Modified frailty index versus risk analysis index for 30-day mortality after noncardiac surgery: a systematic review and meta-analysis

1 Introduction

Frailty is a state of heightened vulnerability to stressors, but it is not represented by a single universally accepted measurement model. Previous systematic reviews have associated frailty with adverse postoperative outcomes across surgical populations (1, 2). The physical phenotype operationalizes weakness, slowness, low activity, exhaustion, and weight loss, whereas the deficit-accumulation model summarizes the proportion of health deficits present (3–5). These approaches overlap clinically but are conceptually distinct, and neither should be conflated with multimorbidity alone (3, 6).

The modified Frailty Index (mFI) was developed within the deficit-accumulation tradition by mapping available surgical-registry variables to deficits and has been widely used in surgical outcomes research (4, 7, 8). Its compact versions, particularly mFI-5, are readily derived from routinely collected data. That economy has an important limitation: the retained content is heavily weighted toward comorbidity and functional dependence. Haren et al. argued that such reduced indices may measure multimorbidity or general vulnerability more directly than the multidomain frailty syndrome (9). This construct-validity concern is substantive. The present review does not treat mFI as a criterion-standard or comprehensive measure of frailty and does not attempt to revalidate it as such.

The Risk Analysis Index (RAI) was developed for surgical risk assessment and combines age, function, cognitive or living-context information, nutrition or weight loss, malignancy, dyspnea, and selected comorbidities, with different operationalizations across clinical and administrative versions (10, 11). The Clinical Frailty Scale (CFS) is a global clinical judgment scale based on overall fitness and frailty (12). Validation of any of these instruments for frailty-related classification or surgical risk stratification does not imply equivalent discrimination for a specific outcome. Construct validity and outcome discrimination are distinct measurement properties.

Thirty-day mortality provides a common, clinically consequential outcome for evaluating rank discrimination. Area under the receiver operating characteristic curve (AUC) quantifies whether a tool assigns higher risk to patients who experience the outcome than to those who do not; it does not establish construct validity, calibration, causal effects, or clinical net benefit (13–15).

We therefore used a two-level evidence framework. First, separate tool-specific meta-analyses characterized average discrimination and its distribution across the available independent surgical cohorts. Second, studies evaluating mFI and RAI in the same underlying population supplied the primary comparative context. Restricting eligibility to head-to-head studies would discard substantial evidence about absolute discrimination and transportability across surgical settings; conversely, separate pooled AUCs across different case mixes cannot serve as a formal indirect superiority test. We also retained CFS as secondary, post hoc exploratory evidence.

Our objective was to estimate the average standalone discrimination of mFI and RAI for 30-day all-cause mortality across independent noncardiac surgical cohorts and to interpret comparative performance primarily through dual-tool studies. Secondary objectives were to characterize heterogeneity and data quality, test sensitivity to variance provenance, overlap, risk of bias, and version dependence, and provide an explicit exploratory synthesis of CFS.

2 Methods

2.1 Protocol, reporting, and review scope

This systematic review and meta-analysis followed PRISMA 2020 guidance and the prospectively registered PROSPERO protocol (CRD420261378778) (16). Database-specific strategies, dates, and screening records are reported in Supplementary Appendix 1. Eligibility was not restricted to head-to-head studies; single-tool and dual-tool evidence were retained because they address complementary questions.

2.2 Eligibility criteria and two-level evidence framework

Eligible studies enrolled adults undergoing noncardiac surgery, assessed mFI (mFI-5 or mFI-11), RAI (original RAI, RAI-rev, or RAI-A), or CFS preoperatively or from baseline variables available before surgery, and reported 30-day all-cause mortality. Cohort and registry studies were eligible. Studies using only postoperative or intensive-care frailty assessment, cardiac surgery, severe burns, major polytrauma, noncomparable mortality windows, or only a frailty-containing multivariable model without a standalone tool estimate were not eligible for quantitative AUC synthesis.

Level 1 comprised separate tool-specific syntheses of standalone AUCs. Its purpose was to characterize the distribution and average discrimination of each instrument across independent cohorts, including evidence from studies evaluating only one tool. Level 2 comprised studies evaluating mFI and RAI within the same underlying study population. Single-tool evidence informed absolute discrimination and transportability; dual-tool evidence informed comparative interpretation.

The separately pooled tool-specific AUCs were intended to characterize overall discrimination across the available independent cohorts and were not interpreted as a formal indirect test of superiority. Comparative inference was therefore supported primarily by studies evaluating both instruments within the same study population.

2.3 Study selection and source verification

Two reviewers screened records and assessed full texts against prespecified criteria, resolving disagreements by consensus. Study-level data were reverified against the original source reports before the final quantitative synthesis. Quantitative inclusion required a source-confirmed standalone 30-day mortality AUC or c-statistic with a reported confidence interval or standard error, or exact sample size and mortality events sufficient for the specified variance approximation. The study selection process is summarized in Figure 1.

2.4 Data extraction, case mix, and cohort independence

The extraction captured study identity, surgical population and specialty, data source and years, sample size, mortality events when reported, instrument family and version, standalone AUC and uncertainty, direct-comparison information, variance provenance, overlap group, and QUIPS judgments. Surgical population, specialty, registry or health-system source, and study years were retained to characterize prognostic case mix and spectrum effects.

Each quantitative observation was mapped to an independent patient cohort. Nonoverlapping procedure cohorts within a publication could remain separate. When multiple versions of one tool family were evaluated in the same patients, the primary analysis used the source-designated or preferred version and did not average correlated AUCs. Alternative-version and cohort-clustered all-version analyses were used to assess sensitivity to this choice.

2.5 Outcome and effect measure

The outcome was 30-day all-cause mortality. The primary effect measure was the standalone tool-level AUC (equivalently, c-statistic). AUC was chosen because it was the most consistently available common measure of rank discrimination across heterogeneous populations. It was not used to infer calibration, threshold-specific performance, decision utility, construct validity, or a causal effect of frailty.

2.6 Risk of bias

Risk of bias was assessed with QUIPS across study participation, attrition, prognostic-factor measurement, outcome measurement, study confounding, and statistical analysis/reporting (17). Two reviewers adjudicated domain and overall ratings. For consistency of overall judgments, domain ratings were summarized using the following review-level rule: all domains Low, or only one Moderate outside D3/D4, was classified Low; 2–3 Moderate domains, or one Moderate in D3/D4, was classified Moderate; any High domain or at least four Moderate domains was classified High. Overall risk of bias was Low in 2 studies, Moderate in 52, and High in 23, with Hao 2025 rated High.

For standalone AUCs, case-mix or spectrum variation is conceptually different from confounding of an exposure–outcome association. AUC is a descriptive performance measure rather than a causal effect estimate, so conventional covariate adjustment is not directly analogous to adjustment of an odds ratio or hazard ratio. We nevertheless used the QUIPS Study Confounding domain to assess whether important prognostic structure was adequately addressed and interpreted AUCs in relation to study population and setting.

2.7 Statistical analysis

AUCs were transformed as logit(AUC) = ln[AUC/(1−AUC)] and back-transformed for presentation. For a reported 95% confidence interval wholly within 0 and 1, the logit-scale standard error was [logit(upper)−logit(lower)]/(2 × 1.96). A source-reported raw AUC standard error was converted by the delta method. For the two source-reported Wald intervals crossing 1, a raw-scale standard error was first calculated and then delta-converted. When only an AUC point estimate, exact total sample size, and exact mortality events were available, the Hanley–McNeil variance approximation was used; these records were explicitly labelled as having variance estimated from sample size and events (13).

Separate random-effects meta-analyses for mFI and RAI used restricted maximum likelihood on the logit-AUC scale. The pooled estimator was the inverse-variance random-effects mean. Hartung–Knapp inference used the unmodified q factor, with k−1 degrees of freedom (18, 19). The 95% confidence interval was μ ± t(k−1,0.975) × SEHK. The 95% prediction interval was μ ± t(k−1,0.975) × sqrt(τ² + SEHK²), followed by inverse-logit transformation. I² used the conventional fixed-effect Q and is reported with τ² (20).

No formal difference between the separately pooled mFI and RAI estimates was tested. Direct studies were classified at study level as all eligible RAI versions numerically higher, all lower, mixed/equal, or not numerically classifiable. No paired meta-analysis was conducted because complete within-study covariance or a standard error for the AUC difference was unavailable.

2.8 Sensitivity and exploratory analyses

Sensitivity analyses included: an alternative eligible tool version for same-cohort version-sensitive clusters; a cohort-clustered CR1 sandwich analysis retaining all eligible versions; exclusion of every record whose variance was estimated from sample size and events; exclusion of High overall QUIPS studies; and an overlap-adjusted analysis retaining one representative per overlap group according to the documented selection rule. Subtype analyses were exploratory. These analyses characterize robustness but do not remove the underlying limitations.

CFS was retained as secondary evidence. Eligible CFS observations were pooled post hoc using the same conventional one-estimate-per-cohort method. This analysis was explicitly exploratory and was not used to rank CFS against mFI or RAI.

3 Results

3.1 Study universe and case mix

Seventy-seven studies met the systematic-review criteria (21–96): 29 evaluated both mFI and RAI, 18 evaluated mFI only, 16 evaluated RAI only, and 14 evaluated CFS only. The most common surgical categories were spine (17), general or abdominal surgery (13), orthopedic surgery excluding hip fracture (11), other (9), neurosurgery (7), hip fracture (6), vascular surgery (5), mixed surgery (4), urology (4), and colorectal surgery (1). Table 1 summarizes study classifications, populations, data sources, tool versions, cohort contributions, direct-comparison status, and QUIPS judgments.

Table 1. Characteristics of the included studies.

ACS-NSQIP, American College of Surgeons National Surgical Quality Improvement Program; CFS, Clinical Frailty Scale; ICES, Institute for Clinical Evaluative Sciences; IQR, interquartile range; KPNC, Kaiser Permanente Northern California; mFI, modified Frailty Index; NRD, Nationwide Readmissions Database; RAI, Risk Analysis Index; VASQIP, Veterans Affairs Surgical Quality Improvement Program.

Total participant count was not summed because of substantial overlap among registry-derived cohorts. Sample-size distribution is therefore reported as median, IQR, and range. Age was summarized only where a single parseable mean or median age was available.

Case mix varied materially by specialty, selected population, registry or health system, and calendar period. Among the 26 primary-analysis studies with completed case-mix source audit, comorbidity burden or cancer/procedure composition was reported in 26/26, sex in 25/26, functional dependence in 25/26, ASA or a comparable baseline-risk measure in 20/26, and age in 19/26. For the QUIPS Study Confounding domain, 51 studies were Low and 26 were Moderate; none was High. Across all QUIPS domains, overall risk of bias was Low in 2 studies, Moderate in 52, and High in 23. This distribution, together with the heterogeneous population descriptions in Table 1, supports interpretation of pooled AUCs as averages across a broad prognostic spectrum rather than intrinsic constants of the instruments.

3.2 Level 1: tool-specific pooled discrimination

For mFI, 39 independent cohorts contributed to the primary analysis. The pooled AUC was 0.634 (95% CI, 0.618–0.649), with τ² = 0.027706, I² = 98.8%, and a 95% prediction interval of 0.551–0.709. For RAI, 30 independent cohorts contributed. The pooled AUC was 0.767 (95% CI, 0.737–0.795), with τ² = 0.176481, I² = 99.8%, and a 95% prediction interval of 0.579–0.888 (Table 2; Figure 2). These separate pools describe average discrimination within each available evidence base and are not a formal indirect superiority test.

Table 2. Primary and sensitivity analyses of pooled AUC estimates for mFI and RAI.

AUC, area under the receiver operating characteristic curve; CI, confidence interval; PI, prediction interval; mFI, modified Frailty Index; RAI, Risk Analysis Index; QUIPS, Quality in Prognosis Studies; REML, restricted maximum likelihood.

Cohorts (k) denotes independent cohorts within each tool-specific analysis; AUC estimates denotes the number of tool-version observations. Primary analyses use one preferred version per cohort. Separate tool-specific pools are not a formal between-tool superiority test.

Estimates were pooled on the logit-AUC scale using REML with Hartung–Knapp inference and back-transformed. τ² is on the logit-AUC scale. The clustered all-version analysis instead uses cohort-clustered CR1 sandwich inference; within-cohort covariances are unknown.

Estimated-variance exclusion removes seven mFI and one RAI observations whose variances were approximated from sample size and mortality events. Overlap adjustment retains one representative per overlap group (Supplementary Appendix 3). Low/Moderate QUIPS analyses exclude High-risk studies. Alternative-version analysis substitutes eligible alternative versions in version-sensitive cohorts.

3.3 Level 2: direct mFI-versus-RAI evidence

Twenty-nine dual-tool studies informed direct comparative interpretation. At the study level, 21 had source-valid numeric comparisons in which RAI was higher, none had a source-valid numeric comparison in which mFI was higher, 2 supported RAI only qualitatively, and 6 were not numerically classifiable for the comparable standalone 30-day all-cause mortality outcome. Figure 3 presents the direct evidence; no paired meta-analysis was performed because within-study covariance or standard errors for AUC differences were unavailable.

Twenty-two source-valid numeric comparison rows represented 21 studies because Conlon et al. reported two eligible RAI versions. Sabet CJ et al. (2026) contributed the corrected 30-day mortality pair (N = 11,965; 19 deaths): mFI-5 AUC 0.532 (95% CI, 0.441–0.623) and RAI AUC 0.548 (95% CI, 0.456–0.640), for a within-study difference of +0.016. Mau 2025 and Farid 2026 supported RAI only qualitatively because an exact standalone mFI mortality AUC was not printed. Courville 2024 was not numerically classifiable and was excluded from both quantitative pools because its source endpoint was in-hospital death at the index admission or a readmission within 30 days, while the results table used the composite label mortality or comfort care; this did not meet the quantitative-synthesis criterion requiring a source-confirmed standalone AUC for 30-day all-cause mortality.

3.4 CFS evidence

Fourteen CFS studies were present in the review universe. After exact mortality-window verification, six independent cohorts had source-confirmed standalone 30-day mortality AUCs with usable uncertainty. Narula 2020 was excluded because its AUC of 0.699 (95% CI, 0.651–0.747) was explicitly for one-year mortality and no standalone 30-day AUC was reported. Zaib 2023 was excluded because its AUC of 0.710 (95% CI, 0.650–0.760) was reported for unspecified mortality in a study presenting both 30-day and 2-year outcomes; the ROC itself was not time-labeled and therefore did not satisfy the 30-day source-confirmation criterion used for this quantitative synthesis. In the post hoc exploratory meta-analysis, the pooled CFS AUC was 0.757 (95% CI, 0.618–0.857), with τ² = 0.254567, I² = 77.4%, and a 95% prediction interval of 0.421–0.930. No eligible direct same-cohort CFS-versus-mFI or CFS-versus-RAI comparison was identified for the comparable 30-day discrimination outcome; accordingly, CFS was not formally ranked against either tool.

3.5 Sensitivity and subtype analyses

Results remained directionally consistent across the data-quality and dependence analyses, while interval width and heterogeneity continued to show substantial uncertainty (Table 2). Excluding estimated variances yielded pooled AUCs of 0.632 (95% CI, 0.615–0.649; prediction interval, 0.547–0.710) for 32 mFI cohorts and 0.769 (95% CI, 0.738–0.797; prediction interval, 0.578–0.890) for 29 RAI cohorts. Seven mFI and one RAI records were excluded by that rule; source-reported confidence intervals or standard errors were retained when available.

The overlap-adjusted results were 0.628 (95% CI, 0.607–0.648; prediction interval, 0.535–0.713) for 28 mFI cohorts and 0.751 (95% CI, 0.714–0.786; prediction interval, 0.547–0.883) for 23 RAI cohorts. Restriction to Low or Moderate overall QUIPS studies yielded 0.633 (95% CI, 0.617–0.648; prediction interval, 0.551–0.708) for 37 mFI cohorts and 0.767 (95% CI, 0.736–0.795; prediction interval, 0.577–0.888) for 29 RAI cohorts.

Alternative-version analysis did not change the mFI estimate and yielded an RAI AUC of 0.763 (95% CI, 0.733–0.790; prediction interval, 0.579–0.883). The cohort-clustered all-version analysis yielded 0.634 (95% CI, 0.619–0.649; prediction interval, 0.551–0.709) for mFI and 0.763 (95% CI, 0.730–0.793; prediction interval, 0.574–0.885) for RAI. In a one-study sensitivity analysis addressing the confidence-interval source conflict in Sbaih et al., the abstract-reported intervals [mFI-5 0.570 [95% CI, 0.530–0.610]; RAI 0.790 [95% CI, 0.750–0.830]] were substituted for the primary endpoint-linked main-results intervals [mFI-5 0.570 [95% CI, 0.500–0.630]; RAI 0.790 [95% CI, 0.720–0.860]]. This substitution changed the mFI pooled AUC from 0.634 to 0.633 (95% CI, 0.618–0.649; τ² = 0.028431; I² = 98.8%; prediction interval, 0.550–0.710) and left the rounded RAI pooled AUC at 0.767 (95% CI, 0.737–0.795; τ² = 0.175691; I² = 99.8%; prediction interval, 0.579–0.887). This did not alter the interpretation. Complete subtype estimates are provided in Supplementary Appendix 4 and are exploratory, particularly where k was small.

4 Discussion

4.1 Principal findings and two-level interpretation

This synthesis found higher average 30-day mortality discrimination for RAI than for mFI across the available independent surgical cohorts. That pattern was also observed in the direct evidence: 21 of 29 dual-tool studies had source-valid numeric comparisons in which RAI was higher, none had a source-valid numeric comparison in which mFI was higher, 2 supported RAI only qualitatively, and 6 were not numerically classifiable. The direct same-study evidence is the comparative anchor because both instruments are evaluated in the same underlying population. The separate tool-specific pools instead describe average performance and transportability across each tool's evidence base.

This distinction is essential. Surgical specialty, baseline risk, emergency or elective composition, instrument version, registry coding, missingness, and analytic specification differ across studies. Consequently, the difference between the pooled AUCs is not a formal indirect superiority test. Same-study comparisons reduce major between-study case-mix differences, but they do not eliminate differences caused by missing data, version selection, or analysis specification, and the absence of paired covariance precluded a formal paired meta-analysis.

4.2 Construct validity is distinct from prognostic discrimination

An important conceptual consideration is that the mFI label should not be interpreted as evidence that the instrument fully captures the multidimensional frailty construct. mFI arose from deficit accumulation and has pragmatic value in registry research (4, 7, 8), but its reduced content is comorbidity-heavy. Haren et al. argued that many such surgical indices more closely represent multimorbidity than multidimensional frailty (9). Vetrano et al. likewise showed that frailty and multimorbidity commonly coexist but are distinct (6). The Fried phenotype and a standard deficit-accumulation index capture different conceptual traditions (3–5).

Our analysis neither adjudicates whether mFI is a definitive frailty measure nor revalidates RAI or CFS as frailty constructs. We do not use predictive performance to validate the frailty construct. The present analysis evaluates prognostic discrimination for one common outcome. The observed difference in prognostic discrimination is compatible with the possibility that broader vulnerability content contributes additional prognostic information, but the present meta-analysis was not designed to establish construct validity or identify the mechanism underlying the difference.

4.3 Why single-tool evidence and AUC remain informative

A head-to-head-only review would answer the direct comparative question but would discard studies that inform absolute discrimination and transportability across specialties and data environments. Our two-level framework preserves both uses without conflating them. Single-tool cohorts characterize how widely performance varies in practice; dual-tool studies provide the more internally comparable evidence for relative performance.

Similarly, prior validation does not imply that two tools have equivalent discrimination for 30-day mortality. AUC measures rank discrimination for this outcome. It does not assess calibration, thresholds, net benefit, workflow burden, or construct validity (13–15). Tool selection therefore cannot be made from AUC alone, but comparing source-valid standalone AUCs remains clinically meaningful when the estimand and limits are explicit.

4.4 Heterogeneity, case mix, and mitigation

Heterogeneity was very high for both primary pools and cannot be considered resolved. The broad prediction intervals show that discrimination expected in a future population may differ substantially from the average. Likely contributors include specialty, baseline mortality risk, emergency versus elective composition, tool version, registry definitions, coding, and health-system context. AUC also varies with the case-mix spectrum: a more heterogeneous population can show different discrimination even when the underlying model is unchanged.

We characterized rather than eliminated these limitations through random-effects modeling, prediction intervals, independent-cohort handling, a variance-provenance sensitivity analysis, assessment of potential database overlap, Low/Moderate QUIPS restriction, subtype analyses, and direct same-study comparison. These measures reduce avoidable analytic distortion and display uncertainty; they do not make the underlying populations homogeneous.

4.5 CFS

CFS is a global, clinician-judgment measure designed to summarize overall fitness and frailty; unlike registry-derived indices, it can be applied at the bedside without reconstructing a deficit score and may therefore be practical for routine clinical screening (12). The review identified 14 CFS studies, of which 6 provided source-confirmed standalone 30-day mortality AUCs for the post hoc exploratory synthesis. The two other previously considered AUC-ready reports were excluded because Narula 2020 reported a one-year AUC and Zaib 2023 did not identify the mortality time window for its ROC AUC. The pooled estimate suggests potentially useful discrimination, but the available data do not permit a valid ranking of CFS relative to either instrument, because no eligible same-cohort comparison used the comparable 30-day standalone AUC outcome. Its practical bedside characteristics may make it attractive for routine screening, but ease of use and implementation burden were not outcomes evaluated in this review. Future direct studies should compare CFS, mFI, and RAI within the same cohorts and report discrimination, calibration, threshold-specific metrics, missingness, workflow feasibility, and decision utility.

4.6 Strengths and limitations

Strengths include source verification of study-level data, use of independent patient cohorts rather than publication or tool-version rows, explicit handling of multiple versions applied to the same cohort, assessment of potential registry overlap, and QUIPS assessment of all 77 studies. The direct-comparison evidence was assembled at the study level rather than inferred from the quantitative meta-analysis subset.

Several limitations remain. First, between-study heterogeneity and broad prediction intervals limit the portability of average AUCs; random effects, prediction intervals, subtype analyses, and direct comparisons characterize but do not remove this limitation. Second, registry overlap cannot be excluded perfectly; a source-based overlap assessment and sensitivity analysis reduced avoidable duplication. Third, uncertainty was estimated from sample size and events for some records; provenance was recorded and all such records were excluded in sensitivity analysis. Fourth, 23 studies were at High overall risk of bias; QUIPS and the Low/Moderate restriction show the effect of excluding them but do not repair primary-study limitations.

Fifth, multiple tool versions, variable coding, and missingness create dependence and measurement heterogeneity; we used one preferred version per independent cohort and performed alternative-version and cohort-clustered analyses. Sixth, case-mix differences prevent the separate pools from serving as a causal or formal indirect comparison; comparative interpretation therefore relies primarily on same-study evidence. Seventh, AUC does not describe calibration or clinical utility (14, 15). Eighth, the mFI label does not settle construct validity, and this review cannot determine why discrimination differs. Finally, CFS synthesis was post hoc and remains exploratory.

5 Conclusions

RAI showed higher average prognostic discrimination for 30-day mortality than mFI across the available independent cohorts, and most direct comparative studies showed the same numerical direction. Heterogeneous case mix prevents the separate pooled AUCs from being interpreted as a formal superiority test, and the review does not establish that either tool is a criterion-standard measure of multidimensional frailty. mFI remains useful for registry-derived population research, while RAI may provide more informative individual risk stratification when its data requirements and workflow are feasible. CFS evidence is promising but post hoc and exploratory, and is insufficient for formal comparative ranking. Future same-cohort studies should evaluate discrimination together with calibration, thresholds, missingness, implementation burden, and decision utility.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmed.2026.1885620/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Jinghua Wang, Tianjin Neurological Institute, China

Reviewed by: Jonathan Hewitt, Cardiff University, United Kingdom

Mohsin Nazir Butt, Aga Khan University Hospital, Pakistan