Preoperative risk model and nomogram for 30‑Day mortality after TAVR: development and internal validation in a Chinese cohort

Abstract

Background

Accurate prediction of 30‑day mortality after transcatheter aortic valve replacement (TAVR) remains challenging. Existing models often overlook key preoperative laboratory variables. We aimed to develop and internally validate a preoperative risk prediction model using routinely available variables measured before the procedure.We state that predictors were preoperative‑only.

Methods

We retrospectively analyzed 1,673 consecutive patients who underwent TAVR at Fuwai Hospital between 2013 and 2023. Patients were randomly split into a training cohort (n = 1,171) and an internal hold out validation cohort (n = 502). Candidate predictors were prespecified as routinely available preoperative variables. All candidate predictors, including laboratory parameters (troponin, D‑dimer, HbA1c, uric acid, etc.), were collected during the preoperative evaluation, within 3 days prior to the TAVR procedure. Independent predictors of 30-day mortality were identified using multivariable logistic regression. Performance was assessed with receiver operating characteristic (ROC) analysis, calibration assessment (calibration plots and Hosmer Lemeshow test), and bootstrap optimism correction; a nomogram was constructed from the final model.

Results

30-day mortality was 3.4% in both the training (40/1,171) and validation (17/502) cohorts. The final preoperative model retained STS score, HbA1c, D-dimer, and uric acid as independent predictors. Discrimination was good (AUC=0.84 in the training cohort; AUC=0.78 in the internal validation cohort), with acceptable calibration (Hosmer-Lemeshow p = 0.211); optimism corrected performance closely matched the hold out out results. The resulting nomogram provides individualized risk estimates for 30-day mortality based on these preoperative predictors.

Conclusions

We developed and internally validated a parsimonious preoperative risk model and nomogram for 30-day mortality after TAVR using routinely available clinical and laboratory variables. This tool may support individualized risk stratification in elective/non salvage TAVR candidates. External validation particularly beyond East-Asian populations is essential before broader application.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12893-025-03297-7.

Key points

What is already known

Risk models for TAVR commonly rely on the STS score and clinical comorbidity profiles; nomogram-based tools are widely used. Evidence for incremental value from routinely available preoperative biomarkers has been growing but remains variably incorporated across models.

What this study adds

We developed and internally validated a parsimonious, preoperative-only model for 30-day mortality after elective/non-salvage TAVR in a Chinese (East-Asian) cohort, using STS score, HbA1c, D-dimer, and uric acid. The model showed good discrimination and acceptable calibration in a hold-out set, with concordant bootstrap optimism-corrected performance. A nomogram and full coefficients are provided to facilitate external validation and potential recalibration in other settings.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12893-025-03297-7.

Introduction

Transcatheter aortic valve replacement (TAVR) has become a cornerstone therapy for patients with severe aortic stenosis who are at increased surgical risk [1, 2]. Accurate preoperative risk prediction is essential for optimal patient selection and clinical decision-making [3]. However, conventional risk models, such as the Society of Thoracic Surgeons (STS) score, primarily rely on demographic and comorbidity data, often overlooking valuable clinical and laboratory parameters obtained during preoperative evaluation [4]. Recent guidelines emphasize the need for comprehensive risk assessment, yet acknowledge the limitations of current models in fully capturing patient complexity and heterogeneity [5]. In addition, epidemiological data indicate that the global burden of aortic stenosis and TAVR procedures continues to rise, underscoring the necessity for improved risk stratification [6].

As TAVR indications expand to include lower-risk and younger patient populations, the accuracy and clinical utility of preoperative risk prediction tools become even more critical [7, 8]. Current research highlights the potential of integrating preoperative laboratory biomarkers with established clinical variables to enhance the prognostic value of risk models [9, 10]. The dynamic landscape of TAVR practice necessitates continuous updating and validation of predictive models to ensure their relevance across diverse patient cohorts and evolving clinical scenarios.

In this study, we aimed to develop and internally validate a multivariable risk prediction model for early mortality following TAVR by incorporating both traditional risk factors and readily available preoperative clinical and laboratory data, including D-dimer and other routinely available biomarkers.

Methods

Study design and population

This investigation was a single-center retrospective cohort study conducted at Fuwai Hospital, Chinese Academy of Medical Sciences. The study encompassed a cohort of 1,835 patients who underwent their initial transcatheter aortic valve replacement (TAVR) between January 1, 2013, and December 31, 2023. All participants were Chinese (East‑Asian) patients treated at a single center in China. The model was prespecified for elective/non‑salvage TAVR candidates in this care setting. After applying inclusion and exclusion criteria, 1,673 patients were included in the final analysis.

Inclusion criteria required:

1. Availability of preoperative clinical and laboratory data.

2. Patients who underwent only TAVR during hospitalization.

3. The main diagnosis was aortic valve disease.

Exclusion criteria included:

1. Patients with incomplete or missing preoperative data.

2. Patients with concomitant malignancies; gouty arthritis or autoimmune diseases that could confound inflammatory marker and nutrition levels.

3. Emergency or unplanned surgery.

4. Loss to follow-up (defined as an inability to confirm vital status with at least two contact methods).

Rationale

We prespecified the exclusion of patients with gouty arthritis to minimize biomarker-related confounding, because acute gout flares are associated with systemic inflammation and transient hypercoagulability that can elevate D-dimer and fibrinogen and alter uric acid levels, potentially distorting the relationship between these preoperative biomarkers and early mortality risk after TAVR.

Endpoints

The primary endpoint was all-cause mortality within 30 days after TAVR. Secondary endpoints included major postoperative complications: ischemic stroke, perivalvular leakage (paravalvular leak, PVL), pacemaker implantation, acute kidney injury (AKI), and conversion to open surgery.

Follow-Up

Patient survival status was verified through dual modalities: telephone follow-ups conducted 1 month after surgery and every 3 months thereafter, and cross-checking with the National Mortality Database. Follow-up data collection concluded on December 31, 2024. Data extraction was independently performed by two researchers, with discrepancies resolved by a third reviewer.

Data source

All clinical and laboratory data were retrieved from the hospital’s electronic medical record system. All candidate predictors, including laboratory parameters (troponin, D‑dimer, HbA1c, uric acid, etc.), were collected during the preoperative evaluation, within 3 days prior to the TAVR procedure.We treated the STS score as a single composite predictor and did not include its individual components to avoid redundancy/collinearity. Component-level fields were not consistently available in structured form in our EHR. If multiple preoperative measurements were available for a given biomarker, the measurement closest in time to the procedure (and prior to arterial access) was selected. Data extraction was independently performed by two researchers, and discrepancies were resolved by a third reviewer.

Missing data handling

The outcome had no missing values. For predictors, continuous variables were single-imputed using the median and categorical variables using the mode, with missingness indicators added for any predictor with > 5% missingness. Imputation parameters were estimated in the training cohort and applied to the validation cohort to avoid information leakage. As a sensitivity analysis, we repeated model development using multiple imputation with chained equations (m = 20) and pooled estimates; discrimination and calibration were materially unchanged relative to the complete‑case/median‑mode approach, indicating robustness of the findings to missing-data handling.

Data quality and standardization

• -Independent duplicate extraction: [Pengfei Chen, MD] and [Haochao Li, MD].
• -Adjudication of discrepancies: [Xu Wang, MD].
• -Standardization: predefined data dictionary; SI units; preoperative time windows.
• -Preoperative rules: last value before arterial access (same/previous day); HbA1c within prior 3 months.
• -Resolution rules: source hierarchy (LIS > structured EHR > clinical notes), unit checks, timestamp verification; unresolved cases adjudicated by [Xinjin Luo, MD].
• -Patient accrual ended on 2023-12-31; follow-up and data lock were 2024-12-31.

Clinical procedure

All patients underwent TAVR in accordance with international guidelines [11]. The procedure was performed via two access routes: transfemoral, involving percutaneous femoral artery puncture and delivery system insertion; or transapical, involving left thoracic incision and direct insertion through the cardiac apex, which was chosen for patients with unsuitable vascular anatomy (e.g., severe calcification, stenosis, or vascular anomalies). All procedures were performed by an experienced cardiovascular team. Postoperative anticoagulant regimens were individualized according to each patient’s bleeding risk.

Cohort grouping

Patients were randomly assigned to a training cohort (70%, n = 1,171) and a validation cohort (30%, n = 502) to evaluate the internal validity of the risk prediction model (Fig. 1).

Fig. 1.

Flow chart of sample screening. Note: Emergency/salvage TAVR was prespecified for exclusion to preserve the construct validity of a preoperative model for elective candidates and to avoid biomarker perturbation related to acute critical illness

Statistical analysis

Continuous variables are expressed as median and interquartile range (IQR) and were compared using the Wilcoxon rank-sum or Kruskal-Wallis test where appropriate. Categorical variables are presented as counts and percentages, and compared using the Fisher’s exact test (for expected frequencies < 5) or the Chi-squared test.

Predictor selection (preoperative-only): We prespecified candidate predictors as routinely available preoperative variables based on clinical plausibility and prior literature. Postoperative variables (e.g., ventilation duration) were not considered. The composite STS score was included as a single predictor; its components were not modeled separately to avoid redundancy/collinearity and because component-level fields were inconsistently available.

Screening and model building: We first ran univariate logistic regressions in the training cohort and carried forward variables with p < 0.20 or strong a priori clinical relevance. We then fit a multivariable logistic regression using backward stepwise selection minimizing the Akaike Information Criterion (AIC) via the stepAIC function in R. To limit overfitting given ≈ 40 events, we targeted an events-per-parameter ratio ≥ 10 and capped the final model at ≤ 4–5 parameters. The final model retained STS score, HbA1c, D-dimer, and uric acid. The nomogram was constructed directly from this final multivariable model; thus, the variables in the nomogram are identical to those in the model. Variable selection was performed strictly within the prespecified preoperative candidate set to preserve clinical plausibility and prevent data leakage from intra/post-procedural factors.

• Model performance: Discrimination was evaluated by receiver operating characteristic (ROC) curve analysis and the area under the curve (AUC) using the pROC package. Calibration was assessed using calibration plots from the rms package and the Hosmer-Lemeshow goodness-of-fit test (10 groups; 8 degrees of freedom; p > 0.05 indicates no evidence of poor fit). Overall accuracy was summarized with the Brier score.

• Sensitivity analysis-bootstrap internal validation: To quantify and correct for potential optimism, we performed bootstrap validation (1,000 resamples) on the development cohort.Model calibration was assessed using 1,000 bootstrap resamples (Supplementary Fig.S1).Repeating the entire modeling process within each resample (i.e., prespecified preoperative candidate set, univariate screening, and AIC‑based backward selection under events‑per‑parameter constraints). For each resample, performance was computed in the bootstrap sample (apparent) and in the original sample (test) to estimate optimism; optimism‑corrected performance was obtained by subtracting the average optimism from the apparent performance in the original data. We report AUC, Brier score, calibration slope, and calibration in the large, and present a bootstrap‑corrected calibration curve.Reporting followed TRIPOD recommendations for model development and internal validation. We additionally report Harrell’s Dxy and Cox–Snell R² as secondary performance summaries.All statistical analyses were conducted using R software (version 4.2.2).

Results

Patient flow and cohort characteristics

After applying prespecified criteria, 1,673 patients were included and randomly assigned to the training cohort (n = 1,171) and the internal hold‑out validation cohort (n = 502) (Fig. 1). The two cohorts were well balanced, with no statistically significant differences in baseline demographic, clinical, procedural, and preoperative laboratory variables (all p > 0.05) (Table 1).

Table 1.

Baseline characteristics of the training and validation cohorts

Characteristic*	Groups		p-value2
	Train N = 1,1711	Validation N = 5021
Ascending Aortic Diameter (mm)	39.0 (35.0, 43.0)	39.0 (35.0, 42.0)	0.601
Amylase (U/L)	10.3 (5.5, 15.9)	10.6 (5.7, 16.5)	0.344
Urea Nitrogen (mmol/L)	15.8 (13.5, 18.8)	15.8 (13.7, 18.8)	0.880
Creatinine (mg/dL)	0.95 (0.77, 1.19)	0.92 (0.75, 1.16)	0.192
Mortality (30 Day)			0.976
No	1,131 (96.6%)	485 (96.6%)
Yes	40 (3.4%)	17 (3.4%)
STS Score	7.02 (5.50, 8.52)	7.19 (5.62, 8.83)	0.213
Gender			0.669
Female	493 (42.1%)	217 (43.2%)
Male	678 (57.9%)	285 (56.8%)
Age	73 (69, 78)	74 (69, 78)	0.737
BMI	24.3 (21.9, 26.7)	24.2 (21.5, 26.4)	0.289
Cause			0.148
AI	256 (21.9%)	94 (18.7%)
AS	915 (78.1%)	408 (81.3%)
Smoking			0.642
No	803 (68.6%)	350 (69.7%)
Yes	368 (31.4%)	152 (30.3%)
Cerebrovascular Disease			0.826
No	927 (79.2%)	395 (78.7%)
Yes	244 (20.8%)	107 (21.3%)
Approach			0.805
Transapical	135 (11.5%)	60 (12.0%)
Transfemoral	1,036 (88.5%)	442 (88.0%)
Hypertension			0.337
No	472 (40.3%)	215 (42.8%)
Yes	699 (59.7%)	287 (57.2%)
Diabetes			0.992
No	896 (76.5%)	384 (76.5%)
Yes	275 (23.5%)	118 (23.5%)
CAD			0.604
No	625 (53.4%)	261 (52.0%)
Yes	546 (46.6%)	241 (48.0%)
Atrial Fibrillation			0.257
No	992 (84.7%)	436 (86.9%)
Yes	179 (15.3%)	66 (13.1%)
Hyperlipemia			0.470
No	694 (59.3%)	307 (61.2%)
Yes	477 (40.7%)	195 (38.8%)
COPD			0.264
No	791 (67.5%)	325 (64.7%)
Yes	380 (32.5%)	177 (35.3%)
CRF			0.958
No	1,119 (95.6%)	480 (95.6%)
Yes	52 (4.4%)	22 (4.4%)
Valve History			0.567
No	1,131 (96.6%)	482 (96.0%)
Yes	40 (3.4%)	20 (4.0%)
WBC (10^9/L)	6.06 (5.04, 7.24)	5.91 (5.04, 7.16)	0.390
Creatine Kinase (IU/L)	2 (2, 10)	2 (2, 7)	0.341
hCT(%)	34 (0, 41)	34 (0, 41)	0.463
TnI (ng/ml)	0.02 (0.01, 0.09)	0.02 (0.01, 0.12)	0.141
Serum Albumin (g/L)	39.8 (36.7, 42.9)	39.6 (36.9, 42.6)	0.660
Uric Acid (µmol/L)	481 (375, 583)	475 (356, 577)	0.231
Triglyceride (mmol/L)	1.96 (1.35, 2.48)	1.97 (1.43, 2.48)	0.612
CRP (mg/L)	2.8 (1.8, 4.8)	2.7 (1.7, 4.8)	0.517
Platelet (10^9/L)	180 (150, 218)	181 (144, 217)	0.510
Sinus Diameter (mm)	33.0 (30.0, 37.0)	34.0 (31.0, 38.0)	0.199
Reflux Area (cm²)	3.74 (1.85, 5.60)	3.60 (1.88, 5.41)	0.368
Total cholesterol(mmol/L)	4.26 (2.65, 5.83)	4.29 (2.95, 5.73)	0.605
HbA1c (%)	5.34 (4.23, 6.32)	5.43 (4.13, 6.43)	0.734
Hb (g/L)	129 (118, 141)	130 (118, 141)	0.424
NT-BNP (pg/ml)	1,423 (463, 3,193)	1,213 (468, 3,079)	0.497
D-dimer (mg/L)	0.48 (0.29, 0.98)	0.46 (0.29, 0.90)	0.970
EF (%)	60 (50, 66)	60 (48, 65)	0.709
LVEDD (mm)	55 (48, 64)	56 (48, 64)	0.239
Stroke			0.567
No	1,131 (96.6%)	482 (96.0%)
Yes	40 (3.4%)	20 (4.0%)
Perivalvular Leakage			0.586
No	1,150 (98.2%)	491 (97.8%)
Yes	21 (1.8%)	11 (2.2%)
Pacemaker Implantation			0.421
No	1,148 (98.0%)	495 (98.6%)
Yes	21 (1.8%)	9(1.8%)
Renal Injury			0.414
No	1,144 (97.7%)	487 (97.0%)
Yes	27 (2.3%)	15 (3.0%)
Open surgery			0.898
No	1,137 (97.1%)	488 (97.2%)
Yes	34 (2.9%)	14 (2.8%)

*preoperative, hCT Hematocrit, BMI Body Mass Index, TnI Troponin I, LVEDD Left Ventricular End-diastolic Dimension, CRP C-reactive protein, EF Ejection Fraction, HbA1c Glycated Haemoglobin, Hb Hemoglobin, NT-BNP N-terminal pro-brain natiruretic peptide, WBC White Blood Cell, CAD Coronary Artery Disease, COPD Chronic Obstructive Pulmonary Disease, CRF Chronic Renal Failure

¹Median (Q1, Q3), n (%) ²Wilcoxon rank sum test, Pearson’s Chi-squared test

Incidence of outcomes

The primary outcome (30‑day all‑cause mortality) occurred in 57/1,673 (3.4%) overall, including 40/1,171 (3.4%) in the training cohort and 17/502 (3.4%) in the validation cohort. Secondary outcomes within 30 days/index hospitalization were: ischemic stroke 60/1,673 (3.6%), paravalvular leak (PVL) 32/1,673 (1.9%), pacemaker implantation 30/1,673 (1.8%), acute kidney injury (AKI) 42/1,673 (2.5%), and conversion to open surgery 48/1,673 (2.9%).

Univariate and multivariate analysis

Univariate analysis

In the training cohort, several preoperative variables were associated with 30‑day mortality on univariate logistic regression (Table 2), including higher STS score (OR = 1.43; 95% CI:1.25–1.64; p < 0.001), HbA1c (OR = 1.56; 95% CI:1.28–1.92; p < 0.001), D‑dimer (OR = 1.14; 95% CI:1.03–1.26; p = 0.008), uric acid (OR = 1.00 per µmol/L; 95% CI:1.00–1.00;p = 0.002), CRP (OR = 1.01; 95% CI:1.00–1.02; p = 0.047), and TnI (OR = 1.57; 95% CI:1.04–2.37; p = 0.031).

Table 2.

Univariate logistic regression analysis for 30-day mortality in the training cohort

Characteristic#	N	Event N	OR	95% CI	p-value1
STS Score	1,171	40	1.43	1.25, 1.64	< 0.001***
Gender
Female	493	15	—	—
Male	678	25	1.22	0.64, 2.34	0.549
Age	1,171	40	1.00	0.96, 1.05	0.847
BMI	1,171	40	1.02	0.94, 1.10	0.656
Hypertension
No	472	17	—	—
Yes	699	23	0.91	0.48, 1.72	0.774
Diabetes
No	896	26	—	—
Yes	275	14	1.79	0.92, 3.49	0.084
CAD
No	625	17	—	—
Yes	546	23	1.57	0.83, 2.98	0.164
Atrial Fibrillation
No	992	33	—	—
Yes	179	7	1.18	0.51, 2.72	0.692
Hyperlipemia
No	694	15	—	—
Yes	477	25	2.50	1.31, 4.80	0.006**
CRF
No	1,119	37	—	—
Yes	52	3	1.79	0.53, 6.01	0.346
Cerebrovascular Disease
No	927	28	—	—
Yes	244	12	1.66	0.83, 3.32	0.151
COPD
No	791	22	—	—
Yes	380	18	1.74	0.92, 3.28	0.088
Valve History
No	1,131	38	—	—
Yes	40	2	1.51	0.35, 6.51	0.577
Cause
AI	256	7	—	—
AS	915	33	1.33	0.58, 3.04	0.498
Smoking
No	803	23	—	—
Yes	368	17	1.64	0.87, 3.11	0.128
Approach
Transapical	135	6	—	—
Transfemoral	1,036	34	0.73	0.30, 1.77	0.486
WBC (10^9/L)	1,171	40	1.08	0.94, 1.24	0.253
TnI (ng/ml)	1,171	40	1.57	1.04, 2.37	0.031*
Hb (g/L)	1,171	40	0.99	0.97, 1.01	0.310
NT-BNP (pg/ml)	1,171	40	1.00	1.00, 1.00	0.560
D-dimer (mg/L)	1,171	40	1.14	1.03, 1.26	0.008**
Creatinine (mg/dL)	1,171	40	1.09	0.76, 1.56	0.627
Urea Nitrogen (mmol/L)	1,171	40	1.00	0.98, 1.02	0.921
Amylase (U/L)	1,171	40	0.96	0.91, 1.01	0.151
hCT(%)	1,171	40	1.00	0.98, 1.01	0.555
Creatine Kinase (IU/L)	1,171	40	1.00	1.00, 1.00	0.559
Triglyceride (mmol/L)	1,171	40	1.03	0.67, 1.57	0.895
CRP (mg/L)	1,171	40	1.01	1.00, 1.02	0.047*
Uric Acid (µmol/L)	1,171	40	1.00	1.00, 1.00	0.002**
Total cholesterol(mmol/L)	1,171	40	1.02	0.86, 1.21	0.796
HbA1c (%)	1,171	40	1.56	1.28, 1.92	< 0.001***
Serum Albumin (g/L)	1,171	40	0.99	0.92, 1.06	0.729
Platelet (10^9/L)	1,171	40	1.00	1.00, 1.00	0.997
Preoperative EF (%)	1,171	40	1.00	0.98, 1.03	0.791
LVEDD (mm)	1,171	40	1.01	0.98, 1.04	0.547
Reflux Area (cm²)	1,171	40	1.03	0.91, 1.17	0.635
Ascending Aortic Diameter (mm)	1,171	40	1.03	0.98, 1.09	0.236
Sinus Diameter (mm)	1,171	40	1.01	0.96, 1.07	0.730

¹*p < 0.05; **p < 0.01; ***p < 0.001

Abbreviations: CI Confidence Interval, OR Odds Ratio

#preoperative, hCT Hematocrit, BMI Body Mass Index, TnI Troponin I, LVEDD Left Ventricular End-diastolic Dimension, CRP C-reactive protein, EF Ejection Fraction, HbA1c Glycated Haemoglobin, Hb Hemoglobin, NT-BNP N-terminal pro-brain natiruretic peptide, WBC White Blood Cell, CAD Coronary Artery Disease, COPD Chronic Obstructive Pulmonary Disease, CRF Chronic Renal Failure

“Events” for the primary outcome denote 30‑day deaths within the training (model development) and validation (internal validation) cohorts

Multivariate analysis

On multivariable logistic regression, independent predictors of 30‑day mortality were STS score (adjusted OR = 1.51; 95% CI:1.29–1.76; p < 0.001), HbA1c (adjusted OR = 1.46; 95% CI:1.18–1.80; p < 0.001), D‑dimer (adjusted OR = 1.26; 95% CI:1.12–1.42; p < 0.001), and uric acid (adjusted OR = 1.00 per µmol/L; 95% CI:1.00–1.00; p = 0.031) (Table 3). The final model included these four preoperative variables.

Table 3.

Multivariate logistic regression analysis for 30-day mortality in the training cohort

Characteristic#	N	Event N	OR	95% CI	p-value1
STS Score	1,171	40	1.51	1.29, 1.76	< 0.001***
Hyperlipemia
No	694	15	—	—
Yes	477	25	1.93	0.96, 3.88	0.065
Smoking
No	803	23	—	—
Yes	368	17	1.81	0.91, 3.61	0.092
Amylase (U/L)	1,171	40	0.95	0.89, 1.00	0.058
Uric Acid (µmol/L)	1,171	40	1.00	1.00, 1.00	0.031*
HbA1c (%)	1,171	40	1.46	1.18, 1.80	< 0.001***
D-dimer (mg/L)	1,171	40	1.26	1.12, 1.42	< 0.001***

¹*p < 0.05; **p < 0.01; ***p < 0.001

^#preoperative, HbA1c Glycated Haemoglobin

Discrimination

The final preoperative model showed good discrimination for 30‑day mortality: AUC = 0.84 (95% CI :0.788–0.892) in the training cohort and 0.78 (95% CI:0.684–0.877) in the validation cohort (Fig. 3). When the STS score was excluded, AUC declined to 0.762 (95% CI 0.683–0.842) in training and 0.688 (95% CI 0.553–0.822) in validation.

Fig. 3.

Receiver Operating Characteristic (ROC) curves for 30-day mortality prediction. A ROC curve for the multivariable logistic regression model including all statistically significant variables. B ROC curve for the model excluding the STS score. The solid lines represent the ROC curves for the training cohort, while the dashed lines represent the validation cohort. The shaded areas indicate the 95% confidence intervals.

Calibration curve

In the validation cohort, the model showed acceptable calibration. The calibration plot demonstrated close agreement between predicted and observed 30‑day mortality, and the Hosmer–Lemeshow test indicated no evidence of poor fit (χ² = 10.8, df = 8; p = 0.211). The Brier score was 0.031, suggesting good overall accuracy. Logistic recalibration yielded values close to ideal (intercept ≈ 0 and slope ≈ 1), consistent with visual inspection of the calibration curve (Fig. 2).

Fig. 2.

Calibration curve for the 30-day mortality prediction model in the validation cohort. Loess‑smoothed calibration curve with 95% confidence bands. The dashed diagonal denotes perfect calibration. Hosmer–Lemeshow χ² = 10.8 (df = 8), p = 0.211

Sensitivity analysis-bootstrap internal validation

As a sensitivity analysis, we performed bootstrap validation with 1,000 resamples of the development cohort, repeating the entire model-building process within each resample (i.e., the prespecified preoperative candidate set, univariate screening, and AIC-based backward selection under events-per-parameter constraints). The optimism-corrected performance closely matched the internal hold-out results: the C-statistic (AUC) was 0.84 and the Brier score was 0.03. Calibration was overall good, with the calibration-in-the-large close to 0 (≈ 0.00) and the calibration slope close to 1 (≈ 1.00). The average calibration error was minimal (E_avg = 0.006; mean absolute error = 0.008), although the maximum error reached 0.147, indicating slight miscalibration in certain probability ranges. The apparent and bootstrap bias‑corrected calibration curves are shown in Supplementary Fig.S1. and detailed metrics are provided in Supplementary Table S1.

Receiver operating characteristic (ROC) curve

The multivariable logistic regression model incorporating statistically significant predictors demonstrated good discrimination for 30-day mortality. The area under the ROC curve (AUC) was 0.84 (95% CI: 0.788–0.892) in the training cohort and 0.78 (95% CI: 0.684–0.877) in the validation cohort. When the STS score was excluded from the model, the discriminatory performance declined, with the AUC dropping to 0.762 (95% CI: 0.683–0.842) in the training cohort and 0.688 (95% CI: 0.553–0.822) in the validation cohort. These findings indicate that the STS score is a major contributor to the model’s predictive accuracy. (Fig. 3.)

Nomogram

A nomogram was derived directly from the final multivariable model to facilitate individualized risk estimation (Fig. 4). Each predictor (STS score, HbA1c, D‑dimer, and uric acid) is assigned points proportional to its regression coefficient; total points map to the predicted 30‑day mortality probability. To avoid ambiguity noted by reviewers, the risk probability axis is displayed from 0 to 1.0 in the main figure. In our data, predicted risks ranged from < 1% to 72%, with the upper bound determined by the observed covariate combinations rather than by any limitation of the model.

Fig. 4.

Nomogram for predicting 30-day mortality after TAVR. Predictors included in the final multivariable model are shown (STS score, HbA1c, D‑dimer, and uric acid). Points assigned are proportional to regression coefficients; total points translate to a predicted probability of 30‑day mortality. The probability axis is plotted from 0 to 1.0; in this cohort, the maximum predicted risk was 0.72. Abbreviations: STS = Society of Thoracic Surgeons (score); HbA1c = glycated hemoglobin

Discussion

In this study, we developed and internally validated a multivariable risk prediction model for 30‑day mortality following transcatheter aortic valve replacement (TAVR), using only preoperative information and incorporating both traditional risk factors and routinely available laboratory data, including uric acid, HbA1c, and D‑dimer. Key predictors identified were STS score, uric acid, HbA1c, and D‑dimer, which were consistently associated with early mortality. The model demonstrated good discrimination and acceptable calibration in the hold‑out validation and in bootstrap optimism‑corrected analyses, and we provide a nomogram to support individualized risk estimation. Our results align with large‑scale and multicenter cohort studies underscoring the need for comprehensive risk stratification in TAVR populations [12, 13]. Consistent with the PARTNER and FRANCE‑TAVI registries, we found that integrating clinical variables with routinely available biomarkers can improve prediction of short‑ and long‑term outcomes after TAVR [3, 12, 14, 15]. Similar to our findings, Condado JF et al. and Leha A et al. identified STS score and markers of inflammation as independent predictors of mortality [16, 17]. Importantly, our study incorporated uric acid and HbA1c into a validated preoperative model, reflecting emerging evidence that metabolic and inflammatory pathways are relevant for risk assessment in structural heart interventions; our observations on uric acid are concordant with reports linking hyperuricemia to heightened cardiovascular risk and mortality after TAVR [18].

Compared with previous models, our approach offers several advantages [19, 20]. First, we focused on preoperative, routinely obtained laboratory parameters (e.g., uric acid and D‑dimer) that are not part of conventional surgical risk scores but show prognostic value in contemporary literature. Second, internal validity was supported by both a prespecified hold‑out cohort and bootstrap optimism‑correction, with closely matched performance estimates. Third, the nomogram provides a user‑friendly tool for bedside estimation of individualized risk. In response to reviewer feedback, we clarified that the model outputs probabilities on the full 0–1 scale; the main figure now makes the axis limits explicit, and the maximum predicted 30‑day mortality observed in our cohort was approximately 0.72.Not all previously reported predictors retained significance in our multivariable analysis. Not all variables reported in prior literature could be evaluated in our dataset. Formal frailty indices and standardized preprocedural imaging markers were not consistently available and were therefore not included in model development. This limitation may contribute to differences from studies that incorporated these domains [21–23]. These differences may reflect variation in patient selection, variable definitions and availability, center experience, and the limited number of primary events, all of which can influence power to detect modest associations.

This study has limitations. First, it is a single‑center retrospective analysis with internal (hold‑out and bootstrap) rather than external validation; external temporal and multicenter validation, with potential recalibration, is needed prior to broader application. Second, some potentially relevant factors (e.g., formal frailty indices and comprehensive imaging parameters) were not consistently available and could not be incorporated. Third, to protect the construct validity of a preoperative model, we prespecified the exclusion of emergency/salvage procedures and patients with gout, as acute pathophysiology and biomarker perturbations in these settings may distort preoperative risk relationships; this may limit generalizability to these subgroups, warranting targeted evaluation in future cohorts. Finally, while we retained the STS score as a composite predictor to avoid redundancy with its individual components, residual confounding from unmeasured factors remains possible.

Our model is intended for preoperative counseling and risk communication in elective/non‑salvage TAVR. Given the constraints of sample size and the desire to minimize overfitting, we prespecified a parsimonious four‑predictor model and confirmed its stability via hold‑out testing and bootstrap optimism‑correction. Future studies should focus on prospective, multicenter external validation across diverse populations, assess transportability to emergency/salvage presentations, include subgroups such as patients with recent gout, and explore the incremental value of additional biomarkers and imaging features. In summary, we provide a parsimonious, internally validated preoperative risk model and nomogram for early mortality after TAVR that is broadly consistent with prior large‑scale evidence while offering practical advantages for individualized risk assessment [12, 13, 19, 20].

Conclusion

This single‑center study developed and internally validated a parsimonious preoperative risk prediction model and nomogram for 30‑day mortality after elective/non‑salvage TAVR in a Chinese (East‑Asian) cohort. Using readily available preoperative variables (STS score, HbA1c, D‑dimer, and uric acid), the model showed good discrimination and acceptable calibration in an internal hold‑out cohort, with concordant bootstrap optimism‑corrected performance. The tool is intended for preoperative counseling and procedural planning in elective candidates; emergency/salvage TAVR and gouty arthritis were prespecified exclusions. External temporal and multicenter validation—particularly in non‑Asian populations—and, if needed, recalibration are required prior to broader implementation.

Abbreviations

TAVR

Transcatheter aortic valve replacement

STS

Society of thoracic surgeons (score)

HbA1c

Glycated hemoglobin

AUC

Area under the (ROC) curve

ROC

Receiver operating characteristic

IQR

Interquartile range

BMI

Body mass index

CRP

C-reactive protein

WBC

White blood cell

TnI

Troponin I

LVEDD

Left ventricular end-diastolic dimension

EF

Ejection fraction

NT-BNP

N-terminal pro-brain natriuretic peptide

CAD

Coronary artery disease

COPD

Chronic obstructive pulmonary disease

CRF

Chronic renal failure

hCT

Hematocrit

Hb

Hemoglobin

AS

Aortic stenosis

AI

Aortic insufficiency

OR

Odds ratio

CI

Confidence interval

Authors’ contributions

H.C. Li and L.Q. Wang: Contributed to the conception and design of the study, drafted and revised the manuscript, and approved the final version. P.F. Chen, X.J. Luo, and X. Wang were responsible for clinical or laboratory data acquisition. C.Y. Liu and S.Y. Wang: Performed data analysis and interpretation. X.S. Li and Y.Q. Xie: Provided critical revisions to the manuscript and approved the final version. Y.T. Wang: Reviewed and approved the final manuscript.

Funding

Chinese Academy of Medical Sciences Fund for Medical Sciences (No. 2021-1-I2M-016).

Data availability

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

The study adhered to the principles of the Helsinki Declaration and the Measures for Ethical Review of Biomedical Research Involving Humans. The protocol was reviewed and approved by the Ethics Committee of Fuwai Hospital (Approval No. 2023-2005). Informed consent was waived because of the retrospective nature of the study and anonymized data processing.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.