This research was designed as a randomized controlled and double-blind trial with 3 different intervention groups: a tai chi exercise and strength training hybrid group (TCSG), a strength training group (STG), and a CG. This study was approved by the Ethics Committee of the Northeast Normal University (NC2019030702) and registered at ClinicalTrials.gov (NCT05694117). The random number table method was used to randomly assign participants to the 3 groups. CT provides precise estimates of muscle quality [45], and quantitative CT (QCT) can reveal the level of muscle edema and steatosis. It can also accurately analyze the form and structure of skeletal muscle by removing measurement errors caused by unstable CT readings [46]. Therefore, QCT scans were used to assess the area and density of skeletal muscle at the third lumbar (L3) in each group of participants at the imaging department of Baishan Central Hospital. The training was conducted 3 times a week for 90 minutes, for a total of 24 weeks of intervention from June 1, 2019, to December 30, 2019. The experimental protocol design is shown in Figure 1. During the experiment, each participant was requested to refrain from taking part in any other sort of physical exercise.

In total, 2 experimental groups and 1 CG were set for this trial. The intervention regimens included tai chi exercise hybrid strength training and strength training. The TCSG and the STG were the experimental groups. All participants in the experimental groups started with a 20-minute warm-up exercise and concluded with a 10-minute cool-down exercise. The following is a comprehensive outline of the intervention programs:

We recruited volunteers from the Baishan Ciming Health Screening Center and the Health Screening Center of Baishan Central Hospital in Baishan, China, for this randomized controlled experiment. A form requesting informed permission was completed by each participant before the beginning of the trial. The screening process involved using the following inclusion and exclusion criteria to choose participants. Inclusion criteria: (1) patients between the ages of 60‐75 years and (2) patients who meet the screening criteria for sarcopenia in Asia established by Asian Working Group for Sarcopenia (AWGS) 2016 [47]. Exclusion criteria: (1) participants taking medications that significantly impact musculoskeletal function, (2) participants enduring respiratory failure or other bodily problems, (3) participants with mental disorders or neurological disorders, and (4) patients who participate in other training programs on a regular basis. Prior research served as the basis for the calculation of the required sample size [48], which was at an α level of .01; we used 80% power and a 0.53 effect, and we assumed a 20% dropout rate. This resulted in a required sample size of 106 participants.

The AWGS screening criteria [47] were used for each individual participant’s evaluation. For a participant to be diagnosed with sarcopenia, they must satisfy each criterion listed below:

In order to make the intervention runs as accurate as possible, we used a machine learning method to predict whether older people had sarcopenia after 24 weeks of intervention based on the participants’ initial status before the intervention and the intervention protocol they received as features. The data used to train the model consisted of each participant’s age, gender, and initial grip strength, RSMI, L3 SMA, and L3 SFMA as features and the presence or absence of sarcopenia after the intervention as a label. The performance of classification models was assessed per the F₁-score, precision, area under the curve (AUC), recall, and accuracy. These metrics were averaged across all validation or test sets during the repeated cross-validation process, with SD used to provide uncertainty estimates for these averages. Each model was trained 100 times individually using a 10-fold cross-validation. Hyperparameter tuning was performed using a grid search with cross-validation within each training fold to identify the optimal parameters for each model. For example, for the random forest (RF) model, we varied the number of trees (50 to 200), maximum depth (5 to 20), and minimum samples split (2 to 10). The best model was defined as the one with the highest average AUC across all folds. We also calculated the precision, recall, and F₁-score using a threshold of 0.5 for class predictions, ensuring consistency in performance evaluation. First, we trained 9 machine learning classification models using a k neighbors classifier [56], logistic regression (LR) [57], a gradient boosting classifier [58], linear discriminant analysis [59], an extra tree classifier [60], an RF classifier [61], a decision tree classifier [62], an XGBoost classifier [63], and a LightGBM classifier [64]. Then, the 3 best-performing models were selected for the stacking model in this dataset. The first layer consisted of the LightGBM, XGBoost, and RF classifiers. The second-level classifier, used to combine the outputs of these 3 models, was LR. In this research, a stacking technique was used in order to integrate numerous classifiers that were produced by various algorithms L₁,...,L_n and applied to a singular dataset S. This dataset included instances of the type S_i = (x_i, y_i), where x_i indicates the characteristic vectors and y_i indicates the classifications. In the beginning of this procedure, a group of base-level classifiers—C₁, C₂, and C₃, together with Ci = L_n(S)—were developed. Then, we had predictions for S_i, as per Equation 2:

The meta-level data were made up of several illustrations of the form ((y^i1,..., y^in), yi), in which features represent the output of the first-level classifier and categories are the appropriate labels for this sample.

The dataset was divided into 10 folds for cross-validation. Each model was trained on 9 folds and validated on the remaining fold, repeated 10 times for a total of 100 repetitions. For each fold, the model was trained and predictions were made on the validation fold. These predictions were saved. For each cross-validation run, the precision, recall, accuracy, and F₁-score were calculated. The means and SDs of these metrics over the 100 repetitions are reported in Table 1, providing an estimate of model performance and its variability. Predictions from all 100 cross-validation runs were aggregated to form a single prediction set. This aggregated prediction set was then used to generate the normalized confusion matrix presented in Figure 2, ensuring a comprehensive evaluation of the model’s performance.

Moreover, the receiver operating characteristic curves for the top 5 models are shown in . The horizontal axis depicts false-positive rate, while the vertical axis depicts true-positive rate. A greater AUC indicates superior performance. The LightGBM classifier model had the highest AUC (0.910). The normalized confusion matrix offers a more perceptible view of the accuracy of each category’s predictions made by the model, while the prediction result confusion matrix clearly displays all model predictions, with the real labels in the horizontal direction and the outcomes predicted by the model in the vertical direction. It is clear from that the model showed great performance in predicting whether or not participants reversed sarcopenia after the intervention (94% and 75%, respectively).

In addition, we calculated SHAP (Shapley additive explanations) values, because it could assign an important value to each feature that represents the effect on the model prediction of including that feature [65]. Any machine learning model’s output can be interpreted using SHAP, a game-theoretic approach. The SHAP value quantifies the contribution of each feature to the model’s predictions, represented by Equation 3:

in which S_F stands for all of the potential subsets that do not contain characteristic j, x stands for the values of the input characteristics, |S_F| stands for the dimension of S_F, and j stands for a specific characteristic. To calculate this effect, 2 models, fSF∪j and fSF, were trained, conditioned on the presence or absence of feature j throughout the training process. This allows the contribution of each feature to be calculated. We used the SHAP TreeExplainer algorithm to identify important features that could accurately predict whether participants could reverse muscle loss after the intervention.

All data analysis, statistical analysis, and visualization procedures for this study were completed using Python (version 3.7.1; Python Software Foundation) and SPSS (version 25.0; IBM Corp) software. Before performing ANOVA, data distribution was assessed using the Shapiro-Wilk test to check for normality, and Levene test was used to check for homoscedasticity. Nonnormally distributed data were log-transformed. Further, 1-way ANOVA and χ² tests were responsible for assessing baseline differences in demographic data among groups. Repeated-measures ANOVA was used to examine the impact of the 3 different intervention programs on the participants. Simple effects were used for the factors that had further interaction effects. The Bonferroni post hoc test was used to further investigate group differences when ANOVA indicated statistical significance. The Bonferroni post hoc test was chosen for its control of the family-wise error rate, making it suitable for pairwise comparisons. A P value of <.05 was used as the threshold for statistical significance

This study was conducted in accordance with the Declaration of Helsinki and approved by the Ethics Committee of Northeast Normal University (approval NC2019030702). A written informed consent form was obtained from each participant. Measures were implemented to safeguard participants’ privacy during data analysis.

A total of 164 individuals were recruited (Figure 3); however, only 106 (48 females and 58 males) were found to meet the inclusion and exclusion criteria. Finally, 93 (50 females and 43 males) participants finished the research. Dropping out of the training (n=8) and being unable to continue due to illness (n=5) were the reasons for not finishing the program. The demographic characteristics of the participants at baseline are summarized in Table 2.

Further, a week after the experiment finished, we reassessed participants for sarcopenia. In the TCSG, we observed the reversal of sarcopenia in 15 (45.5%) participants, and in the STG, we observed the reversal of sarcopenia in 12 (40%) participants, for a total of 27 (29%) with reversal of sarcopenia. Expectedly, sarcopenia was not reversed in the CG participants. Furthermore, Figure 4 shows the QCT measurements of representative participants in each group before and after the intervention.

Figure 5 depicts the main results, and Table 3 presents results of repeated-measures ANOVA for each observation indicator. All variables did not differ between groups at baseline. It was found that grip strength (P=.008), RSMI (P=.002), L3 SMD (P<.001), L3 SMA (P=.005), and MFI (P=.008) had a statistically significant interaction effect. A significant improvement in grip strength, RSMI, L3 SMA, and L3 SMD was seen in the TCSG and STG at 24 weeks. There was a significant difference in postintervention L3 SMA in the TCSG compared to the STG and CG, as determined by a post hoc test. However, we did not find any significant interactions among L3 SMFA and L3 SMFD.

The average performance evaluation of the machine learning model after 100 rounds of training is shown in Table 1. We found that the LightGBM classification model exhibited the best performance in terms of average accuracy (88.4%, SD 2%), average recall (74%, SD 4.5%), and average F₁-score (76.1%, SD 5.3%). In addition, our stacking model with the first layer consisting of LightGBM classification model, XGBoost classification model, and RF classification model, and the second layer consisting of LR exhibited the best average precision (83.9%, SD 5%).

Moreover, the receiver operating characteristic curves for the top 5 models are shown in Figure 2A. The horizontal axis depicts false-positive rate, while the vertical axis depicts true-positive rate. A greater AUC indicates superior performance. The LightGBM classifier model had the highest AUC (0.910). The normalized confusion matrix offers a more perceptible view of the accuracy of each category’s predictions made by the model, while the prediction result confusion matrix clearly displays all model predictions. This was with the real labels in the horizontal direction and the outcomes predicted by the model in the vertical direction. It is clear from Figure 2B that the model showed great performance in predicting whether or not participants reversed sarcopenia after the intervention.

The SHAP values of features for contribution are presented in Figure 6. We used SHAP values specifically derived from the LightGBM model, as it demonstrated the best performance among all models. The feature’s shape value indicates the degree to which the feature contributes to the overall model. The greatest contributions were shown in grip strength and L3 SMA across all variables. It is clear from Figure 6B that the characteristics may have an effect on one another. This implies that by evaluating all possible combinations of features, the optimal combination of features that would increase the performance of the model can be identified.

This study suggests a novel hybrid training strategy for older individuals with sarcopenia that combines SET and tai chi. In addition, AI was used to foresee if the intervention in older adults will reverse their sarcopenia. The results of this study showed that at the end of the 24-week intervention, participants in both the TCSG and STG improved their grip strength, RSMI, and L3 SMA, albeit to varying degrees. In terms of L3 SMA, the TCSG outperformed the others and exhibited the most substantial improvement. The strongest results in reversion of sarcopenia, however, were seen with the hybrid exercise program combining the TCSG and STG. As well as this, we were among the first to use XAI to predict sarcopenia in a Chinese sample of older people with an average accuracy rate of 88.4% (SD 2%).

We revealed that older people with sarcopenia in the STG and TCSG showed improvements in hand grip strength, RSMI, L3 SMA, and L3 SMD after 24 weeks of intervention. In comparison to both the STG and CG, the L3 SMA of the patients in the TCSG significantly improved. In accordance with the present results, previous studies have demonstrated that SET intervention has the ability to reverse sarcopenia and improve physical health [66]. This result may be explained by the fact that the contraction of strength training’s mechanical stress has a variety of positive consequences that increase intracellular protein accumulation net-positively, leading to the remodeling of the extracellular matrix and the growth of muscle fibers [67]. Specifically, SET stimulates PGC-1α expression in a working muscle, which could induce IGF-1 and suppress myostatin, leading to significant skeletal muscle growth [68], and IGF-1 expression increased, which might prevent atrophy and encourage hypertrophy; this might be a contributing factor to the amelioration of skeletal muscle area. Meanwhile, mechanical loading encourages the expression of growth hormone, as the growth hormone/IGF-1 axis and enhanced IGF-1 gene expression are thought to be the primary mechanisms by which growth hormone regulates postnatal bone development.

Perhaps the most important finding is that in terms of L3 SMA and RSMI, the TCSG considerably outperformed the STG after 24 weeks. These findings corroborate those of prior research suggesting that hybrid exercise may have a more potent effect on muscles [39,69-71]. Chinese traditional exercise, an assortment of low-intensity cardio mind-body exercises such as tai chi, Yijin Jing, Baduanjin, etc, has been used in China for centuries to both prevent and treat illness. As an aerobic training with low intensity, tai chi is commonly regarded as that it can improve physical condition, including improving endurance ability, increasing insulin sensitivity, and modulating fat metabolism. This impact is principally linked to the skeletal muscle’s markedly increased mitochondrial volume and density. PGC-1α/FNDC5/UCP1 signaling pathway activation and PGC-1α overexpression are the 2 factors responsible for this increase [72]. This effect increases the amount and activity of mitochondrial enzymes while also increasing the rate of fat burning in muscles both at rest and during low-intensity exercise [73-76]. Another possible explanation for this is that the theoretical principles of traditional Chinese medicine are helpful in maintaining a calm and composed state of mind. They also help the body’s innate self-regulatory or self-repair mechanisms to trigger and help the body to release endogenous neurohormones in a balanced manner [77,78]. Thus, the hybrid exercise training protocol of SET and tai chi had better performance and greater advantages than other intervention strategies.

Further, 1 unexpected finding was the extent to which neither the hybrid program nor SET could prevent intramuscular fat buildup. It is commonly known that as people age, their intramuscular fat content gradually increases [79,80]. The L3 level of the skeletal muscle intramuscular fat area showed a growing tendency in all 3 groups following the exercise intervention, but no statistical significance was identified. The findings of this study do not support the previous research finding that strength training and aerobic activity might assist older persons in avoiding age-related intramuscular fat accumulation [81]. This finding may be attributable to differences in the intensity and frequency of aerobic activity between studies.

The discovery that a total of 29% (n=27) of the participants were able to reverse their sarcopenia is perhaps the most clinically important finding. This included 40% (n=12) of the STG and 45.5% (n=15) of the TCSG. Our study links recovery from geriatric disease with a mixed exercise intervention, and the findings of this study could provide a new option for the treatment of sarcopenia and broadly support other research efforts in similar areas linking geriatric disease recovery to hybrid exercise [69,71].

Several limitations need to be noted regarding this study. First, given that there has been a wealth of research performed on the benefits of tai chi on skeletal muscle, we did not set up a distinct tai chi intervention group. Second, despite its effectiveness in enhancing the skeletal muscle area and RSMI in geriatric individuals with sarcopenia, the hybrid exercise program has limitations in terms of skeletal muscle intramuscular fat area. Third, we believe individual factors, such as the severity of sarcopenia and the activity habits of older people with sarcopenia, may influence the reversibility of sarcopenia. This study’s sample size offers valuable insights, yet further validation in larger cohorts is recommended in alignment with the sample size guidelines proposed [82] for clinical prediction models prior to clinical implementation. By using initial sarcopenia in older individuals, reversal of sarcopenia, and interventions as foundational data, a more comprehensive system with diverse features may be developed in the future to provide specialized medical assistance to this patient population.

According to the results of our research, a combination exercise program consisting of SET and tai chi is capable of increasing muscle mass and reversing sarcopenia in older people effectively. In addition, the LightGBM classifier model performed better in determining if sarcopenia in older people can be reversed.