Pragmatic Screening for Sarcopenia in Non‐Small Cell Lung Cancer: Development and Internal Validation of a Prospective Risk Model Integrating T12 CT Indices and Routine Clinical Variables

ABSTRACT

Background

Sarcopenia or low muscle mass is common in non‐small cell lung cancer and predicts poorer outcomes, but its assessment typically relies on CT‐based skeletal muscle index at the L3 vertebral level. Standard chest CTs often omit the L3 level, leaving many patients without a straightforward muscle mass measure, so sarcopenia frequently goes undetected. We therefore evaluated whether thoracic CT measurements could reliably substitute for L3 SMI and developed a simple, accurate clinical tool to identify patients at risk of CT‐defined low muscle mass.

Methods

In retrospective (n = 192) and prospective (n = 177) cohorts of NSCLC patients, cross‐sectional muscle area was quantified on CT at L3, T12, and T4. Deming regression and Bland–Altman analysis were used to assess correlations and agreement between T4‐ or T12‐derived SMI and reference L3 SMI. Candidate clinical predictors were selected based on univariate associations and clinical plausibility, and machine‐learning methods identified five routine variables, which were combined into the Lung Cancer Patients' Sarcopenia Risk Model (LSRM). Model performance was evaluated in terms of discrimination, calibration, and risk stratification and compared with established clinical indices like BMI and the advanced lung cancer inflammation index (ALI).

Results

T12‐derived SMI showed a strong correlation and minimal bias versus L3, outperforming T4‐derived SMI. A conversion equation was established to estimate L3 SMI from T12. The LSRM, incorporating age, body mass index, carcinoembryonic antigen, C‐reactive protein, and lymphocyte count, demonstrated good discrimination, satisfactory calibration, and clear three‐tier risk stratification. It outperformed BMI and ALI in identifying CT‐defined low muscle mass.

Conclusion

T12 measurements on routine chest CT can replace L3 for muscle assessment. The LSRM provides a practical bedside tool for screening patients at risk of CT‐defined low muscle mass in NSCLC without additional imaging, supporting earlier risk identification and integration into routine care.

Abbreviations

ALI

advanced lung cancer inflammation index

AUC

area under the roc curve

AWGS

Asian Working Group for Sarcopenia

BMI

Body mass index

CEA

carcinoembryonic antigen

CHD

coronary heart disease

CRP

C‐reactive protein

DXA

dual‐emission X‐ray absorptiometry

EWGSOP2

European Working Group on Sarcopenia in Older People 2

FEV1

forced expiratory volume in 1 second

FVC

forced vital capacity

Hb

hemoglobin

HU

hounsfield unit

LASSO

least absolute shrinkage and selection operator

LSRM

Lung Cancer Patients' Sarcopenia Risk Model

LYM

lymphocyte count

NEU

neutrophil count

NSCLC

non‐small cell lung cancer

PHS

postoperative hospital stay

PLT

platelet count

RFECV

recursive feature elimination with cross‐validation

ROC

receiver operating characteristic

SMA

skeletal muscle area

SMI

skeletal muscle index

SMRA

skeletal muscle radiation attenuation

WBC

white blood cell count

1. Introduction

Sarcopenia, the pathological loss of skeletal muscle mass and function, is common in patients with NSCLC and materially affects treatment tolerance, perioperative risk, and long‐term survival [1, 2]. In contemporary cohorts, roughly one in three to one in two patients with NSCLC meet CT‐based criteria for low muscle mass, and sarcopenia has been repeatedly associated with increased postoperative complications, chemotherapy toxicity, and worse overall survival [3, 4]. Accurate and practical identification of sarcopenia is therefore critical for risk stratification and for selecting patients who may benefit from timely, targeted prehabilitation, nutritional interventions, or tailored oncologic approaches [5].

Computed tomography (CT) offers an objective, opportunistic approach to quantify skeletal muscle, as cross‐sectional images from routine oncologic staging can be repurposed to measure muscle area [6]. The skeletal muscle index at the third lumbar vertebra (L3‐SMI), normalized to height, has emerged as a widely used reference because it correlates well with whole‐body lean mass and with clinical endpoints across tumor types [7]. However, abdominal imaging that includes L3 is not always available for NSCLC patients who undergo chest CT without contiguous abdominal coverage [8]. As a result, reliance on L3‐SMI restricts the practical applicability of CT‐based sarcopenia screening in routine thoracic oncology workflows. Although several studies have explored thoracic substitutes (notably T4, T12, and others) for L3‐SMI and reported promising correlations [9, 10], few investigations have provided rigorously derived conversion equations or thoracic cutoffs calibrated to established L3 definitions. Consequently, uncertainty persists about whether and how thoracic measurements can be translated into the L3 framework for accurate, individual‐level patient classification in clinical practice.

Even when imaging identifies patients who already manifest low muscle mass, CT alone does not facilitate the early detection of those at risk of progressive muscle loss before it becomes radiologically apparent [5]. There is, therefore, a complementary clinical demand for simple, bedside prediction tools that flag patients at high risk of sarcopenia using data routinely collected during preoperative assessment—demographics, anthropometry, and standard laboratory tests. A parsimonious, transparent model that relies on routinely available covariates and demonstrates robust performance would enable scalable screening, improve accuracy beyond crude markers such as BMI or ALI, prompt multidisciplinary intervention, and better allocation of supportive resources [11].

Given the above limitations, we aim to develop a clinically actionable tool for preoperative screening that performs adequately when L3 imaging is unavailable. Specifically, we sought to integrate opportunistic thoracic CT measurements with data‐driven feature selection to construct a model that is both more convenient, based on routinely obtainable variables, and more accurate than existing simple indices. The approach combines opportunistic imaging with machine‐learning‐informed variable selection to produce a reproducible, bedside‐friendly instrument for identifying patients at risk of CT‐defined low muscle mass and facilitating subsequent comprehensive sarcopenia assessment.

2. Materials and Methods

2.1. Study Design and Participants

We retrospectively identified patients with pathologically confirmed NSCLC who underwent clinically indicated chest and abdominal CT scans at Sichuan Provincial People's Hospital between January 2024 and February 2025. Of 305 patients initially screened, 192 patients remained after application of prespecified exclusion criteria and were included in the retrospective imaging‐validation cohort. Eligible subjects had contemporaneous height and weight measurements recorded, as well as both a chest CT scan that included the fourth (T4) and twelfth (T12) thoracic vertebral levels and an abdominal CT scan that included the third lumbar (L3) level. Exclusion criteria were as follows: prior major thoracoabdominal surgery that could alter muscle anatomy, CT studies with severe motion or reconstruction artifact precluding muscle segmentation, missing key anthropometric or imaging data, or absence of either the chest or abdominal CT.

For the prospective cohort study, a total of 261 patients were enrolled between March 2025 and November 2025, comprising 188 patients from Sichuan Provincial People's Hospital and 73 from Sichuan Cancer Hospital. After applying prespecified exclusion criteria—including prior major thoracoabdominal surgery that could alter muscle anatomy, CT studies with severe motion or reconstruction artifact precluding muscle segmentation, missing key anthropometric or imaging data, or absence of chest CT—177 patients remained, consisting of 122 from Sichuan Provincial People's Hospital and 55 from Sichuan Cancer Hospital. Baseline demographic data, clinical characteristics, routine laboratory tests, and chest CT‐derived body composition measures were collected at enrollment. Postoperative hospital stay (PHS) was additionally recorded for each patient and defined as the number of days from the date of surgery to the date of hospital discharge. T12‐based sarcopenia status, derived from sex‐specific T12‐SMI cutoffs (see below), served as the primary outcome for model development, with PHS analyzed as an exploratory clinical endpoint.

The flowchart of this study is shown in Figure 1. This study was conducted in accordance with the Declaration of Helsinki and approved by the Medical Ethics Committees of Sichuan Provincial People's Hospital (Approval No. 2025‐565) and Sichuan Cancer Hospital (Approval No. SCCHEC‐02‐2018‐043).

FIGURE 1.

Flowchart of participant inclusion. Hospital A refers to Sichuan Provincial People's Hospital and hospital B refers to Sichuan Cancer Hospital. Patients who met multiple exclusion criteria were assigned to the primary reason based on a hierarchical priority rule (missing data > CT artifacts > lack of CT coverage > prior surgery) to ensure mutually exclusive categorization.

2.2. CT Image Acquisition and Analysis

Preoperative staging at both Sichuan Provincial People's Hospital and Sichuan Cancer Hospital followed a harmonized imaging and analysis protocol. All examinations were performed on 64‐slice helical CT scanners (SOMATOM Force, Siemens Healthineers, at both centers) with the following acquisition parameters: tube voltage, 120 kVp; automatic tube current modulation (reference mAs); slice thickness, 2.0 mm; reconstruction interval, 2.0 mm; pitch, 0.8. Both centers used the same contrast protocols for thoracic imaging. To minimize inter‐scanner and inter‐site variation, scanner models and reconstruction parameters were recorded, and image Hounsfield unit (HU) calibration was checked during study setup; any systematic offsets were addressed according to a prespecified harmonization procedure.

Thoracic (T4, T12) and lumbar (L3) levels were identified on axial images using a standardized anatomical protocol (pedicle level landmarks) with explicit rules for slice selection when vertebral morphology was ambiguous; these rules were applied identically at both centers. Regions of interest (ROIs) were delineated using 3D Slicer (v4.11) with a semi‐automated workflow: initial segmentation used HU thresholds of −29 to +150 HU [12, 13], followed by manual refinement to exclude non‐muscle structures (vascular structures, intramuscular fat islands, and reconstruction artifacts). Mean skeletal muscle radiation attenuation (SMRA, HU) was calculated as the mean HU within the final ROI and skeletal muscle area (SMA, cm²) was computed from segmented voxels. SMA was normalized to measured standing height squared (SMI, SMA/height², cm²/m²) [14]; height and weight were measured at enrollment using a standardized protocol at each site.

To ensure reliability, two experienced readers at each center independently performed all segmentations while blinded to clinical outcomes and other reader measurements. To assess inter‐reader reliability of CT‐based body composition measurements, agreement between the two independent readers was evaluated using intraclass correlation coefficients (ICCs). A two‐way random‐effects model with absolute agreement and single‐measure reliability [ICC(A,1)] was applied for skeletal muscle radiation attenuation (SMRA), T4‐SMI, T12‐SMI, and L3‐SMI. ICC estimates were calculated for the overall cohort and stratified by sex. ICC values were interpreted according to commonly used thresholds, with values > 0.90 considered excellent, 0.75–0.90 good, 0.50–0.75 moderate, and < 0.50 poor reliability. Finally, sarcopenia classification used the predefined L3‐SMI cutoffs (men 40.3 cm²/m²; women 30.8 cm²/m²) [15, 16] to permit comparability with prior studies.

For the retrospective imaging‐validation cohort, we quantified the relationships between thoracic and lumbar SMI using Spearman's rank correlation (T4‐SMI vs. L3‐SMI and T12‐SMI vs. L3‐SMI). Deming regression was then applied with L3‐SMI as the dependent variable and T4‐SMI or T12‐SMI as the independent variable to derive linear conversion equations that account for measurement error in both variables. These equations were used to transform established L3‐SMI sarcopenia cutoffs into sex‐specific thresholds at T4 and T12. Agreement between thoracic‐derived and lumbar‐derived SMI values was evaluated with Bland–Altman analysis by calculating mean bias and 95% limits of agreement for T4 → L3 and T12 → L3 conversions. Based on these analyses, T12‐SMI was selected as the preferred thoracic surrogate and its derived cutoffs were used to define sarcopenia in the prospective cohort.

2.3. Machine‐Learning

Missing data were handled using multiple imputation for all variables. Multiple imputation was performed using the R package mice, with 10 imputations (m = 10). Two complementary feature‐selection methods were applied in parallel to each imputed dataset. Penalized logistic regression (LASSO) was performed using glmnet::cv.glmnet (family = binomial, α = 1) with fivefold cross‐validation; predictors with nonzero coefficients at λmin were recorded for each imputation and retained as LASSO candidates if selected in ≥ 60% of imputations. Recursive feature elimination with cross‐validation (RFECV) was implemented via caret::rfe using random forest as the internal learner and repeated cross‐validation (fivefolds × 10 repeats); predictors appearing in ≥ 60% of imputations were retained as RFECV candidates. The intersection of LASSO and RFECV selections defined the parsimonious predictor set used for multivariable modeling. Continuous predictors were entered into the feature‐selection procedures and multivariable logistic regression using their original measurement scales. No logarithmic transformation, standardization, or normalization was applied before model development.

2.4. Model Development and Performance Evaluation

This study follows the TRIPOD guidelines for prediction model development, focusing on predictive performance rather than inference or identification of causal risk factors. Multivariable logistic regression models were fitted across multiply imputed datasets and pooled using Rubin's rules (mice::with and mice::pool); pooled β estimates, standard errors, p values, and odds ratios were reported. An individual LSRM score was computed per subject in each imputation using the pooled coefficients; LPs were averaged across imputations and transformed to probabilities (logistic function) for subsequent evaluation. Discrimination was assessed by the area under the ROC curve (AUC) (using pROC) computed from the averaged predicted probabilities; internal validation and optimism correction were performed by bootstrap resampling (B = 500). Calibration was evaluated graphically by decile groups with a loess‐smoothed calibration curve and tested by Hosmer–Lemeshow. For the clinical presentation, the final model was translated into a nomogram using the rms package on a single, completed, and imputed dataset.

Discrimination of the LSRM for imaging‐defined sarcopenia was assessed and compared with three simple benchmarks: BMI alone, ALI alone, and a two‐variable model including BMI and age (BMI + age). For the LSRM, we used the published multivariable logistic regression coefficients to compute a linear predictor (LP) for each patient and converted the LP to a predicted probability by the logistic transform. For the comparator models, we fitted univariable or bivariable logistic regressions on the same dataset and used the modeled predicted probabilities for ROC analysis. Discrimination was quantified by the area under the receiver‐operating characteristic curve with 95% confidence intervals obtained by bootstrap resampling (2000 replicates). Pairwise AUC comparisons were performed with DeLong's test for correlated ROC curves. Where multiple pairwise comparisons were reported, p values were Bonferroni‐adjusted.

LSRM predicted probabilities were used to derive three clinically interpretable risk strata (low, intermediate and high). The lower cutoff (cut1) was selected to maximize the Youden index on the ROC curve for LSRM. The upper cutoff (cut2) was chosen to identify a high‐specificity threshold: specifically, the smallest LSRM probability that achieved at least 90% specificity on the ROC curve. Patients were assigned to Low (LSRM ≤ cut1), intermediate (cut1 < LSRM ≤ cut2), or high (LSRM > cut2) strata. Between‐group differences in continuous LSRM scores were tested with the Kruskal–Wallis test; pairwise post hoc comparisons were performed using Wilcoxon rank‐sum tests with Bonferroni correction. Trend across ordered categories was tested using the Cochran–Armitage trend test; a standard χ ² test or Fisher's exact test was used to provide an alternative omnibus comparison. Odds ratios comparing intermediate and high versus low strata were obtained from logistic regression models. All computations and figures were produced in R (using packages including mice, pROC, and ggplot2) with a fixed random seed to ensure reproducibility.

2.5. Statistical Analysis

Data are presented as mean ± SD for continuous variables and n (%) for categorical variables. Comparisons between males and females were performed using Student's t‐test for continuous variables and the chi‐square test for categorical variables. Groups between non‐sarcopenia and sarcopenia used independent‐samples t‐tests for normally distributed variables (Welch correction applied if variances were unequal) and Mann–Whitney U tests for non‐normal variables. Categorical variables were presented as counts and percentages and compared using Pearson's chi‐square test; Fisher's exact test was used when any expected cell count was < 5.

Missing data were recorded, and the sample size (n) used for each comparison was explicitly reported; analyses were performed on available cases for each variable. For multivariable prediction modeling and machine‐learning feature selection, missing predictor values were handled by multiple imputation as described above, rather than by complete‐case analysis, to reduce bias and loss of precision. All tests were two‐sided, and a significance threshold of p < 0.05 was used. Statistical analyses were performed in R (version 4.3.1) using standard packages for data import and testing. Exploratory analyses of PHS used linear regression, with model assumptions assessed appropriately.

2.6. Software and Reporting Standards

All analyses were performed in R (version 4.3.1). Multiple imputation and pooling used mice, penalized regression used glmnet, random‐forest RFECV used caret, model development and nomogram construction used rms, and ROC/correlation plots were generated with pROC and ggplot2. Study design, analysis, and reporting were aligned with STROBE and TRIPOD recommendations where applicable.

3. Results

3.1. Retrospective Imaging‐Validation Cohort: Thoracic Surrogates for Lumbar Skeletal Muscle Index

3.1.1. Cohort Characteristics

A total of 192 participants were included in the retrospective imaging‐validation cohort. Baseline demographics, anthropometric measures, SMRA, and SMI at T4, T12, and L3 levels are summarized in Table 1. Except for BMI, all variables showed significant sex‐specific differences. Inter‐reader reliability of CT‐based measurements was also conducted; CT‐derived muscle attenuation showed the highest reproducibility, while T12‐SMI demonstrated the best inter‐reader agreement among thoracic measurements (Table S1).

TABLE 1.

Baseline demographics and imaging‐derived characteristics of the study population.

Variables	Overall (n = 192)	Male (n = 79)	Female (n = 113)	p
Demographics
Age, years	56.2 ± 11.6	59.5 ± 11.0	53.9 ± 11.6	< 0.001
Sex, n (%)	192 (100%)	79 (41.1%)	113 (58.9%)
Anthropometrics
Height, cm	160.4 ± 7.5	166.3 ± 6.3	156.3 ± 5.3	< 0.001
Weight, kg	59.9 ± 9.5	65.6 ± 9.2	56.0 ± 7.6	< 0.001
BMI, kg/m2	23.2 ± 3.1	23.7 ± 3.1	22.9 ± 3.0	0.086
CT‐derived
Skeletal muscle radiation attenuation (SMRA), HU	32.5 ± 9.8	36.3 ± 9.6	29.7 ± 9.0	< 0.001
Skeletal muscle index (SMI), cm2/m2
L3‐SMI	36.7 ± 6.1	39.9 ± 6.3	34.5 ± 4.8	< 0.001
T12‐SMI	31.7 ± 6.7	35.2 ± 7.1	29.2 ± 5.3	< 0.001
T4‐SMI	67.1 ± 11.1	74.9 ± 9.2	61.7 ± 8.8	< 0.001

Note: Continuous variables are presented as mean ± standard deviation, and categorical variables as number (percentage). L3, T12, and T4 represent the third lumbar, twelfth thoracic, and fourth thoracic vertebral levels, respectively. SMI was calculated as the cross‐sectional skeletal muscle area normalized by height squared (cm²/m²).

Abbreviations: BMI, body mass index; SMI, skeletal muscle index; SMRA, skeletal muscle radiation attenuation.

3.1.2. Correlation and Deming Regression Between Thoracic and Lumbar SMI

We next assessed how well thoracic muscle measurements at T4 and T12 approximated the lumbar reference standard (L3‐SMI). Spearman^'s rank correlation demonstrated strong positive associations between thoracic and lumbar SMI. T4‐SMI correlated with L3‐SMI (Spearman^'s ρ = 0.71, p < 0.001, n = 192), whereas T12‐SMI showed an even stronger correlation with L3‐SMI (Spearman^'s ρ = 0.86, p < 0.001, n = 192).

To quantify these relationships while accounting for measurement error in both variables, we fitted Deming regression models (Figure 2A,B). The fitted conversion equations were:

FIGURE 2.

Agreement and conversion performance of T4‐ and T12‐derived SMI compared with L3‐SMI. Panels (A) and (B) show the Deming regression between L3‐SMI and T4‐SMI and between L3‐SMI and T12‐SMI, respectively. The solid line represents the Deming regression fit accounting for measurement errors in both variables, and the dashed line indicates the line of identity (y = x). Panels (C) and (D) display the corresponding Bland–Altman plots after mapping T4‐SMI and T12‐SMI onto the L3 scale using the Deming regression equations. The solid horizontal line represents the mean bias, and the dashed lines indicate the 95% limits of agreement (mean bias ±1.96 SD).

L3‐SMI = 12.88 + 0.43 × T4‐SMI (slope 0.43, intercept 12.88).

L3‐SMI = 12.37 + 0.92 × T12‐SMI (slope 0.92, intercept 12.37).

The slope for T12 being closer to unity than that for T4 indicates a greater proportional agreement between T12‐SMI and the lumbar standard. Applying the Deming‐derived conversion equations to the established L3‐SMI cutoffs produced equivalent thoracic thresholds: for T4, 63.8 cm²/m² (male) and 41.7 cm²/m² (female); for T12, 30.4 cm²/m² (male) and 20.0 cm²/m² (female) (Table S2). These converted thresholds may be used to classify sarcopenia on thoracic slices when lumbar measurements are unavailable.

3.1.3. Agreement and Selection of T12 as the Representative Thoracic Level

Agreement between thoracic‐derived and lumbar‐derived SMI values was further assessed using Bland–Altman analysis (Figure 2C,D). The 95% limits of agreement (LoA) were narrower for T12 → L3 than for T4 → L3 (Table S3), demonstrating superior interchangeability of T12‐SMI with L3‐SMI at the individual‐patient level.

In summary, across correlation, Deming regression, and Bland–Altman analyses, T12‐SMI demonstrated stronger correlation and tighter agreement with L3‐SMI than T4‐SMI. On the basis of these concordant findings, T12 was selected as the representative thoracic level, and its SMI thresholds were used as the imaging definition of sarcopenia in the subsequent prospective study in NSCLC patients.

3.2. Prospective Study: Development and Internal Validation of the LSRM

3.2.1. Patient Characteristics by Sarcopenia Status

A total of 177 patients were prospectively enrolled and classified as sarcopenic or nonsarcopenic according to the prespecified T12‐SMI threshold. Baseline characteristics are summarized in Table 2.

TABLE 2.

Baseline characteristics of the study population according to Sarcopenia status.

Variables	Overall	Non‐sarcopenia	sarcopenia	p
Demographics
Age, years	54.0 ± 13.7	54.0 ± 13.2	54.0 ± 15.7	0.998
Sex, n (%)				0.580
Male	61 (34.5%)	50 (35.5%)	11 (30.6%)
Female	116 (65.5%)	91 (64.5%)	25 (69.4%)
Anthropometrics
Height, cm	160.7 ± 7.0	160.7 ± 7.2	160.9 ± 6.4	0.576
Weight, kg	60.4 ± 11.1	62.1 ± 11.4	53.7 ± 6.6	< 0.001
BMI, kg·m−2	23.3 ± 3.4	24.0 ± 3.4	20.7 ± 1.8	< 0.001
SMI, cm2/m2	27.7 ± 7.4	29.5 ± 6.9	20.9 ± 5.0	< 0.001
Comorbidities
Hypertension, n (%)				0.168
Yes	28 (15.8%)	25 (17.7%)	3 (8.3%)
No	149 (84.2%)	116 (82.3%)	33 (91.7%)
CHD, n (%)				1.000
Yes	8 (4.5%)	7 (5.0%)	1 (2.8%)
No	169 (95.5%)	134 (95.0%)	35 (97.2%)
Arrhythmia, n (%)				1.000
Yes	3 (1.7%)	3 (2.1%)	0 (0%)
No	174 (98.3%)	138 (97.9%)	36 (100%)
Biomarkers
Creatinine, μmol/L	66.0 ± 15.8	67.0 ± 16.2	62.3 ± 13.5	0.138
Cystatin C, mg/L	0.9 ± 0.2 (n = 122)	0.9 ± 0.2 (n = 101)	0.8 ± 0.2 (n = 21)	0.576
Albumin, g/L	43.7 ± 3.5	43.5 ± 3.1	44.5 ± 4.7	0.437
WBC, × 109/L	5.6 ± 3.1	5.6 ± 3.4	5.7 ± 1.7	0.380
NEU, × 109/L	3.4 ± 1.5	3.4 ± 1.5	3.4 ± 1.5	0.811
LYM, × 109/L	1.6 ± 0.5	1.5 ± 0.5	1.7 ± 0.6	0.048
Hb, g/L	131.5 ± 23.2	132.3 ± 23.6	128.5 ± 21.4	0.194
PLT, × 109/L	196.6 ± 61.4	192.3 ± 58.9	213.1 ± 68.4	0.069
CRP, mg/L	1.2 (0.5–2.1) (n = 170)	1.3 (0.5–2.1) (n = 138)	0.9 (0.5–1.8) (n = 32)	0.488
CEA, ng/mL	1.7 (1.7–2.6) (n = 161)	1.8 (1.7–2.6) (n = 128)	1.7 (1.2–1.9) (n = 33)	0.023
FEV1/FVC, %	76.5 ± 13.7 (n = 119)	76.8 ± 12.3 (n = 96)	75.2 ± 18.6 (n = 23)	0.893
ALI, unitless	54.0 ± 13.7	53.0 ± 23.3	53.8 ± 24.4	0.901
Lifestyle
Smoking status				0.384
Yes	39 (22.0%)	33 (23.4%)	6 (16.7%)
No	138 (78.0%)	108 (76.6%)	30 (83.3%)
Drinking status				0.055
Yes	41 (29.1%)	37 (26.2%)	4 (11.1%)
No	136 (70.9%)	104 (73.8%)	32 (88.9%)

Note: Continuous variables are presented as mean ± standard deviation for normally distributed data and as median (interquartile range) for non‐normally distributed data, categorical variables are expressed as number (percentage). p value < 0.05 was considered statistically significant. Missing baseline laboratory values were addressed using multiple imputation prior to multivariable prediction modeling.

Abbreviations: ALI, advanced lung cancer inflammation index; BMI, body mass index; CEA, carcinoembryonic antigen; CHD, coronary heart disease; CRP, C‐reactive protein; FEV1/FVC, forced expiratory volume in one second/forced vital capacity; Hb, hemoglobin; LYM, lymphocyte count; NEU, neutrophil count; PLT, platelet count; SMI, skeletal muscle index; WBC, white blood cell count.

There were no significant between‐group differences in age, sex distribution, height, most comorbidities, or routine laboratory indices. In contrast, sarcopenic patients had substantially lower body weight, BMI, and T12‐SMI compared with nonsarcopenic patients. Modest but statistically significant between‐group differences were also observed for lymphocyte count (LYM) and carcinoembryonic antigen (CEA).

3.2.2. Candidate Predictors and Feature Selection

To develop a parsimonious clinical model for predicting imaging‐defined low muscle mass, candidate predictors were first defined as variables showing univariate associations with sarcopenia (p < 0.05), supplemented by prespecified demographic and nutrition‐related covariates (sex, age, height, platelet count, C‐reactive protein [CRP], advanced lung cancer inflammation index [ALI], albumin [ALB], and creatinine). Missing values were addressed using multiple imputation.

Two complementary feature‐selection strategies were performed in parallel: LASSO regression (Figure 3A) and random forest recursive feature elimination with cross‐validation (RFECV), both applied to the full set of candidate variables, including sex and all demographic, anthropometric, and biochemical parameters. LASSO retained five predictors (age, BMI, CEA, CRP, and LYM), while RFECV identified nine predictors (age, ALI, BMI, CEA, CRP, LYM, platelet count, sex, and weight) (Figure 3B). The intersection of the two methods yielded five predictors (age, BMI, CEA, CRP, and LYM). In contrast, variables such as sex were not retained because of their limited relevance to the outcome in the LASSO procedure. These five predictors were subsequently entered into the multivariable logistic regression model. Complete model estimates are presented in Table 3.

FIGURE 3.

Variable‐selection results used to derive the LSRM. (A) LASSO coefficient paths showing shrinkage of predictors across log (λ) and the cross‐validated binomial deviance (right) with the vertical line indicating λ_min from 5‐fold cross‐validation (glmnet, alpha = 1). (B) Venn plot for variants between the machine learning methods LASSO and RFECV.

TABLE 3.

Multivariable logistic regression analysis for factors associated with sarcopenia.

Term	β	SE	p	OR	OR_low	OR_high
Intercept	8.577	2.658	0.002
Age	0.054	0.019	0.006	1.055	1.016	1.096
BMI	−0.658	0.137	< 0.001	0.518	0.396	0.677
CEA	−0.463	0.293	0.125	0.630	0.354	1.119
CRP	0.039	0.033	0.251	1.040	0.974	1.109
LYM	1.542	0.527	0.004	4.672	1.663	13.125

Note: β represents the regression coefficient, and SE denotes the standard error. Odds ratios (ORs) and 95% confidence intervals (OR_low to OR_high) were calculated to estimate the strength of association between each variable and sarcopenia.

3.2.3. Construction and Performance of the LSRM

Subsequently, the regression coefficients (β) derived from the final multivariable logistic model were used as weighted risk parameters to construct a quantitative prediction model for the odds of sarcopenia, namely the LSRM. The resulting LSRM incorporates five variables retained in the final prediction model identified by the combined feature‐selection strategy: age, BMI, CEA, CRP, and LYM. For each patient, an individualized sarcopenia risk score can be calculated as a linear combination of these variables weighted by their corresponding regression coefficients, and then transformed into a predicted probability using the logistic function. The linear predictor can be written as:

$$\begin{array}{cc}\mathit{\text{LSRM}}=\hfill & 8.577+0.054\times \mathit{Age}-0.658\times \mathit{BMI}\hfill \\ \hfill & -0.463\times \mathit{CEA}+0.039\times \mathit{CRP}+1.542\times \mathit{LYM}\hfill \end{array}$$

and the predicted probability of sarcopenia is obtained by logistic transformation:

$$P=\frac{e^{\mathit{\text{LSRM}}}}{1+e^{\mathit{\text{LSRM}}}}$$

The full set of β coefficients is reported in Table 3. To facilitate bedside implementation, the multivariable logistic model was further translated into a nomogram that maps each predictor to a point score and corresponding overall probability of sarcopenia (Figure 5A).

FIGURE 5.

Nomogram and distribution of LSRM scores. (A) Point‐based nomogram constructed from the pooled multivariable logistic regression. (B) Boxplot comparing LSRM scores between patients with and without imaging‐defined sarcopenia.

The LSRM demonstrated good discriminative ability for imaging‐defined sarcopenia, with an apparent AUC of 0.870 (Figure 4A). Internal validation was performed by bootstrap resampling; calibration assessment showed close agreement between predicted and observed probabilities across risk strata, with the calibration curve lying near the 45°line (Figure 4B). A boxplot comparing LSRM scores between patients with and without sarcopenia demonstrated a clear separation of risk distributions (two‐sample comparison, p = 7 × 10⁻¹²) (Figure 5B). Finally, we explored the relationship between T12‐defined sarcopenia, LSRM scores, and PHS. In this cohort, the continuous LSRM score did not show a statistically significant association with PHS (p > 0.05; Figure S1). These findings indicate that, within the current cohort, both T12‐based low muscle mass and the LSRM are primarily associated with body composition status rather than short‐term postoperative outcomes such as hospital stay.

FIGURE 4.

Discrimination and calibration of the LSRM for imaging‐defined sarcopenia. (A) Receiver‐operating characteristic (ROC) curve for the LSRM using averaged predicted probabilities across imputations; apparent area under the ROC curve = 0.87. (B) Calibration assessment: Calibration curve with a 45° reference line.

3.2.4. Compare the Performance of LSRM With Simple Benchmarks

In the prospective cohort (n = 177), the LSRM demonstrated superior discrimination for imaging‐defined sarcopenia compared with simple anthropometric and inflammation‐based benchmarks (Figure 6A). Compared with common clinical benchmarks, the LSRM demonstrated higher discriminative performance. Corresponding AUCs were 0.812 (95% CI 0.742–0.881) for BMI alone, 0.493 (95% CI 0.382–0.605) for ALI, and 0.819 (95% CI 0.751–0.887) for the BMI + age model, and DeLong tests confirmed the LSRM's statistically superior discrimination relative to each of these simpler indices.

FIGURE 6.

Performance of the Lung cancer patients' Sarcopenia Risk Model (LSRM) and risk‐stratified results. (A) Receiver‐operating characteristic (ROC) curves comparing the discrimination of LSRM and simple benchmarks for imaging‐defined sarcopenia in the prospective cohort. (B) Distribution of LSRM scores across risk strata. (C) Prevalence of imaging‐defined sarcopenia across LSRM risk strata.

3.2.5. LSRM‐Based Risk Categories

Using the LSRM‐predicted probability and a two‐cutpoint strategy (Youden index for the lower cutpoint and a high‐specificity threshold for the upper cutpoint), patients were classified into low‐, intermediate‐, and high‐risk strata. LSRM score distributions differed markedly across the three strata (Kruskal–Wallis p = 2.05 × 10⁻³¹) (Figure 6B); pairwise post hoc comparisons (Wilcoxon rank‐sum with Bonferroni correction) were all highly significant. The trend across ordered risk strata was highly significant (Cochran‐Armitage trend test p = 4.2 × 10⁻¹⁴); a conventional χ ² test produced consistent results (χ ² p = 4.02 × 10⁻¹³) (Figure 6C). These findings indicate that the LSRM provides a clinically interpretable stratification of sarcopenia risk, with progressively higher predicted risk and model scores across the low‐, intermediate‐, and high‐risk groups.

4. Discussion

Sarcopenia is a multifactorial syndrome characterized by loss of skeletal muscle mass and function, for which ALM remains a widely accepted quantitative anchor in both clinical research and guideline‐driven diagnostic algorithms [17]. Mechanistically and clinically, skeletal muscle mass is determined by the interplay of nutritional intake, systemic inflammation, and the reserve capacity of organ systems; these pathways integrate to drive progressive muscle wasting in vulnerable patients [18, 19, 20]. We considered various nutrition‐ and inflammation‐related indicators that may influence muscle mass in clinical settings.

Contemporary consensus (EWGSOP2 [17] and AWGS 2019 [21]) defines sarcopenia by low muscle strength as the primary criterion, with confirmation by low muscle quantity or quality; EWGSOP2 recognizes dual‐emission X‐ray absorptiometry (DXA) for definitive diagnosis, but explicitly states that opportunistic imaging, such as CT, may identify low muscle mass in clinical practice. Computed tomography at the L3‐SMI has become the most widely used CT reference because it correlates strongly with whole‐body lean mass and with clinical endpoints across tumor types [7]. Importantly, the clinical relevance of CT‐defined muscle depletion has been demonstrated beyond lung cancer. Across several solid tumors, reduced skeletal muscle mass assessed by CT has been associated with adverse oncologic outcomes. In colorectal cancer, particularly rectal cancer, recent evidence from Rakici et al. demonstrated that CT‐defined sarcopenia was associated with cancer‐related survival, further supporting the prognostic value of imaging‐derived muscle assessment across different malignancies [22]. These findings suggest that skeletal muscle depletion represents a common manifestation of cancer‐associated metabolic dysfunction rather than a tumor‐specific phenomenon and highlight the potential role of opportunistic CT‐based body composition assessment as a broadly applicable oncologic biomarker. However, many patients with NSCLC undergo chest CT without contiguous abdominal coverage, which limits routine access to L3 [9, 23]. Several prior investigations have therefore evaluated thoracic vertebral levels (notably T4 and T12) as practical substitutes for L3 [9, 24]; while correlations with L3 are often high, few studies have provided rigorously derived conversion equations or thresholds that are calibrated to established L3 cut‐points. This persistent gap motivated us to systematically evaluate thoracic levels, derive L3‐calibrated, sex‐specific T12 thresholds, and then build a model that can identify patients at risk of CT‐defined low muscle mass even when lumbar imaging is unavailable.

Our study demonstrates across two centers that axial CT‐derived muscle measurements at the T12 vertebra can substitute for the conventional L3‐based sarcopenia index. T12‐SMI showed the strongest correlation and agreement with L3‐SMI, significantly outperforming the T4 level. The Deming regression analysis yielded conversion equations, enabling translation of well‐established L3 cutoffs into equivalent T12 thresholds. By providing explicit, L3‐calibrated T12 cutoffs in a two‐center NSCLC population, our work adds quantitative precision to prior descriptions of thoracic surrogates. Our findings reinforce and extend prior work in CT‐defined sarcopenia: Kaltenhauser et al. found that thoracic‐level (T10) SMI closely tracks L3 measures and that sarcopenia defined by thoracic imaging predicts poorer survival in NSCLC patients [24]. Similarly, Montero‐Benitez et al. reported that T12‐derived indices outperform T4 for detecting low muscle mass, with T12 imaging proving “especially helpful” for spotting sarcopenia in NSCLC patients [10]. Together, the evidence supports T12 as a practical surrogate for L3 when abdominal imaging is unavailable, allowing opportunistic sarcopenia screening on routine chest CTs without extra scanning.

We enrolled a prospective, two‐center cohort of 177 patients to further evaluate the clinical applicability of the T12‐derived measurements and to develop a clinically useful prediction model, which was subsequently internally validated using bootstrap resampling. Interestingly, some traits considered a priori as determinants of lean mass, such as albumin and pulmonary function, did not retain independent associations with CT‐defined sarcopenia in our cohort, a finding that mirrors the work by Cao et al. [25]. This may be because albumin concentrations decrease with systemic inflammation, capillary leak and disease burden, and therefore do not reliably index chronic muscle stores in inflammatory states. Likewise, pulmonary function tests are influenced by a range of both pulmonary and extrapulmonary factors, and prior studies report mixed associations between lung function metrics and muscle mass depending on population and measurement methods [26]. These discrepancies highlight that genetically proxied lifelong exposures and cross‐sectional clinical measurements capture related but not identical constructs, and underscore the importance of empirically derived, context‐specific clinical models.

Because few earlier studies have produced calibrated conversion thresholds or systematic, well‐validated prediction models, we combined machine‐learning techniques (multivariable LASSO and RFECV) to choose predictors and minimize overfitting. Using the selected predictors, we constructed the LSRM, a nomogram based on five routine variables: age, BMI, CEA, CRP, and LYM. All five variables were selected through a data‐driven feature selection process and jointly contributed to the predictive performance of the model. However, statistical significance at the individual variable level was not a prerequisite for inclusion, and therefore these variables should be interpreted as model components rather than independent biological predictors. The LSRM showed good discrimination (AUC≈0.87) and accurate calibration in internal validation. Age and BMI unsurprisingly emerged as important factors, reflecting the known age‐related decline and the “protective” effect of higher body mass on muscle reserves. Notably, CEA and CRP were retained in the final prediction model through the combined feature‐selection strategy, suggesting that they contributed to overall model performance despite not achieving statistical significance as individual coefficients in the final multivariable model. While not traditional muscle markers, these are readily available in oncology practice. Their inclusion highlights the intersection of tumor biology and systemic inflammation with muscle wasting. While informative for prediction, these markers do not establish causality. It is crucial to emphasize that these biomarkers are correlates rather than proven causes of sarcopenia—elevated CEA may simply reflect aggressive cancer that also drives cachexia [27], and high CRP denotes systemic inflammation that accompanies many chronic diseases [18]. Likewise, the positive coefficient observed for lymphocyte count should not be interpreted as evidence that higher lymphocyte levels increase the risk of low muscle mass. Rather, this finding likely reflects the multivariable nature of the prediction model and the complex interrelationships among routinely collected clinical variables within the present dataset. Our findings align with recently published work that likewise used routine clinical and laboratory variables to predict sarcopenia in NSCLC: for example, Gao et al. developed a model in NSCLC patients with malignant pleural effusion and identified age, BMI, albumin, and CYFRA21‐1 as important predictors (AUC≈0.89) [28]. Together, these studies support the feasibility of parsimonious, routine‐data‐based tools for sarcopenia screening in thoracic oncology.

Although sex was included among the candidate predictors in our research, its exclusion from the final LSRM model is not surprising. LASSO tends to select a single representative variable from highly correlated predictors, which may exclude sex if related variables like BMI or weight capture similar information. In addition, the use of sex‐specific SMI cutoffs in our outcome definition already incorporated sex differences, diminishing the marginal value of sex as an explicit covariate. A meta‐analysis by Yin et al. [29] similarly found that among 70 machine learning models, only 13 included sex as a predictor, suggesting that many models prioritize other features, particularly when sample size or sex imbalance limits the power to detect an independent effect of sex.

Importantly, the present study demonstrates not only that T12‐derived measures can serve as a practicable thoracic surrogate for L3‐SMI, but also that a compact, clinic‐ready prediction tool meaningfully improves on simple, commonly used benchmarks. In the prospective cohort, the LSRM achieved an AUC of 0.87, exceeding BMI alone (AUC≈0.812), ALI (AUC≈0.493), and the BMI + age model (AUC≈0.819); pairwise comparisons confirmed the model's superior discrimination. Translating predicted probabilities into a two‐cutpoint, three‐tier schema produced low‐, intermediate‐, and high‐risk strata with sharply different LSRM distributions and sarcopenia prevalences, supporting the model's potential for clinically meaningful triage. The LSRM's added value likely reflects its combination of anthropometry with tumor‐ and inflammation‐related signals, thereby capturing both reserve capacity and disease‐associated systemic perturbation—pathways plausibly linked to muscle wasting [18, 19]. In practical terms, the present model was developed to identify patients with CT‐defined low muscle mass rather than to establish a consensus diagnosis of sarcopenia; such a three‐tier schema may help clinicians identify patients who could benefit from further comprehensive assessment or supportive care considerations, such as nutritional evaluation or prehabilitation referral, pending clinical judgment. Therefore, the LSRM should be considered a pragmatic screening tool to identify individuals who may benefit from further functional assessment rather than a standalone diagnostic tool for sarcopenia.

The modest relationship between the LSRM and immediate perioperative indices, such as postoperative hospital stay, in this sample does not undermine the model's utility for screening and stratification. This apparent discrepancy may be explained by several factors. First, postoperative hospital stay is influenced by a wide range of perioperative and institutional variables, including surgical standardization, enhanced recovery protocols, and discharge practices, which may attenuate the measurable effect of baseline muscle status in relatively homogeneous surgical cohorts. Previous meta‐analyses have similarly highlighted that length of stay is a multifactorial endpoint with substantial heterogeneity across institutions and surgical pathways [30]. Second, the present cohort predominantly consisted of patients undergoing elective surgical evaluation, in whom perioperative management is highly standardized, potentially reducing variability in short‐term outcomes. Third, short‐term recovery is shaped by many acute, procedural, and contextual factors that can dilute the influence of baseline muscle reserves [31]; by contrast, sarcopenia has a stronger and more consistent association with medium‐ and long‐term surgical and oncologic endpoints, where early identification and targeted prehabilitation are most likely to change outcomes [32, 33]. Moreover, our analyses of postoperative hospital stay were exploratory and not specifically powered to detect modest differences in length of stay. Taken together, these findings do not contradict prior literature but rather suggest that the predictive value of sarcopenia‐related indices may be outcome‐dependent, with stronger effects observed for long‐term endpoints compared with short‐term recovery metrics.

Several limitations merit acknowledgement. First, this study concentrated on imaging‐defined sarcopenia and the initial validation of the LSRM rather than demonstrating that LSRM‐guided interventions improve long‐term clinical outcomes; establishing clinical benefit will require prospective interventional or longitudinal studies. Moreover, because muscle strength and physical performance were not assessed, the primary outcome in the present study represents CT‐defined low muscle mass rather than consensus‐defined sarcopenia. Although the model employed bootstrap resampling for internal validation, which is a recommended approach for prediction model development under the TRIPOD framework, the sample size remains modest. Therefore, independent external validation in geographically and demographically diverse cohorts remains an essential next step before widespread clinical implementation. Finally, genetic analyses were not incorporated in the present report; ancestry‐related differences in genetic architecture and body composition may limit the direct generalizability of genetic findings and should be explored in future work.

Collectively, these findings strengthen the rationale for integrating a parsimonious clinical prediction tool with opportunistic thoracic imaging: where L3 imaging is unavailable, T12‐based assessment plus the LSRM can enable earlier, more precise triage for nutritional assessment, prehabilitation, or targeted surveillance, while external validation and implementation studies remain important next steps. By offering both an L3‐equivalent thoracic metric and a simple, accurate bedside risk score based on five routine variables, our approach provides a pragmatic pathway for embedding sarcopenia screening into everyday lung cancer care.

5. Conclusion

In this study, we show that thoracic CT‐derived SMI at T12 can serve as a pragmatic surrogate for L3 SMI when lumbar imaging is unavailable, exhibiting strong correlation and acceptable agreement with L3‐based measurements after linear calibration. Building on these imaging findings, we developed and internally validated a pragmatic prediction model (LSRM) based on routine clinical and laboratory variables to identify patients with CT‐defined low muscle mass in NSCLC. The LSRM demonstrated good discrimination and calibration on internal validation during internal validation using bootstrap resampling and provides a simple tool for risk screening in thoracic oncology.

Author Contributions

Wei Dai: visualization. Zhuorui You: conceptualization, methodology, software, visualization, validation, writing – review and editing, investigation. Cheng Lei: methodology, software. Jieming Cao: validation. Minxian Li: validation. Mengqi Shao: resources, writing – review and editing, funding acquisition, methodology. Xing Wei: investigation. Gang Li: funding acquisition, resources. Jia Wang: software, methodology. Jia Liao: formal analysis, visualization. Hechen Shen: investigation, validation.

Funding

This work was supported by Sichuan Science and Technology Program (Grant No.: 2026NSFSC1948), and The Health Development Promotion Project—Xinghuo Research Plan Project (Grant No.: XHJH0086).

Ethics Statement

The study was approved by the Ethics Committee of Sichuan Provincial People's Hospital, Chengdu, China (Approval No. 2025–565) and Sichuan Cancer Hospital, Chengdu, China (Approval No. SCCHEC‐02‐2018‐043). All procedures were conducted in accordance with the principles of the Declaration of Helsinki.

Consent

Written informed consent was obtained from all participants enrolled in the prospective component of the study. For the retrospective component, the requirement for written informed consent was waived by the Sichuan Academy of Medical Sciences and Sichuan Provincial People's Hospital Ethics Committee for Basic and Clinical Research due to the retrospective nature of the study.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: Association between LSRM score and PHS. Scatterplot of individual LSRM linear predictors (LP) versus postoperative hospital stay (days) with a fitted trend line.

Table S1: Inter‐reader agreement between two independent readers was assessed using intraclass correlation coefficients (ICC) based on a two‐way random‐effects model with absolute agreement and single‐measure reliability [ICC(A,1)]. 95% confidence intervals (CIs) are provided for each estimate. ICC values were interpreted according to conventional thresholds, with values > 0.90 indicating excellent reliability, 0.75–0.90 good reliability, 0.50–0.75 moderate reliability, and < 0.50 poor reliability.

Table S2: Sex‐specific equivalent thoracic skeletal muscle index (SMI) cutoffs at the T4 and T12 levels converted from established L3‐SMI thresholds using Deming regression. L3‐SMI, skeletal muscle index at the third lumbar vertebra; T4‐SMI and T12‐SMI, skeletal muscle index at the fourth and twelfth thoracic vertebrae, respectively; M, male; F, female. Equivalent thoracic cutoffs were derived using Deming regression equations converted from established L3‐SMI thresholds. All SMI values are expressed in cm²/m².

Table S3: Bland–Altman analysis of agreement between thoracic‐derived and lumbar‐derived skeletal muscle index (SMI) values following conversion using Deming regression. Mean bias represents the average difference between thoracic‐derived and lumbar‐derived SMI values. The 95% limits of agreement (LoA) were calculated as mean bias ±1.96 × SD. SD, standard deviation of the paired differences. T4 → L3 and T12 → L3 indicate thoracic SMI values at the T4 and T12 levels converted to the lumbar (L3) scale using Deming regression equations prior to Bland–Altman analysis. Narrower LoA indicate better interchangeability at the individual patient level. All SMI values are expressed in cm²/m².

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.