Development and validation of a nomogram predicting all-cause mortality for patients with an implantable cardioverter defibrillator Authors: Bin Zhou, Lei Ding, Xuerong Sun, Shuang Zhao, Keping Chen, Yangang Su, Jiefu Yang, Zhaoguang Liang, Wei Xu, Daoliang Zhang, Teng Li, Jun Liu, Wei Hua, Min Tang, Zhiyong Liao, Shu Zhang Published: 2026 ABSTRACT AbstractBackgroundFor patients with a high risk of sudden cardiac death, despite the benefits of an implantable cardioverter defibrillator (ICD), some patients are still at high risk of death.AimThe purpose of this study was to develop and validate a nomogram predicting all-cause mortality for patients with an ICD.MethodsWe retrospectively analysed the data of multicentre ICD registration study from 2010 to 2014 in China. A total of 617 ICD patients formed a development cohort. The physical activity monitored by ICD and clinical data was collected. Univariate and multivariate Cox regression analyses were used to screen mortality predictors and construct the nomogram. The performance of the nomogram was evaluated by the consistency index (C-index) and the calibration curve. Additionally, extensive subgroup and sensitivity analyses were conducted to evaluate the model’s robustness. A total of 196 ICD patients formed a validation cohort.ResultsIn the development cohort, physical activity, diabetes and left ventricular end-diastolic diameter were selected as independent prognostic factors. The nomogram was constructed by these three factors. The C-index of the nomogram was 0.80 (95% CI 0.75 to 0.84). The calibration curve showed that the predicted survival probability of the nomogram was in good agreement with the actual survival probability. In the validation cohort, the C-index of the nomogram was 0.74 (95% CI 0.64 to 0.84), and the calibration curve still maintained good consistency. Crucially, the nomogram maintained stable and excellent discriminative capacity across primary and secondary prevention subgroups, as well as for predicting specific cardiac and sudden cardiac death.ConclusionsOur study develops and validates a nomogram predicting all-cause mortality for patients with an ICD by integrating the physical activity monitored by ICD and clinical data. The nomogram performs well and can provide personalised death risk assessment for ICD patients.Trial registration numberChiCTR-ONRC-13003695. WHAT IS ALREADY KNOWN ON THIS TOPIC Current clinical practice guidelines recommend an implantable cardioverter defibrillator (ICD) for patients at high risk of sudden cardiac death. However, despite advances in pharmacological therapy and device programming, a significant proportion of ICD recipients remain highly vulnerable to overall mortality, highlighting a critical unmet need for precise, individualised risk stratification tools. Current clinical practice guidelines recommend an implantable cardioverter defibrillator (ICD) for patients at high risk of sudden cardiac death. However, despite advances in pharmacological therapy and device programming, a significant proportion of ICD recipients remain highly vulnerable to overall mortality, highlighting a critical unmet need for precise, individualised risk stratification tools. WHAT THIS STUDY ADDS This study develops and validates a highly parsimonious prognostic nomogram that seamlessly integrates dynamic device-measured physical activity with two accessible clinical parameters (diabetes and left ventricular end-diastolic dimension). The model demonstrates robust discriminative capacity for all-cause mortality across both primary and secondary prevention cohorts, while consistently identifying those at highest risk for specific cardiac death. This study develops and validates a highly parsimonious prognostic nomogram that seamlessly integrates dynamic device-measured physical activity with two accessible clinical parameters (diabetes and left ventricular end-diastolic dimension). The model demonstrates robust discriminative capacity for all-cause mortality across both primary and secondary prevention cohorts, while consistently identifying those at highest risk for specific cardiac death. HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY By translating objective, device-derived functional data and clinical indices into a practical bedside tool, this nomogram enables rapid and personalised risk stratification. This facilitates more tailored follow-up strategies, optimises targeted interventions and improves holistic clinical decision-making for vulnerable ICD populations. By translating objective, device-derived functional data and clinical indices into a practical bedside tool, this nomogram enables rapid and personalised risk stratification. This facilitates more tailored follow-up strategies, optimises targeted interventions and improves holistic clinical decision-making for vulnerable ICD populations. Introduction Sudden cardiac death (SCD) claims an estimated 4–5 million lives annually worldwide.1 With survival rates frequently below 10%, SCD accounts for over half of all cardiovascular deaths and up to 20% of overall global mortality—a burden exceeding that of any individual cancer.1 In the USA, SCD accounts for approximately half of all coronary heart disease deaths.2 In China, according to epidemiological surveys, given that China’s population is 1.3 billion people, approximately 544 000 people die of SCD each year.3 Current clinical practice guidelines recommend the implantation of an implantable cardioverter defibrillator (ICD) in patients at high risk of SCD. However, despite advances in pharmacological therapy and device programming, many patients at high risk of SCD remain at high risk of death after ICD implantation.4 5 Predicting which patients are at high risk of adverse effects is important for the clinical management of patients and for the adoption of interventions to reduce mortality after ICD implantation. There are a number of studies on risk factors for mortality in ICD patients,6 but most of them focus on traditional clinical data.7 Parameters that can be monitored by ICD devices have received increasing attention in recent years. Among these, physical activity (PA), which is automatically captured by the ICD, provides a quantitative and readily accessible measure of an individual’s functional status8 and has been shown to be an independent predictor of prognosis in several studies.9–12 Several studies have proposed death prediction models4 5 13 14 for ICD patients, but most of them have synthesised only clinical indicators. The focus of prediction models that integrate PA information has been the heart failure population.9 Moreover, the development and validation of these models have focused mostly on European and American populations, and prediction models for the Chinese population are lacking. Moreover, in prediction models, nomograms allow individualised and evidence-based risk estimation that can directly assess the probability of death for each patient and can effectively facilitate management-related clinical decisions.15 In this study, we analysed multicentre ICD data from China, and through the use of PA information automatically collected by ICD, combined with clinical information, we aimed to construct a prediction model for ICD patients’ deaths based on Chinese data and present a model with a nomogram to assess the risk of death directly for each patient. Materials and methods Study population and design The SUMMIT study (Study of Home Monitoring System Safety and Efficacy in Cardiac Implantable Electronic Device (CIED) implanted patients (SUMMIT)) is a prospective, observational national multicentre registry used to evaluate the safety and efficacy of Biotronik pacemakers (Biotronik, Germany) with home monitoring (HM) systems in China (registration No. ChiCTR-ONRC-13003695). The academic reliability and robust data quality of this comprehensive registry have been successfully demonstrated in prior peer-reviewed publications by our research group.16 We conducted a retrospective cohort study based on data from the SUMMIT registry. Between May 2010 and April 2014, a total of 1008 patients who underwent ICD or cardiac resynchronisation therapy defibrillator (CRT-D) (all here and after being harmonised with the ICD) implantation were registered. The inclusion criteria for this study were (1) age greater than or equal to 18 years, (2) met the indications for either primary or secondary prevention of SCD according to the clinical practice guidelines (primary prevention of SCD is defined as the use of an ICD in the presence of sustained ventricular tachycardia (VT) and ventricular fibrillation (VF) or resuscitative cardiac arrest in individuals at risk for but not yet experiencing an episode; secondary prevention is the prevention of SCD in patients who have survived a prior cardiac arrest or VT/VF), (3) survival for at least 3 months after implantation and (4) PA data documented by the HM system. The exclusion criteria were (1) age younger than 18 years (n=7), (2) missing PA data (n=168), (3) missing systolic blood pressure (SBP) and diastolic blood pressure (DBP) data (n=7), (4) missing left ventricular ejection fraction (LVEF) data (n=5) and (5) missing left ventricular end diastolic dimension (LVEDD) data (n=26). Considering these inclusion and exclusion criteria, a total of 813 patients were included in this study. The enrolment numbers by province or municipality are shown in online supplemental table S1. Based on the enrolment rankings, patients from the top five enrolled regions were categorised into the development cohort (n=617); the remaining patients were categorised into the independent validation cohort (n=196). The specific study flow chart is shown in figure 1. All the reports followed the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD) statement.17 Flowchart of the study population. Abbreviations are shown in table 1. Data acquisition PA is measured by an integrated circuit accelerometer embedded in the ICD pulse generator.18 Previous studies have ensured the accuracy of PA measurements.19 The PA resolution is 2 s, and the data are converted to percentages every 24 hours. For example, 5% PA represents 1.2 hours of PA per day, which translates to 72 min in minutes. Biotronik’s remote monitoring system automatically transmits the data stored in the implantable device to the service centre on a daily basis. Because PA is expected to be lower than normal in the early postdischarge period, according to a previous study,10 11 data were collected during the first 30–60 days of device implantation, and the average 30-day PA was calculated as the baseline PA for each patient, with minutes used as the unit. Other baseline clinical characteristics, including age at implantation, sex, body mass index, SBP, DBP, indication for primary or secondary prevention, New York Heart Association (NYHA) classification, implantation of an ICD or CRT-D, heart valve disease, ischaemic cardiomyopathy, dilated cardiomyopathy, hypertrophic cardiomyopathy, hypertension, diabetes, stroke, atrial fibrillation, preimplantation syncope, history of coronary artery bypass grafting, history of coronary artery stent implantation, previous infarction, LVEF, LVEDD, β-blocker, amiodarone, ACE inhibitor or angiotensin receptor blocker (ACEI/ARB), diuretics and spironolactone, were obtained from the patient’s charts prior to ICD implantation. The LVEF was calculated with the Modified Simpson’s Biplane Rule. Definition of outcomes and follow-up The endpoint of this study was all-cause mortality. Routine follow-up was performed, and patient status was confirmed by telephone in case of transmission interruption. Once a patient died, family members or witnesses were contacted to assess the date of death carefully. Statistical methods Data are expressed as mean and SD or proportion. The χ² test (categorical variables) or one-way analysis of variance combined with the Bonferroni post hoc test (continuous variables) was used to calculate differences between groups. Survival of the two cohorts was demonstrated via Kaplan-Meier curves and compared using the log-rank method. To develop and validate a predictive model for all-cause mortality, it was first necessary to identify the variables to which the model would be fitted. The risk factors for all-cause mortality were explored using a one-way Cox risk proportion model. Variables with a univariable p<0.10 were included in the multivariable Cox model. A backward stepwise selection was performed, retaining variables with p<0.05 in the final model. The models were presented using a nomogram, which was developed using the rms package for the R language. The performance of the nomogram was measured by the consistency index (C-index), and the calibration was evaluated by comparing the predicted and observed Kaplan-Meier survival probability estimates of the nomogram. The larger the C-index is, the more accurate the prognostic prediction. During external validation of the nomogram, total scores were calculated for each patient in the validation cohort based on the established nomogram, Cox regression was performed in that cohort using the total score as a factor, and finally, the C-index and calibration curves were derived based on the regression analysis. Subgroup and sensitivity analysis To evaluate the robustness and clinical applicability of our model, we conducted extensive subgroup and sensitivity analyses. First, we performed a prespecified subgroup analysis stratified by ICD implantation indication (primary vs secondary prevention). In accordance with modern prediction modelling guidelines (such as the TRIPOD statement), which strongly caution against re-estimating coefficients or rescreening variables in smaller subcohorts, we directly applied our fully developed three-factor nomogram to both subgroups to evaluate its discriminative capacity and clinical validity. Second, to address competing risks and specific modes of death, causes of death were adjudicated to distinguish between cardiac mortality (including SCD) and non-cardiac mortality. We employed Cause-Specific Cox (CS-Cox) proportional hazards models for these specific endpoints, treating competing events as censored, which allows for the precise identification of intrinsic prognostic factors linked to the specific biology of cardiac death. For non-cardiac death, given the limited number of events, only univariable analysis was performed to prevent statistical overfitting. All analyses were performed using R V.4.0.0 (R Foundation for Statistical Computing, Vienna, Austria). All p values <0.05 (two-sided) were considered statistically significant. Results Baseline population characteristics After applying the inclusion and exclusion criteria, 617 patients were included in the development cohort and 196 were included in the validation cohort. The baseline characteristics of the two cohorts are shown in table 1. The baseline characteristics of the two cohorts showed some degree of difference, with the development cohort showing a higher percentage of hypertension, a greater SBP, a lower percentage of primary prevention, a lower percentage of CRT-D, more syncope, more dilated cardiomyopathy and a greater percentage of amiodarone use than the validation cohort (p<0.05). Table 1Baseline characteristics of the population in the development and validation cohortsTotal population(n=813)Development cohort(n=617)Validation cohort(n=196)P valueMale, %597 (73.4)462 (74.9)135 (68.9)0.098Age, years60.6±13.860.6±13.860.6±14.00.975PA, min155.1±85.8153.8±84.3159.1±90.40.454BMI, kg/m223.6±3.523.6±3.323.8±4.10.444SBP, mmHg124.1±17.5127.9±17.4122.9±17.4<0.001DBP, mmHg76.6±11.276.3±11.577.7±10.30.142NYHA (III-IV), %402 (49.4)303 (49.1)99 (50.5)0.732Primary prevention, %427 (52.5)302 (48.9)125 (63.8)<0.001CRT-D, %216 (26.6)149 (24.1)67 (34.2)0.006Ischaemic cardiomyopathy, %271 (33.3)201 (32.6)70 (35.7)0.417Syncope, %215 (26.4)185 (30.0)30 (15.3)<0.001Atrial fibrillation95 (11.7)73 (11.8)22 (11.2)0.818History of coronary artery bypass grafting, %10 (1.2)10 (1.6)0 (0.0)0.073History of coronary artery stent implantation, %71 (8.7)56 (9.1)15 (7.7)0.539Old infarction, %128 (15.7)105 (17.0)23 (11.7)0.077Heart valve disease, %20 (2.5)18 (2.9)2 (1.0)0.135Hypertrophic cardiomyopathy, %34 (4.2)27 (4.4)7 (3.6)0.624Dilated cardiomyopathy, %196 (24.1)159 (25.8)37 (18.9)0.049Stroke, %16 (2.0)12 (1.9)4 (2.0)0.933Diabetes, %78 (9.6)66 (10.7)12 (6.1)0.058Hypertension, %257 (31.6)209 (33.9)48 (24.5)0.014LVEDD, mm58.6±13.358.3±13.259.8±13.50.156LVEF, %42.5±14.942.9±15.141.5±14.40.240Diuretics, %348 (42.8)264 (42.8)84 (42.9)0.986Spironolactone, %295 (36.3)221 (35.8)74 (37.8)0.741ACEI/ARB, %323 (39.7)251 (40.7)72 (36.7)0.325Amiodarone, %247 (30.4)199 (32.3)48 (24.5)0.040β-blocker, %506 (62.2)377 (61.1)129 (65.8)0.236Data are presented as mean±SD or n (%), as appropriate.ACEI/ARB, ACE inhibitor or angiotensin receptor blocker; BMI, body mass index; CRT-D, cardiac resynchronisation therapy defibrillator; DBP, diastolic blood pressure; LVEDD, left ventricular end diastolic dimension; LVEF, left ventricular ejection fraction; NHYA, New York Heart Association; PA, physical activity; SBP, systolic blood pressure. Baseline characteristics of the population in the development and validation cohorts Data are presented as mean±SD or n (%), as appropriate. ACEI/ARB, ACE inhibitor or angiotensin receptor blocker; BMI, body mass index; CRT-D, cardiac resynchronisation therapy defibrillator; DBP, diastolic blood pressure; LVEDD, left ventricular end diastolic dimension; LVEF, left ventricular ejection fraction; NHYA, New York Heart Association; PA, physical activity; SBP, systolic blood pressure. Comparison of clinical outcomes and survival curves between the two cohorts The median follow-up for the development cohort was 29 months (min to max 3–60 months), with a total of 100 patients experiencing all-cause deaths (16.2%) and for the validation cohort, the median follow-up was 31 months (min to max 3–60 months), with a total of 29 patients experiencing all-cause deaths (14.8%). There was no statistically significant difference in the percentage of deaths between the two cohorts (p=0.638). Figure 2 shows the survival Kaplan-Meier curves for the two cohorts, with no statistically significant difference in the overall probability of survival between the two groups (log-rank, p=0.51). The cumulative probabilities of survival at 1, 2 and 5 years in the development cohort were 93.2%, 87.3% and 78.0%, respectively; the cumulative probabilities of survival at 1, 2 and 5 years in the validation cohort were 98.5%, 90.2% and 70.1%, respectively. Kaplan-Meier curves for the two cohorts. 1 represents the development cohort; 2 represents the validation cohort. Prognostic factor screening and nomogram construction Univariable Cox proportional hazards analysis was performed for the variables in table 1. Subsequently, candidate variables with a univariable p value <0.10 were entered into the multivariable Cox proportional hazards model. Following a backward stepwise selection process, variables maintaining a two-tailed p value <0.05 were retained in the final model. Consequently, PA (HR 0.99, 95% CI 0.98 to 0.99), diabetes (HR 1.76, 95% CI 1.03 to 3.02) and LVEDD (HR 1.02, 95% CI 1 to 1.04) emerged as independent prognostic factors significantly associated with all-cause mortality (table 2). The model was constructed using these three prognostic factors and displayed as a nomogram (figure 3). To use the nomogram, first, the actual value of the patient’s PA can be located on the corresponding axis. Next, a straight line is drawn upward to the top point axis to obtain a score based on the PA counterpart. This process is then repeated for the other variables, and the total score value is calculated by summing all the scores obtained from each variable. The final sum is located on the axis of the total score value, and a straight line is drawn downward to obtain the patient’s probability of survival at 1, 2 and 5 years. Nomogram construed by PA, diabetes, LVEDD. The probability of survival at 1, 2 and 5 years can be determined for each patient based on the total score corresponding to the three prognostic factors. Table 2Risk factors for all-cause death assessed by univariate and multivariate Cox proportional hazards regression analysisUnivariate CoxMultivariate CoxVariablesHR (95% CI)P valueHR (95% CI)P valueFemale vs male1.33 (0.87 to 2.03)0.186 Age1.02 (1.00 to 1.03)0.016 PA0.99 (0.98 to 0.99)<0.00010.99 (0.98 to 0.99)<0.0001BMI0.95 (0.89 to 1.01)0.097 SBP0.99 (0.98 to 1.00)0.087 DBP0.98 (0.96 to 1.00)0.040 NYHA (III–IV) vsNYHA (I–II)2.54 (1.66 to 3.90)<0.0001 Second prevention vs primary prevention0.83 (0.56 to 1.22)0.343 CRT-D vs ICD1.60 (1.06 to 2.43)0.023 Ischaemic cardiomyopathy1.60 (1.07 to 2.37)0.021 Syncope0.91 (0.59 to 1.41)0.678 Atrial fibrillation1.24 (0.71 to 2.19)0.448 History of coronary artery bypass grafting2.80 (1.03 to 7.61)0.044 History of coronary artery stent implantation1.03 (0.54 to 1.99)0.920 Old infarction1.86 (1.20 to 2.88)0.0054 Heart valve disease1.43 (0.53 to 3.89)0.483 Hypertrophic cardiomyopathy0.62 (0.20 to 1.96)0.416 Dilated cardiomyopathy1.32 (0.86 to 2.02)0.2 Stroke1.17 (0.37 to 3.70)0.792 Diabetes1.94 (1.17 to 3.24)0.0111.76 (1.03 to 3.02)0.039Hypertension1.36 (0.91 to 2.02)0.129 LVEDD1.03 (1.01 to 1.04)0.00011.02 (1 to 1.04)0.036LVEF0.98 (0.96 to 0.99)0.0006 Diuretics1.98 (1.33 to 2.96)0.0008 Spironolactone2.05 (1.42 to 3.04)0.0002 ACEI/ARB1.31 (0.88 to 1.94)0.182 Amiodarone1.10 (0.72 to 1.66)0.666 β-blocker1.29 (0.85 to 1.96)0.239 ACEI/ARB, ACE inhibitor or angiotensin receptor blocker; BMI, body mass index; CRT-D, cardiac resynchronisation therapy defibrillator; DBP, diastolic blood pressure; ICD, implantable cardioverter defibrillator; LVEDD, left ventricular end diastolic dimension; LVEF, left ventricular ejection fraction; NHYA, New York Heart Association; PA, physical activity; SBP, systolic blood pressure. Risk factors for all-cause death assessed by univariate and multivariate Cox proportional hazards regression analysis ACEI/ARB, ACE inhibitor or angiotensin receptor blocker; BMI, body mass index; CRT-D, cardiac resynchronisation therapy defibrillator; DBP, diastolic blood pressure; ICD, implantable cardioverter defibrillator; LVEDD, left ventricular end diastolic dimension; LVEF, left ventricular ejection fraction; NHYA, New York Heart Association; PA, physical activity; SBP, systolic blood pressure. Assessment of the nomogram In the development cohort, the C-index for the risk of death predicted by the nomogram was 0.80 (95% CI 0.75 to 0.84). Calibration plots of 1-year, 2-year and 5-year survival with ICD implantation showed good agreement between the estimated and actual values (figure 4A–C). Using the nomogram in the validation cohort, the C-index for the risk of death predicted by the nomogram was 0.74 (95% CI 0.64 to 0.84). In the validation cohort, a calibration plot of 2-year survival with ICD implantation showed good agreement between the estimated and actual values (figure 4D). Calibration curves for predicting all-cause death in the development and validation cohorts. The blue line represents the correlation between actual and predicted values of all-cause death. The diagonal dashed line represents the ideal perfect prediction. (A) Calibration curve for 1-year all-cause death in the development cohort; (B) calibration curve for 2-year all-cause death in the development cohort; (C) calibration curve for 5-year all-cause death in the development cohort; (D) calibration curve for 2-year all-cause death in the validation cohort. Subgroup and sensitivity analysis We first evaluated the model’s robustness across different clinical indications. When directly applied to these subcohorts, the integrated three-factor model maintained outstanding and stable discriminative capacity, achieving an excellent C-index of 0.82 (95% CI 0.78 to 0.86) in the primary prevention subgroup and 0.77 (95% CI 0.71 to 0.84) in the secondary prevention subgroup. This demonstrates that our prespecified combination of functional, metabolic and structural variables can reliably stratify mortality risk across different clinical settings, effectively facilitating better and more individualised candidate selection for ICD implantation (online supplemental table S2). Furthermore, to address competing risks and specific modes of death, causes of death were adjudicated. Among the 100 all-cause deaths in the development cohort, 81 were cardiac deaths (including 28 SCDs and 53 non-SCDs) and 19 were non-cardiac deaths. In the cause-specific multivariable CS-Cox analyses, the three core variables (PA, diabetes and LVEDD) consistently remained independent prognostic factors for both overall cardiac death and SCD. The three-factor model demonstrated excellent discriminative ability for cardiac mortality (C-index: 0.83 (95% CI 0.78 to 0.87) in the development cohort; 0.75 (95% CI 0.64 to 0.86) in the validation cohort) and SCD (C-index: 0.81 (95% CI 0.73 to 0.89) and 0.87 (95% CI 0.78 to 0.95), respectively), accompanied by optimal calibration (online supplemental figures S1 and S2 and online supplemental tables S3 and S4). Conversely, an exploratory univariable analysis of the 19 non-cardiac deaths revealed that while systemic disease burden (history of coronary artery bypass grafting) and general functional frailty (lower PA) were associated with non-cardiac mortality, intrinsic cardiac structural parameters (eg, LVEDD) were not (online supplemental table S4). Discussion Although several clinical trials have shown that ICD implantation in patients with high-risk SCD reduces death,20 21 overall mortality in these trials remains high. An improved assessment of the risk of death despite ICD implantation would be beneficial, and identifying patients at high risk of death would allow physicians to better target interventions to improve survival. The nomogram constructed in this study, which integrates PA data and clinical information, showed good discrimination and calibration in both the development and validation cohorts and is a convenient and clinically valuable tool with some extrapolation applicability. In clinical practice, once the three types of information, namely, PA, LVEDD and diabetes mellitus, are known for an ICD patient, the probability of survival at 1, 2 and 5 years can be estimated by the nomogram, which can be personalised to assess the risk of death, and appropriate interventions can be carried out. If a patient has a high risk of death in the short term, the interval of follow-up can be shortened, and the control of risk factors can be strengthened. Incremental value and clinical implications of the proposed model Bilchick et al5 proposed the ‘SHOCKED’ model for predicting death in a population with an ICD implanted for primary prevention of SCD; seven predictors, namely, age, NYHA classification, atrial fibrillation, chronic obstructive pulmonary disease, chronic kidney disease, ejection fraction and diabetes mellitus, are integrated into this model and the authors reported a C-index of 0.75 (95% CI 0.75 to 0.76) in the development cohort and 0.74 (95% CI 0.74 to 0.75) in the validation cohort. Barreveld et al14 proposed a similar mortality prediction model for a population with ICD implantation for the primary prevention of SCD, which integrates five predictors, namely, age, diuretics, blood sodium, N-terminal pro-B-type natriuretic peptide (NT-proBNP) and ACEI/ARB. The C-index of the model was 0.74 (95% CI 0.70 to 0.78) in the development cohort and 0.74 (95% CI 0.70 to 0.78) in the validation cohort. Borleffs et al13 proposed different death prediction models for the subgroup of ischaemic heart disease and non-ischaemic heart disease in SCD primary prevention patients implanted with an ICD. In the ischaemic heart disease subgroup, the death model included four predictors, namely, renal function, ejection fraction, atrial fibrillation and age, with a C-index of 0.81 (95% CI 0.76 to 0.87) and in the non-ischaemic heart disease subgroup, the death model included six predictors, namely, renal function, cigarette smoking, diabetes mellitus, QRS duration, ejection fraction and age, with a C-index of 0.76 (95% CI 0.69 to 0.82). The model was not tested for calibration or externally validated. Additionally, constructing different predictive models for different subgroups would increase the clinical workload and would not be conducive to clinical application. The present study used the least number of predictors, but the discriminatory power of the model was not inferior to that of the above models. Moreover, the ICD population included in our study included those with primary and secondary prevention of SCD, which has wider population applicability. Kramer et al4 proposed a hazard score for the risk of death at 1 year for ICD patients, which integrates four predictors, namely, age, creatinine level, ejection fraction and peripheral arterial disease. In the study, scores were assigned only for risk stratification through the scoring model, and a specific C-index was not provided, nor was calibration performed. In our study, the differentiation and calibration of the model were reported in strict accordance with the requirements of the TRIPOD statement, and the effect of the model was presented in a more standardised way. Additionally, this study had a mean follow-up of nearly 3 years, spanning from 3 months to 5 years and it can be personalised to determine the risk of death from the short-term to the long-term for each patient at 1, 2 and 5 years based on the nomogram. Conraads et al9 constructed a prediction model integrating PA in ICD patients with heart failure, with predictors that included age, SBP and diuretic dosage, in addition to PA and a C-index of 0.75 (95% CI 0.68 to 0.83). However, the model was not externally validated, and no analysis of calibration was performed. The scope of the population included in our study was not limited to ICD patients with heart failure, and the model has a range of applicability that extends beyond this population. In summary, compared with traditional scoring systems, our nomogram provides a more dynamic, parsimonious and broadly applicable tool, seamlessly translating objective data into immediate clinical risk stratification. Clinical rationale and mechanisms of the core prognostic markers In this study, three prognostic factors were screened, namely, PA, diabetes mellitus and LVEDD. Among them, PA was negatively correlated with death, which is consistent with the findings of previous studies.9–12 22 The mechanism of action by which PA presents protection against death is not clear at present. Some studies suggest that good PA contributes to the homeostasis of autonomic function.23 PA is positively correlated with heart rate variability (HRV),24 25 and HRV is a non-invasive index that has been used to assess autonomic function, which is positively correlated with prognosis.26 27 Furthermore, rather than acting solely through direct physiological pathways, device-measured PA may also serve as a comprehensive surrogate marker reflecting better overall global fitness, preserved systemic functional capacity and lower physical frailty. The mortality-promoting effect of diabetes is also similar to previous findings.5 28 29 Compared with patients without diabetes, patients with diabetes gain similar benefits from ICD therapy.30 However, conventionally treated patients with diabetes tend to be sicker and have a higher mortality rate than patients without diabetes, possibly because patients with diabetes have more comorbidities. Thus, it is not surprising that diabetes is associated with increased mortality in ICD patients. LVEDD, a commonly used index in cardiac ultrasound to assess cardiac structure and function, has been associated with poor prognosis in the hypertrophic cardiomyopathy population, the aortic stenosis population and the chronic heart failure population.31 However, it is puzzling that LVEDD, which is commonly used in clinical practice, was not included in the baseline characteristics of numerous studies on prognostic risk factors for death or death prediction models in ICD patients,4 5 10 13 14 28 29 32 and it was not possible to analyse the effect of LVEDD on the prognosis of death in ICD patients. Our study suggests that LVEDD is independently associated with death in ICD patients and that attention should be given to LVEDD in ICD patients. Model robustness across clinical subgroups and competing risks Our extensive subgroup and sensitivity analyses further confirmed the remarkable robustness and clinical versatility of the proposed model. When stratified by ICD indication, the three-factor nomogram maintained excellent discrimination in both primary and secondary prevention cohorts. Although the primary drivers of mortality may shift—with advanced pump failure dominating in primary prevention and arrhythmic substrates dominating in secondary prevention—the integrated combination of functional (PA), metabolic (diabetes) and structural (LVEDD) variables reliably stratifies risk without the need for subgroup-specific recalculations. Furthermore, our cause-specific analyses confirmed that these three core variables are intrinsically linked to the biology of cardiac death and SCD. Conversely, exploratory analysis revealed that non-cardiac mortality was primarily driven by a heavier burden of systemic atherosclerosis (eg, history of coronary artery bypass grafting) and general functional frailty (lower PA). This powerfully reinforces our hypothesis that device-measured PA acts not merely as a localised cardiac metric, but as a universal surrogate marker for global health vulnerability. Limitations This study has the following limitations: (1) This was a retrospective study, and we did not obtain information such as electrocardiograms and blood indices, which would have prevented us from integrating the possible risk factors for death. It should be noted, however, that it is impossible and unnecessary for any predictive model to include all possible predictors. In the current study, our model, which was constructed with only three prognostic factors, performed comparably to the predictive models constructed with more factors from previous studies. (2) Although our development cohort and validation cohort were from different regions, both cohorts were from the same registry study and a more rigorous external study cohort is needed to further validate our model. The baseline characteristics of the two cohorts in this study differed significantly, but the model modelled with the development cohort still maintained good performance in the validation cohort, which had greater heterogeneity in population characteristics, indicating better extrapolation of the model. (3) Due to the retrospective observational nature of our study, we can only establish associations rather than definitive causality. The variables identified in our model should be interpreted with caution. For instance, device-measured PA likely serves as a broad surrogate marker for better overall fitness and less systemic frailty, rather than a direct mechanistic driver of survival. Therefore, these variables represent robust prognostic markers in the current dataset, but they need to be further validated in large, prospective cohorts before being definitively classified as causal predictors of mortality. Conclusions In this retrospective observational study, we developed and validated a practical nomogram incorporating PA, diabetes and LVEDD. These three factors were significantly associated with all-cause mortality in ICD recipients. While this model demonstrates excellent discriminative capacity for risk stratification, these variables likely serve as composite surrogate markers for underlying structural, metabolic and functional frailty. Further prospective studies are warranted to validate these associations before broadly implementing them as definitive clinical predictors. Supplementary material