Introduction

Severe fever with thrombocytopenia syndrome virus (SFTSV), a novel tick-borne bunyavirus, is the etiologic agent of severe fever with thrombocytopenia syndrome (SFTS). Human infection with SFTSV and its pathogenicity were first identified and confirmed in China [1]. SFTS is predominantly a tick-borne zoonosis with geographically focal endemicity in China and other East Asian regions. Clinically, SFTS typically presents with acute fever, thrombocytopenia, and leukopenia, often accompanied by multiorgan dysfunction; a subset of patients develops coagulopathy, including disseminated intravascular coagulation [2,3]. HLH is a syndrome of pathologic immune overactivation characterized by prolonged high-grade fever, cytopenias, hyperferritinemia, and coagulopathy. The syndrome progresses rapidly and, if untreated, is associated with high mortality [4,5]. HLH encompasses genetically determined primary (familial) disease and secondary forms triggered by infection, malignancy, or autoimmune/rheumatologic disorders [4,6]. SFTSV infection can precipitate sHLH, and existing data indicate that HLH is a life-threatening complication of SFTS associated with poor clinical outcomes and increased mortality [7]. However, prospective data to support risk stratification for HLH in the setting of SFTSV infection are scarce, and available early prediction tools and objective biomarkers remain inadequately developed and validated [5,8]. HScore is a composite clinical scoring system developed to estimate the probability of reactive hemophagocytic syndrome on the basis of features such as fever, organomegaly, cytopenias, ferritin, triglycerides, and fibrinogen; however, it was not specifically developed for SFTS and may rely on data that accumulate over time. Accordingly, early recognition of patients at high risk for SFTS-associated HLH using variables obtainable soon after virologic confirmation could facilitate closer monitoring, earlier diagnostic reassessment, and more individualized management. In this study, we retrospectively evaluated clinical characteristics and laboratory parameters obtained within 3 days after virologic confirmation in hospitalized patients with SFTS, identified factors associated with HLH, and developed a multivariable model to assess its utility for early risk stratification of SFTS-associated HLH [8].

Materials and methods

Variables and definitions

Clinical data for both cohorts were retrieved from the hospital electronic medical record system, including interval from symptom onset to admission, peak preadmission body temperature, duration of fever, lymphadenopathy, splenomegaly, and neurologic manifestations (e.g., altered mental status, headache, vomiting). The interval from symptom onset to admission was recorded and was used as an operational summary of symptom duration at the time early data were captured for model development. In addition, laboratory parameters obtained within 3 days of virologic confirmation were recorded, including hemoglobin, leukocyte and neutrophil counts, platelet count, triglycerides, creatine kinase, alanine aminotransferase, aspartate aminotransferase, lactate dehydrogenase, coagulation indices, and SFTSV blood viral load. This same predefined early-evaluation window was used for descriptive early HScore accounting. The exact laboratory report-return time was not consistently extractable from the retrospective record; however, in routine clinical practice, positive nucleic acid results were typically returned within approximately half a day of virologic confirmation. Accordingly, the recorded time of virologic confirmation was used as the operational temporal anchor for early predictor ascertainment and early HScore assessment. Quantitative values exceeding the upper limit of detection were truncated to that limit. For the small proportion of missing continuous data, imputation followed established methods: approximately normally distributed variables were imputed using the mean, whereas non-normally distributed variables were imputed using the median. Continuous variables with large order-of-magnitude variation were converted into categorical variables according to published evidence and international practice [10–16]. For laboratory variables summarized in Table 1, the most abnormal value within this 3-day window was used (maximum or minimum as appropriate).

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. The demographic and clinical characteristics of patients with SFTS with and without HLH.

https://doi.org/10.1371/journal.pntd.0014247.t001

Results

A total of 249 hospitalized patients with confirmed SFTS were included in the analysis (Table 1). The median age was 63 years (IQR, 55–69), and 47% were male. The median interval from symptom onset to admission was 7 days (IQR, 5–8), which served as an operational summary of symptom duration at the time data were captured for model development. Most patients had persistent fever (median duration, 7 days; IQR, 6–9). Lymphadenopathy and splenomegaly were present in 53% and 6% of patients, respectively. Neurologic manifestations—most commonly headache, psychomotor slowing, and altered mental status—were recorded. Gastrointestinal involvement was frequent, with vomiting or diarrhea reported in 49% and concurrent vomiting and diarrhea in 13%. Hematologically, thrombocytopenia was prominent, with a median platelet count of 50 × 10⁹/L (IQR, 34–68). Hypertriglyceridemia, marked hyperferritinemia, and coagulation abnormalities were frequently observed. Most patients had detectable SFTSV viremia ≥10⁴ copies, and elevations in creatine kinase, aminotransferases, and lactate dehydrogenase were common. Outcome accounting based on HScore and HLH-2004 criteria, including distributions within 3 days after virologic confirmation and over the entire clinical course, is provided in Supporting Information S1–S3 Tables. Notably, at the early evaluation point, 83 of 97 patients (85.6%) in the final HLH group already met the HScore threshold for HLH (S1 Table), indicating that the model should be interpreted as a tool for early risk stratification within a clinically relevant window rather than as proof of preclinical prediction. In the overall cohort, 83 of 249 patients (33.3%) met the HScore threshold within 3 days after virologic confirmation, whereas 97 of 249 (39.0%) met the threshold over the entire clinical course.

The Non-HLH cohort comprised 152 patients. The median interval from symptom onset to admission was 7 days (IQR, 5–7). Marked hyperthermia was uncommon; 73% had a maximum temperature ≤39.0°C. The median duration of fever was 7 days (IQR, 6–9). Lymphadenopathy was present in 47%, whereas splenomegaly was rare (2%). Neurologic involvement was infrequent, with psychomotor slowing in 13% and altered mental status in 2%. The median platelet count was 56.5 × 10⁹/L (IQR, 40.75–75). Triglyceride levels were lower, with a median of 2.03 mmol/L (IQR, 1.50–2.67), whereas fibrinogen was relatively preserved (mean 2.38 g/L). Serum ferritin was < 2000 ng/mL in most patients (84%), and only 7% had levels >6000 ng/mL. Viral load also tended to be lower, with 14% of patients having <10⁴ copies. Moderate-to-severe elevations in creatine kinase, AST, and LDH were less frequent than in the HLH cohort. Coagulation profiles were comparatively preserved, with predominantly normal PT and fewer marked prolongations of APTT and TT.

Relative to the Non-HLH cohort, the HLH cohort had a modestly longer onset-to-admission interval (P = 0.02), higher peak temperature and more prolonged fever duration (both P < 0.001), more frequent lymphadenopathy and splenomegaly (P = 0.018 and P = 0.004, respectively), and more frequent neurologic manifestations (P = 0.002). In the HLH cohort, laboratory abnormalities included lower platelet counts (P < 0.001); higher triglycerides and ferritin with a rightward shift toward ≥2,000 and >6,000 ng/mL (P < 0.001); lower fibrinogen (P < 0.001); higher SFTSV viral load (P < 0.001); greater elevations in CK, AST, ALT, and LDH with a higher AST/ALT ratio (all P ≤ 0.013); and more frequent prolongation of APTT and TT, with no significant difference in PT (P < 0.001 for APTT/TT; P = 0.39 for PT). Overall, the HLH cohort exhibited more profound thrombocytopenia, a heightened inflammatory phenotype, higher levels of viremia, and more severe coagulation derangement. Areas under the receiver operating characteristic curve (AUCs) were calculated to quantify the discriminatory performance of individual early clinical and laboratory variables for HLH complicating SFTS (Fig 1).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Comparison of the discriminatory performance of individual early clinical and laboratory variables for HLH complicating SFTS.

Comparison of the areas under the receiver operating characteristic curves (AUCs) for early clinical and laboratory variables measured within 3 days after virologic confirmation. Each bar represents the AUC with its 95% confidence interval for an individual predictor distinguishing SFTS patients who developed HLH from those who did not. Variables are ordered from highest to lowest AUC, with markers at the top indicating the strongest discriminatory ability.

https://doi.org/10.1371/journal.pntd.0014247.g001

LASSO regression was applied to 28 candidate predictors, followed by refinement with the Boruta algorithm, yielding six variables—peak preadmission temperature, splenomegaly, fever duration, triglycerides, fibrinogen, and ferritin—for inclusion in the multivariable logistic regression model. A multivariable logistic regression model incorporating the six selected predictors was developed and presented as a nomogram (Fig 2). Calibration was assessed using calibration curves (Fig 3). In the overall cohort (n = 249), the likelihood ratio chi-square was 214.82 (P < 0.0001) with R² = 0.784. The model demonstrated excellent discrimination, with a C-index of 0.962 and a Dxy of 0.923, indicating robust separation between low- and high-risk patients.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Nomogram for bedside early risk stratification of secondary hemophagocytic lymphohistiocytosis in patients with severe fever with thrombocytopenia syndrome.

This nomogram is intended for use during the early hospital evaluation of patients with laboratory-confirmed SFTS. After locating the patient’s value for each predictor on the corresponding axis, draw a vertical line upward to the Points axis to assign points, sum the points across all predictors, and project the total vertically to the Predicted Probability of HLH scale to estimate individual risk. Higher total points indicate a higher probability of HLH and may justify closer monitoring, repeated HLH assessment, and earlier specialist consultation. Predictor coding: For Tmax, categories 1–3 correspond to ≤38.0 °C, 38.1–39.0 °C, and ≥39.1 °C, respectively. For ferritin, categories 1–3 correspond to <2000 ng/mL, 2000–6000 ng/mL, and >6000 ng/mL, respectively. For splenomegaly, 0 indicates absence and 1 indicates presence.

https://doi.org/10.1371/journal.pntd.0014247.g002

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Discrimination, calibration, and clinical utility of the SFTS–HLH prediction model in the derivation and internal validation cohorts.

Panel (A) shows the receiver operating characteristic (ROC) curve of the multivariable model in the training cohort, with the area under the ROC curve (AUC) quantifying discrimination; panel (B) displays the corresponding calibration plot comparing predicted and observed risks of secondary HLH in the training cohort; panel (C) shows the ROC curve of the model in the internal validation cohort; and panel (D) shows the calibration plot in the internal validation cohort after intercept-and-slope recalibration. Panels (E) and (F) display decision-curve analyses in the training and internal validation cohorts, illustrating the net clinical benefit of using the prediction model across a range of threshold probabilities compared with “treat-all” and “treat-none” strategies.

https://doi.org/10.1371/journal.pntd.0014247.g003

To evaluate the performance of the combined prediction model, the cohort of 249 patients was randomly split in an 85:15 ratio into derivation and internal validation sets. Discrimination in the derivation cohort was excellent (Fig 3A), and apparent calibration was near-ideal (Fig 3B), with an intercept close to 0, a calibration slope of 1.00, and an average absolute calibration error (Eavg) of 0.0025. In the internal validation set, discrimination remained nearly perfect (Fig 3C; AUC 0.997, 95% CI 0.989–1.000), but calibration was suboptimal, as reflected by a positive intercept (+4.14) and an inflated slope (2.53), indicating systematic underestimation and inadequate dispersion of predicted risk. Intercept-and-slope recalibration substantially improved concordance between predicted and observed risk (Fig 3D). Decision curve analysis demonstrated a positive net clinical benefit across almost the entire range of reasonable decision thresholds (Fig 3E and 3F).

The external validation cohort was derived from The First Affiliated Hospital with Anhui Medical University, where eligible patients were retrospectively identified using the same criteria and procedures. In this cohort (n = 60), the model (Fig 4A) demonstrated excellent discrimination (AUC 0.907, 95% CI 0.835–0.980). Calibration indicated a small positive intercept (+0.233), consistent with slight underestimation of absolute risk, and a calibration slope of 0.896, suggesting mild overfitting and overdispersion of predicted probabilities. Following intercept-and-slope recalibration, agreement between predicted and observed risk improved substantially (Fig 4B). Consistent with the internal validation, decision curve analysis in the external cohort showed that the model conferred a positive net clinical benefit across nearly the entire range of clinically relevant threshold probabilities (Fig 4C).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. External validation of the SFTS–HLH prediction model.

(A) Receiver operating characteristic (ROC) curve of the prediction model in the external validation cohort, with the area under the ROC curve (AUC) quantifying discrimination. (B) Calibration plot comparing predicted and observed probabilities of secondary HLH in the external cohort. (C) Decision-curve analysis in the external validation cohort, showing the net benefit of applying the model across a range of decision thresholds relative to “treat-all” and “treat-none” strategies.

https://doi.org/10.1371/journal.pntd.0014247.g004

Discussion

Since HLH was recognized as a major complication of SFTS, most published data have consisted of case reports or small retrospective series, and robust evidence on population-level incidence and independent risk factors remains scarce [17–21]. A single-center retrospective study from Korea was the first to systematically evaluate the impact of sHLH on outcomes in SFTS, demonstrating a markedly increased case fatality, with mortality reaching 75% (3 of 4 patients) in a small cohort [7]. In Japan, available data are largely limited to case reports, small case series, and reviews. The overall case-fatality rate of SFTS is approximately 26–30%, and the contribution of concomitant HLH to adverse outcomes has attracted increasing attention [22–25]. In adults, sHLH remains associated with poor outcomes; early case-fatality rates of approximately 30–40% have been reported and are even higher in severe cases and ICU populations [5,26,27]. In China, cohort data on SFTS complicated by HLH have emerged in recent years. One cohort of laboratory-confirmed SFTS cases estimated the incidence of HLH at approximately 7.2% and demonstrated that concomitant HLH was associated with an increased risk of death, with an overall case-fatality rate of about 24.5% in that cohort [19,20]. Taken together, data from China parallel observations from Korea and Japan: occurrence of HLH in SFTS is associated with excess mortality and underscores the need for timely recognition, close monitoring, individualized management, and rigorous management of complications to reduce case fatality and improve long-term outcomes. These convergent findings highlight the need to recognize SFTS patients at high risk for HLH at an early stage and to institute targeted immunomodulatory and intensive supportive care without delay [28–30]. Compared with reports from Korea and Japan, which have largely consisted of small retrospective cohorts, case series, or fatal-case descriptions, and with recent studies from China that have focused primarily on incidence, mortality, or general prognostic factors, the present study differs in three respects. First, it specifically targets HLH risk stratification within a cohort of laboratory-confirmed SFTS. Second, it uses six routinely accessible variables available during the early clinical course. Third, it includes external validation in an independent cohort. These features may enhance bedside applicability, although broader geographic validation remains necessary. On this basis, we sought to systematically characterize risk factors for HLH complicating SFTS in a real-world regional cohort and to develop a combined prediction model. The goal was to generate a quantitative tool for early risk stratification of SFTS patients at high risk of sHLH, thereby facilitating closer monitoring, repeated HLH assessment, and optimization of individualized management strategies.

The analysis revealed that greater peak preadmission temperature, splenomegaly, extended duration of fever, hypertriglyceridemia, low fibrinogen, and elevated ferritin are risk factors for HLH complicating SFTS. A multivariable model incorporating these variables provided reliable risk estimation for SFTS-associated HLH within the predefined early-evaluation window and demonstrated robust performance in the derivation cohort. The model showed good overall fit (R² = 0.784), with satisfactory calibration (calibration intercept close to 0 and slope of 1.00) and excellent discrimination (C-index 0.962; Dxy 0.923), indicating near-complete separation between patients at high and low risk of HLH. In both the internal validation set and the external validation cohort, discriminatory performance remained high (AUC 0.997 and 0.907, respectively). Decision curve analysis demonstrated a positive net clinical benefit across nearly the entire range of decision thresholds, supporting the clinical utility of the model and suggesting its generalizability across centers.

Patients with SFTS are at substantial risk for developing sHLH. Once sHLH ensues, the clinical phenotype is characterized by a more intense inflammatory response, more profound thrombocytopenia, more severe coagulopathy, higher circulating viral burden, and worse prognosis. Notably, the six risk indicators incorporated into the model—high peak temperature, splenomegaly, prolonged fever duration, hypertriglyceridemia, hypofibrinogenemia, and marked hyperferritinemia—are readily available early in the hospital course and can be assessed at the bedside with routine clinical and laboratory evaluation. A multivariable model incorporating these parameters may assist early risk stratification during the initial hospital course and could support closer monitoring, repeated HLH assessment, earlier specialist consultation, and timely escalation of supportive management when clinically indicated. Whether use of the model translates into improved survival requires prospective validation.

An important consideration in interpreting the present model is that it was designed to support early risk stratification within a prespecified clinical window rather than to prove that HLH can be identified earlier than with HScore. Notably, 83 of 97 patients (85.6%) in the final HLH group already met HScore criteria at the early evaluation point, indicating that some cases had already evolved to an HLH phenotype by the time early study variables were captured. Accordingly, the incremental advantage of this model over HScore in terms of timeliness cannot be established from the current retrospective dataset. Instead, the principal value of the model may lie in its use of six routinely available variables to facilitate bedside recognition of high-risk SFTS patients, prompt closer reassessment, and support earlier diagnostic escalation. Prospective studies directly comparing this model with HScore in terms of time-to-identification and downstream clinical decision-making are warranted.

From a clinical workflow perspective, the model may be most useful shortly after virologic confirmation and before overt multiorgan failure or unequivocal HLH has been established. Patients flagged as high risk could undergo closer monitoring of ferritin, fibrinogen, triglycerides, cytopenias, coagulation indices, viral load, and organ dysfunction, with earlier hematology consultation and reassessment for evolving HLH. In contrast, once a patient already fulfills established diagnostic criteria for HLH or has progressed to advanced organ failure, the incremental value of a risk-stratification tool is likely to diminish, because management should then be driven primarily by established diagnostic and critical-care frameworks.

Several limitations merit consideration. First, the retrospective design entails risks of selection bias, information bias, and residual confounding. Second, although data were abstracted from electronic medical records using a standardized protocol, some variables were missing and required imputation (mean or median for continuous measures), which may have attenuated true variability and biased effect estimates. Third, the binary outcome was defined using the entire-course HScore, which incorporates laboratory features that overlap with several candidate predictors. To enhance transparency, we provide detailed outcome accounting for both the early window and the entire clinical course (S1–S3 Tables). Importantly, 83 of 97 patients (85.6%) in the final HLH group already met the HScore threshold at the early evaluation point. Therefore, although the present model may be useful for early risk stratification within a clinically relevant window, the current study does not establish that it identifies HLH earlier than HScore. In addition, several HLH-2004 criteria, including NK-cell activity and soluble CD25, were not assessed in this cohort, and hemophagocytosis was not uniformly evaluated; thus, misclassification of borderline or evolving HLH cannot be entirely excluded. The model was derived at a single provincial center and validated in one external hospital, which may limit generalizability. Moreover, treatment variables and outcomes beyond the in-hospital course were not incorporated, precluding direct assessment of whether early risk identification followed by targeted intervention translates into improved survival.

Future studies should address these limitations. Prospective, multicenter validation in diverse endemic settings is needed to confirm robustness, evaluate transportability, and enable site-specific recalibration where appropriate. Incorporating longitudinal laboratory trajectories and organ support requirements may allow development of a dynamic, time-to-HLH or time-to-deterioration prediction model rather than a single baseline binary risk estimate. To demonstrate clinical utility, it will be essential to link predicted risk to standardized early management pathways—including antiviral therapy, organ protection, and anti-inflammatory or immunoregulatory strategies—and to evaluate patient-centered outcomes such as mortality and organ failure–free survival. Ultimately, implementation of this model in routine triage may improve early risk recognition for sHLH in SFTS; however, its clinical use should be accompanied by ongoing performance surveillance, periodic recalibration, and alignment with evolving standards of care.

Supporting information