Hybrid Exercise Program Enhances Physical Fitness and Reverses Frailty in Older Adults: Insights and Predictions from Machine Learning

Abstract

Purpose

The declining physical condition of the older adults is a pressing issue. Wu Qin Xi exercise, despite being low-intensity, is highly effective among older adults. Inspired by its characteristics, we designed a new exercise program for frail older adults, combining strength, endurance, and Wu Qin Xi. Furthermore, we employed machine learning to predict whether frailty can be reversed in older adults after the intervention.

Methods

A total of 181 community-dwelling frail older adults aged 65 years or older participated in this single-center, randomized controlled study, with 54.7% (n=99) being female. The study assessed the effectiveness of several exercise modalities in reversing frailty. The Fried‘s frailty criterion was used to assess the degree of frailty of the subjects. Participants were assigned a three-digit code 001–163 and randomly assigned (1:1:1) by computer to three different groups based on the study participant number: the Wu Qin Xi group (WQX), the strength exercise mixed with endurance exercise training group (SE), and the WQXSE hybrid exercise group incorporated the above two. Body composition and frailty-related physical fitness factors were measured before and after a 24-week intervention. The measurements included Body height, Body mass, Timed Up and Go Test (TUGT), grip strength assessment (GS), 6min walk test (6 min WT), and 10 m maximum walk speed (10 m MWS). Data were analyzed using repeated measures ANOVA to determine group and time interaction effects and machine learning models were used to predict program effectiveness.

Results

A total of 163 participants completed the study, with 53.9% (n=88) of them being female. The two items, 10 m maximum walking speed (10 m MWS) and grip strength, were significantly affected by the interaction of group and time. Compared to the other two groups, the WQXSE group showed the most improvement in the item 10 m MWS. In addition, following 24 weeks of training, 68 (41.7%) of the initially frail older adults had reversed their frailty status. Among them, 19 (36.5%) were in the WQX group, 24 (44.4%) were in the WQXSE group, and 25 (43.9%) were in the SE group. The stacking model exhibited superior performance when compared to other algorithms.

Conclusion

A hybrid exercise regimen comprising the Wu Qin Xi routine and exercises focused on both strength and endurance holds the potential to yield greater improvements in the physical fitness of older adults, as well as reducing frailty. Leveraging a stacking model, it is possible to forecast the likelihood of older adults successfully reversing their frailty status following participation in a prevention exercise program.

Introduction

The global aging process has led to a growing number of frail older adult (1, 2), a complex clinical state that increases the risk of adverse health outcomes (3, 4, 5). Frailty is characterized by five phenotypes: unintentional weight loss, weariness, weakness, inactivity, and sluggishness (6). Older adults who meet three or more of these criteria are considered frail. Frailty significantly impacts the quality of life for older adults and places a burden on the healthcare system (7, 8, 9, 10). However, frailty is often overlooked or undervalued in the hospital's classification system, which excludes special cases such as stroke and frail older adults (11). Therefore, it is important to address the health challenges faced by this vulnerable population..

Fortunately, frailty is not irreversible, and its progression can be slowed and delayed through specific interventional strategies (12). One of the potential strategies to prevent or reverse frailty is physical activity, which has positive effects on immunity, skeletal muscle mass, neuromuscular control, and cardiorespiratory function (13, 14, 15). Among the various forms of physical activity, endurance and strength training have been shown to have a positive impact on cardiorespiratory fitness and specific neuromuscular conditions (16), while traditional Chinese sports such as Wu Qin Xi can improve balance and aerobic endurance, as well as maximal oxygen uptake and mobility, and are more readily accepted by the older population (17, 18). As a traditional Chinese low-intensity exercise that combines physical and mental training, Wu Qin Xi has numerous benefits for the physical and mental health of older adults (18, 19). However, not all frail older adults can benefit from this exercise, and some may still remain frail after a year of training (20, 21). Thus, the question arises as to how low-intensity exercise can effectively enhance the physical condition of older adults individuals who are frail. Fortunately, with fewer adverse events, low-intensity exercise is well tolerated in the older adults (22) and older adults can modify their health behavior with the support of suitable exercise patterns (23). Despite Wu Qin Xi requires lower intensity, it still exhibits significant physiological effects (18, 19). Additionally, endurance and strength training has been shown to positively affect cardiorespiratory adaptations and specific neuromuscular (16): endurance training can induce morphological changes of the heart and increase peak oxygen uptake by 12.0% (24). Meanwhile, to some extent, strength training can improve their balance ability and reduce their risk of falling by 22.0% (25) because of the increased lower limb strength. Therefore, exercise programs that incorporate Wu Qin Xi with strength and endurance exercises may be more effective at improving the motivation, confidence, and adherence of frail older adults.

Another important aspect of our research involves the potential application of explainable artificial intelligence (XAI) in the healthcare industry (26). It is well known that individualized exercise is the key to improving outcomes (27), however, due to the intricate interplay between physical fitness factors, it can often take far longer to identify the suitable exercise program for each individual, and individual nuances may result in completely different exercise programs. XAI, as a branch of artificial intelligence designed to provide transparent, explainable descriptions of the decisions and actions of AI systems, can help improve the credibility, reliability, and validity of personalized treatment plans by providing clear and understandable rationale for the recommendations and outcomes of AI-based systems. In addition, XAI can help identify and correct any errors, biases, or uncertainties in the AI model or data that may affect the quality of personalized treatment planning (28). Thus, we used AI models to simulate clinical treatment procedures, and by using XAI, we can provide more information and more comprehensive feedback about the simulated clinical state, such as the expected outcome (whether or not the frailty can be reversed).

In summary, this study will investigate the following questions: 1) whether Wu Qin Xi combined with strength and endurance exercises can improve or even reverse frailty in frail older adults; 2) whether machine learning models can be used to simulate clinical application scenarios and predict postintervention frailty in the subjects.

Methods

Subjects

From March to June 2018, we recruited volunteers from community older adults service centers in Changchun, China. The study received ethical approval from the Ethics Committee of Northeast Normal University (NC2018041103). The following criteria were used to screen the volunteers. The included criteria were as follows: 1) individuals who scored between 3 to 5 points using the Fried frailty phenotype scale; 2) individuals who were able to ambulate independently for a duration of at least six minutes; 3) individuals who did not engage in any other form of physical training during the past year; 4) individuals whose age was greater than sixty-five years old. The excluded criteria were: 1) individuals who had a past mental illness or serious psychological issues that made it difficult to obey orders; 2) individuals with a history of neurological disorders; 3) individuals with cardiovascular disease and other heart conditions. The sample size for this study was calculated using G*Power software, considering a significance level of 0.05, an effect size of 0.3, and a statistical power of 80% (29). Based on calculations, it was determined that a minimum of 25 subjects per group would be necessary to ensure sufficient statistical analysis. To accommodate potential dropout rates, a dropout rate of 20% was assumed, and a minimum of 30 subjects were recruited for each group to take part in the study.

Study Design

Experimental Program

This study is a single-centre, double-blind randomized controlled trial that aims to compare the effects of various exercise interventions on frail older individuals. Three distinct intervention groups were established for participants: Wu Qin Xi exercise training (WQX), strength training and endurance training (SE), and a hybrid training approach that combines Wu Qin Xi exercise with strength training and endurance training (WQXSE). Initially, each recruited subject was assigned sequential numbers, and randomization was performed using the method of a random number table. Participants were assigned a three-digit code (001–163) and randomly assigned (1:1:1) to three different groups based on their study participant number. The randomization process was conducted independently by an investigator who was not involved in any other tasks of this study, such as patient recruitment, intervention, evaluation, or data analysis. Written informed consent was obtained from all subjects after randomization into groups. The informed consent form provided information about the study's purpose, which was to assess the effects of exercise on frailty symptoms by comparing different exercise programs. The form did not specify whether one of the programs was a control or an intervention. The informed consent forms can be found in the Appendix. Prior to the formal intervention, the subjects were not informed about the specific intervention. After randomization into groups, they only learned about the content of their assigned program from the coaches and were not aware of the interventions in the other groups. To ensure effective masking, the researchers conducting the assessments or analyses were not aware of which group the patients belonged to. Additionally, the patients were prohibited from discussing their intervention protocols during follow-up visits and tests. The subjects underwent frailty assessment and physical fitness testing one week before and one week after the trial's intervention period. The trial was registered with the ClinicalTrials.gov, NCT05832853.

Intervention Programs

In addition to their regular training, the subjects underwent a 20-minute warm-up before the intervention and a 10-minute cool-down at the end. Concurrently, the WQX group participated in a 60-minute session of Wu Qin Xi training, whereas the WQXSE group engaged in 30 minutes of strength and endurance training, followed by an additional 30 minutes of Wu Qin Xi training. The SE group underwent a 60-minute session of strength and endurance training.

• 1)Wu Qin Xi training. The training program of Wu Qin Xi exercise training lasted for a total of 24 weeks, consisting of an 8-week initial phase followed by a 16-week second phase. During the initial stage, subjects in the WQX and WQXSE groups primarily focused on acquiring and reinforcing their proficiency in the Wu Qin Xi exercises. In the second phase, subjects in the WQX group practiced the Five Animals movement with synchronized breathing, three exercises per session. The subjects in the WQXSE group practiced the same movement with synchronized breathing, two times per session. To better demonstrate the methods and techniques of the Wu Qin Xi exercise, we have included a detailed explanation in the form of a video. This video can be found in the Appendix.
• 2)Strength Training. There comprised five exercises that target the major muscle groups in the strength training program (30). The strength training session was composed of a lower-body routine, which included seated calf raises and hip abduction exercises, and an upper-body routine, which consisted of lateral pull downs, biceps curls, and triceps push downs. Color-coded elastic exercise bands were utilized to indicate the level of resistance in all strength training exercises, thereby providing a clear and standardized measure of exercise intensity. The strength training program consisted of three cycles, each lasting 8 weeks, for the goal of optimizing hypertrophy and strength development. The first stage was the initial stage, so the training target of this stage was to give the subjects the ability to adapt to the subsequent high-intensity training. Thus, this study employed a training protocol involving 2–4 sets of exercises, utilizing light loads (40.0%–60.0% of the subject's one-repetition maximum) with 12–20 repetitions. The objective of the second stage of training was to enhance muscle activation, induce muscle hypertrophy, and optimize the muscle-to-adipose tissue ratio. Thus, this study employed a training protocol involving 2–4 sets of exercises, utilizing medium loads (60.0%–70.0% of the subject's one-repetition maximum) with 5–10 repetitions. To further promote hypertrophy, increase strength and optimize strength development, the third phase of the training protocol involved increased loads, ranging from 70.0% to 85.0% of the subject's one-repetition maximum (1RM), for performing 5–8 repetitions with 2–4 sets. The subjects in the WQXSE group completed two sets of each exercise, whereas those in the SE group performed four sets with an inter-set rest period of 2–3 minutes.
• 3)Endurance Training. We conducted endurance training indoors on a distraction-free track by walking continuously. During training, we implemented tracking of subjects' heart rate with a monitor (Heart Rate Monitor, MYZ-3, China). Based on the initial measurements, the subjects' target heart rates were individually adjusted. The target heart rate for each subject was personalized based on their initial measurements. The training intensity progressively increased over the course of the study. The target heart rate was gradually increased from 50% of the initial heart rate reserve during the first 12 weeks to 80% during the final 12 weeks (31). The subjects in SE group underwent 30 minutes of endurance training, whereas the subjects in WQXSE group received 15 minutes of the same training. Safety should be a top priority in any training process. At least one medical personnel were present during each training session, and in the event that the subject experienced any discomfort, the training was promptly terminated.

Primary Outcome Measurement

In the Asia-Pacific region, Fried's frailty criterion is the most commonly utilized frailty evaluation approach (32). A subject was deemed frail when they met three or more Fried frailty criteria, which comprised a set of five phenotypic indicators (6):

• 1)Unintentional weight loss: The subjects who have unintentionally lost over 5% of their body weight or 4.5 kilograms during the previous year met phenotype of frailty.
• 2)The gait velocity: the measurement of 10 m walking time was conducted on the study subjects. The individuals who demonstrated a gait speed below the threshold of 0.8 m/s were deemed to exhibit characteristics consistent with the phenotype of frailty.
• 3)Grip strength: Grip strength was measured by using a calibrated Jamar hydraulic hand dynamometer (model SH5001, Saehan Corp, Masan, Korea, 2017). Three repetitions were performed for each subject and the highest value recorded. Male subjects exhibiting grip strength values below 26 kg and female subjects exhibiting grip strength values below 18 kg were classified as presenting the frailty phenotype.
• 4)Self-reported fatigue: In daily life, the feeling of being unable to walk and doing things very hard was more than twice a week.
• 5)Low level of physical activity: The Physical Activity Scale for the Elderly in China (PASE-C) was utilized to evaluate the physical activity levels of the subjects (33). Male subjects who engaged in less than 383 kcal/week of physical activity and female subjects who engaged in less than 270 kcal/week of physical activity were classified as meeting the frailty phenotype.

  Each participant is awarded one point for each identified phenotype, and the final frailty score serves as the primary outcome measure.

Secondary Outcome Measurement

We evaluated the subjects' physical fitness through a battery of standardized assessments, including the Timed and Go Test (TUGT), grip strength assessment (GS), 6 min walk test (6 min WT), and 10 m maximum walk speed (10 m MWS), which were conducted one week prior to commencement of the intervention and again at one week after the 24-week intervention.

• 1)TUGT. This concise test has been shown to effectively reflect the agility of the tested subjects (r = 0.63) (34). subjects sat on a 45 cm high chair to get ready, got up after hearing the «go» command, then walked as quickly as possible for a distance of 3 meters in a forward direction, and went around the obstacle, and then returned to their seats.
• 2)Grip strength assessment. Grip strength assessment was measured by using a calibrated Jamar hydraulic hand dynamometer (model SH5001, Saehan Corp, Masan, Korea, 2017). Three repetitions were performed for each subject and the highest value recorded.
• 3)6 min WT. We utilized a 6 min WT conducted along a 30-meter indoor promenade with turning points at each end. The subjects were instructed to ambulate with maximal speed during the walking task.
• 4)10 m MWS. Under controlled test conditions with minimal distractions, study subjects performed two 10 m walk speed trials, with instructions to complete each trial as quickly as possible. In order to obtain stable and reliable measurement data, experiments were recorded from 2.5 m to 12.5 m.

Statistical Analyses

Traditional Statistics Analysis

The SPSS 25.0 software was utilized for statistical analysis of the experimental data, and employed the Shapiro-Wilk test to evaluate normality, with a log transformation being applied to non-normal data. Demographic variables at baseline were analyzed between the groups using one-way ANOVA or chi-square tests. The efficacy of the three intervention programs was then assessed using repeated measures ANOVA, with any variables showing interaction effects being further analyzed for simple effects. If significant differences emerged between groups, the Bonferroni test was employed for further analysis. p value below 0.05 was deemed to be statistically significant.

Machine Learning for Classification

To explore the relationship between intervention protocols and initial physical fitness in frail older adults and whether frailty could be reversed, we used XAI to predict whether subjects could reverse frailty. The post-intervention frailty status of the subjects served as the response variable, while the intervention types, along with the pre-intervention physical fitness of the subjects, were utilized as features to establish the dataset. To evaluate the effectiveness of the models, multiple metrics were utilized, including the area under the curve (AUC), precision, accuracy, recall, and F1 score. To ensure the robustness of the model, we conducted 100 iterations of 10-fold cross-validation training. XGBoost (35), Extra Tree (36), LightGBM (37), Gradient Boosting (38), Logistic Regression (39), Random Forest (40) and Linear Discriminant Analysis (41) were utilized in the initial step of the classical machine learning modeling process. Then, we selected the three best models to build the stacking model to attempt to improve the prediction performance of the model. To leverage the strengths of multiple classification algorithms, a stacking approach was utilized in this study to combine the predictions of diverse models L₁ to L_n, all trained on a shared dataset S. The dataset S comprises of instances S_i = (x_i,y_i), where x_i denotes feature vectors and yi represents the corresponding class labels. In the first stage, a collection of foundational models, denoted as M₁, M₂, and M₃, were developed, with M₁ = L_n (S). In the subsequent stage, the output generated by the foundational models was fed into a meta-level classifier. To create a training dataset for the meta-level classifier, we adopted a cross-validation strategy. This involved applying each base-level learning algorithm to the training data and reserving a portion of the dataset for assessing the performance of the model. Subsequently, we use the obtained classifiers to generate forecasts for target S_i, as Eq. (1):

$$\widehat{y_{\iota }^k}=M_k^i(x_i),$$ (1)

The meta-level dataset comprises instances in the format of ${\text{y}}_{\text{i}}^{\wedge \text{k}}$, where the target class of the instance to be evaluated is determined based on the feature predictions of the meta-level classifiers.

We conducted SHAP value calculations to identify the primary contributor among the features in predicting the reversal of frailty (42). Any machine learning model's output can be interpreted using SHAP, a game-theoretic method. The SHAP value offers a numerical measurement of the influence that each feature has on the model's predictions, represented by Eq. (2):

$$(y_i=y_{base}+f(x_{n,1})+f(x_{n,2}))+\dots +f(x_{n,k}),$$ (2)

in which f (x_n,1) represent the Shapley value corresponding to the feature (x_n,m), where n is the index of the observation and m is the index of the feature. This value can be understood as the incremental contribution of the first feature in the nth observation towards the final prediction yn. When the value of f (x_n,1) is positive, it suggests that the corresponding feature has a positive impact on the predicted outcome. Conversely, a negative f (x_n,1) signifies that the characteristic has an adverse impact on the forecast. Python version 3.7.1 was utilized for conducting data analysis.

Results

After 574 applications were put through a series of rigorous screenings, 181 individuals (99 females and 82 men) were selected because they matched the selection criteria. The research was finished with a total of 163 subjects (88 females and 75 males). The subjects who did not finish the programme did so for a variety of reasons, including dropping out (n = 10) and being sick (n = 8). Table 1 shows the initial demographic information and frailty degree. The results shows no significant differences among these subgroups in terms of relevant demographic factors, frail score and frail phenotype. In addition, the subjects' frailty condition was reassessed after the intervention. Following 24 weeks of training, 68 (41.7%) of the initially frail older adults had reversed their frailty status. Among them, 19 (36.5%) were in the WQX group, 24 (44.4%) were in the WQXSE group, and 25 (43.9%) were in the SE group.

Table 1.

Fundamental demographic features of the subjects at baseline

Items	WQX1(n = 52)	WQXSE2(n = 54)	SE3(n = 57)	P-value	Total (n=163)
Sex (male/Female)	22/30	28/26	25/32	0.781	75/88
Age (years)	73.30±3.89	72.13±4.17	70.74±3.52	0.139	72.06±3.86
Body height (cm)	164.72±7.48	167.68±7.82	165.54±8.22	0.100	165.98±7.84
Body mass (kg)	65.24±5.14	62.55±6.99	63.05±6.88	0.055	63.62±6.34
Frail Score	3.73±0.79	3.56±0.66	3.67±0.72	0.453	3.65±0.72
Unintentional weight loss (Yes/No)	38/14	35/19	43/14	0.436	116/47
Gait velocity (Yes/No)	41/11	40/14	42/15	0.789	123/40
Grip strength (Yes/No)	47/5	46/8	48/9	0.604	141/22
Self-reported fatigue (Yes/No)	36/16	42/12	40/17	0.553	118/45
Low level of physical activity phenotype (Yes/No)	32/20	35/19	36/21	0.941	103/60

1. Wu Qin Xi exercise group; 2. hybrid training combining Wu Qin Xi exercise with strength and endurance training group; 3. strength and endurance training group.

Effects of the Interventions

Table 2 presents the statistical outcomes related to the improvement of physical fitness in the three groups. According to the findings from the study, we did not find the statistically significant difference in the pre-intervention physical fitness levels of the subjects in each group. After the intervention, there were significant time and group interaction effects observed in 10 m MWS (partial η2 = 0.218, p < 0.001), and grip strength (partial η2 = 0.065, p < 0.01). After a 24-week intervention, a statistically significant enhancement in grip strength was observed among the subjects in the WQXSE group and SE group as compared to subjects in the WQX group. The post-hoc tests revealed that the participants in the WQXSE group exhibited a significant improvement in TUGT compared to those in the WQX group. On the other hand, subjects in the SE group had a significant enhancement in the 6 min WT. In addition, the study found that significant improvement in 10 m MWS was observed in the subjects of the WQXSE group compared to both the WQX group and the SE group. However, the results did not show any significant group $⇔mes$ time interaction effects in the TUGT and 6 min WT.

Table 2.

. The outcomes of the repeated measures ANOVA for the assessment variables were obtained for each group at pre-intervention and at post-intervention

Parameters	WQX1(n=52)		WQXSE2(n=54)		SE3(n=57)		Group X Time#
	Baseline	24-Weeks	Baseline	24-Weeks	Baseline	24-Weeks	partiali2
10 m MWS (m/s)	0.71 ± 0.13	0.81 ± 0.14†,***	0.74 ± 0.14	1.08 ± 0.14 †,***	0.73 ± 0.13	0.93 ± 0.14 †,***	0.218***
TUGT (s)	12.00 ± 1.32	11.63 ± 1.21†	11.91 ± 0.98	10.98 ± 1.24 †,***	11.90 ± 1.65	11.19 ± 1.29 **	0.016
Grip Strength (kg)	18.32 ± 3.08	19.30 ± 3.46†	18.23 ± 3.35	21.31 ± 3.44 †,***	17.93 ± 3.14	21.63 ± 3.26 †,***	0.065**
6 min WT (m)	350.29 ± 40.27	369.32 ± 36.33†,*	361.69 ± 39.40	388.71 ± 45.68 **	365.07 ± 42.11	392.45 ± 47.49 †,***	0.004
# repeated-measures ANOVA analysis; *p < 0.05, ** p < 0.01, *** p < 0.001 denotes significant difference within the group before and after intervention, † denotes the significant difference between groups. 1. Wu Qin Xi group; 2. hybrid training combining Wu Qin Xi exercise with strength and endurance training group; 3. strength and endurance training group.

Results from Machine Learning Models

To explore the relationship between intervention protocols and initial physical fitness in frail older adults and whether frailty could be reversed, we used XAI to predict whether subjects could reverse frailty. In addition, we chose the classical model with the top-3 performance ranking, Linear Discriminant Analysis, Extra Tree classifier and Random Forest to build stacking model and used Logistic Regression as the second model. The models performance results were shown in Table 3. The stacking model got highest accuracy as 76.4 ± 10.1%, best recall as 81.7 ± 12.9%, and greatest F1-score as 78.8 ± 9.3%. Extra Tree classifier got the highest precision (78.9 ± 11.2%).

Table 3.

. A comparative result on the performance of machine learning Classification Models

Models	Accuracy(%)	Precision (%)	Recall (%)	F1 (%)
XGBoos	71.6 ± 10.7	73.9 ± 11.6	76.3 ± 14.2	73.9 ± 9.7
LightGBM	71.8 ± 10.4	74.5 ± 11.3	75.7 ± 13.7	74.3 ± 10.1
Gradient Boosting	72.3 ± 10.6	74.4 ± 10.7	76.8 ± 14.4	74.8 ± 10.1
Logistic Regression	73.7 ± 10.7	78.4 ± 10.7	73.1 ± 15.2	74.8 ± 10.7
Random Forest	74.5 ± 10.9	77.5 ± 11.1	77.3 ± 14.5	76.8 ± 9.9
Extra Tree classifier	75.8 ± 10.2	78.9 ± 11.2	77.3 ± 13.5	77.5 ± 9.6
LDA1	75.9 ± 10.2	77.4 ± 10.8	79.7 ± 13.6	78.1 ± 10.2
Stacking	76.4 ± 10.1	77.5 ± 10.4	81.7 ± 12.9	78.8 ± 9.3

1. Linear Discriminant Analysis.

Furthermore, the prediction result confusion matrix clearly shows all predictions of the model, and the normalization of the confusion matrix provides a more intuitive representation of the model's predictive accuracy across all categories. Figure 2(a) represents the confusion matrix and Figure 2(b) represents the normalized confusion matrix with the true labels in the horizontal direction and the results predicted by the model in the vertical direction. The superior predictive performance (82.0%) of the model in identifying subject reversal of frailty is evident from Figure 2(b). Figure 2(c) illustrates the receiver operating characteristic (ROC) curves for the top five performance models. The most higher area under the curve (AUC) of ROC indicate better results. The stacking model showed the most higher AUC value of 0.853. In summary, multiple model evaluation metrics indicate that the stacking model performs well in predicting the post-intervention frailty state of the subjects.

Figure 2.

The receiver operating characteristic (ROC) curves of the five most performant models and the stacking model's associated confusion matrix

(a) Confusion Matrix, (b) Normalized Confusion Matrix, (c) receiver operating characteristic curves. LDA is an abbreviation for Linear Discriminant Analysis; LR is an abbreviation for Logistic Regression; ETC is an abbreviation for Extra Tree Classifier; RF is an abbreviation for Random Forest Classifier.

Figure 1.

Flowchart Illustrating the Experimental Design

WQX denotes Wu Qin Xi exercise training, WQXSE denotes hybrid training combining Wu Qin Xi exercise with strength training and endurance training, SE denotes strength training and endurance training; 10 m MWS denotes 10 m maximum walking speed; GS denotes grip strength, 6 min denotes 6 min walk test and TUGT denotes time and up go test.

Feature Contribution

Figure 3 shows the contribution of the features to the model prediction performance, presented by the SHAP values. The magnitude of the SHAP values reflects the contribution of the features to the model. From Figure 4d, it can be seen that grip strength shows the largest contribution among all features, followed by 10 m MWS, while the other features contribute little. This suggests that when aiming to improve the physical health of frail older adults, particular attention should be given to enhancing their grip strength and 10 m MWS performance. Furthermore, It can be seen from 3 (b) that the features can interact with each other, which means that through exhaustive feature combination testing, the optimal feature combination can be identified to enhance the performance of the model.

Figure 3.

This diagram depicts the SHAP values attributed to each individual feature

The vertical axis corresponds to distinct characteristics. (a) shows the SHAP values for each feature; the abscissa denotes the temporal order of the sample sequence; chromatic hues of red indicate positive effects, while shades of blue signify negative effects; the degree of contribution is determined by the color shade; the output before the activation function is represented by the function f(x); (b) shows the interaction relationship between the each feature. (c) describes the prediction process for each sample and the accumulation process for each feature, and the horizontal axis denotes the predictive output of the model on individual samples (d) shows the mean absolute SHAP values for individual features are presented on the x-axis; grip strength has the largest contribution among all features. (e) demonstrates impact of individual features in a given sample on the model, as depicted by the SHAP value along the horizontal axis. Based on the information depicted in the graph, it is observed that grip strength has the most largest positive effect on the model's prediction performance, while 10 m MWS has the most largest negative effect.

Discussion

This study incorporates the conventional Chinese treatment approach «Wu Qin Xi» with a strength and endurance training regimen. It is the first to our knowledge to suggest a hybrid exercise strategy for frail older adults based on traditional Chinese medicine theory. At the same time, artificial intelligence was used as a clinical assistant to predict the degree of frailty. The results revealed that all groups had improved to varying degrees in endurance strength, and speed performance after 24 weeks of training. These findings are in line with previous studies that have demonstrated the benefits of exercise interventions on frailty prevention and reversal among older adults (43). Compared to the other two groups, this hybrid exercise programme improved the greatest performance in 10 m MWS. Furthermore, the hybrid exercise group had the most significant positive effect in reversing frailty, with results showing that 44.4% of subjects had recovered from their frail state. Meanwhile, our artificial intelligence application achieved innovative progress, with 76.4% average accuracy rate in simulating clinical scenarios. The input features for the model were data collected for each motor ability test index prior to the intervention, as well as for the three groups in the intervention program. The outcome variable used was status of frailty reversal. For the results from XAI, grip strength may be the most influential feature, while 10 m MWS may be the second, which means these two were highly identifiable and contributed the most to the characteristics; frailty was more likely to reverse among older adults with stronger grip strength and a faster 10 m MWS.

Following 24 weeks of training, the subjects in the WQXSE group improved the most progress in 10 m MWS, as 0.34 m/s (WQXSE) > 0.20 m/s (SE) > 0.10 m/s (WQX), and TUGT, as 0.93 s (WQXSE) > 0.71 s (SE) > 0.37 s (WQX). Our results indicate that a multicomponent training program consisting of Wu Qin Xi, strength training, and endurance training produced the most optimal results, which is consistent with the results of the study by Sa et al. (44). Since it allows fragile subjects to mix several exercise modalities to make up for their limitations, multi-component exercise programme is beneficial because it enables frail subjects to combine several exercise modalities to compensate for their shortcomings. The result is consistent with (15, 45), who also showed that multi-component exercise programme is more likely to have a positive effect. Second, considering that Wu Qin Xi can help practitioners improve core muscle group strength, flexibility, hip and knee mobility, as well as spine stability. It has additional advantages for quadriceps and hamstring strength, including increased knee stability when ascending and descending stairs (46, 47). Thus, after including walking as an exercise tool with endurance training, this continuous dynamic movement can make up for the relatively slow movement characteristics of Wu Qin Xi, which is more beneficial for enhancing walking ability in a real-life setting. Finally, because Wu Qin Xi has the advantage of enhancing the walking speed and neuromuscular sensitivity (47); therefore, we designed a combined strength and endurance exercise that includes lower body strength exercises, which resulted in additional improvements in 10m maximal walking speed and TUGT for subjects in the WQXSE group.

According to our findings, the SE group's increase in grip strength was slightly higher than that of the WQXSE group (4.95 kg (SE) > 3.08 kg (WQXSE)). The WQX group experienced a moderate but insignificant increase in grip strength. Since Wu Qin Xi is a low-intensity exercise with a limited ability to improve grip strength, the WQXSE group exercised their upper limbs less vigorously than the SE group while undergoing the same training load. Therefore, the WQXSE group exhibited slightly slower progress in grip strength compared to the SE group. Previous studies have shown that Wu Qin Xi can improve grip strength (19). However, Wu Qin Xi is a low-intensity exercise with limited ability to improve grip strength. Even though both the WQXSE and SE groups showed improvements in grip strength, we suspect that this may have been a localized side effect rather than a general gain in overall strength among the subjects.

The subjects in the SE group demonstrated greater improvement in the 6 mi WT relative to the WQX group and WQXSE group. Specifically, the SE group showed the greatest improvement with a distance increase of 43.04 meters, followed by the WQXSE group with a distance increase of 26.72 meters, and finally the WQX group with a distance increase of 19.03 meters. However, we did not find a significant time × group interaction effect. This conclusion is in line with those of several studies (48, 49). Although the subjects' endurance greatly increased, there was no real distinction between the three groups. We noticed that after doing the Wu Qin Xi exercise, some subjects in WQXSE group displayed intolerance and negative attitudes toward the subsequent endurance training. This could be the cause of the outcome.

Despite the encouraging findings of our study, there are several limitations. Our study did not control the intensity of training; only its duration was monitored. Second, the muscles involved in grip strength were indirectly exercised in our intervention programs. Thirdly, subjects with stronger grips may have a coincidental localized side effect. Fourthly, we think further research will reveal a deeper possible relationship between frail older adults and frailty reversal. We intend to incorporate data from the same trial design in the future to construct a complete functional system for more accurate medical support based on the frail older adults' beginning state, and the interventional strategy. Finally, the lack of a control group limited the ability to effectively compare the interventions for which we made direct comparisons.

Conclusion

The results of our study suggest that the enhancement of athletic ability in the older adults is largely influenced by the type of training offered. Hybrid exercise that combines Wu Qin Xi, strength training, and endurance exercise shows more effectiveness in promoting physical health and reversing frailty in this population. In addition, our stacking model accurately predicted frailty in the older adults.

Funding:

Guang Yang obtained «The Fundamental Research Funds for the Central Universities» (Number: 135222026).

Availability of data and materials:

The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.

Competing interests:

All authors certify that they have no affiliations with or involvement in any organization or entity with any financial interest or non-financial interest in the subject matter or materials discussed in this manuscript.

Ethics approval:

This study was granted ethical approval by the Ethics Committee of Northeast Normal University (NC2018041103).

Copyright comment

corrected publication 2023

Clinical trials registration:

The trial was registered in ClinicalTrials.gov (NCT05832853).

Authors' contributions:

Experimental design, data collection were conducted by GY, WM and SH. Data analysis was performed by ZW and DM. The first draft of the manuscript was written by WM, SH and DM, and all authors have reviewed the manuscript. GY and ZW revised the manuscript. All authors read and approved the final manuscript.