its mortality remains up to 20%, even after appropriate antifungal treatment. [3][4][5][6] Several past studies have assessed the risk factors associated with death in patients with talaromycosis. 4,7 However, there remains no clinical risk scoring tool for people living with HIV who suffer from talaromycosis. Providing a risk scoring system to identify high-risk patients with talaromycosis is useful in the clinical care of patients with AIDS. The objective of our study was to establish a clinical risk assessment system for patients with talaromycosis in order to provide appropriate, effective and life-saving interventions in this population at an early stage of their illness.

| Study design and patient selection
This study was a multicentre, retrospective cohort study con-

| Definitions
A confirmed diagnosed case of talaromycosis met one of following criteria: (a) T. marneffei found in histological biopsy specimens, and (b) T. marneffei cultured from clinical specimens.

| Predictor variables
We explored demographic factors, comorbid illnesses, symptoms, signs and laboratory test results at admission, including haemoglobin levels, platelet counts, white blood cell counts, aspartate aminotransferase/alanine transaminase (AST/ALT) ratios, albumin levels, total bilirubin levels, creatinine levels, lactate dehydrogenase (LDH), CD4 + T-cell counts and blood urea nitrogen (BUN) levels, in order to select surrogate variables for risk of death in patients with talaromycosis.

| Statistical analysis
In the derivative cohort, the demographic, clinical and laboratory variables of patients were analysed at the time of admission by standard descriptive statistics. Continuous variables were described using median with interquartile ranges (IQR). Categorical variables were described as frequency rates and percentages. Univariate analysis was performed using logistic regression to explore factors associated with death.
Predictors with continuous variables in the univariate analysis (p ≤ .1) were converted to categorical variables by grouping values using cutoff points based on either the median, or clinically relevant values, and each point score was assigned for each categorical variable. To achieve the scoring sums, age, AST/ALT ratio, albumin level were first included separately in a logistic regression model. Factors that were significant in the step were simultaneously placed into a single logistic regression model. We used the variance inflation factor (VIF) to make assessments of multicollinearity among the independent variables, but there were not any. From the logistic regression models, the co-efficient scores were assigned for each factor with the respective β coefficients. The co-efficient scores in the results were rounded off to the nearest 0.5. A specific risk score is derived by a calculation of each co-efficient score multiplied by each point score. The score for an individual was obtained by summing the scores for the risk score of each of the risk factors.
To verify the validity of the scoring system, we plotted receiver operating characteristic (ROC) curves for patients in the derivative cohort and those in the validation cohort respectively, and compared the areas under these ROC curves.

| RE SULTS
A total of 1137 cases of HIV-positive patients with talaromycosis were screened in this study from the 12 included hospitals. Seven patients were excluded for being under 18 years of age, and 513 patients were excluded for data loss. Figure 1 provides a flowchart for patient inclusion and available data for analysis.

| Construction of the Predictive Model using the derivation cohort
We assigned 'yes' a value of 1, and 'no' a value of 0, as point scores to the dichotomous variables in Table 1. All continuous variables in Table 1 associated with mortality in the univariate analysis (p ≤ .1) were converted into dummy variables. The assignment of these values is shown in Table 2.
All dummy variables associated with mortality in the univariate analysis (p ≤ .1), were included in a logistic regression model. Forward stepwise multivariate analysis indicated that age, AST/ALT ratios, and albumin levels were independent predictors of the risk of death in HIV-infected patients with talaromycosis. The co-efficient scores in the results were rounded off to the nearest 0.5. The details of all significant surrogates in the final stepwise logistic regression model are listed in Table 3.

| Model assessment in the derivative cohort
Our scoring system is valid for the identification of HIV-infected patients with talaromycosis at high risk of death on hospital admission (p < .001; AUC = 0.860; Figure 2). We chose a total score greater than 8.5 as the predictive value for high-risk patients. When using a score of 8.5 as the cut-off value to predict the mortality risk of patients with talaromycosis, the sensitivity of our scoring system was 84.0% (95% CI, 0.639-0.955), its specificity was 71.6% (95% CI, 0.666-0.762), its positive predictive value (PPV) was 17.1% (95% CI, 0.140-0.207), its negative predictive value (NPV) was 98.5% (95% CI, 0.96.3-0.994), and its Youden index, which measures the performance of our scoring system, was 55.59% (Additional file 1: Table S1).
To facilitate the application of this predictive model in the clinic, we categorized the patients with these risk factors into the following 2 groups: a 'low'-risk group (≤8.5 score) and a 'high'-risk group (>8.5 score). The mortality rate increased rapidly as the scores increased. In the derivation cohort, the mortality rate was significantly higher in the 'high'-risk group (21/123, 20.2%) than that in the 'low'-risk group (4/261, 1.1%) (p < .001), as shown in Table 5.

| DISCUSS ION
The present multicentre, retrospective study included cohorts from endemic and non-endemic areas of China. In the present study, the overall in-hospital mortality in the three cohorts was 6.4%, which is relatively lower than that found in other studies, 3 transaminases are specifically associated with mortality in patients with talaromycosis. Another study from Vietnam found that fatal talaromycosis cases have a higher average AST level than nonfatal cases. 7 We included AST/ALT in our forward stepwise multivariate logistic regression analysis, and the results indicate that as the ratio increases, the risk of death for patients with talaromycosis also increased. Hypoproteinemia is often a manifestation of a patient's progression to cachexia. Our study found that lower albumin levels are an indicator of poor prognosis.
Many risk factors for death have been reported in patients with talaromycosis. A retrospective study conducted in Shanghai (unfortunately with a relatively small sample size) indicated that low C0D4 + T-cell counts and low haemoglobin levels were associated with higher mortality of talaromycosis among HIV-infected patients. 11 A recent study indicated that CD4 + T-cell counts and CD4+/ CD8 + ratios are critical determinants of prognosis in T. marneffei-infected patients with respiratory system lesions. 12 However, most patients in these cohorts had CD4 + T-cell counts less than 50 cells/ μL, and there was no statistically significant difference in CD4 + T-cell counts between the survival group and the non-survival group. Even though some patients had initiated ART before hospitalization, their mortality rates were calculated to be not statistically distinct from ART-naive patients. Thus, we speculated that the majority of talaromycosis cases occur in ART-naive patients with low immune function, and the vitality of immune function of these patients does not seem to contribute to mortality during hospitalization in a clinically appreciable manner.
The risk assessment scale designed in this study can distinguish high-risk patients from low-risk patients to a certain extent, especially for low-risk patients. Since the definitive diagnosis of talaromycosis mainly relies on the identification of the aetiologic organism by culture, and the return of the culture results usually takes a few days, physicians often need to decide on whether to initiate empirical antifungal treatment on not. Initiating empirical antifungal treatment for mortality low-risk patients would increase their pharmaceutical and economic burden. With the help of our scoring assessment model, it will be easier for clinicians to a more appropriate decision.
We included all potential risk factors with p-values < . In conclusion, the prognostic scoring system for mortality assessment developed in the present study is an easy-to-use tool designed to assist attending clinicians in the rapid and accurate stratification between low-risk and high-risk patients with talaromycosis, thereby assisting clinicians to formulate an expeditious and effective treatment strategy based on the patients' risk of mortality. The novel clinical risk scoring system is likely to facilitate closer clinical monitoring of talaromycosis patients at higher risk of mortality, and to provide support for empirical antifungal treatment in high-risk patients, and also to support clinical decisions to prescribe precise antifungal treatment in low-risk patients with a definitive diagnosis.

ACK N OWLED G EM ENTS
We express our sincere thanks and gratitude to all medical personnel including clinical fellows, nurses, and cleaners of 12 hospitals from

CO N FLI C T O F I NTE R E S T
All authors declare that they have no competing interests.

DATA AVA I L A B I L I T Y S TAT E M E N T
The data that support the findings of this study are available from the corresponding author upon reasonable request