Dynamic sarcopenia transitions in older Chinese: physical activity and cognitive insights Authors: Yutao Li, Yuxin Tang, Weifeng Pan, Hengguo Song Published: 2025 ABSTRACT ObjectivesTo investigate sarcopenia state transitions (non-sarcopenia, possible sarcopenia, and sarcopenia) and their determinants among older Chinese adults, emphasizing the roles of physical activity, cognitive status, and other risk factors. MethodsA longitudinal study utilizing data from the China Health and Retirement Longitudinal Study (CHARLS) spanning 2011–2015, integrating Multi-State Markov (MSM) models and Transformer-based deep learning approaches. We examined 5,756 participants (including 3,373 for deep learning) across three waves, with a mean age of 67.9 years (SD = 6.5). MSM models estimated transition intensities and probabilities between sarcopenia states. Deep learning with SHAP analysis identified key determinants of transitions and mortality. Covariates included age, sex, BMI, smoking, physical activity, mild cognitive impairment (MCI), and functional disability.ResultsMSM models indicated a high transition rate from non-sarcopenia to possible sarcopenia (intensity: 0.383, 95% CI: 0.355–0.411) and a 38.6% five-year recovery probability from possible sarcopenia to non-sarcopenia. Physical activity reduced deterioration risk (HR: 0.916, 95% CI: 0.842–0.997) and mortality in possible sarcopenia (HR: 0.565, 95% CI: 0.339–0.944). MCI increased deterioration risk (HR: 1.724, 95% CI: 1.268–2.346). Age > 80 significantly elevated deterioration (HR: 3.007, 95% CI: 1.992–4.538) and mortality risks (HR: 7.400, 95% CI: 2.542–21.544). Sex, BMI, smoking, and functional disability also influenced transitions.ConclusionsSarcopenia exhibits bidirectional progression, with physical activity serving as a key protective factor. MCI attenuates this benefit, highlighting the need for tailored interventions that address cognitive status and other risk factors in older adults.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12877-025-06752-5. Introduction Sarcopenia, a progressive age-related condition marked by the decline of skeletal muscle mass and function, poses a growing threat to public health as populations worldwide age [1]. This disorder heightens the risk of falls, fractures, and dependency, often culminating in increased mortality and placing significant burdens on societal and economic resources [2]. With the accelerating trend of global aging, the prevalence of sarcopenia is anticipated to rise sharply [3]. The Asian Working Group for Sarcopenia (AWGS) 2019 consensus offers a tailored diagnostic framework for Asian populations, classifying sarcopenia into three states—non-sarcopenia, possible sarcopenia, and sarcopenia—based on muscle mass, strength, and physical performance [4]. The concept of “possible sarcopenia,” a transitional phase between health and disease, suggests a dynamic potential for either recovery or progression, a feature shared with conditions like Alzheimer’s disease [5]. Evidence increasingly points to physical activity as a pivotal lifestyle factor capable of halting Sarcopenia’s advance and fostering improvement [5]. Concurrently, mild cognitive impairment (MCI), a prevalent cognitive condition among older adults, may interact with physical activity to shape sarcopenia’s trajectory, though the nature of this interplay remains largely uncharted [6]. The course of sarcopenia is far from uniform, varying widely across individuals and reflecting a complex, evolving process [2]. Multi-state Markov (MSM) models have proven effective in mapping state transitions in chronic diseases, quantifying the pace and likelihood of shifts between states, yet their use in sarcopenia research remains underexplored [7]. Meanwhile, deep learning approaches have emerged as a robust complement, adept at uncovering non-linear relationships and intricate interactions among factors [8]. Drawing on data from the China Health and Retirement Longitudinal Study (CHARLS), a nationally representative longitudinal survey of Chinese adults aged 45 and older, this study investigates sarcopenia state transitions in older Chinese populations [9]. By merging MSM models with cutting-edge deep learning techniques, we seek to illuminate the dynamic progression from non-sarcopenia through possible sarcopenia to sarcopenia, placing special emphasis on the central role of physical activity and its interaction with MCI. The MSM framework enables us to measure transition intensities and probabilities between states and assess covariate influences, while deep learning reveals the nuanced interplay between physical activity, MCI, and sarcopenia outcomes. This dual-method approach not only enriches our understanding of sarcopenia’s developmental pathways but also lays a foundation for evidence-based interventions tailored to aging populations [10–12]. Method Data sources This study utilizes data from the China Health and Retirement Longitudinal Study (CHARLS), a nationally representative longitudinal survey of individuals aged 45 and older in China, managed by Peking University’s National School of Development and Institute of Social Science Survey. CHARLS collects comprehensive microdata on demographics, health, socioeconomic status, and lifestyle to support aging research [9]. Initiated in 2011, the baseline survey employed multistage PPS sampling across 28 provinces, 150 counties, and 452 communities, enrolling over 17,000 participants. Biennial follow-ups via CAPI have since tracked the cohort. This analysis draws on Waves 1 (2011), 2 (2013), and 3 (2015) from the five-wave dataset (up to Wave 5, 2020), selected for their anthropometric and functional performance data essential for sarcopenia assessment. Participants aged 60 + with adequate sarcopenia evaluations across these waves were included, yielding 5,756 individuals for MSM analysis, of whom 3,373 also qualified for Deep Learning analysis (see Fig. 1). Approved by Peking University’s Biomedical Ethics Committee, CHARLS obtained written informed consent from all participants, and the dataset is publicly available. Fig. 1Flowchart of Participant Selection Flowchart of Participant Selection Diagnosis of sarcopenia states Sarcopenia was classified based on the AWGS 2019 criteria [4]. Non-sarcopenia was defined by normal muscle strength and physical performance. Possible sarcopenia was identified by low muscle strength or physical performance, while sarcopenia required low muscle mass along with low muscle strength or reduced physical performance. Muscle strength was assessed using handgrip strength, measured with a Yuejian TM WL-1000 dynamometer by trained examiners. Participants squeezed both hands maximally at a 90° angle, and the average of two trials was recorded. Cutoff values for low muscle strength were < 18 kg for women and < 28 kg for men. Physical performance was assessed through gait speed, the five-time chair stand test, and the short physical performance battery (SPPB). Gait speed was measured as the faster of two 2.5-m walks at normal pace. The five-time chair stand test recorded the time to rise five times from a seated position. SPPB included three 10-s positions (side-by-side, semi-tandem, and tandem), scored out of 12 points (4 per test). Low physical performance was defined as gait speed < 1.0 m/s, five-time chair stand test ≥ 12 s, or SPPB score < 9, according to AWGS 2019. Muscle mass was estimated using appendicular skeletal muscle mass (ASM), validated against dual-energy X-ray absorptiometry (DXA) in the Chinese population [13], using the following formula: 1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}\:ASM=&0.193\times\:weight\left(kg\right)+0.107\times\:height\left(cm\right)\\&-4.147\times\:sex\left(1\:for\:male,2\:for\:female\right)\\&-0.036\times\:age\left(yr\right)-2.631 \end{aligned}$$\end{document} Height and weight were measured using a SecaTM213 stadiometer and Omron HN-286 scale, respectively, with gender coded as 1 for males and 0 for females. The lowest 20% of height-adjusted ASM (ASM/height²) defined low muscle mass, set at < 4.89 kg/m² for women and < 6.71 kg/m² for men [14]. Based on the AWGS 2019 criteria, participants were categorized into three groups: non-sarcopenia, possible sarcopenia, and sarcopenia. Physical activity PA was quantified using an indexed scoring system based on daily duration: <0.5 h = 0.5 points, 0.5–2 h = 1 point, 2–4 h = 3 points, > 4 h = 4 points. Weekly scores were calculated by multiplying participation days by the daily duration index per activity type. Total PA intensity was standardized with MET multipliers: PA score = 8.0 × weekly vigorous activity time score + 4.0 × weekly moderate activity time score + 3.3 × weekly walking time score, allowing comparison of energy expenditure across intensities. Mild cognitive impairment MCI was defined using the Telephone Interview for Cognitive Status (TICS), assessing memory (immediate/delayed recall of 10 words, 0–10 points each), visuoconstruction (copying a pentagon, 0–1 points), mathematical ability (serial subtraction, 0–5 points), and orientation (0–5 points), with a total score of 21. Participants aged 60 + were classified as having MCI if their score was ≥ 1 standard deviation below the age-specific norm (AACD criteria) [15]. Covariates Covariates included age group, gender, education, partner status, smoking, drinking, ADL disability, IADL disability, life satisfaction, health status, social activity, trust, chronic disease, depression, BMI, and expenditures. ADL disability was assessed based on difficulties with six tasks (e.g., dressing, bathing, eating), while IADL disability focused on instrumental tasks (e.g., housework, shopping, medication management); scores > 0 indicated impairment. Life satisfaction, health status, and expenditures were dichotomized at the median (1 = satisfied/healthy/higher, 0 = dissatisfied/unhealthy/lower). Social activity was scored by assigning 1 point for each organized activity (e.g., volunteering, classes) and dichotomized at the median (1 = more active, 0 = less active). Trust was measured as perceived support from family/friends (1 = available, 0 = none). Depression was assessed using a validated scale (e.g., CESD-10), with ≥ 10 indicating symptoms. BMI was classified per Chinese standards: <18.5 kg/m² (underweight), 18.5–24 kg/m² (normal), ≥ 24 kg/m² (overweight/obese). Chronic diseases (e.g., hypertension, diabetes, cardiac disease) were categorized as 0, 1, or ≥ 2 conditions. Deep learning model Feature standardization was performed using Min-Max scaling to normalize the data before model implementation. Four machine learning algorithms—Long Short-Term Memory (LSTM), LSTM with Attention (LSTM-Attention), Temporal Convolutional Networks (TCN), and Transformer—were evaluated to identify the most suitable model, with detailed descriptions in the supplementary files. Hyperparameter tuning was conducted via Bayesian optimization with Tree-structured Parzen Estimators (TPE), exploring potential configurations to optimize model parameters; further details on TPE are provided in the supplementary files. Model performance was assessed using five standard metrics: accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC-ROC), calculated according to standard definitions (Eqs. 2, 3, 4, 5, 6). The best-performing model was analyzed for global feature importance, offering insights into its behavior. SHAP (SHapley Additive exPlanations) was used for instance-level, detailed explanations of model predictions, with a comprehensive explanation in the supplementary files. All analyses were conducted in Python 3.9, with scikit-learn for preprocessing, PyTorch for model training, and the SHAP library for interpretability. 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}\:=\frac{\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}}{\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P}+\mathrm{F}\mathrm{N}}$$\end{document} 3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}=\frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}$$\end{document} 4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}=\frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}}$$\end{document} 5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{F}1=\frac{2\times\:\mathrm{P}\times\:\mathrm{R}}{\mathrm{P}+\mathrm{R}}$$\end{document} 6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{A}\mathrm{U}\mathrm{C}={\int\:}_{0}^{1}\mathrm{T}\mathrm{P}\mathrm{R}\left(\mathrm{F}\mathrm{P}\mathrm{R}\right)\mathrm{d}\left(\mathrm{F}\mathrm{P}\mathrm{R}\right)$$\end{document} MSM models We used the MSM package in R 4.2.3 to model sarcopenia state transitions. Four states were defined: non-sarcopenia (state 1), possible sarcopenia (state 2), sarcopenia (state 3), and death (state 4). States 1–3 were transient, allowing transitions to adjacent states or death, while state 4 was absorbing. Transition rates were estimated using maximum likelihood and BFGS optimization. A multivariate MSM model was constructed to examine covariate effects, with covariates selected based on a deep learning model. Full details of the MSM approach are provided in the supplementary files. All analyses were performed in R 4.2.3 with P < 0.05. Results Characteristics of participants A total of 5,756 individuals were included in the MSM analysis, with 3,373 also eligible for the Deep Learning Analysis. At baseline, 1,947 (33.8%) individuals were non-sarcopenia, 2,947 (51.2%) had possible sarcopenia, and 862 (15.0%) had sarcopenia. Baseline demographic and clinical characteristics, stratified by sarcopenia states, are presented in Table 1, with significant differences (p < 0.001) observed across most variables, including age, gender, education, lifestyle factors, health status, and functional disabilities, except for life satisfaction (p = 0.204) and trust (p = 0.252). Table 1Demographic and health characteristics of participants by sarcopeniaVariable_LevelNon-sarcopenia (N = 1947)Possible sarcopenia(N = 2947)Sarcopenia(N = 862)p-valueAge Average65.8 (5.1)67.8 (6.3)73.1 (7.3)< 0.001Age Group≤ 701620 (83.2%)2069 (70.2%)345 (40.0%)< 0.00171–80309 (15.9%)754 (25.6%)377 (43.7%)> 8018 (0.9%)124 (4.2%)140 (16.2%)GenderMale738 (37.9%)1659 (56.3%)477 (55.3%)< 0.001Female1209 (62.1%)1288 (43.7%)385 (44.7%)Education< high school1783 (91.6%)2817 (95.6%)844 (97.9%)< 0.001>high school164 (8.4%)130 (4.4%)18 (2.1%)PartnerYes1619 (83.2%)2233 (75.8%)542 (62.9%)< 0.001No328 (16.8%)714 (24.2%)320 (37.1%)SmokingNo1039 (53.4%)1894 (64.3%)511 (59.3%)< 0.001Yes908 (46.6%)1053 (35.7%)351 (40.7%)DringkingNo1183 (60.8%)2164 (73.4%)636 (73.8%)< 0.001Yes764 (39.2%)783 (26.6%)226 (26.2%)ADL DisabilityNo1639 (84.2%)2111 (71.6%)574 (66.6%)< 0.001Yes308 (15.8%)836 (28.4%)288 (33.4%)IADL DisabilityNo1633 (83.9%)1959 (66.5%)493 (57.2%)< 0.001Yes314 (16.1%)988 (33.5%)369 (42.8%)Life SatisfactionLow1474 (75.7%)2165 (73.5%)645 (74.8%)0.204High473 (24.3%)782 (26.5%)217 (25.2%)Health StatusPoor1566 (80.4%)2506 (85.0%)762 (88.4%)< 0.001Good381 (19.6%)441 (15.0%)100 (11.6%)Social ActivityLow910 (46.7%)1601 (54.3%)524 (60.8%)< 0.001High1037 (53.3%)1346 (45.7%)338 (39.2%)TrustLow576 (29.6%)929 (31.5%)253 (29.4%)0.252High1371 (70.4%)2018 (68.5%)609 (70.6%)Chronic Disease0509 (26.1%)597 (20.3%)236 (27.4%)< 0.0011557 (28.6%)780 (26.5%)246 (28.5%)2 or more881 (45.2%)1570 (53.3%)380 (44.1%)DepressionNo899 (46.2%)997 (33.8%)297 (34.5%)< 0.001Yes1048 (53.8%)1950 (66.2%)565 (65.5%)Body Mass IndexUnderweight139 (7.1%)13 (0.4%)439 (50.9%)< 0.001Normal1270 (65.2%)1972 (66.9%)415 (48.1%)Overweight/Obese538 (27.6%)962 (32.6%)8 (0.9%)ExpendituresLow927 (47.6%)1446 (49.1%)521 (60.4%)< 0.001High1020 (52.4%)1501 (50.9%)341 (39.6%)MCINo1750 (89.9%)2329 (79.0%)617 (71.6%)< 0.001Yes197 (10.1%)618 (21.0%)245 (28.4%)Physical Activity132.4 (122.1)101.0 (112.0)102.6 (115.8)< 0.001 Table 1Demographic and health characteristics of participants by sarcopeniaVariable_LevelNon-sarcopenia (N = 1947)Possible sarcopenia(N = 2947)Sarcopenia(N = 862)p-valueAge Average65.8 (5.1)67.8 (6.3)73.1 (7.3)< 0.001Age Group≤ 701620 (83.2%)2069 (70.2%)345 (40.0%)< 0.00171–80309 (15.9%)754 (25.6%)377 (43.7%)> 8018 (0.9%)124 (4.2%)140 (16.2%)GenderMale738 (37.9%)1659 (56.3%)477 (55.3%)< 0.001Female1209 (62.1%)1288 (43.7%)385 (44.7%)Education< high school1783 (91.6%)2817 (95.6%)844 (97.9%)< 0.001>high school164 (8.4%)130 (4.4%)18 (2.1%)PartnerYes1619 (83.2%)2233 (75.8%)542 (62.9%)< 0.001No328 (16.8%)714 (24.2%)320 (37.1%)SmokingNo1039 (53.4%)1894 (64.3%)511 (59.3%)< 0.001Yes908 (46.6%)1053 (35.7%)351 (40.7%)DringkingNo1183 (60.8%)2164 (73.4%)636 (73.8%)< 0.001Yes764 (39.2%)783 (26.6%)226 (26.2%)ADL DisabilityNo1639 (84.2%)2111 (71.6%)574 (66.6%)< 0.001Yes308 (15.8%)836 (28.4%)288 (33.4%)IADL DisabilityNo1633 (83.9%)1959 (66.5%)493 (57.2%)< 0.001Yes314 (16.1%)988 (33.5%)369 (42.8%)Life SatisfactionLow1474 (75.7%)2165 (73.5%)645 (74.8%)0.204High473 (24.3%)782 (26.5%)217 (25.2%)Health StatusPoor1566 (80.4%)2506 (85.0%)762 (88.4%)< 0.001Good381 (19.6%)441 (15.0%)100 (11.6%)Social ActivityLow910 (46.7%)1601 (54.3%)524 (60.8%)< 0.001High1037 (53.3%)1346 (45.7%)338 (39.2%)TrustLow576 (29.6%)929 (31.5%)253 (29.4%)0.252High1371 (70.4%)2018 (68.5%)609 (70.6%)Chronic Disease0509 (26.1%)597 (20.3%)236 (27.4%)< 0.0011557 (28.6%)780 (26.5%)246 (28.5%)2 or more881 (45.2%)1570 (53.3%)380 (44.1%)DepressionNo899 (46.2%)997 (33.8%)297 (34.5%)< 0.001Yes1048 (53.8%)1950 (66.2%)565 (65.5%)Body Mass IndexUnderweight139 (7.1%)13 (0.4%)439 (50.9%)< 0.001Normal1270 (65.2%)1972 (66.9%)415 (48.1%)Overweight/Obese538 (27.6%)962 (32.6%)8 (0.9%)ExpendituresLow927 (47.6%)1446 (49.1%)521 (60.4%)< 0.001High1020 (52.4%)1501 (50.9%)341 (39.6%)MCINo1750 (89.9%)2329 (79.0%)617 (71.6%)< 0.001Yes197 (10.1%)618 (21.0%)245 (28.4%)Physical Activity132.4 (122.1)101.0 (112.0)102.6 (115.8)< 0.001 Demographic and health characteristics of participants by sarcopenia Deep learning model performance comparison Supplementary Table 1 presents the evaluation results of four deep learning models on the test set. Among these, Transformer achieved the highest performance across most metrics, with an Accuracy of 0.7150, Recall of 0.7210, F1-Score of 0.7175, and the highest AUC of 0.8681. LSTM followed closely, with an Accuracy of 0.7143, F1-Score of 0.7086, and an AUC of 0.8647. LSTM-Attention and TCN showed slightly lower performance, with LSTM-Attention achieving an AUC of 0.8550 and TCN recording the lowest F1-Score (0.6990) and AUC (0.8603). Given its superior performance, particularly in Recall and AUC, Transformer was selected for further analysis. The dataset was split into 80:20 training and test sets, with stratification to ensure balanced class distribution, and the test set doubled as the validation set. To ensure convergence and prevent overfitting, the Transformer employed early stopping (monitoring validation F1-score, patience 20 epochs, restoring best weights), learning rate reduction (halving after 10 epochs of stagnation), dropout regularization, and hyperparameter optimization via TPE. Deep learning analysis: feature importance and interactions Figure 2 presents the SHAP analysis results for 3,373 individuals, examining the relationship between Physical Activity, MCI, and sarcopenia states using a Transformer model. Figure 2 (A) left displays feature importance rankings for all classes, with SHAP values on the x-axis (higher values indicate greater contribution) and features on the y-axis. Color gradients (yellow for high feature values, blue for low) reflect non-linear relationships. Figures 2 (B) left, (C) left, and (D) left show feature importance rankings for non-sarcopenia, possible sarcopenia, and sarcopenia, respectively. Figures 2 (A) right, (B) middle, (C) middle, and (D) middle illustrate the interaction between Physical Activity (x-axis) and MCI (red for MCI present, blue for MCI absent) across sarcopenia states, with SHAP values on the left y-axis. For non-sarcopenia, higher SHAP values indicate a greater likelihood of maintaining the state, while for possible sarcopenia and sarcopenia, lower SHAP values indicate a reduced likelihood of being in the state. Figures 2 (B) right, (C) right, and (D) right show temporal changes in feature importance, with color intensity representing SHAP values over time.Fig. 2SHAP analysis of feature importance and Physical Activity–MCI interactions in sarcopenia state transitions. Panel (A) shows the overall results for all sarcopenia states, while panels (B), (C), and (D) present the results for the non-sarcopenia, possible sarcopenia, and sarcopenia states, respectively SHAP analysis of feature importance and Physical Activity–MCI interactions in sarcopenia state transitions. Panel (A) shows the overall results for all sarcopenia states, while panels (B), (C), and (D) present the results for the non-sarcopenia, possible sarcopenia, and sarcopenia states, respectively Figure 2 (A) reveals a strong non-linear relationship between features and sarcopenia states, with the yellow-to-blue gradients in Fig. 2 (A) left indicating complex associations. The interaction between Physical Activity and MCI in Fig. 2 (A) right further confirms this non-linearity, as SHAP values vary across the spectrum of Physical Activity and MCI presence. Figures 2 (B) left, (C) left, and (D) left highlight Age as a significant factor in sarcopenia states: younger age is associated with maintaining non-sarcopenia and possible sarcopenia, while older age increases the likelihood of sarcopenia. Regarding BMI, individuals with normal BMI show minimal impact across all states, while overweight individuals are more likely to maintain non-sarcopenia or possible sarcopenia. However, underweight individuals are more prone to sarcopenia. The absence of IADL disability predicts non-sarcopenia, while its presence increases the likelihood of transitioning to sarcopenia. Similarly, individuals with MCI are more likely to progress to sarcopenia, with its importance rising in later stages. Physical Activity consistently exerts a protective effect, promoting non-sarcopenia and suppressing both possible sarcopenia and sarcopenia. Depression and Smoking are associated with sarcopenia progression, with higher SHAP values in Figs. 2 (C) left and (D) left, indicating their role in exacerbating sarcopenia risk. Figures 2 (B) middle, (C) middle, and (D) middle illustrate the interaction between Physical Activity and MCI across sarcopenia states. In non-sarcopenia (Fig. 2 (B) middle), as Physical Activity increases, SHAP values rise, particularly for blue dots (MCI absent), indicating a higher likelihood of maintaining non-sarcopenia. Red dots (MCI present) also show increased SHAP values with higher Physical Activity, suggesting that Physical Activity can counteract MCI’s adverse effects and promote non-sarcopenia. In possible sarcopenia (Fig. 2 (C) middle), higher Physical Activity is linked to lower SHAP values for blue dots (MCI absent), reducing the likelihood of possible sarcopenia. In contrast, red dots (MCI present) show higher SHAP values, indicating that MCI weakens Physical Activity’s protective effect and increases the risk of possible sarcopenia. In sarcopenia (Fig. 2 (D) middle), higher Physical Activity leads to lower SHAP values, especially for blue dots (MCI absent), suppressing sarcopenia progression. However, red dots (MCI present) show higher SHAP values, suggesting that MCI exacerbates sarcopenia risk despite Physical Activity’s suppressive effect. Overall, Physical Activity mitigates MCI’s adverse impact across all states, with more pronounced benefits in individuals without MCI. Figures 2 (B) right, (C) right, and (D) right show temporal changes in feature importance. In non-sarcopenia (Fig. 2 (B) right), Physical Activity’s protective effect strengthens over time (increasing SHAP values), while MCI’s negative impact intensifies. In possible sarcopenia (Fig. 2 (C) right), temporal changes are minimal. In sarcopenia (Fig. 2 (D) right), MCI’s negative effect grows significantly over time, while Physical Activity’s suppressive effect slightly weakens. MSM analysis results Transition intensities Transition intensities in the Markov Model, reflecting instantaneous rates of state changes, were estimated with 95% confidence intervals (CI) and are presented in Table 2. These are classified as deterioration (worsening states), recovery (improving states), and death (mortality) transitions. Table 2Transition Intensities for Sarcopenia States Table 2Transition Intensities for Sarcopenia States Transition Intensities for Sarcopenia States Table 2 summarizes transition intensities for sarcopenia states. The highest deterioration intensity occurs from Non-sarcopenia to Possible sarcopenia (0.381793, 95% CI: 0.354972–0.410641), indicating a notable short-term progression risk. The intensity drops for the Possible sarcopenia to Sarcopenia transition (0.069864, 95% CI: 0.063503–0.076863). In recovery transitions, the intensity from Possible sarcopenia to Non-sarcopenia (0.301481, 95% CI: 0.280481–0.324054) exceeds that from Sarcopenia to Possible sarcopenia (0.211424, 95% CI: 0.191154–0.233842), suggesting earlier-stage recovery is more likely. Death transition intensities rise progressively: Non-sarcopenia (0.003212, 95% CI: 0.001411–0.007311), Possible sarcopenia (0.010836, 95% CI: 0.008195–0.014326), and Sarcopenia (0.034602, 95% CI: 0.027368–0.043747), underscoring a sharply elevated mortality risk with increasing sarcopenia severity. Probability of sarcopenia state transitions A Markov Model estimated state transition probabilities for individuals with varying sarcopenia states over 1, 2, 3, and 5-year intervals (Fig. 3). For non-sarcopenia, probabilities of remaining non-sarcopenic, progressing to possible sarcopenia, advancing to sarcopenia, or dying were 0.668, 0.318, 0.012, and 0.002 at 1 year (Panel A), shifting to 0.409, 0.509, 0.064, and 0.017 at 5 years (Panel D). For possible sarcopenia, probabilities of recovering to non-sarcopenia, staying as possible sarcopenia, worsening to sarcopenia, or dying were 0.241, 0.704, 0.049, and 0.006 at 1 year (Panel A), adjusting to 0.386, 0.516, 0.075, and 0.023 at 5 years (Panel D). For sarcopenia, probabilities of recovering to non-sarcopenia, improving to possible sarcopenia, remaining sarcopenic, or dying were 0.062, 0.336, 0.595, and 0.008 at 1 year (Panel A), evolving to 0.329, 0.509, 0.132, and 0.030 at 5 years (Panel D). Fig. 3 Transition diagrams of sarcopenia states over 1 to 5 years. Panel (A) shows the 1-year transition diagram, panel (B) the 2-year transition diagram, panel (C) the 3-year transition diagram, and panel (D) the 5-year transition diagram Transition diagrams of sarcopenia states over 1 to 5 years. Panel (A) shows the 1-year transition diagram, panel (B) the 2-year transition diagram, panel (C) the 3-year transition diagram, and panel (D) the 5-year transition diagram Covariate effects on sarcopenia transitions The effects of covariates on the hazard ratios (HRs) for transitions between sarcopenia states were estimated using the Markov Model and are shown in Table 3. Age significantly influenced transitions, with older individuals experiencing higher HRs for deterioration, recovery, and mortality. Compared to the 60–70 years group, the 71–80 years group had HRs of 1.639 (95% CI 1.321–2.035) for transitions from non-sarcopenia to possible sarcopenia and 1.794 (95% CI 1.384–2.324) for transitions from possible sarcopenia to sarcopenia. The > 80 years group had even higher HRs of 2.530 (95% CI 1.338–4.785) and 3.007 (95% CI 1.992–4.538), respectively. The likelihood of recovery decreased with age: HRs for the 71–80 years group were 0.569 (95% CI 0.425–0.761) for transitions from sarcopenia to possible sarcopenia and 0.744 (95% CI 0.590–0.938) for transitions from possible sarcopenia to non-sarcopenia. In the > 80 years group, these HRs further decreased to 0.440 (95% CI 0.282–0.687) and 0.494 (95% CI 0.234–1.045), respectively. Age also increased mortality risk, particularly in the > 80 years group, with HRs of 3.909 (95% CI 2.000–7.640) for possible sarcopenia to death and 7.400 (95% CI 2.542–21.544) for sarcopenia to death. Table 3Hazard Ratios of Covariates for Sarcopenia State Transitions and MortalityBoldface indicates statistical significance (P<0.05) Table 3Hazard Ratios of Covariates for Sarcopenia State Transitions and MortalityBoldface indicates statistical significance (P<0.05) Hazard Ratios of Covariates for Sarcopenia State Transitions and Mortality Boldface indicates statistical significance (P<0.05) Sex also influenced transitions. Males had a lower risk of deterioration from non-sarcopenia to possible sarcopenia (HR 0.731, 95% CI 0.607–0.880) but a higher likelihood of recovery from possible sarcopenia to non-sarcopenia (HR 1.380, 95% CI 1.145–1.664). Males also had a lower risk of death from non-sarcopenia (HR 0.330, 95% CI 0.127–0.857) compared to females. Regarding BMI, normal-weight individuals were less likely to progress from possible sarcopenia to sarcopenia (HR 0.203, 95% CI 0.089–0.463) compared to those with lower BMI. Similarly, overweight/obese individuals exhibited a lower likelihood of advancing from possible sarcopenia to sarcopenia (HR 0.155, 95% CI 0.068–0.356). These findings highlight the protective role of higher BMI against sarcopenia progression. Physical activity had a protective effect on deterioration transitions, with an HR of 0.916 (95% CI 0.842–0.997) for non-sarcopenia to possible sarcopenia and reduced the risk of death from possible sarcopenia (HR 0.565, 95% CI 0.339–0.944). Mild MCI impairment increased the risk of deterioration from non-sarcopenia to possible sarcopenia (HR 1.724, 95% CI 1.268–2.346) and death from sarcopenia (HR 2.120, 95% CI 1.269–3.541). Smoking significantly increased mortality risk, with HRs of 23.256 (95% CI 6.048–89.421) for non-sarcopenia to death and 2.299 (95% CI 1.141–4.632) for possible sarcopenia to death. Two or more chronic diseases reduced the likelihood of recovery from possible sarcopenia to non-sarcopenia (HR 0.767, 95% CI 0.593–0.993) and the risk of death from sarcopenia (HR 0.374, 95% CI 0.192–0.728). ADL disability increased the risk of death from sarcopenia (HR 2.020, 95% CI 1.196–3.413), while IADL disability reduced recovery chances, with HRs of 0.709 (95% CI 0.531–0.945) for sarcopenia to possible sarcopenia and 0.624 (95% CI 0.511–0.760) for possible sarcopenia to non-sarcopenia. Expenditures showed no significant impact on transitions. Discussion This study uses longitudinal data from the CHARLS to investigate dynamic transitions of sarcopenia states and their influencing factors among older Chinese adults, combining MSM and Transformer-based deep learning models. A total of 5,756 participants were included in the MSM analysis, with 3,373 also in the deep learning analysis. The findings highlight bidirectional sarcopenia transitions, the key role of physical activity, its interaction with MCI, and the effects of age, sex, and BMI. The discussion addresses key findings, methodological innovations, comparisons with prior studies, practical implications, and limitations. Dynamic transitions and recovery potential of sarcopenia states Baseline data show that 33.8% of participants were non-sarcopenic, 51.2% had possible sarcopenia, and 15.0% had sarcopenia, consistent with studies using the AWGS 2019 criteria [4]. MSM analysis reveals significant bidirectional transitions in sarcopenia states. The highest deterioration transition intensity was from non-sarcopenia to possible sarcopenia (0.383, 95% CI: 0.355–0.411), with 31.8% of non-sarcopenic individuals progressing to possible sarcopenia within one year, increasing to 50.9% over five years. The transition from possible sarcopenia to sarcopenia had a lower intensity (0.070, 95% CI: 0.064–0.077), with a five-year probability of 7.5%. For recovery, the intensity from possible sarcopenia to non-sarcopenia was 0.301 (95% CI: 0.280–0.324), with a five-year recovery probability of 38.6%, while recovery from sarcopenia to possible sarcopenia had a transition intensity of 0.211 (95% CI: 0.191–0.234) and a five-year probability of 50.9%. Although the latter showed a higher five-year recovery probability, the former demonstrated a stronger transition intensity, suggesting greater potential for early recovery. This recovery potential exceeds that reported by Trevisan et al. (2019) in a Swedish population (five-year recovery probability of 0.107) [16], possibly due to differences in diagnostic criteria and population characteristics. These findings support Cruz-Jentoft and Sayer’s (2019) notion of acute sarcopenia reversibility [2],, highlighting the possible sarcopenia stage as a critical intervention window. The high short-term deterioration rate and recovery potential emphasize the need for early identification and intervention. Role of physical activity in sarcopenia state transitions This study integrates MSM models with Transformer-based deep learning to explore the central role of physical activity in sarcopenia state transitions. SHAP analysis ranks feature importance and shows the positive or negative contributions of physical activity to different state predictions, while MSM quantifies its specific effects on transitions. Together, these approaches highlight the protective effects of physical activity in delaying deterioration, promoting recovery, and reducing mortality risk. SHAP analysis reveals that physical activity ranks just below age, BMI, and IADL disability in importance for predicting sarcopenia states. It positively contributes to maintaining non-sarcopenia and has a suppressive effect on possible sarcopenia and sarcopenia. MSM analysis quantifies these effects, showing that physical activity reduces the risk of deterioration from non-sarcopenia to possible sarcopenia (HR 0.916, 95% CI: 0.842–0.997), facilitates recovery from possible sarcopenia to non-sarcopenia (five-year recovery probability 38.6%, transition intensity 0.301, 95% CI: 0.280–0.324), and lowers mortality risk in the possible sarcopenia state (HR 0.565, 95% CI: 0.339–0.944). Although the HR for recovery in MSM analysis (0.974, 95% CI: 0.901–1.054) did not reach statistical significance, SHAP analysis, by capturing non-linear relationships, underscores its importance in maintaining non-sarcopenia. SHAP temporal effect analysis shows that physical activity’s protective effect strengthens over time, suggesting a cumulative impact and implying its long-term role in sarcopenia prevention. These results highlight that physical activity not only delays sarcopenia progression but also aids in reversing early sarcopenia and improving survival outcomes. Compared to prior research, this study offers deeper insights into the mechanisms of physical activity. A meta-analysis confirmed that resistance training improves muscle function in sarcopenia patients [17], and Cruz-Jentoft and Sayer noted that physical activity may reduce mortality risk [18]. However, these studies mainly viewed physical activity as a static intervention for muscle quality, without addressing its dynamic role in sarcopenia state transitions. Ya-Xi Luo et al. identified bidirectional sarcopenia transitions in older Chinese adults using CHARLS data [19], but traditional MSM models, limited by linear assumptions, struggle to capture the non-linear relationships of physical activity. This study overcomes these gaps by using Transformer models and SHAP analysis, revealing these non-linear associations and expanding the perspective of Ngandu et al., who focused solely on the cognitive benefits of physical activity [20]. Interaction effects of physical activity and MCI MCI, an age-related neurodegenerative condition, imposes significant social and economic burdens and is linked to sarcopenia [21, 22]. Physical activity is also well-established as a key factor influencing cognitive health [23]. Building on this, our study uses SHAP interaction analysis to explore how physical activity and MCI jointly impact sarcopenia state transitions, revealing their synergistic mechanisms. SHAP interaction analysis shows that as physical activity increases, individuals without MCI experience a significant rise in SHAP values, indicating a strong enhancement in maintaining or improving non-sarcopenia states. In contrast, individuals with MCI show a smaller increase in SHAP values, suggesting a diminished protective effect. Regarding the prevention of deterioration from non-sarcopenia to sarcopenia, high physical activity significantly reduces SHAP values in those without MCI, effectively suppressing deterioration risk, while individuals with MCI show higher SHAP values, indicating an increased risk of deterioration. MSM analysis further confirms these findings, showing that MCI significantly raises the risk of deterioration from non-sarcopenia to possible sarcopenia (HR 1.724, 95% CI: 1.268–2.346) and mortality risk in the sarcopenia state (HR 2.120, 95% CI: 1.269–3.541). This interaction effect may result from MCI reducing exercise adherence or impairing neuromuscular coordination, weakening the protective benefits of physical activity. Ngandu et al. (2015) in the FINGER trial suggested that cognitive function and physical capacity interdependence may influence healthy aging [20], a finding echoed by our study, which extends this mechanism to sarcopenia. While prior studies have established links between MCI and sarcopenia, as well as the cognitive benefits of physical activity, the combined effects of physical activity, MCI, and sarcopenia have been underexplored. Our results indicate that MCI not only directly exacerbates sarcopenia risk but also weakens the effects of physical activity on state transitions, suggesting that interventions should be tailored to cognitive status—such as combining cognitive training with physical activity for MCI patients to enhance its benefits. Other key influencing factors This study integrates SHAP analysis from Transformer deep learning models with MSM models to evaluate the impact of key factors—age, sex, BMI, depression, smoking, and functional disability—on sarcopenia state transitions. SHAP ranks feature importance and their contributions, while MSM quantifies the effects of covariates, providing complementary insights into their roles in sarcopenia progression. Age is a primary driver of sarcopenia deterioration. SHAP analysis shows that advanced age (>80 years) significantly increases sarcopenia risk, while younger age helps maintain non-sarcopenia. MSM further quantifies this effect: compared to the 60–70 age group, those >80 years face higher deterioration risk (HR 3.007, 95% CI: 1.992–4.538, possible sarcopenia to sarcopenia), lower recovery potential (HR 0.440, 95% CI: 0.282–0.687, possible sarcopenia to non-sarcopenia), and elevated mortality risk (HR 7.400, 95% CI: 2.542–21.544). These findings align with prior research on age-related muscle decline [24]. Sex affects sarcopenia transitions differently. SHAP analysis suggests that males are better at maintaining non-sarcopenia and suppressing sarcopenia progression, while MSM shows that males have a lower deterioration risk (HR 0.731, 95% CI: 0.607–0.880, non-sarcopenia to possible sarcopenia) and higher recovery potential (HR 1.380, 95% CI: 1.145–1.664, possible sarcopenia to non-sarcopenia). This may be due to greater muscle responsiveness to physical activity in males [25]. BMI has a non-linear effect. SHAP analysis reveals that underweight individuals are at higher risk of sarcopenia, while those overweight are more likely to maintain non-sarcopenia or possible sarcopenia states. MSM analysis shows that, compared to underweight individuals, normal-weight individuals have a lower deterioration risk (HR 0.015, 95% CI: 0.001–0.177, possible sarcopenia to sarcopenia) but limited recovery potential (HR 0.203, 95% CI: 0.089–0.463), a pattern also seen in overweight/obese individuals (HR 0.002, 95% CI: 0.001–0.028). These results align with research linking low BMI to higher sarcopenia risk [26] Other factors, including depression, smoking, and functional disability, also influence sarcopenia transitions. SHAP analysis shows that depression and smoking increase deterioration risk, while MSM quantifies that smoking significantly increases mortality risk (HR 23.256, 95% CI: 6.048–89.421, non-sarcopenia to death; HR 2.299, 95% CI: 1.141–4.632, possible sarcopenia to death). ADL and IADL disabilities reduce recovery chances (HR 0.740, 95% CI: 0.597–0.917; HR 0.624, 95% CI: 0.511–0.760) and increase mortality risk (HR 2.020, 95% CI: 1.196–3.413), with SHAP confirming their negative effects. These findings are consistent with prior research [27–30]. Broader implications for health crises and mechanisms Our findings confirm that PA provides a protective effect against sarcopenia transitions, although this benefit is attenuated in individuals with MCI. These results have direct relevance for health crises such as COVID-19, during which restricted mobility and disrupted routines reduce PA and compromise nutrition, thereby accelerating muscle decline. In such contexts, remote and personalized programs are essential to support older adults, particularly those with cognitive impairment [31]. Beyond behavioral influences, sarcopenia progression is also shaped by hormonal and inflammatory pathways that regulate anabolic function. Given the dynamic nature of sarcopenia, effective interventions should be timely, flexible, and individually tailored, especially for frail populations during health crises [32]. Limitations Despite advancements in analyzing sarcopenia state transitions, this study has notable limitations. First, the focus on older adults from the CHARLS cohort limits generalizability due to regional, cultural, and demographic differences, requiring validation in diverse populations. Second, reliance on self-reported physical activity data, rather than objective measures like accelerometers, may introduce recall bias and skew results. Third, muscle mass was assessed using indirect CHARLS indicators instead of gold-standard methods like DXA, potentially reducing the accuracy of sarcopenia classification and transition risk estimates. Fourth, while MSM and SHAP analyses provide valuable insights, the deep learning analysis included only 3,373 of the 5,756 MSM participants, which may affect representativeness. Fifth, although the study identifies interactions between physical activity and MCI in sarcopenia transitions, causality remains unclear. Further longitudinal research is needed to explore whether MCI diminishes physical activity’s benefits through mechanisms like neuromuscular decline. Sixth, while depression, smoking, and functional disability effects were measured, their mediation or interaction with physical activity was not fully explored, requiring future investigation. In conclusion, future research should improve data collection, use gold-standard muscle assessments, diversify samples, and leverage longitudinal data to clarify causal mechanisms, enhancing the reliability and applicability of findings. Supplementary Information Supplementary Material 1. Supplementary Material 1. Supplementary Material 2. Supplementary Material 2.