Frailty assessment and risk prediction by GRACE score in older patients with acute myocardial infarction

Patients aged ≥65 years old were assessed prospectively within the screening registry of a phase II randomized controlled trial of future care planning in advanced heart disease at the Royal Infirmary of Edinburgh between October 2013 and September 2014 (NCT 02302014) [16]. Patients with moderate/severe dementia or other barriers to informed consent were excluded. All patients in the registry with a clinical diagnosis of type 1 myocardial infarction were included, except the fifty patients who underwent a tailored intervention in the phase II study. The registry protocol was approved by a local research ethics committee (reference 12/SS/0223) and the study was conducted in collaboration with a UKCRC (UK Clinical Research Collaboration) registered Clinical Trials Unit.

The CFS is a structured scale of descriptors to guide selection between nine levels ranging from “very fit: 1” to “terminally ill: 9”. Frailty may be assessed as a continuum, but is considered present at a score ≥ 5 (Supplementary Figure 1) [14]. Assessment criteria include activity, symptoms and assistance usually required with personal activities of daily living (e.g. washing and toileting), and instrumental tasks necessary for independent community living (e.g. managing finances and medications). Clinical nursing staff completed the CFS for each study patient based on their professional assessment and any documentation of premorbid functional status. Nurses received training in the use of the tool from the research team.

The GRACE estimated risk of 12-month mortality after myocardial infarction was determined using the online version 2.0 calculator (http://​www.​gracescore.​org/​website/​WebVersion.​aspx) [7]. Clinical and biochemical measures required for the score (heart rate, systolic blood pressure, Killip class, creatinine, cardiac troponin and electrocardiogram changes) were taken from the time of initial presentation with cardiac symptoms. The tool assigns categories based on the calculated risk: low (< 4% estimated 12-month mortality risk), medium (4–12%) and high (> 12%) risk.

Comorbidity was measured using the Charlson comorbidity index, with a higher score indicating greater comorbidity [17]. Functional status and disability were recorded on the Karnofsky scale, with increasing dependency indicated by a lower score in the range of 0–100 [18]. Research staff completed these scales using all available paper and electronic health records (TrakCare; InterSystems Corporation, Cambridge, MA, USA) together with patient or family history.

Electronic health records were used to determine the primary endpoint of all-cause mortality in the 12 months following index admission. Secondary outcomes were length of index hospital stay, completion of cardiac catheterization (with or without percutaneous coronary intervention), hospital readmissions within 12 months and attendance at cardiac rehabililation.

Findings from the primary analysis were also tested in an independent cohort, comprising 96 patients aged ≥65 years old undergoing cardiac catheterization following myocardial infarction at the South Yorkshire Cardiothoracic Centre (Sheffield, UK), a tertiary referral centre for a population of 1.8 million people in the North of England. GRACE and CFS scores were available for all participants. The recruitment and data collection within this cohort has been previously described in detail [19, 20].

Continuous data are presented as means ± SD or median ± IQR and where appropriate compared by Student’s t-test, Mann-Whitney U-test or Analysis of Variance (ANOVA). Categorical data are presented as absolute numbers and percentages and compared by Chi-squared test. Logistic, linear and Cox proportional hazards regression modelling were used to determine predictors of the primary and secondary outcomes. Differences between frailty groups in survival analysis were assessed by log rank test. Receiver operating characteristic (ROC) curve analysis was performed by standard methods for discrimination of 12-month mortality. Model fit was assessed by Akaike and Bayesian Information Criteria (AIC and BIC respectively). Coefficients derived from a multiple logistic regression model including both GRACE and CFS scores from the study population were applied to the exernal validation cohort.

To calculate Net Reclassification Improvement (NRI), each patient was assigned one of three risk categories from the GRACE calculator output (low, medium or high risk). Using the multiple logistic regression model including both GRACE and CFS, all patients were reclassified and reported against the same GRACE risk thresholds. The analysis was performed separately in those who died and survived to assess the net reclassification of patients, thereby accounting for both appropriate and inappropriate reclassifications. This was calculated as an unweighted or ‘dimensionless NRI’. All analyses were completed with R (version 3.3.3). NRI calculations were completed using the pROC package [21].

The CFS identified 40 (20%) patients with frailty defined by a CFS score ≥ 5 (Fig. 1). Using this established CFS threshold, frail patients were older, more often female and experienced greater comorbidity (mean Charlson Comorbidity Index 3.9 ± 2.2 vs. 2.6 ± 1.6 in those with CFS ≤4, p < 0.001, Table 1). There were notably higher levels of previous heart failure, chronic kidney disease, respiratory illness and dementia amongst frail patients. Age, comorbidity, functional impairment and GRACE estimated mortality risk increased with higher CFS scores (p < 0.001 for all, Fig. 1). Overall, GRACE estimated 12-month mortality was greater in frail patients (28% vs. 16% in those with CFS ≤4, p < 0.001).

Frail patients were prescribed more medications before hospital admission (10.5 ± 2.9 vs. 8.9 ± 3.1 with CFS ≤4, p = 0.003). Index hospitalization was also longer (11 days [IQR 5–24] vs. 4 days [IQR 3–7] with CFS ≤4, p < 0.001), but cardiac catheterization was undertaken less frequently in frail patients (33% vs. 81% with CFS ≤4, p < 0.001, Table 2). Observed mortality was higher in frail patients, with 15% dying during the index hospitalization and nearly half by 12 months (48% vs. 9% with CFS ≤4, p < 0.001). Thus, frailty was associated with spending a shorter period alive and out of hospital in the year following index hospitalization (mean 229 ± 137 vs. 329 ± 68 days with CFS ≤4, p < 0.001). In those that survived to hospital discharge, no patient with CFS ≥5 attended cardiac rehabilitation. In non-frail patients, each point increase in CFS was associated with a reduced likelihood of attendance at rehabiliation, after adjustment for age, sex and baseline comorbidity (adjusted odds ratio [OR] 0.59, 95% confidence intervals [CI] 0.42–0.80 per unit increase in CFS, p = 0.001). However, the CFS did not predict hospital readmissions over the 12 months after index admission following adjustment (adjusted HR 1.10 (0.86–1.40), p = 0.44).

Considering frailty as a continuous variable, each point increase in CFS score predicted a doubling of the hazard of 12-month mortality (HR 1.98, 95% CI 1.60–2.44, p < 0.001), with the effect minimally attenuated after adjustment for age, sex and comorbidity (adjusted HR 1.90, 95% CI 1.47–2.44, p < 0.001, Table 2). The relationship between CFS and mortality remained independent following adjustment for the GRACE score (adjusted HR 1.72, 95% CI 1.37–2.16 per unit increase in CFS, p < 0.001). Similarly the Karnofsky score remained an independent predictor of mortality beyond GRACE (adjusted HR 0.96, 95% CI 0.94–0.98 per unit increase, p < 0.001), but the Charlson comorbidity index did not (adjusted HR 1.21, 95% CI 0.99–1.46, p = 0.06). By linear regression modelling, the number of days alive and out of hospital in the 12 months after index admission decreased by 25 days with each point increase in CFS (95% CI 18–33 days, p < 0.001).

Using the CFS to divide the population into non-frail (CFS 1–3), vulnerable or mildly frail (CFS 4–5) and moderate to severely frail (CFS 6–9) patients demonstrated separation of survival curves throughout the 12 months following index hospitalization (log rank test p < 0.001, Fig. 2).

The area under the ROC curve (AUC) for an outcome of 12-month mortality after myocardial infarction was 0.80 (95% CI 0.71–0.88) for the GRACE score and 0.81 (95% CI 0.72–0.89) for the CFS. These were stronger discriminators than either the Charlson comorbidity index (AUC 0.68, 95% CI 0.58–0.78) or Karnofsy score (AUC 0.76, 95% CI 0.68–0.87, Supplementary Figure 2). Addition of the CFS to GRACE improved the AUC to 0.86 (95% CI 0.78–0.92), which represented a significant gain by DeLong testing (p = 0.04, Fig. 3). Both the AIC and BIC were lower in the model including the CFS score implying improved model fit.

By GRACE estimation, high-risk status (> 12% risk of 12-month mortality) was applied to 112 (57%) patients, medium risk (4–12% risk) to 81 (41%) patients and low risk (< 4% risk) to 5 (3%) patients. Within the high-risk cohort, survival curves demonstrated continued separation by frailty status (log-rank test p < 0.001, Supplementary Figure 3). Addition of the CFS in a multiple logistic regression model with the GRACE score resulted in a NRI of 0.44 (95% CI 0.28–0.60, p < 0.001), with the majority of reclassification events downgrading estimated risk (Supplementary Table 2). Amongst all patients who were still alive at 12 months, 97 (59%) moved to a lower GRACE risk category with the updated model.

Across the study population, the predicted 12-month mortality by GRACE estimate was 18.5%, which compared to an observed mortality of 16.7%. Dividing the population by CFS frailty status identified an overestimation of risk amongst patients with CFS ≤4 (16.2% predicted vs. 8.9% observed), and an underestimation in frail patients (27.8% predicted vs. 47.5% observed with CFS ≥ 5, Supplementary Table 3). Formal calibration testing was not performed due to sample size.

In the external validation cohort, 27 (28%) patients were frail (CFS ≥5). There were 14 (15%) deaths within 12 months. After applying the multiple logistic regression model coefficients derived in the main study to this external validation cohort, the AUC for 12-month mortality was 0.75 (95% CI 0.60–0.87), although this was not a significant improvement on GRACE alone (p = 0.40, Supplementary Figure 4). By NRI, applying the main study model improved classification in the validation cohort (0.46, 95% CI 0.23–0.69, p < 0.001), once more through reductions in estimated risk using the model including CFS (Supplementary Table 4).