Introduction

Aortic stenosis (AS) is a progressive and life-limiting condition that disproportionately affects the elderly. Its prevalence rises sharply with advancing age due to degenerative calcification of the aortic valve. When left untreated, symptomatic AS carries a poor prognosis, with nearly half of patients dying within four years1. This highlights the necessity of timely treatment, as valve replacement—either surgical or transcatheter—remains the only intervention capable of improving survival and quality of life. While data specific to nonagenarians (≥ 90 years) remain sparse, the burden of disease in this age group is considerable. Among individuals over 85 years, nearly half exhibit aortic sclerosis, and approximately 4% are estimated to have severe AS2,3,4. With global populations aging rapidly, the clinical and healthcare impact of AS in nonagenarians is poised to grow5,6.

Reflecting this trend, recent registry data indicate a marked increase in the number of nonagenarians undergoing transcatheter aortic valve replacement (TAVR). For example, in a multicenter Swiss registry, 507 out of 7,097 patients (7.1%) undergoing TAVR were aged over 90 years7. Similarly, other cohorts have reported that nonagenarians account for approximately 10–12% of all TAVR procedures8,9. This reflects both expanding clinical experience with TAVR and the careful selection of healthier nonagenarians for intervention.

Although nonagenarians experience higher short-term mortality compared with younger patients, their one-year outcomes are broadly comparable, supporting the feasibility and potential benefit of TAVR in this very elderly population7,8,9,10. Nevertheless, risk prediction in this group remains challenging. Established pre-procedural models, such as the Society of Thoracic Surgeons (STS) score and EuroSCORE, were largely derived in younger surgical populations, and their applicability to patients of extreme age is uncertain. Similarly, earlier ICU scoring systems, such as APACHE II, were designed in different eras of clinical practice and may not fully capture the complexity of the oldest-old.

By contrast, more contemporary ICU-based scores, including the Sequential Organ Failure Assessment (SOFA) and Acute Physiology and Chronic Health Evaluation III (APACHE III), incorporate a wider range of physiological parameters and have demonstrated prognostic value in broader critically ill or mixed TAVR cohorts11. However, their specific utility in nonagenarians and centenarians undergoing TAVR has not been well studied. Notably, no predictive models currently focus on data derived exclusively from the immediate post-TAVR ICU stay—an interval during which early physiological and biochemical changes may strongly shape clinical outcomes. In addition, biomarker interpretation in older adults is often complicated by age-related alterations in baseline values and response patterns12.

Since the first Australian transcatheter aortic valve replacement in 2008, access has expanded through MBS rebates and increasing national accreditation13. Most procedures now occur in private hospitals, where ICU admission is more routine than in the public sector.

Against this background, our study aims to evaluate the prognostic value of clinical risk scores and biochemical markers measured during the immediate post-operative ICU stay following TAVR in nonagenarians. By focusing on this highly vulnerable yet increasingly represented group, we aim to address a key knowledge gap, building upon existing octogenarian-focused research while recognising the unique physiological profile of patients aged 90 years and above. Ultimately, this work seeks to inform more refined approaches to risk stratification and perioperative management in the oldest-old14.

Methods

Data sources, study population, and variables

De-identified patient data were accessed from the ANZICS Adult Patient Database, which contains patient demographics, clinical information, interventions, and outcomes from 98% of Australian ICUs and 68% of New Zealand ICUs15. The data were collected by the ANZICS Centre for Outcome and Resource Evaluation.

The ANZICS Adult Patient Database incorporates multiple safeguards to ensure data quality, including uniform data definitions, prospective entry by trained staff, automated checks for implausible values, and regular auditing across participating ICUs. These processes help maintain consistency of data collection and reduce the risk of misclassification or reporting bias across centers and over time. Because the registry captures almost all ICU admissions in Australia and a large proportion in New Zealand, the study cohort is broadly representative of routine practice in these countries.

This study included nonagenarian patients admitted to ICUs between 2017 and 2024 after TAVR for aortic stenosis, as identified with APACHE III-J diagnostic codes. The extracted demographic variables included age, sex, admission type, admission source, discharge destination, and dates of ICU and hospital admission. Clinical information assessed at the time of ICU admission comprised frailty status, as measured with the Clinical Frailty Scale (CFS), and illness severity, as evaluated with the APACHE III-J, Glasgow Coma Scale (GCS), and SOFA scores. The highest values within the first 24 h after ICU admission were recorded for key biochemical markers. Relevant comorbidities were also captured, including immunosuppression; malignancy; and chronic respiratory, cardiovascular, hepatic, or renal disease. Elective procedures were defined as scheduled TAVR admissions without acute decompensation, whereas urgent procedures referred to cases performed during the same hospitalization for clinical deterioration such as decompensated heart failure or cardiogenic shock. All variables which had less than 5% missing data were analyzed with the complete-case method without imputation.

Study outcomes

The primary outcomes were all-cause mortality at multiple timepoints, including ICU, hospital, 1-month, 6-month, 1-year, and long-term follow-up. The prognostic association of clinical severity scores and biochemical markers with these outcomes was examined. In addition to evaluating overall model discrimination, we stratified patients into high- and low-risk groups, according to optimal thresholds for each predictor, and assessed associations with mortality across all timepoints. Secondary outcomes included ICU and hospital lengths of stay, defined from ICU admission to respective discharge. This study was aimed at evaluating the prognostic performance and clinical utility of both admission clinical scores and early biochemical markers in predicting mortality and healthcare resource use among nonagenarian and centenarian patients admitted to ICUs after TAVR.

Ethical approval

This study was approved by the Alfred Ethics Committee (project No. 253/24). The requirement for informed consent was waived because of the study’s retrospective design and use of de-identified data. Data analysis commenced only after ethics approval was obtained. The study adhered to the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines16. All methods were performed in accordance with the relevant guidelines and regulations. The study was conducted in accordance with the principles of the Declaration of Helsinki.

Statistical analyses

All analyses were conducted in R version 4.4.3 (R Foundation for Statistical Computing, Vienna, Austria). Data preprocessing and management were performed with the dplyr and tidyr packages. Continuous variables are summarized as medians with interquartile ranges, and categorical variables are reported as counts with corresponding percentages. Between-group comparisons were conducted with the Wilcoxon rank-sum test for continuous variables with non-normal distributions, and the chi-square or Fisher exact test for categorical variables, on the basis of expected cell frequencies. Normality was evaluated with the Shapiro–Wilk test and visual inspection of quantile–quantile plots.

The discriminative ability of clinical scores and biochemical markers for mortality at prespecified timepoints was assessed with receiver operating characteristic (ROC) curve analysis with the pROC package. Discrimination was quantified by the area under the ROC curve (AUROC), and 95% confidence intervals (CIs) and associated P values were derived with DeLong’s method. Pairwise comparisons of AUROCs between predictors were also performed with DeLong’s test17. A Bonferroni adjustment was applied to control for alpha inflation when performing the DeLong test for pairwise ROC curve comparisons. Optimal cut points for binary classification were determined with the Youden index18.

Univariate and multivariable logistic regression models were employed to examine the associations between clinical predictors and binary mortality outcomes. Multivariable models were adjusted for age, sex, comorbidities, elective admission status, and relevant clinical and biochemical covariates. To minimize overfitting, multivariable analyses were restricted to endpoints with sufficient event numbers. Predictor variables were chosen a priori, and model complexity was limited to conform with standard event-per-variable recommendations (10–20 events per predictor). These safeguards were applied consistently across all multivariable models. Continuous predictors were dichotomized using thresholds derived from receiver operating characteristic (ROC) analyses to enhance the interpretability of odds ratios where appropriate. The assumptions of logistic regression were assessed using the car and rms packages. Multicollinearity was evaluated using variance inflation factors, with values below 4 considered acceptable. Linearity in the logit was assessed using the Box–Tidwell test, supplemented by visual inspection of residual plots and partial residual plots. All clinical scoring systems demonstrated satisfactory linear relationships and were entered as continuous covariates. Several biochemical markers exhibited evidence of non-linearity with respect to mortality risk; therefore, restricted cubic splines with four knots were applied to flexibly model these associations without imposing a linear constraint. Adjusted odds ratios and corresponding 95% confidence intervals for each biomarker were visualized using spline plots to display the functional form of these non-linear relationships.

Associations between clinical predictors and the length of stay (LOS) in the intensive care unit (ICU) and hospital were examined using log-transformed linear regression models. Log transformation of LOS was performed to improve normality and homoscedasticity of residuals. Linearity of continuous predictors with log-transformed LOS was assessed visually using residual and partial residual plots. All clinical scoring systems demonstrated satisfactory linear relationships and were thus entered into the models as continuous variables. Several biochemical markers showed evidence of non-linearity; for these, restricted cubic splines with four knots were applied to flexibly model the association without assuming linearity.

Long-term survival was described with Kaplan–Meier curves generated with the survival and survminer packages. Time-to-event was calculated from ICU admission to the date of death or last known follow-up. Curves were stratified by high- and low-risk groups according to ROC-derived thresholds. Survival analyses were exploratory and were not used to generate adjusted hazard estimates.

To assess incremental predictive performance, we calculated the net reclassification improvement (NRI) and integrated discrimination improvement (IDI) using the nricens package in R, comparing clinical scores and biomarkers. Calibration was evaluated using the Brier score, calculated as the mean squared error between predicted probabilities and observed outcomes at each timepoint. To account for class imbalance, the area under the precision–recall curve (AUPRC) was also computed for each predictor and outcome.

Results

During the study period from 2017 to 2024, 587 nonagenarian and centenarian patients were admitted to ICUs after TAVR across Australia and New Zealand. Of these, 562 underwent their TAVR procedure in private hospitals (Supplementary Table 1). All included patients had complete data for clinical scores, biochemical markers, and outcome measures. Clinical scores, including APACHE III, SOFA, GCS, and frailty, were recorded at ICU admission, whereas biochemical markers were captured as the highest values within the first 24 h after ICU admission (Table 1).

Table 1 Baseline characteristics.
Full size table

Baseline characteristics of the patients

As demonstrated in Table 1, a total of 587 nonagenarian patients who underwent TAVR were analyzed: 195 who died during follow-up and 392 who survived. The median age was similar across groups (92.1 years; P = 0.680). Compared with survivors, non-survivors had longer hospital (190 vs. 137 h; P < 0.001) and ICU (24 vs. 23 h; P = 0.017) stays, and exhibited greater illness severity, with higher APACHE III (56 vs. 53; P < 0.001) and SOFA (3 vs. 2; P < 0.001) scores.

Compared to survivors, non-survivors had higher serum lactate (1.05 vs. 0.90 mmol/L; P < 0.001), glucose (7.05 vs. 6.60 mmol/L; P < 0.001), creatinine (98 vs. 89.5 µmol/L; P = 0.001), and urea (8.60 vs. 7.70 mmol/L; P = 0.001), and lower albumin (31 vs. 32 g/L; P < 0.001). Lower minimum mean arterial pressure (65 vs. 66 mmHg; P = 0.004) and bicarbonate levels (23.3 vs. 24.0 mmol/L; P = 0.008) were also observed. Mortality was associated with higher rates of delirium (8.0% vs. 1.2%; P < 0.001). No significant differences between survivors and non-survivors were noted in frailty scores, sex, or chronic comorbidities.

Among electively admitted patients, mortality remained low across early time points, with ICU, hospital, 1-month, 6-month, and 1-year mortality of 0.2%, 0.6%, 1.5%, 4.6%, and 8.1% respectively (Supplementary Table 2).

Early derangements in metabolic, renal, and hemodynamic parameters were more pronounced among non-survivors than survivors, thereby reflecting greater physiological compromise during the initial 24 h after ICU admission.

Effects of clinical scores in predicting mortality

Both APACHE III and SOFA scores demonstrated excellent discrimination for short-term outcomes (Table 2 and Fig. 1). For ICU mortality, the AUROC was 0.993 (95% CI, 0.980–1.000) for APACHE III and 0.969 (95% CI, 0.938–1.000) for SOFA. Hospital mortality was similarly well predicted, with AUROCs of 0.916 (95% CI, 0.834–0.998) for APACHE III and 0.922 (95% CI, 0.859–0.985) for SOFA. The discriminative performance declined for longer-term outcomes: the AUROCs for APACHE III and SOFA were 0.749 (95% CI, 0.604–0.894) and 0.723 (95% CI, 0.563–0.883), respectively, for 1-month mortality; 0.672 (95% CI, 0.574–0.770) and 0.669 (95% CI, 0.572–0.765), respectively, for 6-month mortality; and 0.641 (95% CI, 0.559–0.722) and 0.635 (95% CI, 0.554–0.715), respectively, for 1-year mortality. DeLong’s test indicated no statistically significant differences in AUROC between scores at any timepoint (e.g., P = 0.2746 for ICU mortality; P = 0.7549 for hospital mortality; P = 0.4508 for 1-month mortality), thus indicating comparable performance (Table 3). Clinically, both scores were effective for early mortality risk stratification, particularly in the ICU and hospital settings.

Table 2 Discrimination and calibration metrics of clinical scores for predicting mortalities.
Full size table
Fig. 1
Fig. 1
Full size image

AUROC curves and values of clinical scores and biomarkers at various timepoints.

Table 3 Delong’s test: pairwise comparisons of discrimination (AUROC) between clinical scores at various mortality timepoints.
Full size table

Effects of biochemistry in predicting mortality

Several biochemical markers measured within the first 24 h after ICU admission demonstrated moderate discriminative ability for ICU and hospital mortality (Table 4 and Fig. 1). Glucose showed the strongest predictive performance for ICU mortality (AUROC: 0.858; 95% CI, 0.800 to 0.907), and was followed by potassium (0.795; 95% CI, 0.453 to 0.985), lactate (0.771; 95% CI, 0.360 to 0.990), and nadir mean arterial pressure (0.745; 95% CI, 0.471 to 0.997). For hospital mortality, lactate remained highly predictive (AUROC: 0.860; 95% CI, 0.708 to 0.980), and glucose (0.843; 95% CI, 0.723 to 0.935) and PaO₂ (0.776; 95% CI, 0.641 to 0.884) also performed well. In contrast, albumin, bilirubin, and creatinine-based metrics showed lower AUROCs, typically below 0.70. Similarly to the clinical scores, the discriminative performance of all biomarkers declined progressively over longer timepoints, and the AUROCs for most markers fell below 0.65 by 6 and 12 months.

Table 4 Discrimination and calibration metrics of biomarker scores for predicting mortalities.
Full size table

DeLong’s test comparisons (Supplementary Table 3) did not identify a consistently superior biomarker across all outcomes after Bonferroni alpha correction. However, in the ICU setting, glucose and potassium significantly outperformed several others, including creatinine, urea, and bicarbonate. For example, glucose had a significantly higher AUROC than high creatinine (P = 0.0043), and potassium outperformed urea (P = 0.0070).

Net reclassification improvement analyses comparing APACHE III with SOFA, glucose, and potassium are provided in Supplementary Tables 4, 5, 6. AUPRC and Brier scores are provided in Supplementary Tables 7, 8. A Kaplan–Meier survival comparison between elective and non-elective TAVR cases is presented in Supplementary Fig. 3.

Effects of clinical scores in predicting ICU and hospital outcomes

Patients classified as high-risk according to the APACHE III score had significantly higher odds of 1-month mortality compared to low-risk patients (unadjusted OR, 10.20; 95% CI, 3.28–32.60; P < 0.001), and this association remained strong after multivariable adjustment (adjusted OR, 10.10; 95% CI, 3.22–32.50; P < 0.001) (Table 4). For 1-year mortality, the adjusted odds also remained significant (adjusted OR, 3.85; 95% CI, 1.97–57.26; P < 0.001). A similar pattern was observed with SOFA-defined high-risk classification, which was associated with increased odds of 1-month mortality in both univariate (OR, 4.18; 95% CI, 1.34–15.60; P = 0.019) and multivariable models (adjusted OR, 4.88; 95% CI, 1.55–18.44; P = 0.009). For 1-year mortality, SOFA-defined high-risk status also remained significantly associated after adjustment (adjusted OR, 2.70; 95% CI, 1.54–4.79; P < 0.001). Additionally, when treated as continuous measures, both APACHE III and SOFA scores were significantly associated with 1-month and 1-year mortality, with adjusted odds ratios indicating an increased risk per unit increase in score (Table 5). Each 1-point APACHE III score increase was associated with a 6% increase in the odds of 1-month mortality (OR, 1.06; 95% CI, 1.03 to 1.10; P = 0.001) and a 4% increase for 1-year mortality (OR, 1.04; 95% CI, 1.02 to 1.07; P < 0.001), after adjustment for age, sex, comorbidities, and elective admission (Table 6). For the SOFA score, each additional point was associated with 26% increase in the odds of mortality (OR, 1.26; 95% CI, 1.11 to 1.43; P < 0.001). Kaplan–Meier survival curves (Fig. 2) further illustrate the clear separation in survival between high- and low-risk groups defined by both scores, reinforcing their independent prognostic value.

Table 5 Univariate and multivariate odds ratios for 1-month and 1-year mortality adjusted for demographics, comorbidities, elective admission status and severity scores.
Full size table
Table 6 Clinical scores associated with 1-month and 1-year mortality: multivariate logistic regression.
Full size table
Fig. 2
Fig. 2
Full size image

Kaplan-Meier curve of low vs. high risk clinical scores.

Effects of biomarkers in predicting ICU and hospital outcomes

Several biochemical markers measured in the first 24 h after ICU admission were associated with mortality risk (Table 7). Elevated lactate was strongly predictive of 1-month mortality (adjusted OR, 4.38; 95% CI, 1.28 to 17.20; P = 0.001), and the association persisted at 1 year (adjusted OR, 1.91; 95% CI, 1.06 to 3.44; P = 0.031). Similarly, high glucose was associated with elevated odds of death at both 1 month (adjusted OR, 4.83; 95% CI, 1.40 to 19.10; P = 0.015) and 1 year (adjusted OR, 2.24; 95% CI, 1.30 to 3.85; P = 0.040). Urea showed a consistent relationship with mortality, including an adjusted OR of 3.67 (95% CI, 1.21 to 12.30; P = 0.028) at 1 month and 2.49 (95% CI, 1.07 to 5.01; P = 0.029) at 1 year. Low bicarbonate and diminished PaO₂ were also associated with elevated 1-year mortality (adjusted ORs, 0.25; 95% CI, 0.07 to 0.84; P = 0.030 and 0.45; 95% CI, 0.20 to 1.00; P = 0.009, respectively). Kaplan–Meier survival curves (Figs. 3 and 4) further illustrate the clear separation in survival between high- and low-risk groups defined by biomarkers, reinforcing their independent prognostic value.

Table 7 Univariate and multivariate odds ratios for 1-month and 1-year mortality adjusted for demographics, comorbidities, elective admission status and severity scores.
Full size table
Fig. 3
Fig. 3
Full size image

Kaplan-Meier curve of low vs. high risk biomarkers.

Fig. 4
Fig. 4
Full size image

Kaplan-Meier curve of low vs. high risk biomarkers.

Due to clear evidence of non-linearity, adjusted odds ratios for each biochemical marker were further explored using restricted cubic spline models, adjusting for age, sex, and elective surgical admission. Overall, the spline plots revealed diverse and clinically relevant non-linear risk patterns, including clear reversal or peaking effects for some markers. For 1-month mortality (Fig. 5), lactate showed a steep, monotonic increase in risk above approximately 2 mmol/L, rising sharply beyond 3 mmol/L. Glucose and urea demonstrated progressive upward trends, with risk increasing above about 7–8 mmol/L for glucose and near 8–10 mmol/L for urea. Markers such as heart rate, potassium and bicarbonate displayed typical U-shaped curves, with mortality risk higher at both lower and higher values relative to the median, indicating bidirectional risk. In contrast, creatinine demonstrated an inverted U-shaped association: the odds of death were highest at intermediate creatinine levels, around 120 to 150 µmol/L, but decreased at both lower and higher extremes, suggesting a peak risk within the mid-range rather than at the tails. Albumin showed a predominantly inverse pattern at 1 month, with lower levels clearly associated with increased risk, but without a strong reversal at higher levels in the short term. Long term mortality trends for each biomarkers are illustrated further in Fig. 6.

Fig. 5
Fig. 5
Full size image

RCS 1-month odds ratio to biochemistry markers (reference at median).

Fig. 6
Fig. 6
Full size image

RCS 1-year adjusted odds ratio to biochemistry markers (reference at median).

Effects of clinical scores in predicting ICU and hospital length of stay

In multivariable models adjusting for age, sex, and relevant biochemical parameters (Table 8), both APACHE III and SOFA scores, indicators of illness severity at ICU admission, were significantly associated with prolonged ICU length of stay.

Table 8 Clinical scores associated with ICU LOS and hospital LOS.
Full size table

Each point increase in the APACHE III score was associated with a 1% longer ICU stay (β = 1.01; 95% CI, 1.00–1.02; P = 0.003) and hospital stay (β = 1.01; 95% CI, 1.00–1.02; P = 0.007). The SOFA score was more strongly associated with ICU stay (7% increase per point; β = 1.07; 95% CI, 1.02–1.11; P = 0.002) but was not significantly linked to hospital duration. These results suggest APACHE III offers more consistent predictive value across the inpatient course, supporting its use in resource planning and patient discussions.

Effects of biochemistry markers in predicting ICU and hospital length of stay

Given evidence of non-linearity, restricted cubic spline models adjusted for age, sex, and surgical status were used to explore associations between biochemical markers and length of stay (Figs. 7 and 8). For ICU stay, lactate, glucose, and urea were positively associated with longer durations, while albumin showed an inverse trend. Bilirubin, bicarbonate, and potassium followed U-shaped patterns. For hospital stay, similar trends were observed, with urea showing the most pronounced association. Higher albumin and mean arterial pressure predicted shorter admissions.

Fig. 7
Fig. 7
Full size image

ICU LOS to biochemistry markers.

Fig. 8
Fig. 8
Full size image

Hospital LOS to biochemistry markers.

Discussion

Key findings

In this large binational cohort of nonagenarian ICU patients undergoing TAVR, we identified multiple early predictors of mortality and resource use. Higher illness severity at ICU admission, as measured with the APACHE III and SOFA scores, was strongly associated with ICU and hospital mortality, and had excellent discrimination for short-term outcomes (AUROC > 0.90) but attenuated predictive value at the 6- and 12-month follow-up. Mortality was also independently associated with several routinely measured biochemical markers, including elevated lactate, glucose, and urea levels; however, their discriminative performance similarly declined over time.

Notably, each point increase in both APACHE III and SOFA scores was associated with rising odds of death, and the strongest association was observed for the APACHE III score at 1 month. Although many biomarkers were associated with adverse outcomes in the unadjusted analyses, only lactate, glucose, urea retained significance after adjustment. While absolute differences in biomarker values between survivors and non-survivors were small, their prognostic relevance became apparent through regression modelling and risk stratification rather than magnitude alone, reflecting the reduced physiological reserve in this population. Both clinical scores were associated with prolonged ICU length of stay, although only APACHE III predicted total hospital length of stay.

For ICU length of stay, higher lactate and urea levels were strongly associated with progressively longer stays, while lower albumin concentrations were consistently linked to prolonged ICU admission, emphasizing the role of metabolic stress and nutritional status in intensive care utilization. In contrast, for total hospital length of stay, elevated lactate, higher glucose and urea levels, and lower albumin remained important predictors of extended admission duration, while vital signs and acid–base markers, including increased heart rate, lower mean arterial pressure, and deviations in bicarbonate, demonstrated modest but clinically relevant associations with longer hospitalization. These patterns highlight that both metabolic derangements and subtle hemodynamic or acid–base abnormalities contribute meaningfully to ongoing inpatient care requirements beyond the ICU setting.

Relationship of findings to the literature

Our findings built on those from prior studies assessing mortality prediction after TAVR in older patients. In a 2016 study in octogenarians, the SOFA score showed AUROCs of 0.74 and 0.80 for predicting in-hospital and 30-day mortality, respectively, whereas the APACHE II score demonstrated an AUROC of 0.88 across both outcomes19. In our nonagenarian cohort, the SOFA score achieved even higher discrimination for these short-term endpoints, although the performance declined with increasing follow-up duration. The APACHE III score, which has not been widely evaluated in prior TAVR populations, demonstrated a superior AUROC for in-hospital mortality but fell below the AUROC reported for the APACHE II score at 30 days, thus suggesting potential differences in calibration between score versions.

This study, to our knowledge, is the first to assess lactate’s discriminatory performance via AUROC analysis in nonagenarian’s post-TAVR. Lactate demonstrated strong short-term predictive utility, particularly for in-hospital mortality, in agreement with its established role in acute circulatory failure20. However, the marked decline in its predictive value at 6 and 12 months indicated its limited long-term prognostic ability.

In contrast, resting heart rate exhibited poor discrimination for mortality across all timepoints, in agreement with previous findings indicating no association between heart rate and 2-year post-TAVR outcomes in octogenarians21. Similarly, although augmented mean arterial pressure has been proposed as a mid-term mortality predictor22, our analysis showed that this parameter had only modest short-term predictive value and no meaningful association beyond hospital discharge. These results align with studies indicating no link between mean arterial pressure and functional improvement after TAVR23.

Despite prior evidence linking acute kidney injury to poor post-TAVR outcomes24,25,26,27, we found that creatinine—used herein as a proxy marker—exhibited weak predictive ability at all timepoints. This finding probably reflected confounding from age-related variations in muscle mass, hydration, and nutritional status, which disproportionately affect creatinine levels in nonagenarians28.

Furthermore, our study supported previous findings associating elevated glucose with increased short-term mortality29,30,31,32, thereby reinforcing the importance of metabolic derangement in acute-phase risk stratification. Although the long-term implications of glucose dysregulation post-TAVR remain debated, our results suggest a potential role of glucose in early mortality prediction.

In contrast to a prior ICU-based TAVR study in a younger cohort (median age 82 years), which indicated no significant difference in hospital length of stay between survivors and non-survivors (median 10.0 vs. 12.0 days; P = 0.467), and only a borderline difference in ICU stay (median 4.0 vs. 8.0 days; P = 0.063)16, our findings demonstrated a clearer and statistically significant association between higher illness severity scores and prolonged ICU and hospital stay. Both the APACHE III and SOFA scores were independently associated with ICU duration, and the APACHE III score additionally predicted hospital length of stay. These results suggested that, compared to a younger patient cohort, in a more physiologically vulnerable nonagenarian population, early severity measures might more accurately reflect the complexity of care and recovery trajectory.

Prior studies have identified hypoalbuminemia as a predictor of prolonged ICU and hospital stay in critically ill populations33. In agreement with those findings, our study indicated that lower albumin levels were independently associated with longer hospitalization after TAVR in nonagenarians. In contrast, insufficient published evidence has examined the association between other routinely measured biochemical markers, such as lactate, bilirubin, glucose, or urea, and ICU or hospital length of stay specifically in TAVR populations. Our study contributes novel insights in this area, by identifying bilirubin as an independent predictor of ICU and hospital duration and highlighting the broader potential of early biochemical derangements in informing post-procedural care trajectories.

These results are consistent with evolving international consensus on the role of severity scores and biochemical markers in prognostication in critically ill populations, reinforcing their utility even in the oldest-old after TAVR34.

Clinical implications

This study provides timely and practical guidance for ICU clinicians managing patients who have undergone TAVR. In a clinical environment often marked by prognostic uncertainty, our findings offer a data-driven approach to early post-admission risk stratification, grounded in routinely collected clinical and biochemical parameters.

The APACHE III and SOFA scores, measured on ICU admission, were independently associated with both short-term mortality and ICU length of stay. These scores can assist clinicians in identifying patients likely to require intensive support, to experience slow recovery, or face high risk of early deterioration. Incorporating these objective scores into early ICU assessments can help guide the intensity of therapy, prompt appropriate monitoring, and inform conversations regarding goals of care with patients and families.

Routinely measured biochemical markers such as elevated lactate, glucose, and urea, and lower bicarbonate, demonstrated independent associations with mortality or prolonged hospitalization in regression analyses. Although the absolute differences in these biomarker values between survivors and non-survivors were small, their prognostic relevance emerged when examined within multivariable models rather than through raw group comparisons alone. These laboratory values, which are already part of standard ICU workflows, may therefore contribute to early risk assessment when interpreted alongside clinical context and severity scores. However, the modest absolute differences observed indicate that no clear clinical cut-off values can be defined from this cohort, and these markers should be viewed as supportive indicators rather than standalone triggers for escalation or goals-of-care discussions. Accordingly, the observed associations should not be interpreted as sufficient to mandate changes to bedside management in isolation. At present, these findings are most appropriate for prognostic enrichment and future model development, pending prospective validation and assessment of clinical impact.

Clinicians can improve practice by integrating these predictors into structured decision-making frameworks rather than relying solely on subjective assessment. This aspect is particularly valuable in older adults, in whom baseline frailty and atypical presentations can complicate clinical impressions. Although our findings are not intended to guide pre-procedural decisions, they are highly relevant for early post-operative risk reassessment and might inform multidisciplinary discussions regarding the likely trajectory of ICU care, including rehabilitation potential or transition to palliative pathways.

Although frailty is an established prognostic factor following TAVR, we observed no significant association with outcomes in this study. This likely reflects the narrow distribution of frailty scores4,5,6, driven by strict patient selection and the near-universal presence of some degree of frailty among nonagenarians and centenarians. Clinically, this suggests that conventional frailty measures may have limited discriminatory value in ultra-elderly cohorts undergoing TAVR, and more nuanced or multidimensional approaches to assessing vulnerability may be required to improve risk stratification in this population.

Our findings extend prior studies in octogenarians by focusing on nonagenarians and centenarians, a subgroup in whom existing risk models may be less applicable. STS and EuroSCORE were not designed or validated in this age group and may underestimate risk, while APACHE II lacks the granularity of more contemporary ICU scores. By demonstrating that APACHE III, SOFA, and simple biochemical markers provide excellent short-term discrimination in this population, our study highlights the need for age-specific prognostic approaches to support clinical decision-making in the era of expanding TAVR use among the very elderly.

Our study, as one of the largest and most detailed evaluations of nonagenarian patients in ICUs following a uniform surgical pathway, offers a practical model for translating early post-admission data into tailored clinical management. It supports the delivery of evidence-informed, person-centered care in the ICU, where nuanced prognostication and resource stewardship are increasingly essential.

Directions for future research

Future studies should validate these findings in broader ICU cohorts and develop longitudinal models that incorporate serial changes in physiological and biochemical parameters beyond the first 24 h. The limited performance of creatinine and albumin may reflect sarcopenia, malnutrition, inflammation, and perioperative fluid balance in nonagenarians, while reliance on single measurements likely underestimated the prognostic value of dynamic trends such as lactate clearance or glucose variability. Studies exploring the integration of frailty assessments, pre-morbid functional status, and patient-centered outcomes might further enhance risk stratification and decision-making in the oldest populations.

Additionally, qualitative research involving clinicians and patients’ families could help elucidate how these predictive tools influence real-time decisions regarding care escalation and de-escalation. Because this study focused on patients who underwent a uniform procedural intervention, future investigations should also examine the generalizability of these predictors to non-surgical or mixed ICU cohorts. Prospective studies incorporating risk-based interventions will be essential to determine whether early stratification based on clinical and biochemical data can improve outcomes, optimize resource use, and support value-aligned care in the critically ill older patients. For some biomarkers, such as albumin, the lowest value is clinically more relevant than the highest. As the database captured only the highest values, this may have attenuated the observed prognostic associations. Future studies should incorporate both nadir, preoperative risk scores and peak biomarker values to more comprehensively characterize risk in this population. Finally, although several biomarkers demonstrated statistically significant differences between outcome groups, the absolute magnitude of these differences was small (e.g., albumin 31.0 vs. 32.0 g/L). Such modest differences are unlikely to yield clinically useful cut-off thresholds in isolation. Future research should explore whether combining biomarkers with clinical scores or evaluating longitudinal biomarker changes may provide more practical and clinically relevant risk stratification tools.

Strengths and limitations

This study provides a rare and comprehensive evaluation of critically ill nonagenarians admitted to intensive care after TAVR and represents one of the largest and most detailed cohorts reported internationally. The use of binational, multicenter data and a uniform procedural cohort enhances external validity within ICU populations. A key strength is the combined assessment of clinical severity scores and routinely collected biochemical markers within the first 24 h of ICU admission, offering prognostic tools that are both accessible and directly applicable to clinical practice. The study also examined a broad range of outcomes, including short and longer term mortality and healthcare utilization, providing insight into both early risk and recovery trajectories. Methodologically, the analyses applied appropriate modelling strategies, incorporated both AUROC and AUPRC metrics, and included checks for model assumptions.

Several limitations should be acknowledged. Frailty was assessed using the Clinical Frailty Scale, but more detailed measures of functional capacity, cognition, and quality of life were unavailable. Because the cohort includes only patients who survived to ICU admission after TAVR, survivorship bias is possible, and early periprocedural mortality may be underestimated. Residual confounding is also likely, as premorbid functional status, detailed frailty indices, and treatment limitation data were not available. As with all retrospective observational studies, associations identified should be interpreted as prognostic rather than causal.

Device specific information was not captured, and we were unable to account for temporal improvements in TAVR technology, which may influence outcomes across the study period. Echocardiographic parameters, including left ventricular systolic function before and after TAVR, were also unavailable, despite their known prognostic relevance. Future work incorporating echocardiography alongside clinical and biochemical variables may help refine risk stratification in this population.

The registry captures laboratory values during the first 24 h after ICU admission but does not provide serial biomarker measurements, peri-procedural or immediate post-TAVR values, or repeated sampling across the early postoperative course. As a result, we could not evaluate biomarker trajectories such as lactate clearance, glucose variability, or evolving renal dysfunction, which may have greater clinical and prognostic relevance than single time-point peak values. Future studies incorporating repeated measurements across peri-procedural, immediate postoperative, and early ICU phases are needed to assess whether dynamic trends improve risk stratification in this population.

External validity is further influenced by regional factors. Although the ANZICS registry has near complete ICU coverage, most cases originated from Australian centers, and differences in admission thresholds, case mix, and coding practices may limit generalizability. In addition, thresholds derived from ROC analyses were not externally validated and should be applied cautiously in other settings.

Given the evaluation of multiple predictors across several mortality endpoints and model specifications, there is potential for Type I error inflation. Although the primary analyses and key hypotheses were prespecified and modelling safeguards were applied, some statistically significant associations may reflect chance rather than true effects. Accordingly, p-values should be interpreted cautiously and in conjunction with effect sizes, confidence intervals, and consistency of findings across analyses, particularly for secondary and exploratory outcomes. External validation in independent cohorts is required to confirm the robustness of these associations.

Finally, because the registry does not capture early postprocedural complications such as stroke, major bleeding, or conduction abnormalities, and because the prognostic variables assessed reflect population-level patterns rather than real-time clinical trajectories, the findings should be interpreted as providing contextual prognostic insight rather than directive guidance for bedside decision-making.

Conclusion

In this multicenter study of nonagenarians admitted to intensive care after TAVR, early clinical scores and routinely available biochemical markers demonstrated strong short-term prognostic value and were independently associated with ICU and hospital outcomes. These findings may offer a potential practical framework for early risk stratification that might enable clinicians to individualize care, allocate resources more effectively, and initiate timely goals-of-care discussions. By bridging a critical knowledge gap in geriatric critical care, this study lays the groundwork for future models of ICU practice that are data-informed, age-appropriate, and aligned with patient-centered values.