Machine learning evaluation of the discriminative ability of Castelli Risk Index-I and other non-traditional lipid indices for sarcopenia: a cross-sectional study based on CHARLS

Abstract

Background

Sarcopenia is a syndrome that occurs in older adults, marked by progressive deterioration in muscle mass, strength, and/or functional capacity. Abnormal lipid metabolism has been associated with a higher prevalence of sarcopenia, but evidence regarding the association between the Castelli Risk Index-I (CRI-I) and sarcopenia is still insufficient. The research was designed to assess the association between CRI-I and sarcopenia status among the Chinese population and to determine its incremental discriminative value within a machine learning model.

Methods

This research utilized information from the 2011 CHARLS survey wave. CRI-I was categorized into quartiles and its association with sarcopenia was evaluated through logistic regression and restricted cubic splines (RCS). Seven candidate models were developed using the 2011 data, and the optimal model was identified, followed by temporal external validation with the 2015 CHARLS wave. To assess the additional discriminative value of CRI-I, we evaluated model performance using receiver operating characteristic (ROC) curves, precision-recall curves (PRC), calibration plots, and decision curve analysis (DCA). Finally, the SHapley Additive exPlanations (SHAP) algorithm was used to show the importance of each feature.

Results

A total of 1,332 individuals (15.1%) met the diagnostic criteria for sarcopenia. After full adjustment, higher CRI-I levels were associated with progressively lower odds of sarcopenia. RCS analysis further demonstrated that the association exhibited a non-linear pattern. Incorporating CRI-I into the optimal model improved discriminative performance. The model also demonstrated good calibration and clinical utility. Additionally, the SHAP algorithm was applied to calculate feature importance for the model’s estimated probability of having sarcopenia, which identified age as the most important feature, followed by CRI-I.

Conclusions

CRI-I showed superior discriminative performance for sarcopenia compared with six other non-traditional lipid indices. Elevated CRI-I levels correlated with substantially reduced sarcopenia likelihood. Adding CRI-I to the model improved probability stratification and may help identify individuals more likely to have sarcopenia.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12944-026-02892-8.

Introduction

Sarcopenia is a disorder that affects skeletal muscles and is most common among older adults and individuals with long-term health problems [1]. Globally, sarcopenia affects approximately 10–16% of individuals [2], while a meta-analysis reported 20.7% among Chinese older adults [3], substantially higher than the global average. Sarcopenia impairs activities of daily living and increases the risks of fractures and disabilities, and is closely related to functional decline, metabolic disorders, and cognitive impairment [3]. With population aging, it has become an urgent public health concern. Accumulation of excess lipids and their metabolites in skeletal muscle can induce lipotoxicity and impair muscle health by increasing oxidative stress, weakening mitochondrial activity, aggravating inflammatory responses and reducing insulin sensitivity [4]. However, emerging evidence suggests that this relationship is more complex. Certain lipid metabolites, such as phosphatidylcholines and specific fatty acids, may exert protective effects on skeletal muscle by enhancing fatty acid oxidation, improving mitochondrial function and metabolic status, and attenuating inflammatory responses [5, 6]. Such complexity is further supported by epidemiological studies, in which findings across different populations have even shown opposite directions. Some researchers reported that several lipid indices, such as high-density lipoprotein cholesterol (HDL-C) [7], total cholesterol (TC) [8], triglycerides (TG) [8], the non–high-density lipoprotein to high-density lipoprotein cholesterol ratio (non–HDL-C/HDL-C) [9], and the triglyceride to high-density lipoprotein cholesterol ratio (TG/HDL-C) [10], are positively associated with a higher likelihood of sarcopenia, whereas others have found opposite associations [11, 12]. These inconsistent findings suggest that different lipid metrics may reflect distinct metabolic pathways involved in muscle health and warrant further evaluation.

The Castelli Risk Index-I (CRI-I), calculated as the ratio of TC to HDL-C (TC/HDL-C) [13], primarily reflects the balance between atherogenic and protective cholesterol fractions [14]. As an emerging lipid marker, CRI-I has been used to assess atherosclerotic and metabolic risk. A study in elderly patients with diabetes found that higher CRI-I levels significantly increase the risk of carotid atherosclerosis [15], and other studies have reported positive correlations between CRI-I and both coronary disease severity and lipid metabolism disturbances [16]. However, its potential association with sarcopenia remains unclear and warrants further exploration.

Numerous models for identifying sarcopenia have been developed based on demographic factors, anthropometric indicators, physical performance tests, or comorbidity characteristics. However, several limitations remain. First, the inclusion of sarcopenia diagnostic components, such as body mass index (BMI), handgrip strength, and gait speed, as model variables may overestimate performance and introduce incorporation bias [17, 18]. Moreover, many models rely solely on internal validation without independent external validation, which limits their transportability and generalizability across different populations and clinical settings [17, 19]. Although some models demonstrate good discriminative ability, their reliance on complex or time-consuming functional assessments restricts feasibility in primary care and large-scale screening settings [20]. Notably, lipid-based biomarkers are rarely incorporated into existing sarcopenia assessment models, and the contribution of composite lipid indices remains insufficiently explored. Recently, machine learning technology has been increasingly used in medical research, and has shown great potential in early detection of diseases [21]. Given its established role in reflecting overall cholesterol balance and metabolic risk, we hypothesized that CRI-I could serve as a valuable biomarker for sarcopenia. Hence, we conducted a study to examine the relationship between CRI-I and the possibility of contracting sarcopenia. At the same time, we also developed a machine learning framework, which incorporates CRI-I as a core variable to improve discrimination of sarcopenia status and identification in screening contexts.

Methods

Study participants and data sources

We used data from the 2011 and 2015 waves of the China Health and Retirement Longitudinal Study (CHARLS), which encompass a nationally representative sample of Chinese individuals aged 45 years and older residing in community settings. CHARLS adopts a multistage probability design, and the survey collects harmonized information on demographic information, health status, lifestyle, and socioeconomic factors [22]. In this study, the 2011 wave was used for model development, whereas newly enrolled participants from the 2015 wave served as the temporal external validation cohort. Details on the construction of the external validation cohort are provided in the Supplementary Material 1 under the section “Construction of the External Validation Cohort.” Further information on CHARLS is available on the official website.

A total of 17,705 participants were initially enrolled in CHARLS 2011. After excluding participants younger than 45 years (n = 187), those with missing sarcopenia data (n = 4,587), abnormal BMI values (n = 247), missing lipid indicators (n = 3,579), missing information on major comorbidities (n = 130), and other covariates (n = 170), 8,805 participants were retained for the subsequent analysis and constituted the internal dataset. In addition, 913 participants from the 2015 wave with complete data were included as the external validation set. The participant selection process and dataset arrangement are illustrated in Fig. 1.

Fig. 1.

Flowchart of participant selection. A Participant selection from the CHARLS 2011 wave. B Participant selection from newly enrolled participants in the CHARLS 2015 wave

Determination and calculation of biomarkers

Venous blood samples from the participants were meticulously collected and analyzed by the CHARLS team following standard protocols, including measurements of TC, TG, HDL-C, low-density lipoprotein cholesterol (LDL-C), fasting plasma glucose (FPG), hemoglobin A1c (HbA1c) and high-sensitivity C-reactive protein (hsCRP).

The primary variable was the CRI-I, calculated as TC/HDL-C [13]. For comparison, we also calculated six other non-traditional lipid indices, including lipoprotein combined index (LCI), Atherogenic index of plasma (AIP), Non–HDL-C, Castelli risk index-II (CRI-II), Remnant cholesterol (RC) and RC/HDL-C ratio. The calculation method followed previous studies [9, 13, 23–25], detailed formulas are provided in the Supplementary Material 1 under the section “Calculation of non-traditional lipid indices.”

Definition of sarcopenia

Sarcopenia was defined as a gradual decline in muscle mass, muscle strength, and/or physical capability following the 2019 Asian Working Group for Sarcopenia consensus [1]. Appendicular skeletal muscle mass (ASM) was calculated as follows: ASM = 0.193 × weight (kg) + 0.107 × height (cm) − 4.157 × sex (male = 1, female = 2) − 0.037 × age − 2.631 [26]. The validity of this equation has been demonstrated in Chinese adults and shows acceptable agreement with dual-energy X-ray absorptiometry (DXA)–derived muscle mass measurements [26]. The determination of low muscle mass was based on height-adjusted calculations, derived from dividing muscle mass by height squared. The threshold was established at the 20th percentile, corresponding to values below 7.031 kg/m² for males and 5.398 kg/m² for females [27]. Handgrip strength was assessed by a dynamometer, with two trials performed for each hand and the highest value recorded. Researchers classified diminished muscle strength as handgrip readings at or below 28 kg for male subjects and 18 kg for female subjects [1]. The assessment of physical performance involved a standing test and walking speed evaluation. Individuals were deemed to have suboptimal physical function if they required 12 s or longer to rise from a chair or demonstrated a walking pace of less than 1.0 m per second [1].

Definition of covariates

Covariates included sociodemographic factors, lifestyle factors, health-related factors, and biochemical factors. Sociodemographic factors were age, gender, education, marital status, and residence. Lifestyle factors included smoking, alcohol consumption, socializing, and sleep. Health-related variables comprised hypertension, diabetes, stroke, and BMI. Biochemical factors included TG, TC, HDL-C, LDL-C, and hsCRP. Covariate definitions were available in the Supplementary Material 1, section “Data Collection Methods and Measurement Protocols.”

Performance comparison of seven lipid indices

We examined the performance of seven lipid indices in discriminating sarcopenia status. The receiver operating characteristic (ROC) curves were generated, and performance was evaluated by calculating the area under the curve (AUC), sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Then we used the DeLong test to evaluate the AUC values of these different indicators to see if there was any statistical significance between them.

Model development and evaluation

The 2011 dataset was randomly split into training (70%) and testing (30%) sets for model development and internal validation, with a fixed random seed used to ensure reproducibility of the data partitioning. Given the relatively low prevalence of sarcopenia (15.1%), class weights were incorporated into the training process to reduce bias arising from class imbalance. Feature selection procedures associated with sarcopenia were performed in the training set. Univariate logistic regression analyses were first used for preliminary variable screening, and variables showing meaningful associations were subsequently incorporated into least absolute shrinkage and selection operator (LASSO) regression. To enable a comparative evaluation of model performance, we trained a logistic regression (LR) model alongside six machine learning algorithms, including Adaptive Boosting (AdaBoost), K-Nearest Neighbors (KNN), Extreme Gradient Boosting (XGBoost), Random Forest (RF), Gaussian Naïve Bayes (GNB), and Support Vector Machine (SVM). Hyperparameters were tuned using grid search with 5-fold cross-validation. Model performance was evaluated using ROC curves, AUC, accuracy, sensitivity, specificity, F1 score, precision, NPV and Brier score. Based on a comprehensive evaluation of performance metrics, the optimal model was selected to assess the incremental value of CRI-I. Finally, the CHARLS 2015 dataset, consisting of new samples, was used for external validation.

Incremental discriminative value of CRI-I

The incremental value of incorporating CRI-I into the model was evaluated using ROC curves, precision-recall curve (PRC), calibration curves, and decision curve analysis (DCA), with AUC differences between models with and without CRI-I assessed using the DeLong test. The ROC analysis revealed the model’s discriminative capacity, with a higher AUC indicating better discrimination of sarcopenia status. The PRC was used to evaluate the model’s ability to identify positive cases under class imbalance. A higher area under the PRC (AUPRC) indicated better overall performance across recall thresholds. Model calibration was evaluated using calibration curves to assess the alignment between estimated probabilities and observed outcomes. Platt scaling was applied for probability recalibration. Model goodness of fit was evaluated using the Hosmer–Lemeshow (H–L) test. The DCA was used to depict potential clinical utility through net benefit comparisons across a range of threshold probabilities. Furthermore, we applied the SHapley Additive exPlanations (SHAP) algorithm to interpret the model’s classification output, quantify the contribution of each variable to the model’s estimated probability of having sarcopenia, and rank variables by their importance.

Statistical analysis

Continuous variables were presented as mean ± standard deviation for approximately normally distributed data, and as median (interquartile range, IQR) otherwise; categorical variables were presented as counts with percentages. The association between CRI-I and sarcopenia was examined using three progressively adjusted logistic regression models, and findings were reported as odds ratios (ORs) with 95% confidence intervals (CIs). The first model contained no adjustments, while the second accounted for sociodemographic variables. The third model incorporated additional lifestyle factors, health-related variables, and biochemical indices to control for potential confounding influences. Variance inflation factor (VIF) measurements were used to evaluate multicollinearity, with values above 5 indicating potential multicollinearity. We also employed restricted cubic splines (RCS) with 4 knots to explore the nonlinear relationship between CRI-I and sarcopenia, adjusting for the same covariates as those included in Model 3 of the multivariable logistic regression. We further applied two-piecewise linear regression to identify potential threshold effects. Subgroup and interaction analyses were conducted to examine whether the association between CRI-I and sarcopenia varied across demographic and clinical characteristics. Notably, TC and HDL-C were not included as covariates because they are components of CRI-I (TC/HDL-C), which would introduce redundancy and collinearity. BMI was also omitted. In this study, sarcopenia assessment incorporated muscle mass calculated from body weight and height. Therefore, BMI was structurally related to the outcome and could have introduced overadjustment if included as a covariate. Sensitivity analyses were conducted to evaluate the robustness of the association between CRI-I and sarcopenia. First, for covariates with missing data (below 20%), multiple imputation by chained equations (MICE) was performed with 10 imputed datasets, followed by pooled multivariable logistic regression analysis. Second, sarcopenia was alternatively represented by the sarcopenia index (SI), calculated as serum creatinine/cystatin C × 100, a validated serum biomarker of skeletal muscle mass, with higher SI values indicating better muscle mass [28]. The linear association between CRI-I and SI was examined to further assess the consistency of the primary findings. Covariates in the SI models were the same as those used in the primary analysis. However, given that SI is derived solely from serum biomarkers without involving body weight or height, we additionally adjusted for BMI in Model 3 to examine whether the association remains independent of body habitus. All analyses utilized R 4.4.2, with statistical significance defined by two-tailed P values below 0.05.

Results

Baseline characteristics

This analysis included 8,805 individuals, with a median age of 60 years (IQR: 54.00–67.00), including 4,151 (47.1%) men and 4,654 (52.9%) women. Most participants lived in rural areas (82.4%) and had an education level below high school (90.2%). Sarcopenia was present in 1,332 (15.1%). Baseline characteristics by CRI-I quartiles (Q1 ≤ 3.11; Q2 3.11–3.85; Q3 3.85–4.82; Q4 > 4.82) are shown in Table 1. With increasing CRI-I, participation in social activities and the prevalence of hypertension, diabetes, and stroke increased, as did BMI and biochemical indices including LDL-C, TC, TG, hsCRP, HbA1c, FPG, Non–HDL-C, CRI-II, RC, and RC/HDL-C (all P < 0.001). In contrast, the proportion of participants reporting alcohol drinking and HDL-C levels decreased with increasing CRI-I quartiles (all P < 0.001). The prevalence of sarcopenia decreased significantly, from 25.8% in Q1 to 6.7% in Q4.

Table 1.

Baseline characteristics of CHARLS participants according to quartiles of CRI-I

Variable		CRI-I quartiles				P-value
		Q1 (≤ 3.11)	Q2 (3.11–3.85)	Q3 (3.85–4.82)	Q4 (> 4.82)
Sample size	Overall (n = 8805)	2222	2195	2194	2194
Gender, n (%)						< 0.001
Male	4151 (47.1)	1165 (52.4)	993 (45.2)	978 (44.6)	1015 (46.3)
Female	4654 (52.9)	1057 (47.6)	1202 (54.8)	1216 (55.4)	1179 (53.7)
Age, (median [IQR])	60.00 [54.00, 67.00]	61.00 [54.00, 68.00]	60.00 [53.00, 68.00]	60.00 [54.00, 66.00]	60.00 [54.00, 67.00]	0.019
Residence, n (%)						< 0.001
Agricultural	7252 (82.4)	1928 (86.8)	1832 (83.5)	1789 (81.5)	1703 (77.6)
Non-Agricultural	1553 (17.6)	294 (13.2)	363 (16.5)	405 (18.5)	491 (22.4)
Education, n (%)						0.028
Below high school	7941 (90.2)	2029 (91.3)	1975 (90.0)	1990 (90.7)	1947 (88.7)
High school or above	864 (9.8)	193 (8.7)	220 (10.0)	204 (9.3)	247 (11.3)
Marital status, n (%)						0.159
Yes	7757 (88.1)	1928 (86.8)	1939 (88.3)	1943 (88.6)	1947 (88.7)
No	1048 (11.9)	294 (13.2)	256 (11.7)	251 (11.4)	247 (11.3)
Smoking, n (%)						0.001
Yes	3457 (39.3)	954 (42.9)	839 (38.2)	817 (37.2)	847 (38.6)
No	5348 (60.7)	1268 (57.1)	1356 (61.8)	1377 (62.8)	1347 (61.4)
Alcohol consumption, n (%)						< 0.001
Yes	2876 (32.7)	885 (39.8)	710 (32.3)	653 (29.8)	628 (28.6)
No	5929 (67.3)	1337 (60.2)	1485 (67.7)	1541 (70.2)	1566 (71.4)
Socializing, n (%)						< 0.001
Yes	4339 (49.3)	1006 (45.3)	1093 (49.8)	1174 (53.5)	1193 (54.4)
No	4466 (50.7)	1216 (54.7)	1102 (50.2)	1020 (46.5)	1001 (45.6)
Hypertension, n (%)						< 0.001
Yes	3723 (42.3)	741 (33.3)	875 (39.9)	975 (44.4)	1132 (51.6)
No	5082 (57.7)	1481 (66.7)	1320 (60.1)	1219 (55.6)	1062 (48.4)
Diabetes, n (%)						< 0.001
Yes	1523 (17.3)	236 (10.6)	314 (14.3)	364 (16.6)	609 (27.8)
No	7282 (82.7)	1986 (89.4)	1881 (85.7)	1830 (83.4)	1585 (72.2)
Stroke, n (%)						< 0.001
Yes	188 (2.1)	27 (1.2)	42 (1.9)	44 (2.0)	75 (3.4)
No	8617 (97.9)	2195 (98.8)	2153 (98.1)	2150 (98.0)	2119 (96.6)
BMI, (median [IQR])	23.35 [21.17, 25.97]	21.69 [20.04, 23.79]	22.87 [20.85, 25.05]	24.11 [21.85, 26.58]	25.13 [22.89, 27.61]	< 0.001
Sleep, (median [IQR])	6.00 [5.00, 8.00]	6.00 [5.00, 8.00]	6.00 [5.00, 8.00]	7.00 [5.00, 8.00]	6.50 [5.00, 8.00]	0.404
HDL-C, (median [IQR])	49.10 [40.21, 59.54]	64.18 [56.83, 73.84]	53.35 [47.17, 60.31]	45.62 [40.21, 51.03]	36.34 [31.41, 41.75]	< 0.001
LDL-C, (median [IQR])	114.43 [93.17, 137.63]	93.56 [79.25, 110.18]	113.66 [97.42, 129.90]	126.42 [106.32, 144.98]	131.44 [103.61, 158.51]	< 0.001
TC, (median [IQR])	190.21[167.01, 215.34]	171.46[151.55, 193.30]	185.18[164.69, 206.44]	196.01[173.97, 216.88]	212.63 [187.50, 241.62]	< 0.001
TG, (median [IQR])	106.20 [75.22, 156.65]	70.80 [55.76, 91.15]	92.93 [72.57, 120.36]	121.25 [92.04, 156.65]	184.08 [132.97, 263.73]	< 0.001
hsCRP, (median [IQR])	1.06 [0.56, 2.19]	0.76 [0.45, 1.72]	0.89 [0.49, 1.93]	1.10 [0.62, 2.20]	1.45 [0.82, 3.00]	0.001
AIP, (median [IQR])	0.77 [0.29, 1.30]	0.10 [-0.19, 0.39]	0.56 [0.28, 0.85]	0.96 [0.68, 1.28]	1.59 [1.23, 2.05]	< 0.001
LCI, (median [IQR])	47636.88	18019.71	36075.61	63831.15	133303.02	< 0.001
	[25719.90, 90686.81]	[12836.47, 24980.58]	[27504.81, 48293.21]	[47660.84, 83971.67]	[95769.12, 190537.55]
Non–HDL-C, (median [IQR])	139.18 [115.98, 165.08]	105.93 [92.40, 121.78]	131.44 [116.75, 146.52]	150.00 [133.38, 167.01]	175.71 [153.87, 201.03]	< 0.001
CRI-II, (median [IQR])	2.34 [1.78, 3.01]	1.49 [1.26, 1.70]	2.12 [1.95, 2.32]	2.74 [2.51, 2.98]	3.54 [3.09, 4.06]	< 0.001
RC, (median [IQR])	20.10 [11.60, 32.47]	11.21 [6.19, 17.40]	16.62 [10.44, 24.36]	23.20 [15.46, 32.86]	38.66 [25.52, 58.38]	< 0.001
RC/HDL-C, (median [IQR])	0.40 [0.21, 0.75]	0.17 [0.09, 0.28]	0.32 [0.19, 0.47]	0.50 [0.33, 0.75]	1.04 [0.67, 1.74]	< 0.001
Sarcopenia, n (%)						< 0.001
Yes	1332 (15.1)	573 (25.8)	382 (17.4)	230 (10.5)	147 (6.7)
No	7473 (84.9)	1649 (74.2)	1813 (82.6)	1964 (89.5)	2047 (93.3)

Data are shown as median [IQR] for continuous variables and as percentages (%) for categorical variables. P-values were obtained using Kruskal-Wallis test for continuous variables and Chi-square tests for categorical variables to assess group differences

CRI-I Castelli Risk Index-I, BMI body mass index, HDL-C high-density lipoprotein cholesterol, LDL-C low-density lipoprotein cholesterol, TC total cholesterol, TG triglyceride, hsCRP high-sensitivity C-reactive protein, AIP atherogenic index of plasma, LCI lipoprotein combined index, Non–HDL-C non-high-density lipoprotein cholesterol, CRI-II Castelli Risk Index-II, RC remnant cholesterol, IQR interquartile range

Performance comparison of seven lipid indices

ROC analyses of seven lipid indices, including AIP, LCI, Non–HDL-C, CRI-I, CRI-II, RC, and RC/HDL-C, were conducted to evaluate their ability to discriminate sarcopenia (Fig. 2). The AUC values for all indices exceeded 0.5, indicating acceptable discriminative performance. Among these indices, CRI-I demonstrated superior AUC performance (0.667, 95% CI: 0.652–0.683), indicating better discrimination than the others. DeLong test comparisons indicated that CRI-I differed from most other lipid indices (Supplementary Table 2). Notably, although the AUC difference between CRI-I and AIP was not statistically significant (P = 0.177), CRI-I consistently exhibited higher AUC values and overall performance, and was therefore selected as the primary lipid indicator for further analyses.

Fig. 2.

Discriminatory performance of lipid indices. A ROC curves for the discriminatory value of lipid indices. B Bar chart of the discriminatory value of lipid indices. AIP, atherogenic index of plasma; LCI, lipoprotein combined index; Non–HDL-C, non-high-density lipoprotein-cholesterol; CRI-I, Castelli Risk Index-I; CRI-II, Castelli Risk Index-II; RC, remnant cholesterol; AUC, area under the curve; PPV, positive predictive value; NPV, negative predictive value

Association between CRI-I and sarcopenia

Multivariable logistic regression indicated that CRI-I was inversely correlated with the odds of sarcopenia (Table 2). Using Q1 as the reference, model 1 showed that the ORs for Q2, Q3, and Q4 were 0.61 (95% CI: 0.52–0.70), 0.34 (95% CI: 0.29–0.40), and 0.21 (95% CI: 0.17–0.25), respectively (all P < 0.001). In model 2, this association was further strengthened, with ORs of 0.55 (95% CI: 0.47–0.66) for Q2, 0.31 (95% CI: 0.25–0.37) for Q3, and 0.17 (95% CI: 0.14–0.21) for Q4. In model 3, ORs for Q2–Q4 were 0.58 (95% CI: 0.48–0.70), 0.33 (95% CI: 0.26–0.42) and 0.21 (95% CI: 0.16–0.29), corresponding to 42%, 67%, and 79% lower odds of sarcopenia compared with Q1 (all P < 0.001). Upon treating CRI-I as a continuous measure, the odds of sarcopenia decreased by approximately 42% for each one-unit increase in model 3.

Table 2.

Association between CRI-I and sarcopenia in CHARLS

	OR (95% CI), P-value
	Model 1	Model 2	Model 3
Continuous CRI-I	0.60 (0.57–0.64) < 0.001	0.58 (0.54–0.61) < 0.001	0.58 (0.52–0.64) < 0.001
CRI-I quartiles
Q1 (≤ 3.11)	1.00 (Ref)	1.00 (Ref)	1.00 (Ref)
Q2 (3.11–3.85)	0.61 (0.52–0.70) < 0.001	0.55 (0.47–0.66) < 0.001	0.58 (0.48–0.70) < 0.001
Q3 (3.85–4.82)	0.34 (0.29–0.40) < 0.001	0.31 (0.25–0.37) < 0.001	0.33 (0.26–0.42) < 0.001
Q4 (> 4.82)	0.21 (0.17–0.25) < 0.001	0.17 (0.14–0.21) < 0.001	0.21 (0.16–0.29) < 0.001
P for trend	< 0.001	< 0.001	< 0.001

Annotation: Model 1 adjusted for none. Model 2 adjusted for age, gender, residence, education, marital status. Model 3 adjusted for smoking, alcohol consumption, socializing, sleep, hypertension, diabetes, stroke, TG, LDL-C, hsCRP on the basis of Model 2

CRI-I Castelli Risk Index-I, TG triglyceride, LDL-C low-density lipoprotein cholesterol, hsCRP high-sensitivity C-reactive protein, CI confidence interval, OR odds ratio

We compared VIFs before and after including TC and HDL-C. Substantial multicollinearity was observed when TC and HDL-C were included (Supplementary Table 3), whereas all VIFs were < 5 after their exclusion (Supplementary Table 4). Therefore, to minimize collinearity and potential overadjustment, TC, HDL-C, and BMI were excluded from the final model.

We also observed a nonlinear relationship between CRI-I and sarcopenia in the RCS analysis, with significant overall (P < 0.001) and nonlinearity (P < 0.001) tests (Fig. 3). Evidence for a nonlinear relationship was further supported by threshold effect analysis (Supplementary Table 5). A CRI-I value of 4.84 was identified as a threshold. Below this level, CRI-I was significantly inversely associated with sarcopenia (OR = 0.48, 95% CI: 0.42–0.54), whereas the association attenuated and was no longer statistically significant above the threshold (OR = 0.87, 95% CI: 0.73–1.03). The log-likelihood ratio test indicated that the piecewise linear model provided a significantly better fit than the linear model (P < 0.001).

Fig. 3.

Association between CRI-I and the prevalence of sarcopenia in RCS. The solid line represents an estimate of OR and the shaded area represents 95% CI. The model was adjusted for age, gender, residence, education, marital status, smoking, alcohol consumption, socializing, sleep, hypertension, diabetes, stroke, TG, LDL-C and hsCRP. CRI-I, Castelli Risk Index-I; TG, triglyceride; LDL-C, low-density lipoprotein cholesterol; hsCRP, high-sensitivity C-reactive protein; CI, confidence interval; OR, odds ratio

Subgroup and interaction analyses

An examination of the relationship between CRI-I and sarcopenia within various demographic categories was performed. These subgroup analyses indicated that the link between CRI-I and sarcopenia remained robust, irrespective of the group (Fig. 4). In all stratified variables, participants in Q4 of CRI-I exhibited the most pronounced protective effect against sarcopenia compared with those in Q1, which indicates the strong stability and generalizability of the association. Interaction analyses further revealed a significant sex-based interaction (P for interaction < 0.05), suggesting that the inverse association between CRI-I and sarcopenia was stronger in men. A significant interaction effect was also observed in relation to alcohol consumption. In other subgroups, no such significant interactions came to light.

Fig. 4.

The plot displays the OR and 95% CI for sarcopenia across different subgroups. CRI-I, Castelli Risk Index-I; CI, confidence interval; OR, odds ratio. P-interaction values were calculated using likelihood ratio tests

Sensitivity analyses

Sensitivity analyses revealed the robustness of the inverse association between CRI-I and sarcopenia. After multiple imputation of covariates with missing data, both categorical (quartiles) and continuous analyses of CRI-I yielded results consistent with the primary analysis, indicating that the main findings were not materially influenced by missing data handling (Supplementary Table 6). When the SI was used as an alternative muscle-related indicator, CRI-I was positively associated with SI in linear regression analyses. In the fully adjusted model, this association remained significant (β = 0.50, 95% CI: 0.17–0.84, P = 0.003), demonstrating robustness after accounting for potential confounders. Categorical analyses by CRI-I quartiles further illustrated this relationship, with participants in the highest quartile exhibiting significantly higher SI compared to those in the lowest quartile (β = 5.31, 95% CI: 3.78–6.83, P < 0.001). Since higher SI values reflect better muscle mass, these findings indirectly suggested that elevated CRI-I levels were associated with a lower occurrence of sarcopenia, thereby further validating the primary results through an alternative analytical approach (Supplementary Table 7). Baseline characteristics of participants included in the SI-based sensitivity analysis are summarized in Supplementary Table 8.

Feature selection in machine learning model

The sample was randomly divided into 70% training set (n = 6,163) and 30% testing set (n = 2,642). The baseline characteristics of these sets are summarized in Supplementary Table 9. All feature selection procedures were performed in the training set. Variables identified in univariate logistic regression screening as significantly associated with sarcopenia included age, CRI-I, marital status, TG, education, residence, sleep, socializing, diabetes, smoking, LDL-C, hsCRP, hypertension, and gender (Supplementary Table 10). These candidate variables were subsequently entered into a LASSO regression model, from which eight variables were retained as key features: age, CRI-I, residence, education, socializing, hypertension, diabetes, and sleep (Fig. 5).

Fig. 5.

Feature selection by LASSO regression. A LASSO coefficient profiles of candidate variables plotted against the log(λ) sequence. B Optimal λ value was selected via 5-fold cross-validation

Model development and evaluation

Model performance was compared between a traditional logistic regression model and six machine learning algorithms, with results summarized in Supplementary Table 11. In the training set, XGBoost demonstrated strong discriminative performance with an AUC of 0.864 (95% CI: 0.853–0.875), along with balanced classification metrics, including a sensitivity of 0.840 (95% CI: 0.815–0.862), specificity of 0.744 (95% CI: 0.732–0.755), and accuracy of 0.758 (95% CI: 0.747–0.769). Comparable performance was observed in the testing set. Notably, XGBoost yielded the lowest Brier scores in both the training and testing sets, indicating superior calibration relative to the other models. Although certain models achieved similar or marginally higher AUC values, their calibration or classification balance was inferior. Therefore, considering discrimination, classification metrics, and calibration together, XGBoost was selected as the model with the best overall performance, into which CRI-I was incorporated to assess its incremental discriminative value.

Incremental discriminative value of CRI-I

As shown in Figs. 6, 7 and 8, incorporation of CRI-I into the XGBoost model resulted in consistent improvements in model performance across the training, testing, and external validation sets. The AUC increased to 0.886 (95% CI: 0.876–0.896) in the training set, 0.871 (95% CI: 0.855–0.888) in the testing set, and 0.893 (95% CI: 0.866–0.919) in the external validation set. DeLong tests confirmed statistically significant AUC improvements in the training and testing sets (both P < 0.001), but not in the external validation set (P = 0.466) (Supplementary Table 12). Consistent with these findings, incorporation of CRI-I also led to improvements in precision-recall performance. DCA demonstrated that the XGBoost model incorporating CRI-I yielded a consistently higher net benefit than the model without CRI-I across a wide range of clinically relevant threshold probabilities, supporting the incremental clinical utility of CRI-I. Calibration curves demonstrated excellent agreement between the model-estimated probabilities and the observed outcomes. H–L test indicated good calibration in the training set (χ² = 10.629, df = 8, P = 0.224) and the testing set (χ² = 11.405, df = 8, P = 0.180), whereas the external validation set exhibited calibration deviation (χ² = 26.211, df = 8, P = 0.001). In addition, the SHAP algorithm was applied to quantify feature contributions to the model’s estimated probability of sarcopenia in the XGBoost model (Fig. 9). Age was the most important feature, followed by CRI-I, hypertension, residence, sleep, education, socializing and diabetes. Advancing age, reduced social participation, lower educational level, rural residence and shorter sleep time were related to a higher model-estimated probability of sarcopenia.

Fig. 6.

Comparison of ROC and PR curves between XGBoost models with and without CRI-I. A, B, C ROC curves for the training set, testing set, and external validation set. D, E, F PR curves for the training set, testing set, and external validation set. XGBoost, Extreme Gradient Boosting; CRI-I, Castelli Risk Index-I; ROC, receiver operating characteristic; PR, precision-recall; AUC, area under the curve

Fig. 7.

Comparison of DCA curves between XGBoost models with and without CRI-I. A DCA curve for the training set. B DCA curve for the testing set. C DCA curve for the external validation set. XGBoost, Extreme Gradient Boosting; CRI-I, Castelli Risk Index-I; DCA, decision curve analysis

Fig. 8.

Comparison of calibration curves between XGBoost models with and without CRI-I. A calibration curve for the training set. B calibration curve for the testing set. C calibration curve for the external validation set. XGBoost, Extreme Gradient Boosting; CRI-I, Castelli Risk Index-I

Fig. 9.

SHAP summary and feature importance plots. A The SHAP summary plot illustrates each feature’s impact on the model’s discrimination. Positive SHAP values denote increased prevalence, while negative values indicate decreased prevalence. B The average SHAP value was used to rank feature importance, where higher values indicate greater importance in model discrimination. SHAP, SHapley Additive exPlanations

Discussion

This study, based on a large Chinese cohort, systematically compared seven non-traditional lipid indices and identified CRI-I as the one with the strongest discriminative ability for sarcopenia. In addition, we incorporated CRI-I into the XGBoost model to assess its incremental discriminative value. We observed a robust inverse association between CRI-I and sarcopenia that was generally consistent across subgroups. Integrating CRI-I into a machine learning model, we further improved discrimination and enhanced identification of individuals with a higher likelihood of prevalent sarcopenia.

In our study, muscle mass was estimated using an anthropometric equation derived from DXA-measured ASM. Although this equation showed a high correlation with DXA (R² ≈ 0.9), it relies on body size surrogates and may not accurately distinguish skeletal muscle from adipose tissue. This could lead to misclassification, particularly in individuals with atypical body composition [26, 29]. In addition, the equation was developed and validated in a specific population, and its accuracy may vary across ethnic groups, age structures, and metabolic profiles, which may introduce systematic estimation error in a large heterogeneous cohort. To mitigate this potential measurement bias, we performed a sensitivity analysis using the SI, which has been proposed as a validated biomarker of skeletal muscle mass [28, 30, 31]. As a biochemical indicator, SI is less dependent on body size measures and may provide a complementary assessment of muscle status. In multivariable linear regression, the observed positive association between CRI-I and SI further supported the inverse association between CRI-I and sarcopenia. Nonetheless, some limitations remain. Evidence supporting SI as a reliable muscle mass surrogate remains inconsistent across studies [32], and SI was also initially developed and validated in a specific population, which warrants further validation. Because the sarcopenia assessment in this study incorporated body weight and height, adjusting for BMI as a covariate could lead to overadjustment bias, and BMI was therefore excluded from the primary models. However, this exclusion may leave residual confounding related to adiposity. Since SI is derived entirely from serum biomarkers, BMI can be included without the same concern for overadjustment, allowing us to evaluate whether the association is independent of body habitus. The consistency of findings across these two approaches strengthens our confidence in the observed association.

Previous studies have suggested that lipid metabolic dysfunction may promote the development of sarcopenia by inducing fat accumulation, chronic inflammation, insulin resistance, and oxidative stress, which accelerate muscle protein degradation and inhibit protein synthesis [4, 33]. Several studies have investigated the associations between conventional lipid indices (TC, TG, LDL-C and HDL-C) and sarcopenia [34–36]. However, growing evidence suggests that single lipid measures reflect only a limited aspect of lipid metabolism, and fail to represent overall metabolic load and lipid balance [37], which limits their usefulness for discriminating sarcopenia status. Consequently, attention has shifted toward composite lipid ratios, such as TG/HDL-C, TC/HDL-C, and LDL/HDL-C, which are considered to better reflect overall metabolic status [38, 39]. Nevertheless, findings regarding the relationships between composite lipid markers and sarcopenia remain inconsistent. Several epidemiological studies have reported inverse associations of TG/HDL-C or Non–HDL-C to HDL-C ratio (NHHR) with sarcopenia [27, 40, 41], suggesting that moderate lipid levels may help maintain muscle metabolism. In contrast, studies conducted in Korean and American cohorts have demonstrated opposite trends [9, 42]. These discrepancies may be related to differences in metabolic status, ethnicity, and dietary patterns of the study populations, as well as variations in the assessment of sarcopenia, statistical models, and control of confounding factors [40, 43]. Therefore, it is essential to develop a lipid indicator that accurately reflects overall metabolic balance.

In this study, CRI-I demonstrated superior discriminative performance compared with other lipid indices. This advantage may stem from its integration of two lipid components, TC and HDL-C, with opposing metabolic effects. By combining these, CRI-I reflects the balance between cholesterol burden and HDL-mediated reverse cholesterol transport [14]. This dual-component ratio may therefore provide a more accurate representation of the metabolic interactions among lipid fractions, and might be relevant to metabolic reserve and muscle health. Moreover, because its calculation was simple and less affected by short-term fluctuations in triglycerides, CRI-I exhibited high stability and reproducibility across statistical models. These characteristics not only explained its robust performance in the present analysis but also accounted for its widespread use in studies on metabolic syndrome, insulin resistance, and atherosclerosis [15, 16].

RCS and threshold analyses indicated that the inverse association between CRI-I and prevalent sarcopenia was most prominent at lower CRI-I levels, with the relationship attenuating beyond an approximate turning point in our cohort. Lower CRI-I values could result from lower TC, higher HDL-C, or the coexistence of both features. In this context, extremely low CRI-I is more likely to indicate limited metabolic reserve and health vulnerability rather than reflecting an ideal lipid status [44, 45].

Moreover, lower TC reflects undernutrition and frailty, which indicate insufficient energy and substrate reserves for sustained muscle protein synthesis, coupled with an increased susceptibility to net protein breakdown during chronic illness [44, 46, 47]. This state of metabolic vulnerability also contributes to chronic inflammation and oxidative stress [48, 49]. Similarly, a higher HDL-C concentration in this clinical setting does not necessarily imply protective HDL function [50, 51]. In chronic inflammation and oxidative stress, HDL particles undergo remodeling and oxidative modification. This reduces cholesterol efflux capacity and weakens HDL’s anti-inflammatory and antioxidative functions, limiting the protection typically attributed to high HDL-C [50, 52, 53]. This quantity-function mismatch is consistent with longitudinal evidence reporting a nonlinear relationship, with very high HDL-C being linked to subsequent muscle strength decline and incident sarcopenia [54]. Therefore, the steep decline in prevalent sarcopenia in the lower range of CRI-I reflects a shift away from limited metabolic reserve and potentially dysfunctional HDL.

Beyond the estimated turning point in our cohort, further increases in CRI-I should not be interpreted as a universal biological cut-off. Instead, in this higher range, a higher ratio is more likely to reflect a more atherogenic cholesterol profile, which often accompanies insulin resistance related dyslipidemia [55]. These metabolic disturbances commonly coexist with chronic low-grade inflammation and may contribute to skeletal muscle anabolic resistance, reducing the muscle’s synthetic response to nutritional and physical stimuli [56, 57]. In such a multifactorial setting, a single lipid-based ratio may explain less of the variability in prevalent sarcopenia across individuals, consistent with the curve showing attenuation and plateau at higher CRI-I levels. From a clinical standpoint, this pattern supports using CRI-I as an opportunistic case finding signal for prevalent sarcopenia and reserving guideline-consistent confirmatory assessment of muscle strength, physical performance, and muscle mass, for individuals flagged by the index [1].

Apart from CRI-I, our machine learning model also identified several variables associated with sarcopenia, including age, residence, education, socializing, hypertension, diabetes, and sleep. SHAP analysis revealed that older age, rural residence, reduced social engagement, lower educational attainment, and shorter sleep duration were associated with a higher probability of sarcopenia. These factors have been widely recognized as correlates of sarcopenia in previous epidemiological studies [3, 58]. Notably, hypertension and diabetes showed inverse associations with sarcopenia in our sample, contrary to most previous reports [59]. In our study, hypertension and diabetes status were determined by self-report and medication use. Prior research has demonstrated that antihypertensive and antidiabetic medications may exert protective effects on muscle mass or physical performance [60, 61]. Therefore, we speculated that our finding may be partly attributable to the use of these medications, which could help attenuate muscle loss to some extent. Overall, these findings highlight that sarcopenia results from the complex interplay of multiple biological, behavioral, and environmental factors.

Previous studies showed that existing sarcopenia risk models often suffered from incorporation bias, complexity of variables, and limited external validation [17–20], which affected their reliability and generalizability. In contrast, our model used a simpler set of variables that were easily accessible in routine clinical practice, enhancing its feasibility for primary care. The model also avoided the inclusion of diagnostic components, such as BMI and handgrip strength, which overlapped with outcome definitions under the Asia Working Group for Sarcopenia 2019 sarcopenia assessment criteria, thus reducing incorporation bias. The inclusion of CRI-I, a composite lipid index rarely considered in earlier models, improved the model’s discriminative ability, providing additional value beyond established correlates of sarcopenia. Finally, the temporal external validation conducted in this study enhanced the model’s generalizability to the Chinese middle-aged and elderly population.

Potential clinical implementation and future directions

From a translational perspective, the variables incorporated in this model are routinely captured during community health services, primary care encounters, and health examinations, supporting pragmatic implementation in real-world settings [62]. When embedded in electronic health record systems or health examination platforms, the model could automatically estimate the probability of sarcopenia, enabling opportunistic screening in routine clinical practice [62]. A probability threshold informed by DCA could be used to identify individuals who warrant confirmatory assessment, prioritizing muscle strength and physical performance and adding muscle mass measurement when available in accordance with consensus pathways [1, 63]. Those confirmed could then be referred for exercise and nutrition-focused management, which is most relevant for primary care teams, geriatric clinics, and health examination centers [1]. Prior to deployment, multicenter external validation and local recalibration are needed. Moreover, ongoing surveillance should address data heterogeneity, missingness, workflow integration, the risk of alert fatigue, and assess the model’s performance in different population subgroups [64, 65]. Future work should include prospective impact studies and cost-effectiveness evaluations [66, 67], alongside transparent reporting and critical appraisal guided by TRIPOD + AI and PROBAST + AI [68, 69].

Strengths and limitations

We used a machine learning approach to assess the incremental discriminative value of CRI-I for identifying sarcopenia. Moreover, we compared CRI-I with six other non-traditional lipid indices to evaluate their discriminative capacity. Lastly, comprehensive evaluation of model efficacy involved ROC and PRC, calibration charts, and DCA. At the same time, several limitations warrant consideration. To begin with, although we performed temporal validation using the CHARLS 2015 wave, both datasets came from the same CHARLS platform and primarily represent a Chinese population. The model may therefore be better suited to similar Chinese settings, while its transportability to other populations and healthcare contexts may be limited. Prospective studies in independent cohorts are warranted to further evaluate its broader applicability. Additionally, the cross-sectional design precludes the establishment of temporal sequence and causal inference. The observed associations may be subject to reverse causation, as sarcopenia may influence lipid metabolism and affect CRI-I values. Longitudinal studies are needed to clarify the directionality of these relationships. Moreover, muscle mass was estimated using anthropometric equations rather than DXA, which may introduce measurement error and misclassification of sarcopenia. Lastly, although the analysis was based on a large national dataset, the limited number of sarcopenia cases may increase the risk of overfitting.

Conclusion

This study demonstrated that CRI-I is a stable and effective lipid-based indicator for identifying individuals with a higher likelihood of prevalent sarcopenia. Compared with other non-traditional lipid indices, CRI-I showed superior discriminative capacity and improved the discriminative performance of the XGBoost model. These results suggest that CRI-I, as an easily obtainable and low-cost lipid biomarker, may facilitate opportunistic screening to identify individuals more likely to have sarcopenia in primary care settings and routine health examinations. Nevertheless, prospective cohort studies and multicenter research are warranted to validate these connections and assess their generalizability. Moreover, it is critically important to conduct further research to shed light on the physiological processes that connect lipid metabolism with muscle health.

Abbreviations

CRI-I

Castelli Risk Index-I

BMI

Body mass index

HDL-C

High-density lipoprotein cholesterol

LDL-C

Low-density lipoprotein cholesterol

TC

Total cholesterol

TG

Triglyceride

hsCRP

High-sensitivity C-reactive protein

AIP

Atherogenic index of plasma

LCI

Lipoprotein combined index

Non–HDL-C

Non-high-density lipoprotein cholesterol

CRI-II

Castelli Risk Index-II

RC

Remnant cholesterol

ROC

Receiver operating characteristic

AUC

Area under the curve

LASSO

Least absolute shrinkage and selection operator

AdaBoost

Adaptive Boosting

KNN

K-Nearest Neighbors

RF

Random Forest

SVM

Support Vector Machine

XGBoost

Extreme Gradient Boosting

GNB

Gaussian Naïve Bayes

SHAP

SHapley Additive exPlanations

DCA

Decision curve analysis

AUPRC

Area under the precision-recall curve

VIF

Variance inflation factor

RCS

Restricted cubic splines

CHARLS

China Health and Retirement Longitudinal Study

DXA

Dual-energy X-ray Absorptiometry

ASM

Appendicular skeletal muscle mass

IQR

Interquartile Range

Authors’ contributions

Conceptualization, resources and supervision: WMH and FZ; Formal analysis and investigation: BDX and HLS; Writing-original draft preparation: ZMZ and GQW; Writing-review and editing: BDX, HLS, WMH and FZ.

Funding

This work was supported by the Youth Foundation of National Natural Science Foundation of China (Nos. 82003795), University Natural Science Research Project of Anhui Province (2023AH053318, 2025AHGXZK30594), clinical science foundation (2023xkj193), College Students’Innovation Training Project (S202410366072, X202510366095), Nursing Project of Anhui Institute of Translational Medicine, and Doctor Research Project of the First Affiliated Hospital of Anhui Medical University (Nos. 2024zhyx-hl-B20, and BSKY2019016), and Teaching Project of Anhui Medical University (2024shsjsfkc011, 2024xjxm76).

Data availability

This research utilized information sourced from the CHARLS database. Access to the dataset is available at [ https://charls.pku.edu.cn.]( https://charls.pku.edu.cn) The code used for data cleaning and statistical analyses in the present study is provided in the supplementary material 1.

Declarations

Ethics approval and consent to participate

The study was approved by the Biomedical Ethics Review Board of Peking University (IRB00001052–11015). All participants provided written informed consent, and all procedures followed relevant ethical guidelines and regulations.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.