Predicting mortality risk of severe fever with thrombocytopenia syndrome: A multi-center retrospective cohort study Authors: Xue-Geng Hong, Hong-Han Ge, Ning Cui, Yan-Li Xu, Xin Yang, Jia-Hao Chen, Xiao-Hong Yin, Yi-Mei Yuan, Chao Zhou, Hao Li, Xiao-Ai Zhang, Ming Yue, Ling Lin, Wei Liu Published: 2025 ABSTRACT Severe fever with thrombocytopenia syndrome (SFTS) is an emerging tick-borne disease with high mortality, and clinical practice lacks dynamic tools to assess its rapidly evolving course. This study aims to develop stage-specific machine learning models to predict mortality risk using longitudinal biomarker data. We conducted a retrospective analysis of 5359 laboratory-confirmed SFTS patients from two hospitals in the highly endemic region in China. Serial measurements of 46 clinical and laboratory variables were integrated into a three-stage prognostic model developed using extreme gradient boosting (XGBoost). Within each clinical stage, key predictors and their relative contribution (RC) of mortality risk were assessed. Model performance was assessed based on discrimination, calibration, and decision curve analysis (DCA) in internal and external test sets. XGBoost models were constructed across 10 temporal phases, later consolidated into three clinically distinct stages via hierarchical clustering: early (≤7 days), intermediate (days 8–9), and late (≥10 days). Key predictors included age (dominant in early phase; RC, 18.44%), lactate dehydrogenase (LDH; RC peaking at 60.10% in late phase), and monocyte percentage (RC range from 5.25% to 16.04%). Pathophysiological shifts across clinical stages were revealed: early viral cytopathy (dominated by age and MONO%), intermediate immunopathology (marked by LDH surge), and late hepatic failure (dominated by LDH, AST, and TBA). The model showed strong discrimination (Area under the receiver operating characteristic curve, AUCs: 0.84–0.98 internal; 0.91–0.98 external), calibration (Brier scores: 0.04–0.11), and clinical utility via DCA. This study introduces a dynamic staging system that leverages predictive models and real-time patient data to monitor mortality risk and personalize SFTS care, which enables timely interventions to reduce deaths. Introduction Severe fever with thrombocytopenia syndrome (SFTS) is an emerging tick-borne hemorrhagic fever that was first reported in rural central China in 2009 (Yu et al., 2011). Since its discovery, SFTS cases have been reported across several Asian countries, including South Korea, Japan, Vietnam, and Myanmar (Kim et al., 2013; Takahashi et al., 2014; Tran et al., 2019; Win et al., 2020). The causative pathogen, originally termed SFTS virus (SFTSV), was reclassified in 2022 as Bandavirus dabieense, a member of the genus Bandavirus in the family Phenuiviridae (Walker et al., 2022). The clinical spectrum of SFTS ranges widely from subclinical infection or influenza-like symptoms to severe complications such as multiple organ failure (Casel et al., 2021). Reported case fatality rates among hospitalized patients vary from 12% to 50% (Yu et al., 2011; Kim et al., 2013; Takahashi et al., 2014; Li et al., 2018). Several demographic factors, such as advanced age, male sex and pre-existing comorbidities, have been consistently associated with poorer outcomes across studies (Li et al., 2018; Jung et al., 2019; Wang et al., 2021; Zhao et al., 2021). Laboratory markers indicative of hepatic/renal dysfunction, coagulopathy and metabolic disturbances have also been recognized as prognostic predictors (Li et al., 2018; Ge et al., 2022; Lee et al., 2024; Tang et al., 2024; Zhang et al., 2024). Additionally, hematologic abnormalities, including lymphopenia, monocytopenia, neutrophilia, and elevated inflammatory markers, are significantly associated with disease severity and mortality (Peng et al., 2016; Park et al., 2021; Song et al., 2024). Notably, even among fatal cases, laboratory trends are heterogeneous: although most exhibit progressive deterioration, some show transient normalization near the terminal phase, typically occurring around 12 days after symptom onset (Zhao et al., 2022). The clinical course of SFTS is often short and rapidly evolving, complicating early risk stratification based on static evaluation of the clinical parameters. Although prior studies have identified promising early warning indicators, they predominantly rely on cross-sectional data collected at admission (Xu et al., 2018; Chen et al., 2025), leaving the dynamic clinical trajectory that distinguishes survivors from fatal cases poorly characterized. Moreover, no validated predictive model currently exists for staging-based mortality risk stratification in SFTS. Machine learning (ML), a transformative branch of artificial intelligence, demonstrates exceptional capability in identifying complex patterns in high-dimensional biomedical datasets, often yielding predictive models that outperform traditional statistical approaches in accuracy and generalizability (Greener et al., 2022). Recent clinical applications of ML have demonstrated its revolutionary potential, from predicting sepsis mortality risk (Goh et al., 2021) to forecasting long-term sequelae of COVID-19 (Pfaff et al., 2022). In this study, by integrating longitudinal clinical and laboratory markers, we introduce an interpretable ML framework to predict mortality risk in patients with SFTS. By using extreme gradient boosting (XGBoost), we developed a staged modeling approach that identifies key prognostic variables across distinct disease phases. Our model enables precise, early risk stratification and has the potential to guide timely interventions, thereby improving clinical outcomes in this rapidly progressive infection. Results Demographic and laboratory characteristics of SFTS patients A total of 5359 laboratory-confirmed patients were enrolled in the study, comprising 4128 cases from the 154th Hospital (Xinyang, Henan Province) and 1231 cases from Qishan Hospital (Yantai, Shandong Province) (Fig. 1). Among the patients from the 154th Hospital, 650 experienced fatal outcomes (Table 1), with a median clinical duration (CD) of 13 (interquartile range [IQR] 10–15). In Qishan Hospital, 249 deaths were recorded, with a median CD of 16 (IQR 12–21) (Supplementary Table S1). Comparative analysis revealed significant differences between survivors and non-survivors in both the development and external test sets. Patients with fatal outcomes were significantly older than survivors and exhibited more pronounced laboratory abnormalities, including markedly lower platelet counts (PLT), lymphocyte percentages (LYM%), and monocyte percentages (MONO%), alongside elevated levels of liver injury markers such as total bile acid (TBA) and lactate dehydrogenase (LDH) (all P < 0.001). These prognostic disparities between survivors and fatal cases were consistently observed across both independent cohorts (Table 1).Fig. 1Flowchart of the study design.Fig. 1Table 1Statistical descriptions of survival and death features in the Xinyang development set and Yantai external test set.Table 1Development set (Xinyang)P valueExternal test set (Yantai)P valueFeaturesSurvival (N = 3478)Death (N = 650)Total (N = 4128)Survival (N = 982)Death (N = 249)Total (N = 1231)Individual characteristicsAge (years)63.00 (53.00–70.00)71.00 (64.00–75.00)64.00 (55.00–71.00)<0.00164.00 (55.00–71.00)70.00 (64.00–76.00)65.00 (57.00–73.00)<0.001Sex0.0010.011Female2082 (59.86%)341 (52.46%)2423 (58.70%)533 (54.28%)112 (44.98%)645 (52.40%)Male1396 (40.14%)309 (47.54%)1705 (41.30%)449 (45.72%)137 (55.02%)586 (47.60%)Delay from disease onset (days)5.00 (4.00–6.00)5.00 (4.00–7.00)5.00 (4.00–7.00)<0.0016.00 (4.00–7.00)5.00 (4.00–7.00)6.00 (4.00–7.00)0.012Routine blood parametersMONO (%)7.90 (5.85–10.40)3.50 (2.00–5.65)7.40 (5.00–10.00)<0.00110.40 (8.25–12.75)7.00 (4.50–10.50)9.90 (7.50–12.60)<0.001LYM (%)28.48 (22.10–35.55)20.05 (15.10–26.10)27.10 (20.50–34.45)<0.00130.00 (22.90–37.10)20.10 (14.80–26.60)28.00 (20.58–35.90)<0.001PLT (109/L)60.50 (44.50–80.00)37.25 (28.00–48.00)56.00 (40.75–76.00)<0.00173.25 (54.50–105.00)40.00 (30.00–54.00)65.50 (46.00–95.25)<0.001RBC (1012/L)3.99 (3.64–4.33)4.07 (3.71–4.47)4.00 (3.65–4.36)<0.0014.31 (3.97–4.65)4.53 (4.00–4.92)4.35 (3.98–4.70)<0.001WBC (109/L)4.00 (3.30–5.10)3.15 (2.40–4.60)3.90 (3.10–5.00)<0.0014.39 (3.52–5.43)3.90 (2.38–6.06)4.31 (3.31–5.48)<0.001Liver function parametersTBA (μmol/L)4.55 (3.00–7.50)14.45 (7.60–28.95)5.00 (3.20–9.30)<0.0013.60 (2.50–5.35)17.30 (9.10–32.20)4.20 (2.80–8.05)<0.001LDH (U/L)450.75 (325.50–634.50)1007.50 (652.00–1514.10)492.00 (344.50–729.50)<0.001415.50 (331.00–551.00)1580.00 (867.31–2405.34)470.32 (348.00–724.48)<0.001GLB (g/L)27.60 (24.90–30.60)27.50 (24.70–30.70)27.60 (24.88–30.70)0.80827.43 (24.80–30.20)26.35 (22.92–29.81)27.30 (24.42–30.06)0.001DBIL (μmol/L)3.80 (2.70–5.00)4.90 (3.60–7.60)3.90 (2.80–5.20)<0.0015.34 (3.93–7.46)8.35 (5.40–16.95)5.80 (4.11–8.45)<0.001IBIL (μmol/L)5.50 (4.10–7.80)4.90 (3.60–6.90)5.40 (4.00–7.70)<0.0017.76 (5.88–10.12)6.05 (4.04–8.98)7.50 (5.58–9.98)<0.001TBIL (μmol/L)9.45 (7.15–12.60)10.65 (7.60–13.90)9.60 (7.20–12.80)<0.00113.20 (10.38–17.70)14.60 (9.90–24.37)13.48 (10.30–18.30)<0.001ALB (g/L)34.55 (32.00–37.20)32.20 (30.10–34.70)34.20 (31.65–36.90)<0.00131.11 (29.10–33.45)27.91 (25.50–30.21)30.66 (28.30–33.00)<0.001A/G1.25 (1.10–1.45)1.20 (1.00–1.40)1.25 (1.10–1.42)<0.0011.12 (1.00–1.30)1.08 (0.90–1.25)1.11 (1.00–1.30)<0.001TP (g/L)62.40 (59.20–65.95)60.30 (56.75–63.95)62.10 (58.80–65.70)<0.00159.00 (55.44–62.40)54.33 (51.30–59.00)58.17 (54.50–61.90)<0.001ALP (U/L)72.00 (55.50–93.00)87.50 (62.00–133.00)73.50 (56.00–98.00)<0.00165.10 (52.96–81.10)98.00 (64.15–168.00)68.00 (54.10–90.72)<0.001ALT (U/L)62.50 (38.50–103.00)101.00 (59.00–179.00)67.00 (40.00–112.25)<0.00164.85 (42.50–101.00)122.65 (75.80–224.00)72.00 (44.98–118.52)<0.001AST (U/L)107.00 (64.50–178.50)326.00 (174.00–610.50)122.00 (70.50–223.25)<0.00170.00 (45.90–123.50)482.00 (226.70–892.98)89.09 (51.95–194.00)<0.001GGT (U/L)42.00 (23.50–92.00)53.25 (27.00–131.00)44.00 (24.00–99.00)<0.00146.00 (26.30–92.00)78.00 (37.70–177.00)50.00 (28.00–103.00)<0.001ElectrolyteK (mmol/L)3.51 (3.26–3.80)3.80 (3.46–4.21)3.55 (3.29–3.89)<0.0013.70 (3.48–3.94)4.20 (3.80–4.54)3.75 (3.50–4.05)0.001Laboratory data are presented as the median value of all measurements for each patient during the entire hospitalization.Age, delay from disease onset, and all laboratory testing variables were treated as continuous variables and described using medians with interquartile ranges (IQR) due to non-normality of values, while Gender was treated as a categorical variable and described as counts and proportions. A Shapiro-Wilk test was performed to decide between normal or non-normal distribution for the variables. MONO, Monocytes; LYM, Lymphocytes; PLT, Platelet count; RBC, Red blood cell count; WBC, White blood cell count; TBA, Total bile acid; LDH, Lactate dehydrogenase; GLB, Globulin; DBIL, Direct bilirubin; IBIL, Indirect bilirubin; TBIL, Total bilirubin; ALB, Albumin; A/G, Albumin/globulin ratio; TP, Total protein; ALP, Alkaline phosphatase; ALT, Alanine aminotransferase; AST, Aspartate aminotransferase; GGT, Gamma-glutamyl transferase; K, Potassium. Flowchart of the study design. Table 1Statistical descriptions of survival and death features in the Xinyang development set and Yantai external test set.Table 1Development set (Xinyang)P valueExternal test set (Yantai)P valueFeaturesSurvival (N = 3478)Death (N = 650)Total (N = 4128)Survival (N = 982)Death (N = 249)Total (N = 1231)Individual characteristicsAge (years)63.00 (53.00–70.00)71.00 (64.00–75.00)64.00 (55.00–71.00)<0.00164.00 (55.00–71.00)70.00 (64.00–76.00)65.00 (57.00–73.00)<0.001Sex0.0010.011Female2082 (59.86%)341 (52.46%)2423 (58.70%)533 (54.28%)112 (44.98%)645 (52.40%)Male1396 (40.14%)309 (47.54%)1705 (41.30%)449 (45.72%)137 (55.02%)586 (47.60%)Delay from disease onset (days)5.00 (4.00–6.00)5.00 (4.00–7.00)5.00 (4.00–7.00)<0.0016.00 (4.00–7.00)5.00 (4.00–7.00)6.00 (4.00–7.00)0.012Routine blood parametersMONO (%)7.90 (5.85–10.40)3.50 (2.00–5.65)7.40 (5.00–10.00)<0.00110.40 (8.25–12.75)7.00 (4.50–10.50)9.90 (7.50–12.60)<0.001LYM (%)28.48 (22.10–35.55)20.05 (15.10–26.10)27.10 (20.50–34.45)<0.00130.00 (22.90–37.10)20.10 (14.80–26.60)28.00 (20.58–35.90)<0.001PLT (109/L)60.50 (44.50–80.00)37.25 (28.00–48.00)56.00 (40.75–76.00)<0.00173.25 (54.50–105.00)40.00 (30.00–54.00)65.50 (46.00–95.25)<0.001RBC (1012/L)3.99 (3.64–4.33)4.07 (3.71–4.47)4.00 (3.65–4.36)<0.0014.31 (3.97–4.65)4.53 (4.00–4.92)4.35 (3.98–4.70)<0.001WBC (109/L)4.00 (3.30–5.10)3.15 (2.40–4.60)3.90 (3.10–5.00)<0.0014.39 (3.52–5.43)3.90 (2.38–6.06)4.31 (3.31–5.48)<0.001Liver function parametersTBA (μmol/L)4.55 (3.00–7.50)14.45 (7.60–28.95)5.00 (3.20–9.30)<0.0013.60 (2.50–5.35)17.30 (9.10–32.20)4.20 (2.80–8.05)<0.001LDH (U/L)450.75 (325.50–634.50)1007.50 (652.00–1514.10)492.00 (344.50–729.50)<0.001415.50 (331.00–551.00)1580.00 (867.31–2405.34)470.32 (348.00–724.48)<0.001GLB (g/L)27.60 (24.90–30.60)27.50 (24.70–30.70)27.60 (24.88–30.70)0.80827.43 (24.80–30.20)26.35 (22.92–29.81)27.30 (24.42–30.06)0.001DBIL (μmol/L)3.80 (2.70–5.00)4.90 (3.60–7.60)3.90 (2.80–5.20)<0.0015.34 (3.93–7.46)8.35 (5.40–16.95)5.80 (4.11–8.45)<0.001IBIL (μmol/L)5.50 (4.10–7.80)4.90 (3.60–6.90)5.40 (4.00–7.70)<0.0017.76 (5.88–10.12)6.05 (4.04–8.98)7.50 (5.58–9.98)<0.001TBIL (μmol/L)9.45 (7.15–12.60)10.65 (7.60–13.90)9.60 (7.20–12.80)<0.00113.20 (10.38–17.70)14.60 (9.90–24.37)13.48 (10.30–18.30)<0.001ALB (g/L)34.55 (32.00–37.20)32.20 (30.10–34.70)34.20 (31.65–36.90)<0.00131.11 (29.10–33.45)27.91 (25.50–30.21)30.66 (28.30–33.00)<0.001A/G1.25 (1.10–1.45)1.20 (1.00–1.40)1.25 (1.10–1.42)<0.0011.12 (1.00–1.30)1.08 (0.90–1.25)1.11 (1.00–1.30)<0.001TP (g/L)62.40 (59.20–65.95)60.30 (56.75–63.95)62.10 (58.80–65.70)<0.00159.00 (55.44–62.40)54.33 (51.30–59.00)58.17 (54.50–61.90)<0.001ALP (U/L)72.00 (55.50–93.00)87.50 (62.00–133.00)73.50 (56.00–98.00)<0.00165.10 (52.96–81.10)98.00 (64.15–168.00)68.00 (54.10–90.72)<0.001ALT (U/L)62.50 (38.50–103.00)101.00 (59.00–179.00)67.00 (40.00–112.25)<0.00164.85 (42.50–101.00)122.65 (75.80–224.00)72.00 (44.98–118.52)<0.001AST (U/L)107.00 (64.50–178.50)326.00 (174.00–610.50)122.00 (70.50–223.25)<0.00170.00 (45.90–123.50)482.00 (226.70–892.98)89.09 (51.95–194.00)<0.001GGT (U/L)42.00 (23.50–92.00)53.25 (27.00–131.00)44.00 (24.00–99.00)<0.00146.00 (26.30–92.00)78.00 (37.70–177.00)50.00 (28.00–103.00)<0.001ElectrolyteK (mmol/L)3.51 (3.26–3.80)3.80 (3.46–4.21)3.55 (3.29–3.89)<0.0013.70 (3.48–3.94)4.20 (3.80–4.54)3.75 (3.50–4.05)0.001Laboratory data are presented as the median value of all measurements for each patient during the entire hospitalization.Age, delay from disease onset, and all laboratory testing variables were treated as continuous variables and described using medians with interquartile ranges (IQR) due to non-normality of values, while Gender was treated as a categorical variable and described as counts and proportions. A Shapiro-Wilk test was performed to decide between normal or non-normal distribution for the variables. MONO, Monocytes; LYM, Lymphocytes; PLT, Platelet count; RBC, Red blood cell count; WBC, White blood cell count; TBA, Total bile acid; LDH, Lactate dehydrogenase; GLB, Globulin; DBIL, Direct bilirubin; IBIL, Indirect bilirubin; TBIL, Total bilirubin; ALB, Albumin; A/G, Albumin/globulin ratio; TP, Total protein; ALP, Alkaline phosphatase; ALT, Alanine aminotransferase; AST, Aspartate aminotransferase; GGT, Gamma-glutamyl transferase; K, Potassium. Statistical descriptions of survival and death features in the Xinyang development set and Yantai external test set. Laboratory data are presented as the median value of all measurements for each patient during the entire hospitalization. Age, delay from disease onset, and all laboratory testing variables were treated as continuous variables and described using medians with interquartile ranges (IQR) due to non-normality of values, while Gender was treated as a categorical variable and described as counts and proportions. A Shapiro-Wilk test was performed to decide between normal or non-normal distribution for the variables. MONO, Monocytes; LYM, Lymphocytes; PLT, Platelet count; RBC, Red blood cell count; WBC, White blood cell count; TBA, Total bile acid; LDH, Lactate dehydrogenase; GLB, Globulin; DBIL, Direct bilirubin; IBIL, Indirect bilirubin; TBIL, Total bilirubin; ALB, Albumin; A/G, Albumin/globulin ratio; TP, Total protein; ALP, Alkaline phosphatase; ALT, Alanine aminotransferase; AST, Aspartate aminotransferase; GGT, Gamma-glutamyl transferase; K, Potassium. Clinical staging and temporal dynamics of prognostic variables in SFTS cases The clinical course of SFTS was systematically stratified using an iterative analytical approach. To delineate stage-specific prognostic patterns, we developed distinct XGBoost models for each of 10 temporal intervals, and quantified variable importance using relative contributions (RCs). Hierarchical clustering of RC patterns consolidated these intervals into three clinically distinct stages: Stage 1 (early stage, ≤7 days), Stage 2 (intermediate stage, days 8–9), and Stage 3 (late stage, ≥10 days) (Fig. 2). This staging revealed dynamic shifts in prognostic determinants, with early-stage predictors dominated by demographic factors and later stages increasingly influenced by organ-specific injury markers (Fig. 2).Fig. 2Cluster analysis based on XGBoost models for initial clinical staging. The relative contributions of variables in the initial clinical stage models were determined by XGBoost models and subjected to hierarchical clustering using the complete linkage method. Stages that clustered together, indicative of similar variable-result response relationships, were classified as the same clinical stage. The clustering results presented here represent the final clinical staging. The color intensity from light yellow to dark red corresponds to increasing relative contribution of the variables.Fig. 2 Cluster analysis based on XGBoost models for initial clinical staging. The relative contributions of variables in the initial clinical stage models were determined by XGBoost models and subjected to hierarchical clustering using the complete linkage method. Stages that clustered together, indicative of similar variable-result response relationships, were classified as the same clinical stage. The clustering results presented here represent the final clinical staging. The color intensity from light yellow to dark red corresponds to increasing relative contribution of the variables. In Stage 1 (≤ 7 days post-onset), demographic factors predominated, with age as the strongest predictor (RC of 18.44%), followed by MONO% (RC of 16.04%), and LDH (RC of 11.01%). Stage 2 (days 8–9) represented a transitional phase, with LDH emerging as the most prominent predictor (RC of 38.58%), followed by MONO% (RC of 13.30%). Notably, the contribution of age declined substantially (RC of 12.19%), indicating a pathophysiological shift from host susceptibility to organ dysfunction. By Stage 3 (≥10 days post-onset), hepatic injury markers dominated, with LDH reaching its peak predictive value (RC of 60.10%), accompanied by elevated contributions from aspartate aminotransferase (AST; RC of 15.51%) and TBA (RC of 12.08%). Four variables maintained clinical relevance across all stages (all RC > 5%): AST, LDH, TBA, and MONO%. Among these, LDH exhibited a progressive increase in RC from 11.01% to 60.10% throughout disease progression (Table 2). Variables with RCs exceeding 10% at any stage, including age, LDH, TBA, and AST, were positively correlated with mortality risk, whereas MONO% and white blood cell count (WBC) were inversely correlated with fatal outcomes (Fig. 3).Table 2Mean (standard deviation) relative contributions of major features in the prediction model of fatal outcome using XGBoost.Table 2FeaturesMean relative contributions (%) (Standard deviation)Stage 1Stage 2Stage 3Individual characteristicsAge (years)18.44 (1.61)12.19 (1.29)Routine blood parametersWBC (109/L)10.69 (1.31)8.13 (1.35)LYM (%)4.79 (1.36)6.26 (1.43)3.72 (1.49)MONO (%)16.04 (1.66)13.30 (1.36)5.25 (1.50)PLT (109/L)4.74 (1.06)3.34 (1.55)Liver function parametersAST (U/L)10.94 (2.40)5.45 (2.06)15.51 (7.76)DBIL (μmol/L)4.97 (1.09)5.66 (1.14)TBA (μmol/L)10.96 (1.73)10.43 (2.06)12.08 (5.17)LDH (U/L)11.01 (2.00)38.58 (4.08)60.10 (10.34)ElectrolyteK (mmol/L)7.42 (1.29)WBC, White blood cell count; LYM, Lymphocytes; MONO, Monocytes; PLT, Platelet count; AST, Aspartate aminotransferase; DBIL, Direct bilirubin; TBA, Total bile acid; LDH, Lactate dehydrogenase; K, Potassium.Fig. 3The mean curves and 95% percentiles for the effects of major variables (Relative Contribution, RC > 5%) on the risk of mortality, derived from the XGBoost models for each clinical stage. Stage 1 (A), Stage 2 (B), and Stage 3 (C). Each stage signifies a distinct phase in the progression of the condition under study.Fig. 3 Table 2Mean (standard deviation) relative contributions of major features in the prediction model of fatal outcome using XGBoost.Table 2FeaturesMean relative contributions (%) (Standard deviation)Stage 1Stage 2Stage 3Individual characteristicsAge (years)18.44 (1.61)12.19 (1.29)Routine blood parametersWBC (109/L)10.69 (1.31)8.13 (1.35)LYM (%)4.79 (1.36)6.26 (1.43)3.72 (1.49)MONO (%)16.04 (1.66)13.30 (1.36)5.25 (1.50)PLT (109/L)4.74 (1.06)3.34 (1.55)Liver function parametersAST (U/L)10.94 (2.40)5.45 (2.06)15.51 (7.76)DBIL (μmol/L)4.97 (1.09)5.66 (1.14)TBA (μmol/L)10.96 (1.73)10.43 (2.06)12.08 (5.17)LDH (U/L)11.01 (2.00)38.58 (4.08)60.10 (10.34)ElectrolyteK (mmol/L)7.42 (1.29)WBC, White blood cell count; LYM, Lymphocytes; MONO, Monocytes; PLT, Platelet count; AST, Aspartate aminotransferase; DBIL, Direct bilirubin; TBA, Total bile acid; LDH, Lactate dehydrogenase; K, Potassium. Mean (standard deviation) relative contributions of major features in the prediction model of fatal outcome using XGBoost. WBC, White blood cell count; LYM, Lymphocytes; MONO, Monocytes; PLT, Platelet count; AST, Aspartate aminotransferase; DBIL, Direct bilirubin; TBA, Total bile acid; LDH, Lactate dehydrogenase; K, Potassium. The mean curves and 95% percentiles for the effects of major variables (Relative Contribution, RC > 5%) on the risk of mortality, derived from the XGBoost models for each clinical stage. Stage 1 (A), Stage 2 (B), and Stage 3 (C). Each stage signifies a distinct phase in the progression of the condition under study. Predictive model performance across clinical stages The stage-specific prediction models demonstrated robust discriminative ability throughout the SFTS disease course, with performance improving as the disease progressed (Fig. 4). In the internal validation, Stage 1 achieved an area under the receiver operating characteristic curve (AUC) of 0.84 (95% confidence interval [CI] 0.81–0.88) (Fig. 4A), accompanied by an accuracy of 0.82 (95% CI 0.79–0.85), precision of 0.47, recall of 0.56, and F1 score of 0.51 (Table 3). Model performance improved in subsequent stages: Stage 2 (days 8–9 post-onset) attained an AUC of 0.95 (95% CI 0.93–0.97), accuracy of 0.89 (95% CI 0.86–0.91), precision of 0.53, recall of 0.88, and F1 score of 0.66 (Table 3). Stage 3 reached peak performance, with an AUC of 0.98 (95% CI 0.97–0.99), accuracy of 0.95 (95% CI 0.92–0.96), precision of 0.59, recall of 0.92, and F1 score of 0.72, reflecting enhanced predictive reliability in later phases (Fig. 4A, Table 3). This temporal refinement pattern was corroborated in external validation (Fig. 4B). Although Stage 1 showed higher discrimination in external cohorts (AUC 0.91 vs internal 0.84), Stages 2 and 3 maintained peak performance (AUC 0.95 and 0.98 respectively), supporting model generalizability across heterogeneous populations (Table 3). The reliability of the model was further supported by concordance metrics, with Brier scores for both internal and external test sets ranging from 0.04 to 0.11, indicating strong agreement between predicted and observed risk (Fig. 4C and D, Table 3).Fig. 4Performance and calibration of XGBoost model for clinical staging. A Receiver operating characteristic (ROC) curve for the XGBoost models based on clinical stage in the internal test set. The AUC values and 95% CI for each stage are displayed. B ROC curve for the external test set, with AUC values and 95% CI for each stage. C Calibration curves for the internal test set. The Brier scores and 95% CI for each stage are shown, providing a measure of the accuracy of the probabilistic predictions. D Calibration curves for the external test set, with Brier scores and 95% CI for each stage. ROC, Receiver operator characteristic curve; AUC, Area under the receiver operating characteristic curve; CI, Confidence interval.Fig. 4Table 3Performance metrics of models at different stages on internal and external Test Sets.Table 3Inner test setExternal test setStage 1Stage 2Stage 3Stage 1Stage 2Stage 3AUC (95% CI)0.84 (0.81–0.88)0.95 (0.93–0.97)0.98 (0.97–0.99)0.91 (0.89–0.93)0.95 (0.93–0.97)0.98 (0.96–1.00)Accuracy (95% CI)0.82 (0.79–0.85)0.89 (0.86–0.91)0.95 (0.92–0.96)0.79 (0.76–0.81)0.88 (0.86–0.90)0.97 (0.95–0.98)Precision0.470.530.590.500.590.81Recall0.560.880.920.940.850.84F1 score0.510.660.720.650.700.82Brier score (95% CI)0.11 (0.09–0.13)0.06 (0.05–0.07)0.04 (0.03–0.04)0.11 (0.09–0.12)0.06 (0.05–0.07)0.04 (0.03–0.05)AUC, Area under the receiver operating characteristic curve; CI, Confidence interval. Performance and calibration of XGBoost model for clinical staging. A Receiver operating characteristic (ROC) curve for the XGBoost models based on clinical stage in the internal test set. The AUC values and 95% CI for each stage are displayed. B ROC curve for the external test set, with AUC values and 95% CI for each stage. C Calibration curves for the internal test set. The Brier scores and 95% CI for each stage are shown, providing a measure of the accuracy of the probabilistic predictions. D Calibration curves for the external test set, with Brier scores and 95% CI for each stage. ROC, Receiver operator characteristic curve; AUC, Area under the receiver operating characteristic curve; CI, Confidence interval. Table 3Performance metrics of models at different stages on internal and external Test Sets.Table 3Inner test setExternal test setStage 1Stage 2Stage 3Stage 1Stage 2Stage 3AUC (95% CI)0.84 (0.81–0.88)0.95 (0.93–0.97)0.98 (0.97–0.99)0.91 (0.89–0.93)0.95 (0.93–0.97)0.98 (0.96–1.00)Accuracy (95% CI)0.82 (0.79–0.85)0.89 (0.86–0.91)0.95 (0.92–0.96)0.79 (0.76–0.81)0.88 (0.86–0.90)0.97 (0.95–0.98)Precision0.470.530.590.500.590.81Recall0.560.880.920.940.850.84F1 score0.510.660.720.650.700.82Brier score (95% CI)0.11 (0.09–0.13)0.06 (0.05–0.07)0.04 (0.03–0.04)0.11 (0.09–0.12)0.06 (0.05–0.07)0.04 (0.03–0.05)AUC, Area under the receiver operating characteristic curve; CI, Confidence interval. Performance metrics of models at different stages on internal and external Test Sets. AUC, Area under the receiver operating characteristic curve; CI, Confidence interval. Calibration curves revealed optimal alignment for internal validation. In external validation, a slight conservative bias was observed, with predicted risks marginally underestimating observed outcomes, likely attributable to inter-hospital differences in clinical management (Fig. 4C andD). Decision curve analysis (DCA) demonstrated consistent net benefit across threshold probabilities for both internal and external validations. The XGBoost models outperformed both “treat all” and “treat none” strategies in DCA across all three clinical stages (Fig. 5). This finding underscores the model's potential utility in guiding time-sensitive clinical decisions, such as prioritizing antiviral therapies in early phases and organ support in advanced stages—while avoiding unnecessary clinical escalation.Fig. 5Decision curve analysis (DCA) curves for internal and external test sets across three clinical stages. A–C DCA curves for three clinical stages in the internal test set. D–F DCA curves of three external test sets for clinical stage. We employ DCA to assess the potential clinical utility of the model, that is, the model's net benefit. DCA takes into account the net benefit of the model across a wide range of clinical probability thresholds, considering the trade-offs between true positives and false positives. DCA, Decision curve analysis.Fig. 5 Decision curve analysis (DCA) curves for internal and external test sets across three clinical stages. A–C DCA curves for three clinical stages in the internal test set. D–F DCA curves of three external test sets for clinical stage. We employ DCA to assess the potential clinical utility of the model, that is, the model's net benefit. DCA takes into account the net benefit of the model across a wide range of clinical probability thresholds, considering the trade-offs between true positives and false positives. DCA, Decision curve analysis. Discussion Previous investigations have consistently identified divergent laboratory trajectories between fatal and non-fatal SFTS cases (Xu et al., 2018; Hou et al., 2024). While survivors typically exhibit transient abnormalities that resolve spontaneously, fatal cases often demonstrate progressive deterioration throughout the entire course of disease, culminating in organ failure (Gai et al., 2012; Li et al., 2018). Building on this evidence, our longitudinal analysis of 5359 hospitalized patients delineates three pathophysiologically distinct stages, providing new insights into the dynamic evolution of prognostic factors. Stage 1 (days 0–7) is characterized by demographic vulnerability and monocytic suppression. During this early viremic phase, limited biomarker divergence is observed between different outcomes, with age serving as the dominant predictor. This aligns with prior reports that a 10-year age increase elevates mortality risk by 82% (Li et al., 2018), underscoring age as a critical risk factor. MONO% also emerged as a key early predictor, reflecting both viral targeting and immune regulatory function (Peng et al., 2016; Li et al., 2021). Fatal cases exhibited a 39.1% lower median MONO% than survivors (2.8% [1.6–4.9] vs. 4.6% [2.5–7.6]; P < 0.001), consistent with histopathological evidence of SFTSV-induced monocyte apoptosis (Song et al., 2018; Li et al., 2020). This early depletion of monocyte, driven by viral cytopathy, likely compromises innate immune defenses, predisposing patients to subsequent systemic dysregulation. Stage 2 (days 8–9 post-onset) is marked by an immunopathological transition, highlighted by a 3.5-fold increase in the predictive contribution of LDH and a decline in the influence of MONO%. This shift coincided with previously reported decreases in viral loads and increases in inflammatory cytokine levels during the second stage (Li et al., 2023), suggesting a transition from direct viral cytopathy to immunopathology. The rising prominence of LDH, a marker of tissue injury, closely correlated with the onset of multiple organ dysfunction syndrome (MODS), while age-related contributions decline, reflecting evolving pathophysiological priorities. Stage 3 (≥10 days after disease onset) is characterized by pathophysiological divergence: survivors demonstrated progressive normalization of biomarkers, while fatal cases exhibited sustained hepatocytic injury (evidenced by elevated LDH and TBA); coagulopathy (characterized by declining PLT), and immune exhaustion (featured by reduced MONO%). Notably, LDH retained peak predictive importance, with AST and TBA further underscoring hepatic dysfunction. Additionally, MONO% or WBC maintain inverse correlations with mortality throughout all stages, serving as indicators of residual immune competence. External validation confirmed the robustness of our staging models, with calibration fidelity and decision curve analysis demonstrating superior net benefit over conventional strategies. Although the external calibration curves exhibit a slight conservative bias, likely attributable to inter-hospital difference in clinical practices, the models emphasize optimized recall (84%–94%) to enhance sensitivity and minimize the risk of missing fatal cases. This sensitivity is crucial given the rapid progression of SFTS (median time from symptom onset to outcome: 13 days). Our study holds significant implications for clinical practice. Temporal risk quantification further revealed a decline in the influence of age (RC 18.44% → 12.19%), alongside a marked increase in tissue injury markers (LDH RC 11.01% → 60.10%), clarifying that previously reported “recovery-like” terminal patterns may represent failed compensatory efforts rather than true physiological recovery (Gai et al., 2012; Xu et al., 2018). The dynamic risk stratification system introduced in this study enables time-sensitive triage for patients with SFTS. During the early viremic phase (days 0–7), when age and monocyte dynamics are the predominant prognostic factors, clinicians can prioritize antiviral therapies. As hepatic injury markers escalate from day 8 onward, the model supports a timely transition toward organ-directed supportive care. Furthermore, our models delineate the pathophysiological shift from initial viral cytopathy to subsequent immunopathology and potential multiorgan failure, offering a mechanistic rationale for stage-specific interventions. Notably, the interval between days 8 and 9, characterized by a surge in inflammatory activity, represents a pivotal timepoint where mortality risk becomes closely linked to LDH trajectories. These trajectories not only align with earlier work utilizing the L-index (Wang et al., 2023), calculated as Log10 [LDH × Blood urea nitrogen/LYM], but also uniquely delineate the transition from viremia-induced cytopathic effects to late-phase dominance of tissue injury markers, a dynamic shift undetectable by conventional static models. When contextualized with current prognostic paradigms, our findings complement rather than replace established models. While existing multivariable models demonstrate strong predictive performance, they are limited in detecting phase-specific deterioration and have largely been derived from single-center cohorts, which constrains their generalizability (Jia et al., 2017; Wang et al., 2017). Similarly, although certain ratio-based indicators are simple to use, most have not been externally validated and lack the temporal resolution to track dynamic pathophysiological transitions across disease stages (Wang et al., 2022; Guo et al., 2024). In contrast, our staging system, validated across multiple centers, integrates temporal risk stratification with evolving biomarker profiles, improving generalizability while preserving clinical interpretability. The study has several limitations. First, the variability in inter-hospital calibration likely reflects differences in local treatment protocols rather than a model deficiency, underscoring the need for validation in settings with standardized care. Second, the requirement for phase-complete data led to the exclusion of certain biomarkers—such as viral load, specific cytokines, and coagulation parameters, due to their irregular measurement in routine practice. Although this maintains the model's practicality, it limits mechanistic depth. These limitations point toward future research directions. Evaluating these markers' predictive value using non-phased approaches, followed by integrating into the current temporal framework, is expected to build a more comprehensive predictive system. Conclusions In conclusion, this study introduces a novel clinical staging paradigm that prioritizes dynamic risk adaptation over static assessment methodologies. By continuously integrating longitudinal biomarker information with data-driven modeling, the proposed framework not only recalibrate disease mechanisms but also transform complex datasets into actionable insights, representing a pivotal advancement in efforts to reduce the high mortality burden associated with SFTS. Materials and methods Study design and participants A retrospective observational study was performed at two designated hospitals for SFTS patients in China: the 154th Hospital (Xinyang, Henan Province) and Qishan Hospital (Yantai, Shandong Province) (Fig. 1). Confirmed SFTSV infection required detection of SFTSV RNA via real-time reverse transcriptase-polymerase chain reaction (RT-PCR), according to the standard guidelines issued by the National Health Commission of China (Ministry of Health PRC, 2010). All patients who met the confirmed diagnostic criterion between 2010 and 2023 were included in this study. Patients without definitive clinical outcomes were excluded from the analysis. Data collection and predictive variables selection For each patient, clinical and laboratory data were systematically collected from admission through hospitalization until death or discharge. A structured electronic case report form was used to obtain information on individual characteristics (age, sex, and delay from disease onset —the time interval between the patient's self-reported date of symptom onset and the date of hospital admission), longitudinal laboratory parameters (indicators spanning hematologic, hepatic, renal, and inflammatory profiles), outcomes (mortality, survival, or loss to follow-up), and clinical duration (days from symptom onset to either death for fatal cases or discharge for survivors). To minimize recall bias, symptom onset dates were cross-validated against family interviews and primary care records. The variables selected for the model spanned two categories. Individual characteristics included age, sex, and delay from disease onset. Laboratory indicators included five routine blood parameters: MONO%, LYM%, PLT, red blood cell count (RBC), and WBC; 13 liver function parameters: TBA, LDH, globulin (GLB), direct bilirubin (DBIL), indirect bilirubin (IBIL), total bilirubin (TBIL), albumin (ALB), albumin/globulin ratio (A/G), total protein (TP), alkaline phosphatase (ALP), alanine aminotransferase (ALT), AST, gamma-glutamyl transferase (GGT), and one electrolyte: potassium (K) (Supplementary Table S2). A total of 46 candidate variables were initially screened, including 3 individual-level characteristics and 43 laboratory parameters. To ensure the objectivity of the model, clinical symptoms, signs, and comorbidities were excluded as candidate predictors due to their inherent subjectivity, inconsistent severity grading, and varying documentation standards across hospitals. The selection and processing of predictive variables was performed through a systematic three-stage protocol. First, variables with a missing rate ≥ 25% (n = 22) were excluded to maintain analytical robustness (Supplementary Table S2). Second, multicollinearity was assessed using Spearman correlation analysis, and highly correlated variables (|ρ| > 0.85) were reduced by retaining only the most clinically interpretable variable from each correlated pair (Supplementary Fig. S1). Finally, missing values in the remaining 22 variables were imputed using multiple imputation via the R package “mice” (Multivariate Imputation by Chained Equations, v3.16.0), employing outcome-dependent imputation to preserve associations with mortality risk. Statistical analysis and modeling Continuous variables were assessed for normality using the Shapiro-Wilk test, with non-normally distributed data summarized as median IQR and categorical variables reported as frequencies (proportions). Given the inherent challenges of nonlinear relationships, sparse data structures, and class imbalance in clinical datasets, we employed the XGBoost algorithm—a decision tree-based framework validated for high-dimensional biomedical predictions (Chen and Guestrin, 2016; Cai et al., 2024). To characterize the temporal heterogeneity of prognostic biomarkers across the clinical course of SFTS, we developed a staged modeling framework that integrates dynamic biomarker trajectories with machine learning. The framework began with establishing a refined clinical staging system, in which disease progression was segmented into 10 temporally defined phases: an early consolidation phase (days 0–4), eight consecutive daily intervals spanning days 5–12, and a terminal phase (day 13 onwards). For each phase, we constructed stage-specific XGBoost models to quantify the prognostic relevance of variables. Variable importance was assessed using gain values—a metric reflecting the average improvement in model accuracy (i.e., reduction in loss function) when splitting on nodes in the decision trees on a given variable. To standardize comparisons across stages, gain values were normalized to percentage-based RCs, enabling the identification of variables with dominant predictive roles at specific disease phases. Subsequently, hierarchical clustering was performed using Ward's method, with a Euclidean distance threshold of <0.15 to aggregate initial clinical stages that shared concordant feature-outcome relationships. This process yielded three consolidated clinical stages, each characterized by distinct biomarker dynamics. For each refined clinical stage, the final set of modeling variables was determined from the imputed dataset via a two-step approach to ensure predictor stability and clinical relevance. First, variables exhibiting consistently low importance (mean RC < 3%) across multiple bootstrap training iterations within a given stage were excluded as unstable. Subsequently, the relative importance of the remaining variables was robustly estimated by averaging their contributions across 100 stage-specific XGBoost models. For model input, the representative value of each longitudinal laboratory parameter was taken as the median of all available measurements per patient within the corresponding stage. Finally, the XGBoost models were retrained on these refined stages, ensuring alignment between temporal pathophysiological patterns and model architecture. The detailed modelling process is shown in Supplementary Note 1. Model performance was evaluated using both discrimination and calibration metrics. Discrimination was assessed via the AUC with 95% CI, precision-recall curves, and F1 scores (Lekadir et al., 2025). Optimal classification thresholds were determined using Youden's index via the pROC package (v1.18.4), with class weights adjusted (death:survival = 0.15) to mitigate imbalance. Calibration accuracy was quantified using Brier scores (bootstrapped 95% CIs; 1000 resamples) and visualized through reliability diagrams (Steyerberg et al., 2010). All statistical analyses were performed using R software (version 4.1.2). Data availability The datasets generated and/or analyzed during the current study are available from the corresponding author on reasonable request. E-mail: liuwei@bmi.ac.cn; lwbime@163.com. Ethics statement The study was approved by the Ethics Committee of 154th hospital (154YYLL-2015-10) and Qishan Hospital (202104). Written informed consent was obtained from all participants. Author contributions Xue-Geng Hong: methodology, software, formal analysis, data curation, writing-original draft, visualization. Hong-Han Ge: methodology, software, formal analysis, data curation, writing-original draft, visualization. Ning Cui: methodology, data curation, writing-original draft. Yan-Li Xu: data curation. Xin Yang: data curation. Jia-Hao Chen: data curation. Xiao-Hong Yin: data curation. Yi-Mei Yuan: data curation. Chao Zhou: data curation. Hao Li: methodology, data curation. Xiao-Ai Zhang: methodology, data curation. Ming Yue: conceptualization, methodology, writing-review & editing, supervision, project administration. Ling Lin: conceptualization, methodology, data curation, writing-review & editing, supervision, project administration, funding acquisition. Wei Liu: conceptualization, methodology, writing-review & editing, supervision, project administration, funding acquisition. Conflict of interest The authors declare that no conflict of interest exists.