Skip to main content
NCBI home page
PLOS Medicine logoLink to PLOS Medicine
. 2026 Sep 1;23(9):e1005004. doi: 10.1371/journal.pmed.1005004

Evaluation of the FI-Lab, a laboratory-based, automated frailty index for acute care: A multicohort study

Hugh Logan Ellis 1,2,3, Peter Hanlon 4, Liam Dunnell 5, Martin Whyte 6,7, Daniel H J Davis 8, Josephine Bates 9, Adeel Jafri 10, Diana Shamsutdinova 1, James T Teo 1,11,12, Zina Ibrahim 1, Kenneth Rockwood 2,13,*
Editor: Sanjay Basu14
PMCID: PMC13577487  PMID: 42678992

Abstract

Background

Laboratory-based frailty indices (FI-Lab) have shown promise in geriatric medicine research. We aimed, in diverse samples, to determine the optimal construction of an FI-Lab for acute care and evaluate its validity as a measure of latent health status across the adult life span.

Methods and findings

Our retrospective multicohort study used emergency department encounters in Boston, USA (MIMIC-IV-ED; 2011–2019) and London, UK (King’s College Hospital (KCH); 2017–2020); and a community-based cohort in the UK (UK Biobank; 2006–2010). We evaluated FI-Lab configurations by varying test selection strategies, number of tests, and minimum test thresholds. Sensitivity analyses examined performance across age, ethnicities, sexes, and model types to identify potential blind spots. The primary outcome was 1-year all-cause mortality.

We analysed 227,736 visits (113,032 patients) from MIMIC-IV-ED; 152,305 visits (96,843 patients) from KCH; and 492,703 UK Biobank. An FI-Lab calculated from 25 commonly ordered tests (minimum 15 required) demonstrated hazard ratios for 1-year mortality approaching those of chronological age and exceeding the National Early Warning Score 2 (NEWS2). In combined models, hazard ratios for FI-Lab per standard deviation increase were 2.04 (95 %CI [1.99, 2.09]; p < 0.001) in MIMIC-IV-ED, 2.55 (95 %CI [2.42, 2.69]; p < 0.001) in KCH, and 1.87 (95% CI [1.80, 1.94]; p < 0.001) in UK Biobank. Performance was consistent across subject groupings. Discrimination plateaued at 20–40 tests. This was a retrospective study; future work exploring the impact on everyday use requires prospective evaluation, potentially in a randomised controlled trial.

Conclusions

The FI-Lab provides an automated, scalable measure of patient vulnerability that is robust across healthcare settings and populations. A core set of 20–40 commonly ordered tests is sufficient for signal capture. It offers a pragmatic complement to clinical judgement without requiring manual data entry or additional tests.

Author summary

Why was this study done?

  • Healthcare providers frequently ponder the question, “How sick is this person?”. Even so, there is no simple way to summarise overall illness severity using routinely collected clinical data.

  • The laboratory frailty index (FI-Lab for short)—based on the proportion of abnormal laboratory test results—is an established measure of frailty in older adults.

  • We investigated whether the FI-Lab could be used as a general measure of illness severity across all adult ages and in different healthcare settings.

What did the researchers do and find?

  • We used already-collected healthcare data for almost 900,000 adults from three datasets: emergency department visits in Boston (MIMIC-IV-ED) and London (King’s College Hospital), and a UK community cohort (UK Biobank).

  • We tested multiple FI-Lab models by varying the laboratory tests included and compared their performance with age and standard health deterioration scores.

  • A simple FI-Lab built from approximately 25 commonly-ordered blood tests is as strongly associated as age with risk of dying, hospital admission, and readmission, while standard deterioration risk scores perform much worse.

  • The FI-Lab works regardless of age, sex, ethnicity, or tested healthcare system.

What do these findings mean?

  • FI-Lab provides a fast, automatic, and transparent measure of illness severity, without needing extra testing or bedside assessment.

  • The same approach could be extended beyond blood tests to vital signs and other routinely collected data that can be classified as normal or abnormal.

  • An approach originally developed for older people can provide a broad indicator of health across the adult life span and could aid clinical assessment.

  • Even though this retrospective study using data from two high-income countries shows the FI-Lab reliably tracks poor outcomes, more research is needed to determine whether real-time access to the FI-Lab improves clinical decision-making or patient care.

In this multicohort study, Hugh Logan Ellis and colleagues optimize the laboratory-based frailty index for use as a general measure of illness severity across all adult ages and different healthcare settings.

Introduction

As populations age and medical care successfully extends survival with chronic illness, chronological age has become an increasingly poor proxy for patient vulnerability [1,2]. Two 85-year-olds presenting to an emergency department might share a birth year but possess vastly different physiological reserves. While medicine has adopted the language of frailty to describe this heterogeneity in older adults, clinical assessment tools have not evolved to match the data now available. Electronic health records (EHRs) capture a detailed digital profile for every patient, laboratory panels, imaging sequences, telemetry streams, and medication histories, yet decision-making relies on simpler metrics. The UK National Early Warning Score 2 (NEWS2) [3], for all its merits, uses six vital signs and underestimates risk for older and frailer patients [4,5]. With ever more complex patients, the gap widens between what we measure and what we score.

Why haven’t we closed this gap? The era of big data and machine learning (ML) has produced a proliferation of risk prediction models, yet clinicians worldwide continue to use decades-old risk scores. Though promising, learning from historical data is limited by the absence of counterfactuals: we cannot observe what might have happened with different treatment decisions or under alternative standards of care [6]. Models can only recognise patterns present in their training data. This perpetuates historical inequities and variations in care [7,8]. Beyond these fundamental constraints, practical barriers compound the problem: ML models require mature technological infrastructure and site-specific calibration [9]. Traditional scores like the Acute Physiology and Chronic Health Evaluation (APACHE) score require recalibration [10,11] and complete data collection [12], undermining widespread use. As a result, simple measures like NEWS2 remain the standard, despite their limitations.

Building on the well-established deficit accumulation model of frailty [13], the laboratory-based frailty index (FI-Lab) [14] offers an approach that differs from prediction models and is rooted in the mathematical principles describing complex systems at risk of collapse [15,16]. Rather than predicting specific outcomes, frailty indices such as FI-Lab quantify physiological reserve with a simple calculation: count abnormal results and divide by total tests performed, providing passive, data-driven measures of latent health status. Evidence in older adults shows the FI-Lab can be adapted to effectively measure both acute illness severity and chronic frailty [17], with strong associations to such key outcomes as mortality and readmissions [18,19].

Although FI-Labs have predominantly been used in frailty in older adults, the theory of deficit-accumulation is agnostic to age or setting [17–19]. The proportion of abnormal biological parameters captured in the FI-Lab is an index of underlying dysregulation, making it a latent measure of overall health. In acute settings, this will be a measure of both chronic health status (a combination of overlapping concepts such as frailty, performance status, physiological reserve and baseline) and acute illness severity. If the FI-Lab can capture this underlying construct, its prognostic signal—and with that, its ability to identify multiply- determined risk that increases with age—should generalise across the adult life span, healthcare settings, and outcomes. To realise the potential of the FI-Lab, we aimed to:

  1. Identify the optimal construction method by determining which laboratory test selection strategy performs best across datasets and the optimal number of tests that balances discriminative ability with patient coverage;

  2. Compare performance to established metrics including NEWS2 and patient age; and

  3. Evaluate generalisability by assessing whether the FI-Lab works effectively across diverse populations, ages, sexes, ethnicities, socioeconomic groups, countries, and clinical contexts.

Methods

Study design and data sources

We conducted a retrospective multi-cohort study using routinely collected EHR data from two hospital systems (Beth Israel Deaconess Medical Center [MIMIC-IV-ED; [20]], Boston, 2011–2019; King’s College Hospital [KCH], London, 2017–2020) and one prospective research cohort (UK Biobank, 2006–2010). We selected emergency department encounters in the EHR data to capture unplanned hospital presentations with baseline laboratory testing before intervention.

Study population

We included adults (≥18 years) from MIMIC-IV-ED and from KCH EDs where we could link same-day laboratory data by patient identifier and timestamp. UK Biobank participants (aged 37–73) were recruited from 22 assessment centres and provided blood samples.

Exclusions

Patients without linkable same-day laboratory data (e.g., minor injuries without blood tests; outside referrals with tests not repeated) were excluded. For mortality analyses, only first visits were analysed. At KCH, additional exclusions were: nonbinary gender records, deaths recorded before ED arrival time, and impossible (negative) or implausible (>1,000 days) hospital stays and unparseable datetime records.

Primary and secondary outcomes

The primary outcome was all-cause mortality within 1 year of ED presentation or UK Biobank baseline assessment. Secondary outcomes were in-hospital mortality, hospital length of stay, 30-day hospital readmission, 1-year hospital (re)admission, and hospital admission from the ED and in the UK Biobank mortality up to 10 years.

Laboratory-based frailty index construction

For each patient, an FI-Lab score was calculated following the deficit-accumulation technique for laboratory data established by Howlett and colleagues [14] as the proportion of abnormal test results to the number of tests performed from a pre-specified list. We used the first test result within 24 hours for each parameter. Tests coded abnormal reflected site-specific laboratory reference ranges. As the FI-Lab accommodates varying test numbers by adjusting the denominator to each patient’s performed tests, no imputation is required.

To identify an optimal construction method, we systematically evaluated FI-Lab configurations by varying three parameters:

  1. Test Selection Strategy: We compared lists created from the most ‘Common’ tests ordered versus lists of the most prognostically ‘Important’ tests for 1-year mortality (determined via a random forest model) and with lists created randomly.

  2. Number of Tests: We varied the size of the test list (from 10 to 70).

  3. Feature Floor: We varied the minimum number of tests required for a patient’s score to be calculated, from no imposed minimum to 20 tests.

The Supplementary Appendix details construction and validation methods, the full list of tests considered (Tables C–F in S1 Appendix), and the NEWS2 calculation.

Statistical analysis

We used Cox proportional hazards models to evaluate associations between FI-Lab scores and outcomes. To compare performance, we used nested models: demographics (age/sex) alone, with NEWS2, with FI-Lab, and with both scores. Continuous variables were scaled before analysis. Hazard ratios are reported per standard deviation increase with 95% confidence intervals (CI). One standard deviation of the FI-Lab corresponds to ~0.10 in the ED cohorts and 0.06 in UK Biobank. One standard deviation of age corresponds to ~20 years in the ED cohorts and 8 years in UK Biobank. We use C-statistics to assess discrimination to validate FI-Lab as a measure, not to develop a prediction model. We used AIC to assess model fit.

For mortality and admission outcomes, time-zero was ED presentation. For readmission analyses, time-zero was hospital discharge with censoring at death or study end (30 days or 1 year). Given the exploratory nature of configuration analyses, we did not adjust for multiple comparisons.

To quantify FI-Lab reliability, recognising that similar patients can have different test combinations ordered, we assessed: (1) sampling variation, by calculating 20 random test combinations per patient at various feature set sizes and (2) measurement uncertainty, by adding realistic noise to laboratory values near cut-points.

To further evaluate generalisability, we selected 25 commonly ordered tests with a feature floor of 15 for sensitivity analyses with 1 year mortality as our outcome of interest. We examined incorporation of vital signs as binary features, stratification by 10-year age bands, stratification by sex (gender was used as a proxy in the KCH dataset as sex was not routinely recorded), stratification by socioeconomic status and ethnicity and use of a competing risk model for readmission. Finally, to assess the utility of FI-Lab alongside other measures of frailty, we examined associations between FI-Lab and mortality (1-year and 10-year) across baseline strata of two existing UK Biobank frailty measures: a cumulative deficit frailty index based on 49 self-reported diagnoses and symptoms [21] and an adaptation of the Fried Frailty Phenotype (comprising weight loss, low physical activity, exhaustion, slow-walking speed (all self-reported) and measured low grip strength with 3 or more items indicating frailty, and 1–2 items a pre-frail state [22]). Analyses used R (v4.5.0/v4.2.1).

Our reporting follows the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines [23] and the Reporting of studies Conducted using Observational Routinely-collected health Data (RECORD) extension [24]. A completed STROBE-RECORD checklist is provided in the supplementary appendix (S1 Checklist).

The research questions and conceptual framework were presented at patient involvement sessions organised through King’s College Hospital.

Ethics approval

We utilised deidentified data from three research databases. The MIMIC-IV-ED database [20] is publicly available following completion of required training. UK Biobank analysis was conducted under application 14,151. Use of KCH data was approved by the King’s Electronic Patient Record Interface Committee (approval ID 20230411B). UK Biobank has ethical approval from the North West Multi-centre Research Ethics Committee. Formal consent was not required for this secondary use of deidentified data.

AI statement

Claude Opus 4.0 (Anthropic) was used to refine code as necessary, primarily to improve efficiency, comprehensibility, and to streamline ggplot syntax used for visualisation. The authors manually tested the code and relevant outputs post-iteratively to ensure accuracy.

Results

Study population

From the US-based MIMIC-IV-ED database (2011–2019), we identified 425,011 emergency department visits; after applying eligibility criteria, we analysed 227,736 visits from 113,032 unique patients. From the UK-based King’s College Hospital database, we identified 449,060 records (2017–2020); after exclusions, 152,305 visits from 96,843 unique patients were analysed. From the UK Biobank cohort (2006–2010 baseline recruitment of 502,533 participants), 492,703 with available blood test data were included. (STROBE flow diagram Fig A in S1 Appendix).

The MIMIC-IV-ED cohort was ~8 years older than King’s, with proportionately more patients aged ≥60 years (Table 1). 1-year mortality and hospital admission rates were higher in MIMIC-IV-ED, though in-hospital mortality was lower.

Table 1. Baseline characteristics of patients in the MIMIC-IV-ED and King’s College Hospital Emergency Department Cohorts.

Characteristic MIMIC-IV-ED cohort KCH cohort UK Biobank
Cohort size
 Total ED Visits, n 227,736 152,305
 Unique patients, n 113,032 96,832 492,703
Patient demographics
 Age, years, Mean (SD) 52.07 (20.94) 43.8 (18.8) 56.5 (8.1)
 Age group, n (%)
  18–39 years 37,312 (33.0) 37,734 (39.0)
  40–59 years 31,609 (28.0) 31,195 (32.2) 285,036 (57.9)
  60–79 years 30,881 (27.3) 19,215 (19.8) 217,497 (43.3)
   ≥ 80 years 13,230 (11.7) 8,703 (9.0)
  Female, n (%) 62,087 (54.9) 52,254 (54.0) 273,401 (54.4)
 Ethnic group, n (%)
  White 71,513 (63.3) 43,867 (45.3) 464,105 (94.2)
  Black or Black British/ African American 17,851 (15.8) 27,891 (28.8) 7,801 (1.6)
  Asian or Asian British 5,654 (5.0) 3,907 (4.0) 9,620 (2.0)
  Mixed 2,747 (2.8) 2,885 (0.6)
  Other ethnic group 7,309 (6.5) 8,176 (8.4) 4,430 (0.9)
  Prefer not to answer 9,694 (10.0) 2,335 (0.5)
  Chinese 557 (0.6) 1,527 (0.3)
  Unknown 3,490 (3.1)
  Latino/Hispanic 7,215 (6.4)
  NA 8 (0.0)
 Insurance category
  Medicare 27,090 (24.0)
  Medicaid 10,885 (9.6)
  Private 22,349 (19.8)
  Other/Unknown 3,824 (3.4)
  NA 48,884 (43.2)
 Index of multiple deprivation

decile, Mean, (SD)		 4.00 (1.95)
 Townsend score, mean (SD) −1.29 (3.10)
Total admissions, n (%)
 1 Visit 75,166 (74.8) 71,512 (73.8)
 2 Visits 17,779 (17.7) 14,236 (14.7)
 3+ Visits 7,538 (7.5) 11,099 (11.5)
Presenting acuity and outcomes
 NEWS2 score, mean (SD) 0.94 (1.32) 1.21 (1.58)
 Hospital admission, n (%) 51,115 (45.2) 30,221 (31.2)
 In-hospital mortality, n (%) 1,676 (2.6) 1,182 (3.9)
 30-day readmission, n (%) 8,783 (13.6) 7,721 (8.0)
 1-year all-cause mortality, n (%) 8,202 (7.3) 3,222 (3.3) 965 (0.2)
Laboratory testing patterns
 Patients with ≥10 tests, n (%) 107,991 (95.5) 93,133 (96.2) 490,255 (99.5)
 Patients with ≥20 tests, n (%) 106,991 (94.6) 89,373 (92.3) 489,072 (99.3)
 Patients with ≥30 tests, n (%) 94,996 (84.0) 66,359 (68.5) 457,646 (92.9)
Open in a new tab

Data are presented as mean (SD) for continuous variables and n (%) for categorical variables. KCH denotes King’s College Hospital, NEWS2 National Early Warning Score 2, and SD standard deviation.

The distribution of FI-Lab scores varied by age and by the number of tests (features) used to calculate the score (Fig 1).

Fig 1. FI-Lab score distributions across age groups, feature number, and feature floor in MIMIC without observations (vital signs).

Fig 1

Open in a new tab

Each panel shows the distribution of FI-Lab scores (proportion of abnormal laboratory tests) by age group, varying by the number of included features (columns: 10, 15, 30, and 70) and the minimum feature floor (rows: 0, 5, and 15). The feature floor defines the minimum number of available laboratory tests required for an FI-Lab score to be calculated for a given individual. Age groups are displayed on the y-axis from <40 to 80+ years. The total number of patients per panel is shown at the bottom. As both the number of features and the minimum feature floor increase, the score distributions become more tightly clustered around mid-range values, and the proportion of patients with very high FI-Lab scores (>70%) decreases substantially. This reflects expected statistical properties: as more features are required, extreme scores become less likely, and the index becomes less sensitive to single abnormal results. Panels with insufficient data to meet the minimum feature floor display no patients (e.g., 10 features with a floor of 15).

Configuration analysis

Measurement stability improved progressively with increasing test numbers. Coefficients of variation decreased from 1.65 (15 tests) to 0.34 (70 tests) in MIMIC-IV-ED, with similar patterns in KCH (Figs B, C in S1 Appendix). With measurement noise added to simulate assay variability, coefficients of variation fell from 0.30 (10 tests) to 0.16 (≥40 tests) (Fig D in S1 Appendix).

Model discrimination showed a different pattern: performance improved with initial test inclusion but plateaued beyond 20–30 tests, with minimal further gains or occasional decrements (Figs 2 and E–I in S1 Appendix).

Fig 2. Discrimination (C-statistic) of FI-Lab models across outcomes, datasets, and selection strategies.

Fig 2

Open in a new tab

Rows show outcomes; columns show dataset × strategy (Common, Importance). Lines are FI-Lab models coloured by the minimum-tests floor (0–20); points indicate the number of features included. The y-axis shows Harrell’s C-statistic (scales are fixed within rows but free between rows). Blank panels in UK Biobank indicate outcomes not available in that dataset. UK Biobank lacks readmission; 1-year hospital admission is used as the comparator and plotted under “1-Year Readmission.” Across datasets, the C-statistic typically rises with 10–30 features and then plateaus, with only small differences of modest influence between Common and Importance strategies.

For 1-year mortality in KCH, C-statistics were 0.895 (20 tests), 0.901 (25 tests), and 0.902 (70 tests). In MIMIC, some secondary outcomes showed performance decrements with additional tests: 30-day readmission C-statistics decreased from 0.585 (20 tests) to 0.577 (70 tests). When sample sizes remained constant, varying the number of tests made little difference to model fit as measured by AIC (Fig J in S1 Appendix).

Tests selected by frequency of ordering achieved C-statistics within 0.002 of tests selected by importance for the 25-test configuration (MIMIC: 0.845 versus 0.843; KCH: 0.901 versus 0.901). Both strategies outperformed random selection (mean C-statistic 0.738 in MIMIC, 0.751 in KCH). A feature floor of 15 maintained >90% patient coverage while reducing score variability by 60% compared to no floor requirement (Fig K in S1 Appendix).

In UK Biobank, discrimination did not plateau with either selection strategy: for 1-year mortality, C-statistics rose from 0.776 to 0.781 for Common and from 0.732 to 0.781 for Importance when increasing from 30 to 50 features with a floor of 20.

Table 2 summarises the effects of changes to configuration parameters on performance.

Table 2. Configuration trade-offs in FI-Lab development.

Configuration choice Effect on variance Effect on patient coverage Effect on performance
Increasing feature floor Substantially reduces Moderately reduces Modest; reflects cohort selection (see Discussion with age - should generalise across the adult life span, healthcare settings, and outcomes)
Adding more features Reduces Moderately increases Improves then declines or plateaus
Including vital signs Minimal effect Increases coverage Better for short-term outcomes
Importance vs. common selection Not applicable Dramatically lower for importance in some datasets (dependent on how importance is defined) Differences are very small
Open in a new tab

Performance compared to established measures

We chose 25 common tests with a feature floor of 15 as our reference configuration. In combined models, the FI-Lab was associated with hazard ratios for 1-year mortality of 2.04 (95% CI [1.99, 2.09]; p < 0.001) in MIMIC-IV-ED, 2.55 (95% CI [2.42, 2.69]; p < 0.001) in KCH, and 1.87 (95% CI [1.80, 1.94]; p < 0.001) in UK Biobank, with C-statistics of 0.847, 0.911, and 0.771, respectively. In the same models, the hazard ratios for age were 2.69 (95% CI [2.60, 2.78]; p < 0.001) in MIMIC-IV-ED, 2.70 (95% CI [2.54, 2.87]; p < 0.001) in KCH, and 1.72 (95% CI [1.59, 1.85]; p < 0.001) in UK Biobank. The hazard ratios for NEWS2 were 1.08 (95% CI [1.06, 1.10]; p < 0.001) in MIMIC-IV-ED and 1.03 (95% CI [0.99, 1.08]; p = 0.13) in KCH; NEWS2 was unavailable in UK Biobank. Figure 3 demonstrates FI-Lab’s (sex-stratified) associations approach or exceed age and consistently exceed NEWS2 across all datasets.

Fig 3. Hazard ratios for 1-year mortality by variable type and gender across three datasets.

Fig 3

Open in a new tab

Columns show variable types (FI-Lab, Age, and NEWS2); rows show gender (Female, Male). Lines represent hazard ratios by dataset (KCH in green, MIMIC-IV-ED in purple, UK Biobank in blue) across numbers of laboratory features (used to compute the FI-Lab) with the Common feature-selection strategy (minimum-tests floor = 15). All variables are shown with solid lines, points, and shaded 95% confidence intervals. NEWS2 was not available in UK Biobank. The y-axis is log2-scaled; the dashed horizontal line marks a hazard ratio of 1.

Secondary outcomes

For 30-day readmission FI-Lab was associated with hazard ratios of 1.32 (95% CI [1.30, 1.33]) in MIMIC and 1.35 (95% CI [1.33, 1.36]) in KCH and 1.36 (95% CI [1.35, 1.37]) in MIMIC and 1.33 (95% CI [1.32, 1.34]) in KCH for 1-year readmission. Age showed the strongest association with mortality, FI-Lab showed the strongest association for readmission, ED admission, and length of stay. NEWS2 consistently showed the weakest association across outcomes (Figs L–P in S1 Appendix).

Sensitivity analyses

FI-Lab performance was consistent across sexes, with greater hazard ratios in males in ED cohorts (MIMIC: HR 2.10 versus 1.96; KCH: HR 2.74 versus 2.37) and in females in UK Biobank (HR 1.99 versus 1.83), with overlapping CI in all comparisons.

FI-Lab was associated with mortality across all age bands. Hazard ratios peaked at ages 30–60 years (range 2.11–3.16) and declined to 1.77–2.11 in those aged ≥80 years, but remained significant across all groups (p < 0.001; detailed HRs and 95% CIs in Table A in S1 Appendix).

FI-Lab was significantly associated with mortality across all ethnic groups except Chinese in UK Biobank (fewest events). Hazard ratios were broadly similar, though ethnicity interactions were statistically significant in all cohorts (p < 0.001, Table B and Fig Q in S1 Appendix).

Adding vital signs to the FI-Lab improved discrimination for in-hospital mortality (C-statistics increased from 0.750 to 0.782 in MIMIC, p < 0.001 (Fig B in S1 Appendix); and from 0.797 to 0.801 in KCH, p < 0.001) (Fig C in S1 Appendix) but not for 1-year mortality (C-Statistics 0.845–0.845 in MIMIC; 0.901–0.899 in KCH).

Using Fine–Gray subdistribution hazards (SHR) attenuated-hazard model for readmission on a subset of patients (MIMIC-IV-ED; 25 Common tests; feature floor 15; 50,000-patient subset), accounting for death as a competing risk, attenuated ratios by ~2%–3% relative to cause-specific estimates (30-day readmission SHR 1.29, 95% CI [1.27, 1.32]; p < 0.001 versus cause-specific HR 1.32; 1-year readmission SHR 1.32, 95% CI [1.31, 1.34]; p < 0.001 versus cause-specific HR 1.36).

When stratifying UK Biobank data by baseline levels of other frailty measures (a deficit accumulation FI based on diagnoses/symptoms, and an adapted Frailty Phenotype), a higher FI-Lab was associated with increased risk of 1- and 10-year all-cause mortality across each strata of baseline frailty status by both frailty measures (Tables G, H in S1 Appendix). When including the FI based on diagnoses/symptoms and the FI-Lab in the same model, both were associated with higher 10-year mortality with broadly comparable hazard ratios (Table I in S1 Appendix).

Discussion

In this large, multicohort study, we found that a laboratory-based frailty index constructed from routine blood tests provided prognostic information with signal strength approaching that of chronological age and superior to NEWS2 (representative of the current paradigm for acute severity assessment). The FI-Lab demonstrated robust associations with outcomes ranging from emergency department admission to mortality. Critically, this signal remained consistent across the full adult age spectrum, all ethnic groups, and different healthcare systems, and was detectable up to 10 years after measurement. This consistency suggests that the FI-Lab captures something fundamental about patient vulnerability, independently of where, when, or in whom it is measured. An optimal FI-Lab balances performance with measurement stability and patient coverage, typically achieved with a core set of 20–40 commonly ordered laboratory tests. We are persuaded that FI-Lab captures a latent property of health, combining both elements of chronic health status and acute illness severity. In this sense, the deficit accumulation model could serve as a general-purpose instrument for measuring latent health status, extending the application of FI-Labs from geriatrics into routine clinical care.

FI-Lab performance has an early peak with 20–40 features and for some outcomes does not improve indefinitely. This provides a crucial, evidence-based answer to the question of how to best construct an FI-Lab. It resolves a key uncertainty in the literature, where FI-Lab configurations vary widely [18]. The performance plateau suggests that a core set of common tests captures sufficient physiological derangement for prognostication. The decline seen when including rare tests might reflect selection bias inherent in clinical care data [25,26]. That these tests are ordered when pathology is already suspected, gives them a high pre-test probability of being abnormal (Fig R in S1 Appendix).

The FI-Lab showed consistent performance across all adult age groups. While frailty is traditionally associated with older adults, our data show that the physiological dysregulation captured by the FI-Lab is strongly associated with adverse outcomes in younger adults as well. This aligns with other work suggesting the accumulation of deficits reflects underlying vulnerability regardless of age [22,27–33]. As shown in Fig 1, with a feature floor of 15 the score distribution broadens with increasing age while maintaining a submaximal limit well short of 1.0—the same characteristic behaviour described for frailty indices of varying composition. In this respect, the FI-Lab is a frailty index in the formal, properties-based sense. Furthermore, as in Tables B and E in S1 Appendix, we see that, in both the UK-Biobank’s two measures of frailty (index [21] and phenotype [22]), the FI-Lab is associated across grades of fitness/robustness/prefrailty/frailty with an increasing dose-response in both the 1-year and 10-year hazard ratios. In short, the FI-Lab correlates with and adds to the information provided by these two measurement approaches. We recommend a pragmatic approach to FI-Lab construction for clinical implementation: use the most commonly ordered tests at a given institution with a minimum of 15 tests required for score calculation. This strategy optimises the trade-off between associative strength, score stability, and patient coverage. Importance-based selection offered no meaningful benefit and reduced coverage. Include any routinely performed investigation with binary normal/abnormal classification that shows variation, but exclude clinician-determined variables such as diagnoses or medications that can introduce subjective bias [25,34].

The inherent variability of the FI-Lab, stemming from both binomial sampling effects and laboratory measurement noise, underscores that it should be interpreted as an estimate of general vulnerability, not a precise measurement. We recognise that, employing values from ED visits, in many cases the FI-Lab might represent the severity of acute illness rather than frailty. Note however that the UK Biobank FI-Lab remained associated with mortality at hazard ratios per SD of 1.87 (95% CI [1.80, 1.94]) and at time horizons of up to 10 years. Were the ED cohorts purely signalling acute illness severity, it should be absent in UK Biobank; it is not. It is likely a combination of both. Finally, we position the FI-Lab not to replace clinical judgment, but as a data-driven frailty (or fitness) measure to help clinicians rapidly contextualise a patient’s underlying health in light of their chronological age. In practice, this could identify the 50-year-old whose EHR data suggest the vulnerability of someone decades older, or the 80-year-old with remarkably preserved physiology. Though age remains relevant, FI-Lab captures vulnerability not reflected by age alone. It does so without requiring any additional testing, as the items that compose it can be pulled from test results as they become available.

The retrospective design establishes associations but not causation. Data from two high-income countries with mature EHR infrastructure might not generalise to other healthcare settings or resource-constrained environments. Despite adjusting for age and NEWS2, unmeasured confounders remain. We used standard Cox models rather than competing risk models potentially underestimating readmission incidence, though our sensitivity analysis suggests model choice only marginally impacts effect size. Computational cost prevents us from running these analyses across all cohorts and configurations, so cause-specific hazards are reported as primary. We tested multiple configurations without adjustment for multiple comparisons and did not test proportional hazard assumptions. Our linkage methodology likely underestimated laboratory coverage, particularly for pre-emergency department tests. We recognise that using the same cohort for feature selection and evaluation is a potential source of overfitting for the ‘Importance’ strategy in MIMIC-IV-ED. Most importantly, whether FI-Lab measurement improves clinical outcomes requires prospective evaluation, preferably extended to a range of countries. Laboratory reference ranges were standard adult ranges and were not age-stratified. The smaller effect size observed in UK Biobank could plausibly reflect, in part, a higher proportion of deaths from causes that are not physiologically predictable in a community cohort compared with patients presenting to hospital. Additional methodological considerations are discussed in the Supplementary S1 Appendix.

The laboratory-based frailty index is a simple, powerful, transparent, and scalable tool for quantifying patient vulnerability. An optimal FI-Lab can be constructed using a core set of the most commonly available laboratory results, with a minimum of 15 tests suggested for stability. As a passive, data-driven measure of latent health status, it performs robustly across the settings and populations studied. The FI-Lab thereby offers a pragmatic complement to current clinical measures and human assessment.

Supporting information

S1 Appendix. Supplementary methods, figures, and tables.

(DOCX)

pmed.1005004.s001.docx (9.6MB, docx)
S1 Checklist. STROBE-RECORD Checklist.

STROBE: von Elm E, Altman DG, Egger M, Pocock SJ, Gøtzsche PC, Vandenbroucke JP, et al. (2007). The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) Statement: Guidelines for Reporting Observational Studies. PLoS Med. 4(10): e296. https://doi.org/10.1371/journal.pmed.0040296. RECORD: Benchimol EI, Smeeth L, Guttmann A, Harron K, Moher D, Petersen I, Sørensen HT, von Elm E, Langan SM; RECORD Working Committee. (2015). The REporting of studies Conducted using Observational Routinely-collected health Data (RECORD) statement. PLoS Med. 12(10):e1001885. https://doi.org/10.1371/journal.pmed.1001885. Fig A: STROBE flow diagram of the study cohort selection process. Fig B: FI-Lab Stability Analysis in the MIMIC-IV Database. This plot shows the median coefficient of variation (CV) of FI-Lab scores against the number of requested features, categorised by the availability of vital signs and varying feature floor thresholds (0, 5, 10, 15, and 20) within the MIMIC-IV database. The number of patients included at each specific number of features requested is indicated by the labels on the plot. Fig C: FI-Lab Stability Analysis in the King’s College Hospital (KCH) Database. This plot illustrates the median coefficient of variation (CV) of FI-Lab scores as the number of requested features increases, stratified by the presence or absence of vital signs and different feature floor thresholds (0, 5, 10, 15, and 20) within the KCH dataset. Labels indicate the number of patients contributing to each data point. Fig D: FI-Lab Score Stability Against Simulated Analytical Noise. To test robustness against laboratory measurement uncertainty, a simulation was performed on 5,000 randomly sampled patients from the KCH database. For each patient, the FI-Lab score was recalculated 20 separate times, with noise (for example, increasing or decreasing a recorded Sodium by 2 mmol/l) introduced to their laboratory values prior to each calculation. The plot shows the median coefficient of variation (CV) of these 20 scores. The results demonstrate that score stability improves (i.e., CV decreases) as the number of features increases, with the effect plateauing after 30 features. Fig E. Model Performance (C-Statistic) by FI-Lab Configuration Without Vital Signs in the MIMIC-IV-ED Cohort. The figure displays the predictive performance (Concordance Index/C-Statistic) of the FI-Lab across six clinical outcomes, based on laboratory data alone. Performance is shown for the ‘Common’ and ‘Importance’ test selection strategies (panels). Within each plot, the x-axis indicates the number of laboratory features used, and colours denote the minimum feature floor required to calculate a score. A key pattern, particularly for the ‘Common’ strategy, is that performance does not increase indefinitely with more features. Instead, it often peaks when using 20–40 tests and then plateaus or declines, a trend visible across outcomes like in-hospital death, ED admission, and readmissions. The ‘Importance’ strategy tends to show more consistent improvement, though performance gains often level off at higher feature counts. Fig F. Model Performance (C-Statistic) by FI-Lab Configuration Without Vital Signs in the UK Biobank Cohort. Predictive performance (Concordance Index/C-Statistic) of FI-Lab across four clinical outcomes, based on laboratory data alone. Performance is shown for the ‘Common’ and ‘Importance’ test-selection strategies (panels); within each, the x-axis indicates the number of laboratory features used and colours denote the minimum feature floor required to calculate a score. Unlike the ED cohorts, where the ‘Common’ strategy peaks and then plateaus or declines, performance in UK Biobank generally increases with additional features—consistent with near-universal baseline testing that is largely independent of patient acuity. Fig G. Predictive Performance of FI-Lab Configurations in the MIMIC-IV Database. This figure illustrates the predictive performance (Concordance Index/ C-Statistic) of various FI-Lab configurations for six clinical outcomes within the MIMIC-IV cohort. Performance is plotted against the number of features used (x-axis). Panels compare feature selection based on ‘Common’ prevalence versus ‘Importance’ ranking. Different lines represent varying minimum feature prevalence thresholds (‘Feature Floor’). The six outcomes evaluated are 1-year mortality, 1-year readmission, 30-day readmission, ED admission, in-hospital death, and length of stay. Fig H. Performance of FI-Lab Configurations in the King’s College Hospital (KCH) Database (with Vitals). This figure shows the predictive performance (C-Statistic) of FI-Lab configurations for six outcomes in the KCH cohort, including vital signs as potential features. Performance is shown across a range of feature counts (x-axis), stratified by the feature selection method (‘Common’ versus ‘Importance’) and the minimum feature prevalence (‘Feature Floor’). The six outcomes evaluated are 1-year mortality, 1-year readmission, 30-day readmission, ED admission, in-hospital death, and length of stay. Fig I. Predictive Performance of FI-Lab Configurations in the King’s College Hospital (KCH) Database (without Vitals). This figure shows the predictive performance (C-Statistic) of FI-Lab configurations for six outcomes in the KCH cohort when vital signs have been excluded from the potential feature set. Performance is shown across a range of feature counts (x-axis), stratified by the feature selection method (‘Common’ versus ‘Importance’) and the minimum feature prevalence (‘Feature Floor’). The six outcomes evaluated are 1-year mortality, 1-year readmission, 30-day readmission, ED admission, in-hospital death, and length of stay. Fig J. Model Fit (AIC) for FI-Lab Configurations Across Outcomes and Datasets. AIC values for FI-Lab models across outcomes and datasets. Lower AIC indicates better model fit. Rows correspond to outcomes; columns to dataset–strategy combinations. Blank UK Biobank panels indicate unavailable outcomes; UK Biobank ‘1-Year Admission’ is shown under ‘1-Year Readmission’. Fig K. Average Patient Coverage by FI-Lab Configuration Across Datasets. Panels show dataset and feature-selection strategy (KCH, MIMIC-IV-ED, UK Biobank; Common versus Importance). Lines show the mean proportion of eligible patients for whom an FI-Lab score could be computed (averaged across available outcomes) as the number of features increases; colours denote the minimum-tests “floor” (0, 5, 10, 15, and 20). Coverage generally increases with more features and decreases with stricter floors. With floors ≤10, coverage approaches ~95%–100% by ~30–40 features in all datasets; a 20-test floor yields lower coverage at small feature counts most notably in UK Biobank under the Importance strategy, but rises as additional features are included. Fig L. Hazard Ratios for 1-Year Mortality Across Datasets and Feature-Selection Strategies. Columns show datasets (KCH, MIMIC-IV-ED, UK Biobank); rows show strategies (Common, Importance). Coloured lines represent FI-Lab hazard ratios by number of laboratory features and minimum-tests “floor”; shaded bands are 95% confidence intervals (CI). Red triangles show the median Age hazard ratio with whiskers indicating the range across floors/models; grey squares show the median NEWS2 hazard ratio with range (not available in UK Biobank). The y-axis is log2-scaled; the dashed line marks a hazard ratio of 1. Fig M. Hazard Ratios for In-Hospital Death Across FI-Lab Configurations, Age, and NEWS2 in KCH and MIMIC-IV-ED. Hazard ratios for in-hospital death are shown for FI-Lab (coloured lines, shaded 95% CI) across varying numbers of features and minimum tests required, for both the common and importance-based feature selection strategies. Red triangles indicate the median hazard ratio for age with range (minimum–maximum) across feature floors; grey squares indicate the median hazard ratio for NEWS2 with range. FI-Lab estimates represent the effect within the combined model. Hazard ratios are displayed on a log2 scale; the dashed horizontal line indicates a hazard ratio of 1.0. Fig N. Hazard Ratios for Length of Stay Across FI-Lab Configurations, Age, and NEWS2 in KCH and MIMIC-IV-ED. Hazard ratios for time to hospital discharge are shown for FI-Lab (coloured lines, shaded 95% CI) across varying numbers of features and minimum tests required (“feature floor”), for both the common and importance-based feature selection strategies. Red triangles indicate the median hazard ratio for age with range (minimum–maximum) across feature floors; grey squares indicate the median hazard ratio for NEWS2 with range. FI-Lab estimates represent the effect within the combined model. Hazard ratios are displayed on a log2 scale; the dashed horizontal line indicates a hazard ratio of 1.0. In this Cox model, hazard ratios < 1 indicate a lower likelihood of discharge at any given time (i.e., longer stay). FI-Lab showed a stronger association with prolonged stay than age. Fig O. Hazard Ratios for 1-Year Readmission Across FI-Lab Configurations, Age, and NEWS2 in KCH, MIMIC-IV-ED, and UK Biobank. Hazard ratios for 1-year readmission are shown for FI-Lab (coloured lines, shaded 95% CI) across varying numbers of features and minimum tests required (“feature floor”), for both the common and importance-based feature selection strategies. In the UK Biobank dataset, the outcome “1-Year Admission” is equivalent to “1-Year Readmission” in KCH and MIMIC-IV-ED. Red triangles indicate the median hazard ratio for age with range (minimum–maximum) across feature floors; grey squares indicate the median hazard ratio for NEWS2 with range. FI-Lab estimates represent the effect within the combined model for KCH and MIMIC-IV-ED, and within the FI-Lab model for UK Biobank; NEWS2 was unavailable in UK Biobank. Hazard ratios are displayed on a log2 scale; the dashed horizontal line indicates a hazard ratio of 1.0. Fig P. Hazard Ratios for ED Admission Across FI-Lab Configurations, Age, and NEWS2 in KCH and MIMIC-IV-ED. Hazard ratios for emergency department (ED) admission are shown for FI-Lab (coloured lines, shaded 95% CI) across varying numbers of features and minimum tests required (“feature floor”), for both the common and importance-based feature selection strategies. Red triangles indicate the median hazard ratio for age with range (minimum–maximum) across feature floors; grey squares indicate the median hazard ratio for NEWS2 with range. FI-Lab estimates represent the effect within the combined model; NEWS2 was unavailable in the UK Biobank dataset. Hazard ratios are displayed on a log2 scale; the dashed horizontal line indicates a hazard ratio of 1.0. Higher hazard ratios indicate a greater likelihood of admission from the ED. Table A. FI-Lab Hazard Ratios for 1-Year Mortality by Age Group. Hazard ratios for 1-year mortality across age bands in three cohorts. Models adjusted for age (within band) and gender. FI-Lab calculated using 25 most common laboratory tests with minimum 15 tests required for score calculation. All hazard ratios p < 0.001. HR, hazard ratio; CI, confidence interval. Table B. Hazard ratios for 1 year mortality by ethnic group. Hazard ratios for 1-year mortality stratified by ethnicity and race across three cohorts. Models adjusted for age and gender. FI-Lab calculated using 25 most common laboratory tests with minimum 15 tests required for score calculation. All hazard ratios p < 0.001. Interaction p-values test for differences in FI-Lab effect across ethnic/racial groups within each cohort. HR, hazard ratio; CI, confidence interval. Fig Q. FI-Lab Hazard Ratios for 1-Year Mortality Stratified by Ethnicity Across Three Cohorts. Forest plot showing hazard ratios (per standard deviation increase in FI-Lab) for 1-year mortality across ethnic groups in MIMIC-IV-ED (purple), King’s College Hospital (blue), and UK Biobank (green for 1-year, yellow for 10-year mortality). Points represent hazard ratio estimates with horizontal lines indicating 95% CI. The vertical dashed line at hazard ratio = 1.0 indicates no effect. Sample sizes and event counts are displayed in the table above. Results are grouped by ethnicity (White, Asian, Black, Chinese, Mixed, Other, Hispanic/Latino, Not stated) to facilitate comparison across datasets. The x-axis is truncated at 0.68–3.7 for clarity; one confidence interval with very few events extends beyond this range and is indicated by an arrow (full numeric results are provided in the table above). UK Biobank displays both 1-year and 10-year mortality results due to low event rates in some ethnic groups at 1 year. FI-Lab demonstrates significant associations with mortality across nearly all ethnic groups, with broadly consistent hazard ratios across datasets despite variation in baseline mortality risk and population composition. The Black and Chinese ethnic groups in UK Biobank 1-year follow-up did not reach statistical significance, likely reflecting low statistical power with 8 and fewer than 5 events, respectively. Table C. Descriptive statistics and feature importance for the top 70 tests (including vitals) derived from the MIMIC database. Abbreviations: I, Icterus Index; H, Haemolysis Index; L, Lipaemia Index. Table D. Descriptive statistics and feature importance for the top 63 tests (excluding vitals) derived from the MIMIC database. Abbreviations: I, Icterus Index; H, Haemolysis Index; L, Lipaemia Index. Table E. Descriptive statistics and feature importance for the top 63 tests (excluding vitals) derived from UK Biobank. Table F. Descriptive statistics and feature importance for the top 50 tests (excluding vitals) derived from KCH. Fig R. Relationship between test frequency and abnormality rate in emergency department patients. Scatter plot showing the proportion of abnormal results versus total test frequency (log scale) for blood tests from emergency department attendances in the MIMIC-IV-ED cohort. Each point represents a unique laboratory test, colored by abnormality rate band (grey: no abnormal flags; dark purple to green: increasing proportions of abnormal results). Vertical dashed lines indicate frequency cutoffs for the top 30 and top 70 most commonly ordered tests. Commonly ordered tests (right side) demonstrate a range of abnormality rates, typically 10%–50%. Notably, very high-abnormality rates (>70%) are observed exclusively among infrequently ordered tests (left side), such as CD4/CD8 ratio (76% abnormal), Free Kappa (75%), and NTproBNP (70%). These high-abnormality, low-frequency tests represent investigations ordered only when clinicians have high pre-test probability of a specific condition, resulting in substantial enrichment for abnormal results compared to the baseline prevalence in the general emergency department population. Table G. Hazard ratios (95% CI) for mortality per SD-increase in FI-Lab within strata of the Williams and colleagues frailty index at baseline. Table H. Hazard ratios (95% CI) for mortality per SD-increase in FI-Lab across baseline levels of the adapted Frailty Phenotype. Table I. Hazard ratios (95% CI) for 10-year mortality at specified levels of FI-Lab and Williams and colleagues frailty index in UK Biobank—both variables in same model adjusted for age and sex.

(DOCX)

pmed.1005004.s002.docx (22.1KB, docx)

Acknowledgments

The authors thank Dr. Mohammad Al-Agil and Mr. Samora Hunter for their contributions to data extraction from King’s College Hospital, which was essential to this analysis. The authors also acknowledge the crucial help provided by Alexandra van der Valk, MA, in manuscript preparation and submission. Thank you to the many study participants, without whom this work would not be possible. The authors take full responsibility for all content.

Abbreviations

APACHE

Acute Physiology and Chronic Health Evaluation

CI

confidence intervals

COV

coefficient of variation

ED

emergency department

EHRs

Electronic health records

FI-Lab

laboratory-based frailty index

KCH

King’s College Hospital

ML

machine learning

NEWS2

National Early Warning Score 2

RECORD

Reporting of studies Conducted using Observational Routinely-collected health Data

STROBE

Strengthening the Reporting of Observational Studies in Epidemiology.

Data Availability

UK Biobank data are available for use subject to application and approval of a data sharing agreement (https://www.ukbiobank.ac.uk/use-our-data/apply-for-access/). MIMIC-IV-ED data are freely available upon request to the Massachusetts Institute of Technology (https://mimic.mit.edu/docs/faq/how-to-get-access.html). King’s College Hospital data are not publicly available; access is typically limited to approved researchers who meet the relevant data access and ethics requirements. Interested users can contact the King’s Electronic Records Research Interface (KERRI) at kch-tr.cogstackrequests@nhs.net. All analysis code for FI-Lab is available at GitHub (https://github.com/jhfl2/FI-Lab-Recipe) or https://doi.org/10.5281/zenodo.21686996.

Funding Statement

This work was supported by the DRIVE-Health, King’s College London funded Centre for Doctoral Training (CDT) in Data-Driven Health (https://www.drive-health.org.uk/), and a Dalhousie University Internal Medicine Research Foundation Research Fellowship (UIMRF; https://medicine.dal.ca/departments/department-sites/medicine/research/research-funding/uimrf.html), both awarded to H.L.E. Z.I. is supported by the NIHR Biomedical Research Centres at South London and Maudsley NHS Foundation Trust and University College London Hospitals, King’s College London (https://www.maudsleybrc.nihr.ac.uk/), and Innovate UK (grant 10104845; https://www.ukri.org/councils/innovate-uk/). KR is supported by Dalhousie Medicine Advancement as Professor of Frailty & Aging. UK Biobank was established by the Wellcome Trust, Medical Research Council, Department of Health, Scottish Government, and the Northwest Regional Development Agency, with additional funding from the Welsh Assembly Government, British Heart Foundation, Cancer Research UK, Diabetes UK, and National Institute for Health and Care Research (NIHR) (https://www.ukbiobank.ac.uk/about-us/our-funding/). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

References

  • 1.Faitna P, Bottle A, Klaber B, Aylin PP. Has multimorbidity and frailty in adult hospital admissions changed over the last 15 years? A retrospective study of 107 million admissions in England. BMC Med. 2024;22(1):369. doi: 10.1186/s12916-024-03572-z [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2.Office for National Statistics (ONS). National population projections: 2022-based. 2025. Available from: https://www.ons.gov.uk/peoplepopulationandcommunity/populationandmigration/populationprojections/bulletins/nationalpopulationprojections/2022based#changing-age-structure
  • 3.Royal College of Physicians. National Early Warning Score (NEWS) 2 standardising the assessment of acute-illness severity in the NHS. 2017. Available from: https://www.rcp.ac.uk/media/a4ibkkbf/news2-final-report_0_0.pdf
  • 4.Vardy ER, Lasserson D, Barker RO, Hanratty B. NEWS2 and the older person. Clin Med (Lond). 2022;22(6):522–4. doi: 10.7861/clinmed.2022-0426 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5.Nissen SK, Candel BGJ, Nickel CH, de Jonge E, Ryg J, Bogh SB, et al. The impact of age on predictive performance of national early warning score at arrival to emergency departments: development and external validation. Ann Emerg Med. 2022;79(4):354–63. doi: 10.1016/j.annemergmed.2021.09.434 [DOI] [PubMed] [Google Scholar]
  • 6.Logan Ellis H, Palmer E, Teo JT, Whyte M, Rockwood K, Ibrahim Z. The early warning paradox. NPJ Digit Med. 2025;8(1):81. doi: 10.1038/s41746-024-01408-x [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7.Harris S. I don’t want my algorithm to die in a paper: detecting deteriorating patients early. Am J Respir Crit Care Med. 2021;204(1):4–5. doi: 10.1164/rccm.202102-0459ED [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8.Ellis H L, Rockwood K. Does predicting who will fall tomorrow tell you who is frail today? Age Ageing. 2025;54(10):afaf309. doi: 10.1093/ageing/afaf309 [DOI] [PubMed] [Google Scholar]
  • 9.Sendak M, Gao M, Nichols M, Lin A, Balu S. Machine learning in health care: a critical appraisal of challenges and opportunities. EGEMS Wash DC. 2019;7(1):1. doi: 10.5334/egems.287 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10.Knaus WA, Draper EA, Wagner DP, Zimmerman JE. APACHE II: a severity of disease classification system. Crit Care Med. 1985;13(10):818–29. [PubMed] [Google Scholar]
  • 11.Zimmerman JE, Kramer AA, McNair DS, Malila FM. Acute Physiology and Chronic Health Evaluation (APACHE) IV: hospital mortality assessment for today’s critically ill patients. Crit Care Med. 2006;34(5):1297–310. doi: 10.1097/01.CCM.0000215112.84523.F0 [DOI] [PubMed] [Google Scholar]
  • 12.Hayakawa K, Uchino S, Endo H, Hasegawa K, Kiyota K. Impact of missing values on the ability of the acute physiology and chronic health evaluation III and Japan risk of death models to predict mortality. J Crit Care. 2024;79:154432. doi: 10.1016/j.jcrc.2023.154432 [DOI] [PubMed] [Google Scholar]
  • 13.Kim DH, Rockwood K. Frailty in older adults. N Engl J Med. 2024;391(6):538–48. doi: 10.1056/NEJMra2301292 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14.Howlett SE, Rockwood MRH, Mitnitski A, Rockwood K. Standard laboratory tests to identify older adults at increased risk of death. BMC Med. 2014;12:171. doi: 10.1186/s12916-014-0171-9 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15.Rutenberg AD, Mitnitski AB, Farrell SG, Rockwood K. Unifying aging and frailty through complex dynamical networks. Exp Gerontol. 2018;107:126–9. doi: 10.1016/j.exger.2017.08.027 [DOI] [PubMed] [Google Scholar]
  • 16.Blodgett JM, Theou O, Howlett SE, Rockwood K. A frailty index from common clinical and laboratory tests predicts increased risk of death across the life course. Geroscience. 2017;39(4):447–55. doi: 10.1007/s11357-017-9993-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17.Ellis LH, Dunnell L, Whitney J, Jennings C, Wilson D, Tippett J. Can laboratory test-based frailty indices contribute to frailty screening in emergency departments? Age Ageing. 2025;54(7):afaf192. doi: 10.1093/ageing/afaf192 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18.Sapp DG, Cormier BM, Rockwood K, Howlett SE, Heinze SS. The frailty index based on laboratory test data as a tool to investigate the impact of frailty on health outcomes: a systematic review and meta-analysis. Age Ageing. 2023;52(1):afac309. doi: 10.1093/ageing/afac309 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19.Hakeem FF, Maharani A, Todd C, O’Neill TW. Development, validation and performance of laboratory frailty indices: A scoping review. Arch Gerontol Geriatr. 2023;111:104995. doi: 10.1016/j.archger.2023.104995 [DOI] [PubMed] [Google Scholar]
  • 20.Johnson A, Bulgarelli L, Pollard T, Celi LA, Mark R, Horng S. MIMIC-IV-ED. PhysioNet. Available from: https://physionet.org/content/mimic-iv-ed/1.0/. Accessed 2025 November 19.
  • 21.Williams DM, Jylhävä J, Pedersen NL, Hägg S. A frailty index for UK Biobank participants. J Gerontol A Biol Sci Med Sci. 2019;74(4):582–7. doi: 10.1093/gerona/gly094 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 22.Hanlon P, Nicholl BI, Jani BD, Lee D, McQueenie R, Mair FS. Frailty and pre-frailty in middle-aged and older adults and its association with multimorbidity and mortality: a prospective analysis of 493 737 UK Biobank participants. Lancet Public Health. 2018;3(7):e323–32. doi: 10.1016/S2468-2667(18)30091-4 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 23.von Elm E, Altman DG, Egger M, Pocock SJ, Gøtzsche PC, Vandenbroucke JP, et al. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies. Lancet. 2007;370(9596):1453–7. doi: 10.1016/S0140-6736(07)61602-X [DOI] [PubMed] [Google Scholar]
  • 24.Benchimol EI, Smeeth L, Guttmann A, Harron K, Moher D, Petersen I, et al. The REporting of studies Conducted using Observational Routinely-collected health Data (RECORD) statement. PLoS Med. 2015;12(10):e1001885. doi: 10.1371/journal.pmed.1001885 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 25.Hegedus EJ, Moody J. Clinimetrics corner: the many faces of selection bias. J Man Manip Ther. 2010;18(2):69–73. doi: 10.1179/106698110X12640740712699 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 26.Naugler CT, Guo M. Mean abnormal result rate: proof of concept of a new metric for benchmarking selectivity in laboratory test ordering. Am J Clin Pathol. 2016;145(4):568–73. doi: 10.1093/ajcp/aqw041 [DOI] [PubMed] [Google Scholar]
  • 27.Morales DR, Guthrie B, Downes TJ, McAllister DA, Hanlon P. Applicability of the electronic frailty index in younger and older adults in England: a population-based cohort study. Lancet Healthy Longev. 2025;6(8):100752. doi: 10.1016/j.lanhl.2025.100752 [DOI] [PubMed] [Google Scholar]
  • 28.Tirkkonen A, Mak JKL, Eriksson JG, Halonen P, Jylhävä J, Hägg S, et al. Predicting cardiovascular morbidity and mortality with SCORE2 (OP) and Framingham risk estimates in combination with indicators of biological ageing. Age Ageing. 2025;54(4):afaf075. doi: 10.1093/ageing/afaf075 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 29.Blodgett JM, Theou O, Mitnitski A, Howlett SE, Rockwood K. Associations between a laboratory frailty index and adverse health outcomes across age and sex. Aging Med (Milton). 2019;2(1):11–7. doi: 10.1002/agm2.12055 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 30.Bai G, Wang Y, Mak JKL, Ericsson M, Hägg S, Jylhävä J. Is frailty different in younger adults compared to old? Prevalence, characteristics, and risk factors of early-life and late-life frailty in samples from Sweden and UK. Gerontology. 2023;69(12):1385–93. doi: 10.1159/000534131 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31.Wightman H, Quinn TJ, Mair FS, Lewsey J, McAllister DA, Hanlon P. Frailty in randomised controlled trials for dementia or mild cognitive impairment measured via the frailty index: prevalence and prediction of serious adverse events and attrition. Alzheimers Res Ther. 2023;15(1):110. doi: 10.1186/s13195-023-01260-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32.Blodgett JM, Rockwood K, Theou O. Changes in the severity and lethality of age-related health deficit accumulation in the USA between 1999 and 2018: a population-based cohort study. Lancet Healthy Longev. 2021;2(2):e96–104. doi: 10.1016/S2666-7568(20)30059-3 [DOI] [PubMed] [Google Scholar]
  • 33.Rockwood K, Song X, Mitnitski A. Changes in relative fitness and frailty across the adult lifespan: evidence from the Canadian National Population Health Survey. CMAJ. 2011;183(8):E487-94. doi: 10.1503/cmaj.101271 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 34.FitzGerald C, Hurst S. Implicit bias in healthcare professionals: a systematic review. BMC Med Ethics. 2017;18(1):19. doi: 10.1186/s12910-017-0179-8 [DOI] [PMC free article] [PubMed] [Google Scholar]

Decision Letter 0

Katarzyna Starowicz

4 Mar 2026

Dear Dr Rockwood,

Thank you for submitting your manuscript entitled "The FI-Lab: A Simple, Scalable, Electronic Health Measure of Frailty" for consideration by PLOS Medicine.

Your manuscript has now been evaluated by the PLOS Medicine editorial staff, and I am writing to let you know that we would like to send your submission out for external peer review.

However, before we can send your manuscript to reviewers, we need you to complete your submission by providing the metadata that is required for full assessment. To this end, please login to Editorial Manager where you will find the paper in the 'Submissions Needing Revisions' folder on your homepage. Please click 'Revise Submission' from the Action Links and complete all additional questions in the submission questionnaire.

For clinical studies, please upload a copy of your trial study protocol as a supporting information file. The study protocol should be the version submitted for approval to the institutional review board or ethics committee, should include any amendments to the study protocol, as well as the date of their approval by the institutional review or ethics committee. Please also detail any deviations from the study protocol in the Methods section of your manuscript. The editors will consider the protocol and study conduct prior to a final decision for external review.

Please re-submit your manuscript within two working days, i.e. by Mar 06 2026 11:59PM.

Login to Editorial Manager here: https://www.editorialmanager.com/pmedicine

Once your full submission is complete, your paper will undergo a series of checks in preparation for peer review. Once your manuscript has passed all checks it will be sent out for review.

Feel free to email us at plosmedicine@plos.org if you have any queries relating to your submission.

Kind regards,

Katarzyna Starowicz, PhD

Associate Editor

PLOS Medicine

kstarowicz@plos.org

Decision Letter 1

Katarzyna Starowicz

9 Apr 2026

Dear Dr Rockwood,

Many thanks for submitting your manuscript "The FI-Lab: A Simple, Scalable, Electronic Health Measure of Frailty" (PMEDICINE-D-26-00743R1) to PLOS Medicine. The paper has been reviewed by subject experts and a statistician; their comments are included below and can also be accessed here: [LINK]

After discussing the paper with the editorial team and an academic editor with relevant expertise, I'm pleased to invite you to revise the paper in response to the reviewers' comments. As you will see, the reviewers find your study important for the field however, they have some concerns which you need to address in full. We plan to send the revised paper to some or all of the original reviewers, and we cannot provide any guarantees at this stage regarding publication.

In addition to these revisions, you may need to complete some formatting changes, which you will receive in a follow up email. A member of our team will be in touch with a set of requests shortly. If you do not receive a separate email within a few days, please assume that checks have been completed, and no additional changes are required.

When you upload your revision, please include a point-by-point response that addresses all of the reviewer and editorial points, indicating the changes made in the manuscript and either an excerpt of the revised text or the location (eg: page and line number) where each change can be found. Please also be sure to check the general editorial comments at the end of this letter and include these in your point-by-point response. When you resubmit your paper, please include a clean version of the paper as the main article file and a version with changes tracked as a marked-up manuscript. It may also be helpful to check the guidelines for revised papers at http://journals.plos.org/plosmedicine/s/revising-your-manuscript for any that apply to your paper.

We ask that you submit your revision by Apr 30 2026 11:59PM. However, if this deadline is not feasible, please contact me by email, and we can discuss a suitable alternative.

Don't hesitate to contact me directly with any questions (kstarowicz@plos.org).

Best regards,

Katarzyna

Katarzyna Starowicz, PhD

Associate Editor

PLOS Medicine

kstarowicz@plos.org

-----------------------------------------------------------

Comments from the reviewers:

Reviewer #1: 1. Using standard Cox models for readmission outcomes ignores the strong competing risk of high 1-year mortality (e.g., 7.3% in MIMIC-IV-ED), which likely overestimates readmission risks. Despite citing "computational constraints," the authors should employ Fine-Gray or cause-specific hazard models (at least in sensitivity analyses) for methodological rigor

2. FI-Lab relies on standard laboratory reference ranges, which are typically for healthy adults and lack age stratification. The authors should discuss how this binary thresholding might misclassify normal age-related physiological changes as "deficits" and affect construct validity in older adults.

3. ED lab tests are ordered with high clinical suspicion. The authors should further discuss whether FI-Lab thresholds validated in this high-pretest-probability environment might overestimate its predictive performance when applied to routine community screening (e.g., UK Biobank).

4. The claim that FI-Lab is robust across diverse settings is overstated. The cohorts only represent high-income countries (US and UK) with mature EHR systems. Its applicability in resource-constrained environments lacks data support.

Minor: the finding that predictive discrimination plateaus at 20-40 features is highly valuable. This specific numerical recommendation should be explicitly added to the Abstract's conclusion to improve clinical utility.

Reviewer #2: This is a retrospective multicohort study that explored the optimal configuration of the Laboratory-based frailty indices (FI-Lab) in two emergency (MIMIC-IV-ED; KCH) and one community (UK Biobank) setting. This is an interesting study and the methods used appear largely robust. I have several comments:

1. Several exclusion criteria were applied (lines 107-110). Please specify the number excluded for each criteria.

2. The 'Important' tests were determined based on a random forest model but no details were provided. How was it defined based on the model? what was the underlying data? if using the same data for evaluation, would it be a cause for overfitting?

3. Feature floor - presumably this will affect the numbers of patients included in analysis. Please specify this and whether the analyses comparing different feature floors were based on the common set of patients. If the sample size differs, the comparison won't be fair; otherwise there'd be a selection bias skewed towards the sicker patients.

4. Table 2 is a good summary of the findings, though I'm not sure I agree that increasing feature floor 'generally improves' performance. with a few exceptions, the differences in C indices appeared quite small.

5. I also am not sure I fully understand Figure 3. Why the age panels have number of features as x-axis? And the comparison with age would dependent on how age is parametrerised. If age is expressed as per 5 years increase the HR would be greater, for example.

6. One notable limitation i think is that each dataset would have own 'normal' definition, which made them not comparable. This should at least be acknowledged. Additionally the setting are also not comparable between A&E and UK Biobank, which could explain the lower performance in UKB.

Reviewer #3: The authors seek to develop a Lab-FI for use in a clinical setting from commonly available biomarkers. This is an important and necessary step in the field of frailty medicine and may have real meaningful outcome for helping to triage patients at the point of care.

The manuscript is well written and the authors are commended for the extensive analyses done to validate the Lab-FI, which are quite sound. I have a few fundamental considerations for the authors to consider.

1.Given existing Lab-FI data from other cohorts, strong rationale is needed for a new framework, particularly one that appears to be validated on mortality specifically (eg Methods line 125 "'Important' tests for mortality) when frailty is distinctly a syndrome of vulnerability that predicts mortality but is not defined by it. Better clarification for how the Lab-FI components were selected is needed to frame this as a frailty tool. I suggest an anchor on the original FI-Lab developed by Howlett, Rockwood et al. Otherwise, this becomes yet another FI that may or may not be useful.

2.Moreover, comparison to NEWS2 makes this a very UK specific analysis. It would strengthen the results considerably to compare the Lab-FI to an eFI which is more broadly used for frailty assessment (vs mortality risk, as noted above).

3.Given the text in the manuscript, I was surprised to see that the Lab-FI includes biometric measures such as BP and HR which vary at times minute to minute. Additionally, inclusion of potentially toxic exposures (ethanol, acetaminophen) will be rarely measured outside of the emergency department making these less generalizable. Also, as above, how are these measures of frailty?

Reviewer #4: The authors present a thorough manuscript testing approaches to developing a laboratory marker-based frailty index using EHR data. This approach complements other approaches to assessing frailty in EHR data using diagnoses, procedures, and prescription data. The authors' work carefully explores the hyper-parameters of such an index using data from different EHR systems in two different countries and with different clinical contexts. On the whole this is a well-done study. There are a few points the authors may wish to clarify.

The first point is conceptual and should be at least mentioned in the discussion, though there is no clear resolution. Frailty is meant to capture reduced ability to respond to physiologic stressors. This suggests that values for acute phase reactants, such as WBC or lactate, might be inappropriately normal in acute illness in frail people. By relying largely on values from ED/A&E visits, this index may represent severity of acute illness rather than frailty. This may explain some of the differences in discrimination among the UK Biobank, KCH, and MIMIC-IV. This might also explain some of the decline in AUC with additional labs in MIMIC-IV.

Second, the details of how frailty was measured and used in the Cox regressions are unclear. The makeup of the 25-items FI used in the Cox regressions should be clearly presented in the appendix. The scale should also be made clear: an FI is typically a proportion between 0 and 1, however Figure 1 suggests it was measured on a percentage point scale. Cox regressions yielded HRs of ~2, a very strong relationship on the percentage point scale, but on the 0-1 scale implies an HR of 1.07 for a 0.1-increment change in FI, similar to the NEWS score. Finally, the FI distribution is a non-normal and asymmetric; bootstrapped confidence intervals may be more appropriate than analytic ones.

Third and finally, this is a paper investigating a different approach to assessing frailty and it is worth including a comparison with a more traditional FI and possibly with physical frailty. This is likely most feasible in the UK Biobank data. Both FIs and physical frailty have been developed in this cohort (see Williams DM et al. J Gerontol A Biol Sci Med Sci. 2019;74(4):582-587. and Hanlon P et al. Lancet Public Health. 2018;3(7):e323-e332.).

Minor Comments:

Figure S2 is not referenced in the manuscript and looks identical to Figure 1. It should be removed.

Any attachments provided with reviews can be seen via the following link: [LINK]

--------------------------------------------------------- ---

General editorial requests:

(Note: not all will apply to your paper, but please check each item carefully)

*We ask every co-author listed on the manuscript to fill in a contributing author statement, making sure to declare all competing interests. If any of the co-authors have not filled in the statement, we will remind them to do so when the paper is revised. If all statements are not completed in a timely fashion this could hold up the re-review process. If new competing interests are declared later in the revision process, this may also hold up the submission. Should there be a problem getting one of your co-authors to fill in a statement we will be in contact. Please do not add or remove authors without first discussing this with the handling editor. You can see our competing interests policy here: http://journals.plos.org/plosmedicine/s/competing-interests.

*Please upload any figures associated with your paper as individual TIF or EPS files with 300dpi resolution at resubmission; please read our figure guidelines for more information on our requirements: http://journals.plos.org/plosmedicine/s/figures. While revising your submission, we strongly recommend that you use PLOS's NAAS tool (https://ngplosjournals.pagemajik.ai/artanalysis) to test your figure files. NAAS can convert your figure files to the TIFF file type and meet basic requirements (such as print size, resolution), or provide you with a report on issues that do not meet our requirements and that NAAS cannot fix.

After uploading your figures to PLOS's NAAS tool - https://ngplosjournals.pagemajik.ai/artanalysis, NAAS will process the files provided and display the results in the "Uploaded Files" section of the page as the processing is complete.

If the uploaded figures meet our requirements (or NAAS is able to fix the files to meet our requirements), the figure will be marked as "fixed" above. If NAAS is unable to fix the files, a red "failed" label will appear above.

When NAAS has confirmed that the figure files meet our requirements, please download the file via the download option, and include these NAAS processed figure files when submitting your revised manuscript.

*Please ensure that the paper adheres to the PLOS Data Availability Policy (see http://journals.plos.org/plosmedicine/s/data-availability), which requires that all data underlying the study's findings be provided in a repository or as Supporting Information. For data residing with a third party, authors are required to provide instructions with contact information (web or email address) for obtaining the data. Please note that a study author cannot be the contact person for the data. PLOS journals do not allow statements supported by "data not shown" or "unpublished results." For such statements, authors must provide supporting data or cite public sources that include it.

*We expect all researchers with submissions to PLOS in which author-generated code underpins the findings in the manuscript to make all author-generated code available without restrictions upon publication of the work. In cases where code is central to the manuscript, we may require the code to be made available as a condition of publication. Authors are responsible for ensuring that the code is reusable and well documented. Please make any custom code available, either as part of your data deposition or as a supplementary file. Please add a sentence to your data availability statement regarding any code used in the study, e.g. "The code used in the analysis is available from Github [URL] and archived in Zenodo [DOI link]" Please review our guidelines at https://journals.plos.org/plosmedicine/s/materials-software-and-code-sharing and ensure that your code is shared in a way that follows best practice and facilitates reproducibility and reuse. Because Github depositions can be readily changed or deleted, we encourage you to make a permanent DOI'd copy (e.g. in Zenodo) and provide the URL.

*Please ensure that the study is reported according to the relevant guideline and include the completed [XXXX] checklist as Supporting Information. When completing the checklist, please use section and paragraph numbers, rather than page numbers. Please add the following statement, or similar, to the Methods: "This study is reported as per [XXXX] guideline (S1 Checklist)."

FORMATTING - GENERAL

*Abstract: Please structure your abstract using the PLOS Medicine headings (Background, Methods and Findings, Conclusions). Please combine the Methods and Findings sections into one section.

*At this stage, we ask that you include a short, non-technical Author Summary of your research to make findings accessible to a wide audience that includes both scientists and non-scientists. The Author Summary should immediately follow the Abstract in your revised manuscript. This text is subject to editorial change and should be distinct from the scientific abstract. Ideally each sub-heading should contain 2-3 single sentence, concise bullet points containing the most salient points from your study. In the final bullet point of 'What Do These Findings Mean?', please include the main limitations of the study in non-technical language. Please see our author guidelines for more information: https://journals.plos.org/plosmedicine/s/revising-your-manuscript#loc-author-summary.

*Please express the main results with 95% CIs as well as p values. When reporting p values please report as p<0.001 and where higher as the exact p value p=0.002, for example. Throughout, suggest reporting statistical information as follows to improve clarity for the reader "22% (95% CI [13%,28%]; p</=)". Please be sure to define all numerical values at first use.

*Please include page numbers and line numbers in the manuscript file. Use continuous line numbers (do not restart the numbering on each page).

*Please cite the reference numbers in square brackets. Citations should precede punctuation.

FIGURES AND TABLES

*Please provide titles and legends for all figures and tables (including those in Supporting Information files).

*Please define all abbreviations used in each figure/table (including those in Supporting Information files).

*Please consider avoiding the use of red and green in order to make your figure more accessible to those with color blindness.

SUPPLEMENTARY MATERIAL

*Please note that supplementary material will be posted as supplied by the authors. Therefore, please amend it according to the relevant comments outlined here.

*Please cite your Supporting Information as outlined here: https://journals.plos.org/plosmedicine/s/supporting-information

REFERENCES

*PLOS uses the numbered citation (citation-sequence) method and first six authors, et al.

*Please ensure that journal name abbreviations match those found in the National Center for Biotechnology Information (NCBI) databases (http://www.ncbi.nlm.nih.gov/nlmcatalog/journals), and are appropriately formatted and capitalised.

*Where website addresses are cited, please include the complete URL and specify the date of access (e.g. [accessed: 12/06/2023]).

*Please also see https://journals.plos.org/plosmedicine/s/submission-guidelines#loc-references for further details on reference formatting.

STUDY TYPE-SPECIFIC REQUESTS - please note not all might apply:

OBSERVATIONAL STUDIES

*Abstract: Please include the study design, population and setting, number of participants, years during which the study took place (enrollment and follow up), length of follow up, and main outcome measures.

*Please ensure that the study is reported according to the STROBE (or appropriate STOBE extension) guideline (available from: https://www.equator-network.org/reporting-guidelines/strobe) and include the completed STROBE (or STROBE extension) checklist as Supporting Information. Please add the following statement, or similar, to the Methods: "This study is reported as per the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guideline (S1 Checklist)." When completing the checklist, please use section and paragraph numbers, rather than page numbers.

*[FOR POPULATION HEALTH/REGISTRY STUDIES] Please ensure that the study is reported according to the RECORD guideline (available from https://www.record-statement.org) and include the completed checklist as Supporting Information. Please add the following statement, or similar, to the Methods: "This study is reported as per the Reporting of Studies Conducted using Observational Routinely-Collected Data (RECORD) guideline (S1 Checklist)." When completing the checklist, please use section and paragraph numbers, rather than page numbers.

*[FOR POPULATION HEALTH ESTIMATES] Please ensure that the study is reported according to the GATHER statement (available from https://www.equator-network.org/reporting-guidelines/gather-statement) and include the completed checklist as Supporting Information. Please add the following statement, or similar, to the Methods: "This study is reported as per the Guidelines for Accurate and Transparent Health Estimates Reporting (GATHER) statement (S1 Checklist)." When completing the checklist, please use section and paragraph numbers, rather than page numbers.

*[FOR MEDIATION ANALYSES] We recommend that the study is reported according to the AGReMA statement (https://agrema-statement.org/#:~:text=AGReMA%20is%20an%20evidence%2D%20and,randomised%20trials%20and%20observational%20studies) and include the completed checklist as Supporting Information. Please add the following statement, or similar, to the Methods: "This study is reported as per the Guideline for Reporting Mediation Analyses (AGReMA) statement (S1 Checklist)." When completing the checklist, please use section and paragraph numbers, rather than page numbers.

*For all observational studies, in the manuscript text, please indicate: (1) the specific hypotheses you intended to test, (2) the analytical methods by which you planned to test them, (3) the analyses you actually performed, and (4) when reported analyses differ from those that were planned, transparent explanations for differences that affect the reliability of the study's results. If a reported analysis was performed based on an interesting but unanticipated pattern in the data, please be clear that the analysis was data driven.

*Please state in the Methods section whether the study had a prospective protocol or analysis plan. If a prospective analysis plan (from your funding proposal, IRB or other ethics committee submission, study protocol, or other planning document written before analyzing the data) was used in designing the study, please include the relevant document(s) with your revised manuscript as a Supporting Information file to be published alongside your study and cite it in the Methods section. A legend for this file should be included at the end of your manuscript. If no such document exists, please make sure that the Methods section transparently describes when analyses were planned, and when/why any data-driven changes to analyses took place. Changes in the analysis, including those made in response to peer review comments, should be identified as such in the Methods section of the paper, with rationale.

MODELLING STUDIES

The following list is derived from Geoffrey P Garnett, Simon Cousens, Timothy B Hallett, Richard Steketee, Neff Walker. Mathematical models in the evaluation of health programmes. (2011) Lancet DOI:10.1016/S0140-6736(10)61505-X:

*If pertinent, please provide a diagram that shows the model structure, including how the natural history of the disease is represented, the process and determinants of disease acquisition, and how the putative intervention could affect the system.

*Please provide a complete list of model parameters, including clear and precise descriptions of the meaning of each parameter, together with the values or ranges for each, with justification or the primary source cited and important caveats about the use of these values noted.

*Please provide a clear statement about how the model was fitted to the data, including goodness-of-fit measure, the numerical algorithm used, which parameter varied, constraints imposed on parameter values, and starting conditions.

*For uncertainty analyses, please state the sources of uncertainties quantified and not quantified [can include parameter, data, and model structure].

*Please provide sensitivity analyses to identify which parameter values are most important in the model. Uncertainty estimates seek to derive a range of credible results on the basis of an exploration of the range of reasonable parameter values. The choice of method should be presented and justified.

*Please discuss the scientific rationale for the choice of model structure and identify points where this choice could influence conclusions drawn. Please also describe the strength of the scientific basis underlying the key model assumptions.

*For studies that develop a prediction model or evaluate its performance, please ensure that the study is reported according to the TRIPOD statement (https://www.equator-network.org/reporting-guidelines/tripod-statement) and include the completed checklist as Supporting Information. Please add the following statement, or similar, to the Methods: "This study is reported as per the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis Or Diagnosis (TRIPOD) statement (S1 Checklist)." For studies using machine learning, please use the TRIPOD-AI checklist. When completing the checklist, please use section and paragraph numbers, rather than page numbers.

Decision Letter 2

Katarzyna Starowicz

2 Jun 2026

Dear Dr Rockwood,

Many thanks for submitting your manuscript "The FI-Lab: A Simple, Scalable, Electronic Health Measure of Frailty" (PMEDICINE-D-26-00743R2) to PLOS Medicine. The paper has been reviewed by subject experts and a statistician; their comments are included below and can also be accessed here: [LINK]

Your manuscript has been see by all original reviewers. As you will see, the reviewers #3 and #4 feel that their concerns were not fully addressed. After discussing the paper with the editorial team and an academic editor with relevant expertise, I'm pleased to invite you to revise the paper in response to the reviewers' comments. We think that they have provided valuable feedback and we would like you to address their concerns in full.

We plan to send the revised paper to some or all of the original reviewers, and we cannot provide any guarantees at this stage regarding publication.

In addition to these revisions, you may need to complete some formatting changes, which you will receive in a follow up email. A member of our team will be in touch with a set of requests shortly. If you do not receive a separate email within a few days, please assume that checks have been completed, and no additional changes are required.

When you upload your revision, please include a point-by-point response that addresses all of the reviewer and editorial points, indicating the changes made in the manuscript and either an excerpt of the revised text or the location (eg: page and line number) where each change can be found. Please also be sure to check the general editorial comments at the end of this letter and include these in your point-by-point response. When you resubmit your paper, please include a clean version of the paper as the main article file and a version with changes tracked as a marked-up manuscript. It may also be helpful to check the guidelines for revised papers at http://journals.plos.org/plosmedicine/s/revising-your-manuscript for any that apply to your paper.

We ask that you submit your revision by Jun 16 2026 11:59PM. However, if this deadline is not feasible, please contact me by email, and we can discuss a suitable alternative.

Don't hesitate to contact me directly with any questions (kstarowicz@plos.org).

Best regards,

Katarzyna

Katarzyna Starowicz, PhD

Associate Editor

PLOS Medicine

kstarowicz@plos.org

-----------------------------------------------------------

Comments from the reviewers:

Reviewer #1: No further concern from my side.

Reviewer #2: Thank you for responding to my comments. The revised paper is much clearer and accurate.

Reviewer #3: I appreciate the thoughtful responses to my review.

I am left wondering, however, if the field would be best served by a different name for this instrument? FI-Lab invokes the cumulative deficit approach, whether the authors intend this or not, given how synonymous "FI" has become with eFI/Rockwood FI/cumulative deficit index etc.

I would also counter on the comparison to eFI - while the UK eFI is specific to outpatient there are many other electronic frailty indices that could be used as a comparison. As the US moves towards automated screening in hospitals as part of the CMS Age Friendly Health System measures, using the Rockwood FI approach, such a comparison could be valuable.

Reviewer #4: I appreciate the authors' response and technical thoroughness. I continue to have conceptual issues with this work, however. While the authors, in their response to my comments, attempt to distance what they present from frailty, they plainly call this a "frailty index" and describe its relationship to the deficit accumulation model of frailty. The idea of capturing physiologic dysregulation through lab values without accounting for whether these values could, in an acute setting, be inappropriately normal further suggests that this bears a stronger relationship to frailty and the patient's baseline status than to acute physiologic dysregulation when receiving emergency care. My concern that this is not meaningfully different from frailty conceptually and in terms of its applications is supported by their response to reviewer 3: "However, we do not want to frame FI-lab as solely a frailty tool but as a latent-health-factor measure—an index of underlying physiological dysregulation, of which clinical frailty in older adults is one manifestation and acute illness severity is another." They go on to state "We envisage a broader use to measure some factor of general health status across ages and settings." On some level the lab-based frailty index presented is simply an attempt to go down a level or two of physiology from the usual construction of a frailty index (some of which, such as the Atrium Health eFI and the CGA-FI, do explicitly incorporate lab values).

I think that the best way to address this conceptual issue, to contextualize its association alongside established related measures, and in line with the authors' stated vision for this index, is to explicitly compare the lab-FI with other established measures of frailty that rely more on diagnosis codes, as suggested previously.

Any attachments provided with reviews can be seen via the following link: [LINK]

--------------------------------------------------------- ---

Decision Letter 3

Katarzyna Starowicz

16 Jul 2026

Dear Dr. Rockwood,

Thank you very much for re-submitting your manuscript "The FI-Lab: A Simple, Scalable, Electronic Health Measure of Frailty" (PMEDICINE-D-26-00743R3) for review by PLOS Medicine.

I have discussed the paper with my colleagues and the academic editor and it was also seen again by two reviewers. I am pleased to say that provided the remaining reviewers, editorial and production issues are dealt with we are planning to accept the paper for publication in the journal.

The remaining issues that need to be addressed are listed at the end of this email. As during previous revision please include your rebuttal letter to all editorial and reviewers comments.

Any accompanying reviewer attachments can be seen via the link below. Please take these into account before resubmitting your manuscript:

[LINK]

***Please note while forming your response, if your article is accepted, you may have the opportunity to make the peer review history publicly available. The record will include editor decision letters (with reviews) and your responses to reviewer comments. If eligible, we will contact you to opt in or out.***

In revising the manuscript for further consideration here, please ensure you address the specific points made by each reviewer and the editors. In your rebuttal letter you should indicate your response to the reviewers' and editors' comments and the changes you have made in the manuscript. Please submit a clean version of the paper as the main article file. A version with changes marked must also be uploaded as a marked up manuscript file.

Please also check the guidelines for revised papers at http://journals.plos.org/plosmedicine/s/revising-your-manuscript for any that apply to your paper. If you haven't already, we ask that you provide a short, non-technical Author Summary of your research to make findings accessible to a wide audience that includes both scientists and non-scientists. The Author Summary should immediately follow the Abstract in your revised manuscript. This text is subject to editorial change and should be distinct from the scientific abstract.

In addition to these revisions, you may need to complete some formatting changes, which you will receive in a follow up email. A member of our team will be in touch with a set of requests shortly. If you do not receive a separate email within a few days, please assume that checks have been completed, and no additional changes are required.

We expect to receive your revised manuscript within 1 week. Please email us (plosmedicine@plos.org) if you have any questions or concerns.

We ask every co-author listed on the manuscript to fill in a contributing author statement. If any of the co-authors have not filled in the statement, we will remind them to do so when the paper is revised. If all statements are not completed in a timely fashion this could hold up the re-review process. Should there be a problem getting one of your co-authors to fill in a statement we will be in contact. YOU MUST NOT ADD OR REMOVE AUTHORS UNLESS YOU HAVE ALERTED THE EDITOR HANDLING THE MANUSCRIPT TO THE CHANGE AND THEY SPECIFICALLY HAVE AGREED TO IT.

Please ensure that the paper adheres to the PLOS Data Availability Policy (see http://journals.plos.org/plosmedicine/s/data-availability), which requires that all data underlying the study's findings be provided in a repository or as Supporting Information. For data residing with a third party, authors are required to provide instructions with contact information for obtaining the data. PLOS journals do not allow statements supported by "data not shown" or "unpublished results." For such statements, authors must provide supporting data or cite public sources that include it.

To enhance the reproducibility of your results, we recommend that you deposit your laboratory protocols in protocols.io, where a protocol can be assigned its own identifier (DOI) such that it can be cited independently in the future. Additionally, PLOS ONE offers an option to publish peer-reviewed clinical study protocols. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols

Please review your reference list to ensure that it is complete and correct. If you have cited papers that have been retracted, please include the rationale for doing so in the manuscript text, or remove these references and replace them with relevant current references. Any changes to the reference list should be mentioned in the rebuttal letter that accompanies your revised manuscript.

Please note, when your manuscript is accepted, an uncorrected proof of your manuscript will be published online ahead of the final version, unless you've already opted out via the online submission form. If, for any reason, you do not want an earlier version of your manuscript published online or are unsure if you have already indicated as such, please let the journal staff know immediately at plosmedicine@plos.org.

If you have any questions in the meantime, please contact me or the journal staff on plosmedicine@plos.org.

We look forward to receiving the revised manuscript by Jul 24 2026 11:59PM.

Sincerely,

Katarzyna Starowicz, PhD

Associate Editor, PLOS Medicine

kstarowicz@plos.org

------------------------------------------------------------

Requests from Editors:

GENERAL EDITORIAL REQUESTS

* Please confirm that your title complies with PLOS Medicine's style. Your title must be nondeclarative and not a question. It should begin with main concept if possible. "Effect of" should be used only if causality can be inferred, i.e., for an RCT. Please place the study design ("A randomized controlled trial," "A retrospective study," "A modelling study," etc.) in the subtitle (ie, after a colon).

e.g “Evaluation of FI-Lab, a laboratory-based, automated frailty index for acute care: a multicohort study” this is only an example and we are open for other suggestions.

* Please confirm that your abstract complies with our requirements, including format (three sections: Background, Methods and Findings, and Conclusions) and providing all the information relevant to this study type https://journals.plos.org/plosmedicine/s/submission-guidelines#loc-abstract

* Please ensure that the Introduction ends with a clear description of the study question or hypothesis.

* Please ensure that all abbreviations are defined at first use throughout the text.

* Please confirm that all numbers presented in the abstract are present and identical to numbers presented in the main manuscript text.

GENERAL

* Please disclose any use of AI tools and technologies to the study or article content at the end of the Methods. Please specify the name(s) and version(s) of any tool(s) used and a description of their use, including how the AI-assisted or generated content was evaluated by authors, and a clear statement of which aspects of the study, article contents, data, or supporting files were affected/generated by AI tool usage. Please also include a statement specifying who retains responsibility for the final manuscript. See our guidelines for more information: https://journals.plos.org/plosmedicine/s/ethical-publishing-practice#loc-artificial-intelligence-tools-and-technologies.

* Please review your text for claims of novelty or primacy (e.g. 'for the first time') and remove this language. In addition, please check that any use of statistical terms (such as trend or significant) are supported by the data, and if not please remove them.

* Please remove the 'conclusions' subheading from the discussion. Please also remove any other subheadings from the discussion.

* Statistical reporting: Please revise throughout the manuscript, including tables and figures.

- Please report statistical information as follows to improve clarity for the reader ""22% (95% CI [13,28]; p</=)"".

- Please separate upper and lower bounds with commas instead of hyphens as the latter can be confused with reporting of negative values.

- Please repeat statistical definitions (HR, CI etc.) for each set of parentheses."

* In the last sentence of the Abstract Methods and Findings section, please describe the main limitation(s) of the study's methodology.

* In the author summary, please revise formatting and ensure you use bullet points.

FUNDING STATEMENT

* The funding statement should include: specific grant numbers, initials of authors who received each award, URLs to sponsors’ websites. Also, please state whether any sponsors or funders (other than the named authors) played any role in study design, data collection and analysis, the decision to publish, or preparation of the manuscript. If they had no role in the research, include this sentence: “The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

COMPETING INTERESTS STATEMENT

* Please revise Competing Interest statement and remove first sentence – PhD’s student research is not a conflict.

DATA AVAILABILITY

* PLOS defines the “minimal data set” to consist of the data set used to reach the conclusions drawn in the manuscript with related metadata and methods, and any additional data required to replicate the reported study findings in their entirety. Authors do not need to submit their entire data set, or the raw data collected during an investigation. Please submit the following data:

The values behind the means, standard deviations and other measures reported;

The values used to build graphs;

The points extracted from images for analysis.

* The Data Availability Statement (DAS) requires revision. Please include DAS in your manuscript. For each data source used in your study:

a) If the data are freely or publicly available, note this and state the location of the data: within the paper, in Supporting Information files, or in a public repository (include the DOI or accession number)

b) If the data are owned by a third party but freely available upon request, please note this and state the institutional owner of the data set and contact information for data requests (insitutional web or email address). Note that a study author cannot be the contact for the data.

c) If the data are not freely available, please describe briefly the ethical, legal, or contractual restriction that prevents you from sharing it. Please also include an appropriate institutional contact (insitutional web or email address) for inquiries (again, this cannot be a study author).

* For studies in which a novel model is central to the manuscript's findings, as is the case here, authors are responsible for providing the source code needed to replicate the study's findings in a repository (such as GitHub, SourceForge or Bitbucket) or a cloud computing service (such as Code Ocean). Please include a citable digital object identifier (DOI) or other persistent identifier of the code (such as through linkage with Zenodo). Protection of authors’ intellectual property will not be cause for exception. Please explain in the manuscript’s Data Availability Statement how readers can access the shared code.

* Please include the statement on code availability in the data availability statement. Please see guidance here:https://journals.plos.org/plosmedicine/s/code-sharing-guidance#loc-data-availability-statement-examples

ETHICS AND CONSENT

* Please revise for use of patient-centered language. Please note that patient-centered language is constructed with the use of post-modified nouns (e.g. 'patients with psoriasis’ (or similar) instead of ‘psoriasis patients’) putting the person first in the sentence structure.

FIGURES

* Please show graph axes beginning at zero. If this is not possible, please show a break in the axis.

* When a p value is given, please specify the statistical test used to determine it in the legend.

* Please consider avoiding the use of red and green in order to make your figure more accessible

* Please provide the unadjusted comparisons as well as the adjusted comparisons in all relevant Tables

* Please specify the variables controlled for in all relevant Tables

* Please consider if moving some of the figures shared in the supplementary information into the main text would aid the reader. I am asking specifically for Sup Fig 1 (STROBE flow diagram) – I think moving it to the method section as Figure 1 will add clarity in the method section and provide quick overview of the work.

ACKNOWLEDGMENTS

* Please include an acknowledgment of study participants.

OBSERVATIONAL, COHORT, CROSS-SECTIONAL, AND CASE CONTROL STUDIES

* Thank you for provide RECORD checklist as Supplementary Table S1. We do ask to provide it separately as S1 Checklist file and attach as separate supplementary file.

* Your study is observational and therefore causality cannot be inferred. Please remove language that implies causality and refer to associations instead.

--------------------------------------------------------------------------

Comments from Reviewers:

Reviewer #3: I greatly appreciate the authors' very thoughtful response - and Dr. Rockwood's personal input here. The paper is much clearer and the additional analyses are very helpful. The point of "acute" vs "critical" illness unmasking frailty, often in those younger than expected is very nicely now made. This will be a very important contribution to the frailty literature.

Reviewer #4: Thank you for this very thoughtful response and further analyses, as well as a helpful outline of the history behind this work (and by extension our field).

The stratified results are very interesting, thank you for including them. While the text also mentions them, they do not appear in the supplement that was included in this revision (S4-S7 per line 362 on page 22 of the manuscript). It would also be helpful for readers to present the hazard ratios of the FI-Lab and UK Biobank FI in a regression of the two together to provide some comparison of the magnitude of association. I hope these points are not troublesome; if addressed I have no further suggestions for this manuscript.

Any attachments provided with reviews can be seen via the following link:

[LINK]

Decision Letter 4

Katarzyna Starowicz

24 Jul 2026

Dear Dr Rockwood,

On behalf of my colleagues and the Academic Editor, Sanjay Basu, I am pleased to inform you that we have agreed to publish your manuscript "Evaluation of the FI-Lab, a laboratory-based, automated frailty index for acute care: a multicohort study  " (PMEDICINE-D-26-00743R4) in PLOS Medicine.

Before your manuscript can be formally accepted you will need to complete some formatting changes, which you will receive in a follow up email. Please be aware that it may take several days for you to receive this email; during this time no action is required by you. Once you have received these formatting requests, please note that your manuscript will not be scheduled for publication until you have made the required changes.

In the meantime, please log into Editorial Manager at http://www.editorialmanager.com/pmedicine/, click the "Update My Information" link at the top of the page, and update your user information to ensure an efficient production process.

PRESS

We frequently collaborate with press offices. If your institution or institutions have a press office, please notify them about your upcoming paper at this point, to enable them to help maximise its impact. If the press office is planning to promote your findings, we would be grateful if they could coordinate with medicinepress@plos.org. If you have not yet opted out of the early version process, we ask that you notify us immediately of any press plans so that we may do so on your behalf.

We also ask that you take this opportunity to read our Embargo Policy regarding the discussion, promotion and media coverage of work that is yet to be published by PLOS. As your manuscript is not yet published, it is bound by the conditions of our Embargo Policy. Please be aware that this policy is in place both to ensure that any press coverage of your article is fully substantiated and to provide a direct link between such coverage and the published work. For full details of our Embargo Policy, please visit http://www.plos.org/about/media-inquiries/embargo-policy/.

To enhance the reproducibility of your results, we recommend that you deposit your laboratory protocols in protocols.io, where a protocol can be assigned its own identifier (DOI) such that it can be cited independently in the future. Additionally, PLOS ONE offers an option to publish peer-reviewed clinical study protocols. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols

Thank you again for submitting to PLOS Medicine. We look forward to publishing your paper.

Sincerely,

Katarzyna Starowicz, PhD

Associate Editor

PLOS Medicine

Associated Data

    This section collects any data citations, data availability statements, or supplementary materials included in this article.

    Supplementary Materials

    S1 Appendix. Supplementary methods, figures, and tables.

    (DOCX)

    pmed.1005004.s001.docx (9.6MB, docx)
    S1 Checklist. STROBE-RECORD Checklist.

    STROBE: von Elm E, Altman DG, Egger M, Pocock SJ, Gøtzsche PC, Vandenbroucke JP, et al. (2007). The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) Statement: Guidelines for Reporting Observational Studies. PLoS Med. 4(10): e296. https://doi.org/10.1371/journal.pmed.0040296. RECORD: Benchimol EI, Smeeth L, Guttmann A, Harron K, Moher D, Petersen I, Sørensen HT, von Elm E, Langan SM; RECORD Working Committee. (2015). The REporting of studies Conducted using Observational Routinely-collected health Data (RECORD) statement. PLoS Med. 12(10):e1001885. https://doi.org/10.1371/journal.pmed.1001885. Fig A: STROBE flow diagram of the study cohort selection process. Fig B: FI-Lab Stability Analysis in the MIMIC-IV Database. This plot shows the median coefficient of variation (CV) of FI-Lab scores against the number of requested features, categorised by the availability of vital signs and varying feature floor thresholds (0, 5, 10, 15, and 20) within the MIMIC-IV database. The number of patients included at each specific number of features requested is indicated by the labels on the plot. Fig C: FI-Lab Stability Analysis in the King’s College Hospital (KCH) Database. This plot illustrates the median coefficient of variation (CV) of FI-Lab scores as the number of requested features increases, stratified by the presence or absence of vital signs and different feature floor thresholds (0, 5, 10, 15, and 20) within the KCH dataset. Labels indicate the number of patients contributing to each data point. Fig D: FI-Lab Score Stability Against Simulated Analytical Noise. To test robustness against laboratory measurement uncertainty, a simulation was performed on 5,000 randomly sampled patients from the KCH database. For each patient, the FI-Lab score was recalculated 20 separate times, with noise (for example, increasing or decreasing a recorded Sodium by 2 mmol/l) introduced to their laboratory values prior to each calculation. The plot shows the median coefficient of variation (CV) of these 20 scores. The results demonstrate that score stability improves (i.e., CV decreases) as the number of features increases, with the effect plateauing after 30 features. Fig E. Model Performance (C-Statistic) by FI-Lab Configuration Without Vital Signs in the MIMIC-IV-ED Cohort. The figure displays the predictive performance (Concordance Index/C-Statistic) of the FI-Lab across six clinical outcomes, based on laboratory data alone. Performance is shown for the ‘Common’ and ‘Importance’ test selection strategies (panels). Within each plot, the x-axis indicates the number of laboratory features used, and colours denote the minimum feature floor required to calculate a score. A key pattern, particularly for the ‘Common’ strategy, is that performance does not increase indefinitely with more features. Instead, it often peaks when using 20–40 tests and then plateaus or declines, a trend visible across outcomes like in-hospital death, ED admission, and readmissions. The ‘Importance’ strategy tends to show more consistent improvement, though performance gains often level off at higher feature counts. Fig F. Model Performance (C-Statistic) by FI-Lab Configuration Without Vital Signs in the UK Biobank Cohort. Predictive performance (Concordance Index/C-Statistic) of FI-Lab across four clinical outcomes, based on laboratory data alone. Performance is shown for the ‘Common’ and ‘Importance’ test-selection strategies (panels); within each, the x-axis indicates the number of laboratory features used and colours denote the minimum feature floor required to calculate a score. Unlike the ED cohorts, where the ‘Common’ strategy peaks and then plateaus or declines, performance in UK Biobank generally increases with additional features—consistent with near-universal baseline testing that is largely independent of patient acuity. Fig G. Predictive Performance of FI-Lab Configurations in the MIMIC-IV Database. This figure illustrates the predictive performance (Concordance Index/ C-Statistic) of various FI-Lab configurations for six clinical outcomes within the MIMIC-IV cohort. Performance is plotted against the number of features used (x-axis). Panels compare feature selection based on ‘Common’ prevalence versus ‘Importance’ ranking. Different lines represent varying minimum feature prevalence thresholds (‘Feature Floor’). The six outcomes evaluated are 1-year mortality, 1-year readmission, 30-day readmission, ED admission, in-hospital death, and length of stay. Fig H. Performance of FI-Lab Configurations in the King’s College Hospital (KCH) Database (with Vitals). This figure shows the predictive performance (C-Statistic) of FI-Lab configurations for six outcomes in the KCH cohort, including vital signs as potential features. Performance is shown across a range of feature counts (x-axis), stratified by the feature selection method (‘Common’ versus ‘Importance’) and the minimum feature prevalence (‘Feature Floor’). The six outcomes evaluated are 1-year mortality, 1-year readmission, 30-day readmission, ED admission, in-hospital death, and length of stay. Fig I. Predictive Performance of FI-Lab Configurations in the King’s College Hospital (KCH) Database (without Vitals). This figure shows the predictive performance (C-Statistic) of FI-Lab configurations for six outcomes in the KCH cohort when vital signs have been excluded from the potential feature set. Performance is shown across a range of feature counts (x-axis), stratified by the feature selection method (‘Common’ versus ‘Importance’) and the minimum feature prevalence (‘Feature Floor’). The six outcomes evaluated are 1-year mortality, 1-year readmission, 30-day readmission, ED admission, in-hospital death, and length of stay. Fig J. Model Fit (AIC) for FI-Lab Configurations Across Outcomes and Datasets. AIC values for FI-Lab models across outcomes and datasets. Lower AIC indicates better model fit. Rows correspond to outcomes; columns to dataset–strategy combinations. Blank UK Biobank panels indicate unavailable outcomes; UK Biobank ‘1-Year Admission’ is shown under ‘1-Year Readmission’. Fig K. Average Patient Coverage by FI-Lab Configuration Across Datasets. Panels show dataset and feature-selection strategy (KCH, MIMIC-IV-ED, UK Biobank; Common versus Importance). Lines show the mean proportion of eligible patients for whom an FI-Lab score could be computed (averaged across available outcomes) as the number of features increases; colours denote the minimum-tests “floor” (0, 5, 10, 15, and 20). Coverage generally increases with more features and decreases with stricter floors. With floors ≤10, coverage approaches ~95%–100% by ~30–40 features in all datasets; a 20-test floor yields lower coverage at small feature counts most notably in UK Biobank under the Importance strategy, but rises as additional features are included. Fig L. Hazard Ratios for 1-Year Mortality Across Datasets and Feature-Selection Strategies. Columns show datasets (KCH, MIMIC-IV-ED, UK Biobank); rows show strategies (Common, Importance). Coloured lines represent FI-Lab hazard ratios by number of laboratory features and minimum-tests “floor”; shaded bands are 95% confidence intervals (CI). Red triangles show the median Age hazard ratio with whiskers indicating the range across floors/models; grey squares show the median NEWS2 hazard ratio with range (not available in UK Biobank). The y-axis is log2-scaled; the dashed line marks a hazard ratio of 1. Fig M. Hazard Ratios for In-Hospital Death Across FI-Lab Configurations, Age, and NEWS2 in KCH and MIMIC-IV-ED. Hazard ratios for in-hospital death are shown for FI-Lab (coloured lines, shaded 95% CI) across varying numbers of features and minimum tests required, for both the common and importance-based feature selection strategies. Red triangles indicate the median hazard ratio for age with range (minimum–maximum) across feature floors; grey squares indicate the median hazard ratio for NEWS2 with range. FI-Lab estimates represent the effect within the combined model. Hazard ratios are displayed on a log2 scale; the dashed horizontal line indicates a hazard ratio of 1.0. Fig N. Hazard Ratios for Length of Stay Across FI-Lab Configurations, Age, and NEWS2 in KCH and MIMIC-IV-ED. Hazard ratios for time to hospital discharge are shown for FI-Lab (coloured lines, shaded 95% CI) across varying numbers of features and minimum tests required (“feature floor”), for both the common and importance-based feature selection strategies. Red triangles indicate the median hazard ratio for age with range (minimum–maximum) across feature floors; grey squares indicate the median hazard ratio for NEWS2 with range. FI-Lab estimates represent the effect within the combined model. Hazard ratios are displayed on a log2 scale; the dashed horizontal line indicates a hazard ratio of 1.0. In this Cox model, hazard ratios < 1 indicate a lower likelihood of discharge at any given time (i.e., longer stay). FI-Lab showed a stronger association with prolonged stay than age. Fig O. Hazard Ratios for 1-Year Readmission Across FI-Lab Configurations, Age, and NEWS2 in KCH, MIMIC-IV-ED, and UK Biobank. Hazard ratios for 1-year readmission are shown for FI-Lab (coloured lines, shaded 95% CI) across varying numbers of features and minimum tests required (“feature floor”), for both the common and importance-based feature selection strategies. In the UK Biobank dataset, the outcome “1-Year Admission” is equivalent to “1-Year Readmission” in KCH and MIMIC-IV-ED. Red triangles indicate the median hazard ratio for age with range (minimum–maximum) across feature floors; grey squares indicate the median hazard ratio for NEWS2 with range. FI-Lab estimates represent the effect within the combined model for KCH and MIMIC-IV-ED, and within the FI-Lab model for UK Biobank; NEWS2 was unavailable in UK Biobank. Hazard ratios are displayed on a log2 scale; the dashed horizontal line indicates a hazard ratio of 1.0. Fig P. Hazard Ratios for ED Admission Across FI-Lab Configurations, Age, and NEWS2 in KCH and MIMIC-IV-ED. Hazard ratios for emergency department (ED) admission are shown for FI-Lab (coloured lines, shaded 95% CI) across varying numbers of features and minimum tests required (“feature floor”), for both the common and importance-based feature selection strategies. Red triangles indicate the median hazard ratio for age with range (minimum–maximum) across feature floors; grey squares indicate the median hazard ratio for NEWS2 with range. FI-Lab estimates represent the effect within the combined model; NEWS2 was unavailable in the UK Biobank dataset. Hazard ratios are displayed on a log2 scale; the dashed horizontal line indicates a hazard ratio of 1.0. Higher hazard ratios indicate a greater likelihood of admission from the ED. Table A. FI-Lab Hazard Ratios for 1-Year Mortality by Age Group. Hazard ratios for 1-year mortality across age bands in three cohorts. Models adjusted for age (within band) and gender. FI-Lab calculated using 25 most common laboratory tests with minimum 15 tests required for score calculation. All hazard ratios p < 0.001. HR, hazard ratio; CI, confidence interval. Table B. Hazard ratios for 1 year mortality by ethnic group. Hazard ratios for 1-year mortality stratified by ethnicity and race across three cohorts. Models adjusted for age and gender. FI-Lab calculated using 25 most common laboratory tests with minimum 15 tests required for score calculation. All hazard ratios p < 0.001. Interaction p-values test for differences in FI-Lab effect across ethnic/racial groups within each cohort. HR, hazard ratio; CI, confidence interval. Fig Q. FI-Lab Hazard Ratios for 1-Year Mortality Stratified by Ethnicity Across Three Cohorts. Forest plot showing hazard ratios (per standard deviation increase in FI-Lab) for 1-year mortality across ethnic groups in MIMIC-IV-ED (purple), King’s College Hospital (blue), and UK Biobank (green for 1-year, yellow for 10-year mortality). Points represent hazard ratio estimates with horizontal lines indicating 95% CI. The vertical dashed line at hazard ratio = 1.0 indicates no effect. Sample sizes and event counts are displayed in the table above. Results are grouped by ethnicity (White, Asian, Black, Chinese, Mixed, Other, Hispanic/Latino, Not stated) to facilitate comparison across datasets. The x-axis is truncated at 0.68–3.7 for clarity; one confidence interval with very few events extends beyond this range and is indicated by an arrow (full numeric results are provided in the table above). UK Biobank displays both 1-year and 10-year mortality results due to low event rates in some ethnic groups at 1 year. FI-Lab demonstrates significant associations with mortality across nearly all ethnic groups, with broadly consistent hazard ratios across datasets despite variation in baseline mortality risk and population composition. The Black and Chinese ethnic groups in UK Biobank 1-year follow-up did not reach statistical significance, likely reflecting low statistical power with 8 and fewer than 5 events, respectively. Table C. Descriptive statistics and feature importance for the top 70 tests (including vitals) derived from the MIMIC database. Abbreviations: I, Icterus Index; H, Haemolysis Index; L, Lipaemia Index. Table D. Descriptive statistics and feature importance for the top 63 tests (excluding vitals) derived from the MIMIC database. Abbreviations: I, Icterus Index; H, Haemolysis Index; L, Lipaemia Index. Table E. Descriptive statistics and feature importance for the top 63 tests (excluding vitals) derived from UK Biobank. Table F. Descriptive statistics and feature importance for the top 50 tests (excluding vitals) derived from KCH. Fig R. Relationship between test frequency and abnormality rate in emergency department patients. Scatter plot showing the proportion of abnormal results versus total test frequency (log scale) for blood tests from emergency department attendances in the MIMIC-IV-ED cohort. Each point represents a unique laboratory test, colored by abnormality rate band (grey: no abnormal flags; dark purple to green: increasing proportions of abnormal results). Vertical dashed lines indicate frequency cutoffs for the top 30 and top 70 most commonly ordered tests. Commonly ordered tests (right side) demonstrate a range of abnormality rates, typically 10%–50%. Notably, very high-abnormality rates (>70%) are observed exclusively among infrequently ordered tests (left side), such as CD4/CD8 ratio (76% abnormal), Free Kappa (75%), and NTproBNP (70%). These high-abnormality, low-frequency tests represent investigations ordered only when clinicians have high pre-test probability of a specific condition, resulting in substantial enrichment for abnormal results compared to the baseline prevalence in the general emergency department population. Table G. Hazard ratios (95% CI) for mortality per SD-increase in FI-Lab within strata of the Williams and colleagues frailty index at baseline. Table H. Hazard ratios (95% CI) for mortality per SD-increase in FI-Lab across baseline levels of the adapted Frailty Phenotype. Table I. Hazard ratios (95% CI) for 10-year mortality at specified levels of FI-Lab and Williams and colleagues frailty index in UK Biobank—both variables in same model adjusted for age and sex.

    (DOCX)

    pmed.1005004.s002.docx (22.1KB, docx)
    Attachment

    Submitted filename: Response_to_Reviewers_PMEDICINE-D-26-00743_R2_May4-.docx

    pmed.1005004.s003.docx (45.7KB, docx)
    Attachment

    Submitted filename: PMEDICINE D-26-00743_R3 REPLY TO REVIEWERS JUNE 22.docx

    pmed.1005004.s004.docx (2.1MB, docx)
    Attachment

    Submitted filename: Response to Reviewers - July 23.docx

    pmed.1005004.s005.docx (32.7KB, docx)

    Data Availability Statement

    UK Biobank data are available for use subject to application and approval of a data sharing agreement (https://www.ukbiobank.ac.uk/use-our-data/apply-for-access/). MIMIC-IV-ED data are freely available upon request to the Massachusetts Institute of Technology (https://mimic.mit.edu/docs/faq/how-to-get-access.html). King’s College Hospital data are not publicly available; access is typically limited to approved researchers who meet the relevant data access and ethics requirements. Interested users can contact the King’s Electronic Records Research Interface (KERRI) at kch-tr.cogstackrequests@nhs.net. All analysis code for FI-Lab is available at GitHub (https://github.com/jhfl2/FI-Lab-Recipe) or https://doi.org/10.5281/zenodo.21686996.

    Articles from PLOS Medicine are provided here courtesy of PLOS

    RESOURCES