Predicting fall risk among older adults with sarcopenia in China using machine learning models: a six-year longitudinal study from CHARLS

Abstract

Objective

Falls constitute potentially devastating health events for older adults with sarcopenia, whereas there remains a critical gap in validated fall risk prediction models tailored to this vulnerable population in China. This study aims to develop machine learning algorithms for predicting 6-year fall risk among patients with sarcopenia.

Methods

Data were used from the China Health and Retirement Longitudinal Study (CHARLS) spanning from 2013 to 2018. A total of 110 input variables at the baseline level were regarded as candidate features. Sarcopenia cases were identified according to the Asian Working Group for Sarcopenia 2019 criteria. Six machine learning models were developed through rigorous cross-validation, with model performance evaluated using accuracy, sensitivity, specificity, F1-score, and the area under the receiver operating characteristic curve (AUC), to estimate the 6-year fall risk prediction models for patients with sarcopenia.

Results

Among 1,087 participants with sarcopenia (mean age 71 years, 68.54% female), 246 experienced falls during follow-up. The random forest model demonstrated superior predictive performance among the six models, achieving an AUC of 0.971, sensitivity of 89.31%, specificity of 95.19%, and accuracy of 92.26%. Feature importance analysis identified 48 key predictors, with functional capacity measures, psychosocial factors, and cognitive function emerging as the strongest risk determinants.

Conclusions

The optimized random forest algorithm provides an effective tool for identifying high-risk sarcopenia patients who may benefit from targeted fall prevention strategies. These findings underscore the importance of multidimensional interventions addressing functional decline, cognitive impairment, and psychosocial well-being in sarcopenia management.

Introduction

Sarcopenia is an age-related, progressive, and generalized skeletal muscle disorder characterized by the progressive loss of skeletal muscle mass and decline in function [1]. It is associated with various adverse health outcomes, including an increased risk of falls, fractures, hospitalization, and increased mortality [2]. Among these outcomes, falls are considered the leading cause of adverse health outcomes in older adults with sarcopenia [3]. A meta-analysis and systematic review involving 52,838 individuals with sarcopenia reported that one in every six patients experienced a fall [4]. A study conducted in the United States indicated that the direct medical costs related to sarcopenia-induced falls reached as high as $18.5 billion, accounting for 1.5% of the total annual healthcare expenditure. Moreover, a 10% reduction in sarcopenia prevalence could result in a decrease of $1.1 billion in healthcare costs [5]. Therefore, early identification of risk factors is crucial for developing fall prevention strategies in older adults with sarcopenia to effectively reduce the incidence of falls.

In recent years, studies employing machine learning methods to develop predictive models for identifying risk factors have gained increasing popularity [6, 7]. A study involving 82,580 older adults in Belgium employed a random forest algorithm to construct a fall prediction model that identified 24 fall-related risk factors, the model achieved an average accuracy of 73%, demonstrating substantial improvements in predictive performance. This advancement plays a crucial role in the early identification of older adults at high risk of falling, facilitating timely interventions to reduce fall incidence and related healthcare costs. Additionally, machine learning has been applied to predict conditions such as Parkinson’s disease [8], stroke [9], and multiple sclerosis [10], achieving predictive accuracies of approximately 70%. Based on model predictions, personalized fall prevention strategies can be developed for patients, thereby enhancing the effectiveness of prevention strategies.

Despite recent advancements in fall risk prediction, current models still present several limitations. First, most fall risk prediction studies have been conducted in developed countries and primarily targeted older adults [11, 12]. Current research in China follows a similar trend [13, 14], and no predictive models specifically targeting fall risk factors in older adults with sarcopenia have been established. Second, fall prediction models predominantly focus on individuals who have not yet experienced a fall [15–17]. Nonetheless, older adults with a history of falling remain at risk of recurrent falls [18]. Consequently, longitudinal studies investigating risk factors in older adults with sarcopenia who have experienced prior falls are scarce. Third, conventional fall risk prediction models predominantly rely on regression-based techniques, such as logistic or Cox regression, to establish associations between dependent and independent variables [19, 20]. While these approaches help identify multiple fall risk factors, their ability to quantify the relative contribution of each risk factor to sarcopenia-related falls remains limited [21]. Thus, leveraging machine learning approaches in fall risk prediction is crucial to accurately identify key risk factors in older adults with sarcopenia.

Therefore, this study aims to: (1) utilize nationally representative data from the second to fourth waves of the CHARLS and employ machine learning methods to develop a fall risk prediction model for older adults with sarcopenia in China, and (2) identify key factors influencing falls and formulate corresponding fall prevention strategies. First, a random forest algorithm will be utilized to identify predictive features with an importance score of at least 0.5 in predicting fall risk in older adults with sarcopenia. Subsequently, the selected features will be incorporated into longitudinal data from 2013 to 2018 to construct and evaluate a fall risk prediction model for older adults with sarcopenia.

Methods

Study population

The CHARLS is one of the largest nationally representative longitudinal surveys conducted by the National School of Development and the Institute for Social Science Surveys [22], which monitor the health of individuals aged 45 years and older across 450 villages, 150 counties, and 28 provinces, involving over 17,000 participants from approximately 10,000 households in China. The national baseline survey of CHARLS was conducted in 2011 (wave 1), with follow-up examinations administered every 2 to 3 years. The survey collected basic demographic data, family information, intergenerational transfer payments, respondents’ health status, details of medical care and insurance, employment, income, expenditure, assets, and other relevant variables. Additionally, CHARLS includes 13 physical measurements. At the design stage, CHARLS project has systematically established protocols for obtaining information from special populations. For participants who are unable to complete the survey independently due to dementia or mental health conditions, CHARLS has implemented a formal proxy access mechanism. Specifically, a close family member or primary caregiver, who is most knowledgeable about the respondent’s situation, is designated to complete the questionnaire on their behalf. This procedure is clearly documented in the CHARLS survey manuals and is publicly accessible (https://charls.pku.edu.cn/en/). All CHARLS waves received ethical approval from the Ethical Review Committee of Peking University (IRB00001052–11015), and informed consent was obtained from all participants.

This study aimed to conduct a comprehensive analysis of two clinically meaningful patient trajectories: first, individuals who were free from sarcopenia at baseline (2013) but developed newly identified incident sarcopenia by follow-up (2018); and second, individuals diagnosed with sarcopenia at baseline who maintained persistent sarcopenia status throughout the follow-up period. Therefore, participants were excluded if they met any of the following criteria: (1) age < 60 years; (2) incomplete or invalid data on falls or sarcopenia assessment; (3) participants not diagnosed with sarcopenia in Wave 4; or (4) missing values exceeding 30% for individual variables. Ultimately, a total of 1,087 older adults with sarcopenia were included in the final analysis. Detailed information on the sample selection process is shown in Fig. 1.

Fig. 1.

Flowchart of this study. BPNN Back Propagation Neural Network, ELM Extreme Learning Machine, RF Random Forest, DT Decision Tree, SVM Support Vector Machine, LR Logistic Regression

Assessment of sarcopenia

Sarcopenia was diagnosed based on the criteria established by the Asian Working Group for Sarcopenia (AWGS) 2019 [23], which comprise three components: handgrip strength, physical performance, and appendicular skeletal muscle mass (ASM). Handgrip strength was evaluated with a Yuejian™ WL-1000 dynamometer. Participants were guided by trained researchers and instructed to apply maximum effort during the assessment. Each individual completed two trials using both their dominant and non-dominant hands. Out of the four measurements recorded, the two highest values were chosen, and their mean was computed to determine the participant’s final handgrip strength [22]. This procedure adhered to the standardized physical examination protocol established by the CHARLS study, as opposed to the single-maximum-value approach recommended by the AWGS 2019 [23]. Cut-off values for low grip strength were set at < 28 kg for men and < 18 kg for women. Low physical performance is defined as gait speed < 1.0 m/s, 5-time chair stand test (CST) ≥ 12 s, or Short Physical Performance Battery score (SPPB) ≤ 9 [23].Gait speed was assessed by timing how long it took participants to walk a distance of 2.5 m at their normal walking pace. Participants initiated their walk from a stationary position behind the starting line, beginning only after receiving a verbal cue. While this distance of 2.5 m is less than the ≥ 4 m suggested by the AWGS 2019 guidelines, it serves as a practical adjustment for extensive field surveys in constrained spaces. Significantly, this short-distance method has been previously validated and extensively utilized in studies focusing on sarcopenia that employ the same data set [22], thereby affirming its practicality and acceptability in research conducted within populations [24, 25]. The CST measured the time required for participants to stand up from a 47-cm chair and sit down repeatedly for five cycles with their arms folded across their chest. In addition to the aforementioned tests, the SPPB evaluation included three 10-second balance tests in the following positions: side-by-side, semi-tandem (with the heel of one foot positioned beside the big toe of the other), and tandem (with the heel of one foot placed directly in front of and touching the toes of the other).

For ASM, due to the absence of bioelectrical impedance analysis or dual-energy X-ray absorptiometry (DXA) as recommended by the AWGS 2019 for the direct measurement of ASM in the CHARLS study, muscle mass was estimated using a validated anthropometric equation applied within the CHARLS database. For the Chinese population, ASM was calculated using the following previously reported formula [26]: ASM = 0.193 × body weight (kg) + 0.107 × height (cm) − 4.157 × gender − 0.037 × age (where gender is coded as 1 for males and 2 for females) − 2.631. Several studies have demonstrated that ASM calculated using this formula exhibits a high degree of consistency with measurements obtained via DXA [27, 28]. Among it, height was measured with a SecaTM213 stadiometer (Seca Trading, Hangzhou, China), while weight was recorded using an Omron™ HN-286 scale (Krell Precision, Yangzhou, China). The height-adjusted muscle mass (ASM/Ht²) was computed by dividing the ASM value by the square of height in meters. The cutoff value for defining low muscle mass was determined based on the gender-specific lower 20th percentile of ASM/Ht² within the study population, which is 5.80 kg/m² for females and 6.88 kg/m² for males in this study [28].

Sarcopenia was defined as low muscle mass was present in conjunction with either low muscle strength or low physical performance. A diagnosis of severe sarcopenia was made if all three criteria were met; However, due to the limited number of cases (346 participants, 6.83%) meeting the criteria for severe sarcopenia, they were merged with the broader sarcopenia group [29].

Determination of falls

The primary outcome, the occurrence of falls, was ascertained from self-reported data in the fourth wave of the CHARLS dataset. Participants were asked, “Since your last interview, have you fallen?” Those who responded affirmatively were categorized as having experienced a fall.

Candidate variables and outcome variables

Candidate variables were selected based on a review of relevant literature and the availability of data within the CHARLS dataset. These variables were categorized as follows: (1) Sociodemographic characteristics (e.g., gender, age, ethnicity, education level, occupation, residential status, marital status, economic status, health insurance, and pension insurance). (2) Family characteristics (e.g., grandchild/parent caregiving, number of siblings, number of surviving children, and financial support). (3) Health status and function (e.g., vision, hearing, history of stroke, hypertension, diabetes, lung disease, kidney disease, heart disease, liver disease, gastrointestinal disease, cancer, arthritis, pain, hip fracture, tooth loss, self-rated health, disability status, activities of daily living (ADL) and instrumental activities of daily living (IADL) [30], mobility function, use of assistive devices, systolic blood pressure, diastolic blood pressure, pulse, grip strength, balance, five-time CST, gait speed, BMI(calculated as weight in kilograms divided by height in meters squared (kg/m²), and central obesity (defined as a waist circumference ≥ 102 cm in men and ≥ 88 cm in women, based on the criteria of the National Cholesterol Education Program (NCEP)/ Adult Treatment Panel III (ATP III)) [31]. Among these, ADL encompassed essential self-care tasks such as dressing, bathing, eating, transferring in and out of bed, using the toilet, and managing urination and defecation. IADL included more complex tasks such as housework, meal preparation, shopping, financial management, and adherence to medication schedules. Each activity was evaluated using a four-point scale: no difficulty, difficulty but can still perform, difficulty requiring assistance, and unable to perform. A participant was classified as experiencing difficulty in ADL or IADL if they reported encountering challenges in any item within the respective category (responses 2 to 4). (4) Lifestyle factors (e.g., alcohol consumption, smoking, sleep duration, social participation, and physical activity level). (5) Cognitive and psychological status (e.g., depressive symptoms (assessed using the Center for Epidemiologic Studies Depression Scale, CESD), cognitive function (assessed through the Telephone Interview for Cognitive Status-Modified, TICS-M) and episodic memory tests (assessed through immediate recall and delayed recall with a total score of 31 points) [32–34], and memory function (assessed through immediate recall of 10 interviewer-read words and delayed recall after 4–10 min). (6) Housing conditions (e.g., availability of accessible pathways, indoor cleanliness, and indoor temperature). Notably, although cognition and the CESD are multi-item scales that encompass several sub-dimensions, the total scores were utilized as single features. This approach was adopted to preserve essential information while ensuring the robustness of machine learning algorithms and enhancing the interpretability of model parameters [35, 36].

Data preprocessing

Based on the literature review and data availability in the current database [37, 38], non-temporal variables (e.g., gender, education) were imputed using data from either 2013 or 2018, if corresponding values were available. Temporal variables with more than 30% missing data at the individual level were excluded; for the remaining variables, multiple imputation was conducted using the Markov Chain Monte Carlo (MCMC) method conducted with five iterations. The final imputed values were computed as the mean of the five imputed datasets. To address class imbalance in the model training process, the Synthetic Minority Over-sampling Technique (SMOTE) was applied to the imputed dataset. SMOTE is a widely recognized and effective oversampling method for managing imbalanced data in classification tasks [39]. Finally, input values were standardized using the Z-score algorithm to ensure comparability across features and mitigate model overfitting. Following prior studies, feature selection was performed using the random forest algorithm, and features with an importance score greater than 0.5 were prioritized [40].

Predictive model development and evaluation

Six machine learning algorithms were employed to build models for predicting fall risk: Back Propagation Neural Network (BPNN), Extreme Learning Machine (ELM), Random Forest (RF), Decision Tree (DT), Support Vector Machine (SVM), and Logistic Regression (LR). BPNN exhibits robust nonlinear fitting capabilities, enabling it to automatically learn and adapt to complex data patterns, thereby yielding highly accurate predictions of fall risk [41]. ELM computes output layer weights using an analytical approach, thereby circumventing local optima issues inherent in traditional neural networks and enhancing model stability and predictive performance. RF enhances model generalizability and robustness by integrating multiple decision trees through two key mechanisms: random sampling of training data (bootstrapping) and stochastic feature selection during node splitting. This ensemble approach effectively mitigates overfitting while maintaining predictive accuracy, demonstrates exceptional competence in handling high-dimensional data with missing values, and achieves computational efficiency through parallel construction of constituent trees, making it particularly suitable for contemporary big data applications [42]. DT offers high interpretability by providing an intuitive representation of the decision-making process, thereby facilitating an understanding of risk assessment mechanisms. SVM improves classifier generalizability by projecting input variables into a high-dimensional feature space to construct a linear decision boundary. LR, as a classical statistical method, employs the Logit function to capture relationships between independent and dependent variables, thereby providing a simple model setup, rapid training, and high transparency [43].To evaluate model performance, a repeated random splitting procedure was employed to examine the impact of data partitioning on performance metrics. Specifically, the entire dataset was randomly divided into training and testing sets with a ratio of 7:3. This division was independently trained over 100 epochs to minimize bias associated with the randomness of any single data split. Additionally, during the model training phase, 10-fold cross-validation was incorporated within the training set for hyperparameter tuning and model selection, thereby reducing the risk of overfitting.

In evaluating machine learning models, key metrics such as true positives (TP), false positives (FP), false negatives (FN), and true negatives (TN) were calculated. From these, specificity (Sp), sensitivity (Se), accuracy (Acc), precision, recall, F1-score, Kappa coefficient and its variants (Kappa1, Kappa2), and the AUC of the random forest model (AUCRF) were derived to comprehensively assess classification performance and generalizability. In this study, the ranking of feature contributions was determined using a random forest algorithm based on permutation importance. The fundamental principle involves randomly shuffling the values of a specific feature, which disrupts its association with the outcome variable. A significant decline in the model’s predictive performance on the shuffled data suggests that the feature is important to the model. The contribution of each feature is quantified by comparing the model’s performance before and after shuffling, using metrics such as changes in accuracy, mean squared error, or other relevant indicators. A higher contribution score indicates that the feature serves as a strong predictor within the model [44].

Statistical analysis

For the comparison of patient characteristics, continuous variables were assessed for normality using the Shapiro-Wilk test, supplemented by boxplots. Based on this assessment, normally distributed variables are expressed as mean ± standard deviation and compared using the independent samples t-test, while non-normally distributed variables are expressed as median (interquartile range) and compared using the Mann-Whitney U test. Categorical variables are presented as frequency (percentage), with group comparisons conducted using the Chi-Square test. Raw data cleaning was performed using Stata 27.0 software, while statistical analysis was conducted using SPSS 27.0.

Results

Sample characteristics

A total of 14,277 participants were initially enrolled in the a six year follow-up conducted during the second to fourth waves of the CHARLS. Based on predefined exclusion criteria, 5,426 individuals younger than 60 years, 3,184 with possible sarcopenia or non-sarcopenia, 3,759 with missing sarcopenia diagnosis data, 29 with missing fall information, and 792 with over 30% missing key variables were excluded. Ultimately, 1,087 participants were included in the final analytic sample, of whom 246 experienced falls and 841 did not. Among fallers, 53 (21.55%) were male, compared to 289 (34.36%) among non-fallers. In contrast, 193 (78.45%) of fallers were female versus 552 (65.64%) in the non-faller group. This difference in gender distribution was statistically significant (P < 0.001). Furthermore, significant differences were observed between fallers and non-fallers with respect to cancer history, emotional and mental disorder history, CESD scores, cognitive scores, BMI, and measures of ADL and IADL. The detailed baseline of other characteristics is presented in Table 1.

Table 1.

Baseline characteristics of participants according to sarcopenia status in Chinese adults

Characteristics	Total (n = 1087)	Sarcopenia patients with Nonfalls (n = 841)	Sarcopenia patients with falls (n = 246)	t/Z/	P value
Age a (years)	71.00(65.00–76.00)	71.00(65.00–76.00)	71.00(65.00–77.00)	-0.291	0.771
Gender, n(%)
Male	342(31.50)	289(34.43)	53(21.54)	14.504	<0.001
Female	745(68.50)	552(65.57)	193(78.46)
BMI a(kg/m2, M ± SD)	20.15 ± 2.27	20.03 ± 2.22	20.57 ± 2.35	-3.287	<0.001
ADL score a (M, IQR)	6.00(5.00–6.00)	6.00(5.00–6.00)	6.00(5.00–6.00)	3.896	<0.001
IADL score a (M, IQR)	5.00(4.00–6.00)	5.00(5.00–6.00)	5.00(3.00–6.00)	5.455	<0.001
Social Participation a , n(%)	562(51.70)	435(51.72)	127(51.63)	0.001	0.978
CESD score a (M, IQR)	8.00(5.00–13.00)	8.00(5.00–12.00)	11.00(7.00–14.00)	-5.926	<0.001
Cognition score a (M, IQR)	5.00(2.00–8.00)	5.00(2.00–8.00)	4.00(1.00–7.00)	3.318	<0.001
Residential area, n(%)
Urban	1002(92.18)	779(92.63)	223(90.65)	1.032	0.310
Rural	85(7.82)	62(7.37)	23(9.35)
Education, n(%)				3.467	0.483
Illiterate	564(51.89)	433(51.48)	131(53.25)
Primary school	449(41.11)	348(41.38)	101(41.06)
Middle school	60(5.52)	48(5.72)	12(4.88)
High school/vocational high school	8(0.74)	6(0.71)	2(0.81)
Junior college or above	6(0.55)	6(0.71)	0(0.00)
Marital status, n(%)				2.956	0.228
Married	728(66.97)	573(68.13)	155(63.01)
Separated	31(2.85)	25(2.97)	6(2.44)
Unmarried/divorced/widowed	328(30.17)	243(28.89)	85(33.74)
Ever/current smoke, n(%)				3.089	0.184
No	6(0.55)	6(0.71)	0(0.00)
Yes	1081(99.45)	835(99.29)	246(100.00)
Ever/current alcohol, n(%)				0.006	0.938
No	749(68.91)	579(68.84)	170(69.11)
Yes	338(31.09)	262(31.16)	76(30.89)
Daily sleep time (hours, n(%))				9.001	0.079
<6	498(45.81)	374(44.47)	124(50.41)
6	182(16.74)	135(16.05)	47(19.11)
7	127(11.68)	101(12.01)	26(10.57)
8	170(15.64)	145(17.24)	25(10.16)
≥ 9	110(10.12)	86(10.22)	24(9.76)
Co-morbidities, n(%)
Cancer	477(43.88)	385(45.78)	92(37.40)	5.428	0.020
Chronic lung disease	689(63.38)	535(63.61)	154(62.60)	0.084	0.772
Heart disease	652(59.98)	500(59.45)	152(61.80)	0.433	0.511
Stroke	583(53.63)	447(53.15)	136(55.28)	0.348	0.555
Emotional and mental disorders	399(36.71)	329(39.12)	70(28.45)	9.318	0.002
Arthritis	842(77.46)	643(76.46)	199(80.89)	2.147	0.143
Dyslipidaemia	595(54.74)	451(53.62)	144(58.54)	1.852	0.174
Hepatic disease	634(58.33)	495(58.86)	139(56.50)	0.434	0.510
Digestive system disease	697(64.12)	534(63.50)	163(66.26)	0.632	0.427
Asthma	555(51.06)	420(49.94)	135(54.88)	1.857	0.173
Memory-related disease	565(51.98)	424(50.42)	141(57.32)	3.631	0.057
Hypertension	681(62.65)	516(61.35)	165(67.07)	2.659	0.103
Hyperglycaemia	606(55.75)	464(55.17)	142(57.77)	0.502	0.479

^aContinuous variables with normal distribution and equal variance are presented as mean ± standard deviation (Mean ± SD); Non-normally distributed variables are presented as median (interquartile range); Categorical variables are presented as frequency (percentage); Significant differences are highlighted in bold

Abbreviations: BMI Body mass index, ADL Activities of daily living, IADL Instrumental activities of daily living, CESD Center for Epidemiologic Studies Depression Scale. Cognition score was assessed using the Telephone Interview for Cognitive Status-Modified (TICS-M) and episodic memory tests

Feature selection

Employing random forest feature selection, 48 features were extracted from an initial pool of 110 variables. These features were categorized into: (1) Sociodemographic Characteristics (5 features: marital status, education level, personal assets, personal income, and debt status); (2) Family Characteristics (6 features: number of siblings, number of living children, financial support from children, financial support for parents, whether the father is alive, and whether the father is currently working).(3) Health Status and Function (24 features: overall health status, visual impairment, hearing impairment, 12 comorbidities [e.g., hypertension, hyperlipidemia, memory-related diseases, arthritis or rheumatism], blood pressure and blood glucose control, hip fracture, glaucoma, dentition status, pain, ADL, assistive device usage, IADL, and mobility); (4) Lifestyle Factors (5 features: overall physical activity at various intensities, social participation, recreational social participation, smoking, and nap duration); (5) Cognitive and Psychological Status (3 features: depression score, cognitive score, and memory score); (6) Housing Conditions (2 features: indoor cleanliness and indoor temperature).

Model performance

The 1,087 older adults with sarcopenia were randomly partitioned into a training set (70%) and a test set (30%). Utilizing the selected features, six machine learning models were employed to predict fall risk among older adults with sarcopenia. Their performance was evaluated using accuracy, sensitivity, specificity, precision, recall, the F1 score, the Kappa coefficient, and the area under the receiver operating characteristic curve (AUC). Performance metrics for the six machine learning models, as evaluated on the test set, are presented in Table 2; Fig. 2. The results demonstrated that among the six models, the RF model outperformed the others overall, emerging as the optimal predictive model. The RF model achieved sensitivity of 89.31 ± 2.85, specificity of 95.19 ± 0.92, accuracy of 92.26 ± 1.38, precision of 94.86 ± 1.03, recall of 89.31 ± 2.85, F1 score of 91.98 ± 1.55, a Kappa coefficient of 0.85 ± 0.03, and an AUC of 0.97 ± 0.01, which were significantly higher than those of the other models. Notably, the RF model’s AUC, approaching 1, and its ROC curve located near the top-left corner, underscored its strong classification performance and its ability to accurately distinguish between positive and negative samples. Compared with other models, including ELM (AUC = 0.66 ± 0.01), BPNN (AUC = 0.71 ± 0.04), Decision Tree (AUC = 0.81 ± 0.02), SVM (AUC = 0.71 ± 0.02), and Logistic Regression (AUC = 0.71 ± 0.03), the RF model consistently outperformed them across all metrics, underscoring its superior generalizability and broad applicability. Thus, whether high sensitivity or high specificity is prioritized, the RF model consistently satisfies both requirements, establishing it as the clear optimal choice.

Table 2.

Prediction performance of six classification models

Model	Sp	Se	Acc	Precision	Recall	F1	Kappa	AUCRF
BPNN	59.63 ± 7.11	72.08 ± 1.07	65.84 ± 3.21	64.17 ± 3.97	72.08 ± 1.07	67.83 ± 1.96	0.32 ± 0.06	0.71 ± 0.04
ELM	55.67 ± 1.67	67.82 ± 1.27	61.72 ± 0.91	60.34 ± 1.01	67.82 ± 1.27	63.85 ± 0.64	0.24 ± 0.02	0.66 ± 0.01
RF	95.19 ± 0.92	89.31 ± 2.85	92.26 ± 1.38	94.86 ± 1.03	89.31 ± 2.85	91.98 ± 1.55	0.85 ± 0.03	0.97 ± 0.01
DT	79.48 ± 4.76	71.09 ± 4.84	75.28 ± 0.23	77.73 ± 3.37	71.09 ± 4.84	74.09 ± 1.42	0.51 ± 0.01	0.81 ± 0.02
SVM	61.19 ± 3.03	73.27 ± 4.69	67.19 ± 1.62	65.26 ± 0.88	73.27 ± 4.69	68.97 ± 1.87	0.34 ± 0.03	0.71 ± 0.02
LR	64.70 ± 3.19	69.13 ± 4.17	66.90 ± 2.65	66.10 ± 1.79	69.13 ± 4.17	67.55 ± 2.52	0.34 ± 0.05	0.71 ± 0.03

Abbreviations: RF Random Forest, BPNN Back Propagation Neural Network, ELM Extreme Learning Machine, DT Decision Tree, SVM Support Vector Machine, LR Logistic Regression

Fig. 2.

ROC curve comparison of six classification models. BPNN Back Propagation Neural Network, ELM Extreme Learning Machine, RF Random Forest, DT Decision Tree, SVM Support Vector Machine, LR Logistic Regression

Feature importance

Random forest feature selection was utilized to evaluate each variable’s contribution to predicting fall risk among older adults with sarcopenia, thereby identifying the key predictors. Figure 3 displays the ranked importance of the selected features. Notably, these top 10 predictors—drawn from the 48 retained variables—exhibited the greatest contributions and had a substantial influence on the model’s predictive performance. Specifically, the ten most influential features included: ADL (1.07), social participation (0.91), IADL) (0.89), education level (0.85), depression score (0.84), personal assets (0.75), financial support from children (0.75), recreational social participation (0.74), presence of liver disease (0.73), and cognitive score (0.71).

Fig. 3.

Feature importance ranking (top 10) with contribution rate analysis for the Random Forest model. CESD Center for Epidemiologic Studies Depression Scale, ADL Activities of daily living, IADL Instrumental activities of daily living

Discussion

This study developed and validated a fall risk prediction model for older adults with sarcopenia in China utilizing six years of longitudinal data from the CHARLS and applying machine learning techniques. The results showed that the RF model outperformed alternative algorithms, achieving high sensitivity (89.31 ± 2.85), specificity (95.19 ± 0.92), and accuracy (92.26 ± 1.38), with an AUC of 0.97 ± 0.01, indicating excellent predictive performance. Further analysis revealed that ADL and IADL in health status and function, social participation in lifestyle factors, and depression and cognitive scores in psychological status were the key predictors of falls among older adults with sarcopenia. These findings suggest that targeted interventions addressing these factors may effectively reduce fall risk in older adults with sarcopenia. Additionally, these findings offer robust tools for the timely identification of high-risk individuals among older adults with sarcopenia in clinical settings.

This study confirmed that ADL and IADL exhibited significant predictive power for fall risk among older adults with sarcopenia, aligning with findings from several previous studies [45–49]. These studies further revealed that individuals with sarcopenia commonly experience impairments in ADL and IADL, characterized by progressive declines in essential abilities such as dressing and bathing. Such functional impairments not only reflect underlying physiological deterioration but are also closely associated with an increased risk of falls and greater dependence on caregiving, representing an important clinical manifestation of adverse outcomes in sarcopenia.

Integrating assessments of ADL and IADL into clinical decision-making systems is not only significant for risk stratification but also essential for informing the development of targeted intervention pathways. Specifically, existing evidence supported the feasibility of such approaches from multiple perspectives: following a 2-week mixed exercise intervention, the Barthel Index scores of older adults with sarcopenia improved by 7.8 points (95% CI, 4.0–11.8) [50]; even among low-income older adults with ADL/IADL limitations, a multidisciplinary program that combined environmental and behavioral strategies reduced ADL disability scores by approximately 30% [51]; and a 3-month combined aerobic and resistance training program significantly enhanced core daily skills, including eating, transferring, bathing, and household activities [52]. Collectively, these findings suggest that for patients with sarcopneia who already exhibit ADL/IADL impairments, the early initiation of function-oriented multimodal interventions, encompassing resistance training, nutritional optimization, and environmental modification, may effectively delay functional decline while reducing fall risk. Therefore, the assessment of ADL and IADL serves both theoretical and practical roles in fall prevention among individuals with sarcopenia, providing an empirical foundation for the development of early identification and targeted intervention models.

Notably, our findings identified social participation as an important predictor of fall risk among older adults with sarcopenia, although the direction of its effect appeared to be inconsistent. Supporting evidence suggests that reduced social contact frequency is associated with an elevated fall risk, potentially due to decreased physical activity and diminished psychosocial support, which may exacerbate impairments in balance and overall functional capacity [53]. For instance, a cross-sectional study involving 1,000 participants reported that a lower frequency of social interactions was significantly associated with a higher incidence of falls, implying that enhanced social participation might mitigate fall risk [53]. Additionally, a longitudinal study of 13,061 older adults in the United Kingdom found that limited social contact was independently linked to a 4% increase in self-reported fall risk (HR = 1.04, 95% CI: 1.01–1.07) and a 6% increase in hospital-related falls (HR = 1.06, 95% CI: 1.01–1.10) [54]. Conversely, some studies reported that participation in high-intensity social activities was associated with an increased fall risk (OR = 1.07, 1.04, and 1.09, respectively [55], possibly due to environmental hazards common in urban settings or the elevated physical demands of such activities. These inconsistent findings suggest that the effect of social participation on fall risk is likely moderated by the type, intensity, and frequency of social activities, as well as by contextual environmental factors. Accordingly, fall prevention strategies for older adults with sarcopenia should be personalized, taking into account not only the level of social participation but also the individual’s physical condition and living environment.

The present model also identified education level, depressive symptoms, and cognitive function as significant factors influencing fall risk among older adults with sarcopenia. Previous studies have indicated that older adults with higher educational attainment generally possess greater self-awareness regarding fall prevention. This awareness may stem from their stronger health consciousness and the availability of more time and financial resources to acquire fall-related knowledge through books and online sources after retirement. Conversely, individuals with lower education levels often exhibit limited awareness of fall risks [56]. This finding underscores the necessity of enhancing fall risk awareness among older adults with lower educational backgrounds by utilizing easily comprehensible formats, such as videos and pictorial materials, to communicate potential environmental hazards and improve their vigilance against falls. Epidemiological evidence reveals that many fall-related risk factors, including psychomotor retardation and gait disturbances, are frequently observed among older adults experiencing depressive symptoms [57, 58]. Additionally, individuals with depression tend to demonstrate cognitive deficits that impair attention, executive function, and processing speed, all of which contribute to an increased likelihood of falls [59]. Prior research has shown that executive function, attention, and processing speed play direct roles in gait and postural control; cognitive impairment can lead to gait deterioration, characterized by reduced walking speed, shorter stride length, and prolonged double-support phase, particularly under complex or dual-task conditions, thereby significantly elevating fall risk [60, 61]. Therefore, it is essential to incorporate emotional and cognitive factors into the comprehensive management framework for fall prevention in older adults with sarcopenia. It is advisable to integrate depression scales, such as the CESD, and brief cognitive screening tools, like the Montreal Cognitive Assessment (MoCA) or the Mini-Mental State Examination (MMSE), into routine assessments. This integration will facilitate the early identification of high-risk individuals and enable timely, targeted interventions. Additionally, group-based physical activities, including Tai Chi, dance, and community walking, are highly recommended as primary strategies for fall prevention. These activities not only enhance muscle strength and balance but also promote emotional well-being, cognitive engagement, and social interaction, thereby synergistically reducing fall risk across physical, psychological, and social domains.

In addition to the factors discussed above, the integration of machine learning methods in this study presents a novel approach to understanding fall risk in older adults with sarcopenia. This study integrated machine learning methods with longitudinal cohort data to systematically assess a multidimensional predictive framework for fall risk in older adults with sarcopenia. The random forest model demonstrated clinical utility and provided a robust theoretical foundation for targeted interventions by elucidating the influence of risk factors—namely, ADL, IADL and social participation on sarcopenia-related falls. However, this study has several limitations. Firstly, the predictive variables were derived from longitudinal data, which did not account for dynamic changes in fall risk. Future research should incorporate real-time data, such as gait patterns, heart rate, and other wearable device metrics, to enable more precise and timely risk assessments. Secondly, the assessment of fall exposure in this study relied entirely on participant self-reporting. This method is inherently susceptible to recall bias, which may lead to an underestimation of the true incidence of falls and potentially attenuate the observed associations between risk factors and fall outcomes. Future studies would benefit from incorporating objective monitoring data or conducting external validation using independent cohorts to further corroborate the robustness of these findings. Thirdly, To ensure comparability with previous similar studies in this field based on the CHARLS database [62, 63], this study adopted the average of the two highest readings from two repeated tests of each hand as the grip strength value. Evidence indicated that averaging repeated grip strength trials may yield more reliable results than relying on a single measurement [64]. However, this operational definition diverges from the AWGS 2019 recommendation of taking the maximum value from four tests [23], which introduces a potential methodological limitation. Future studies are encouraged to further compare the effects of these two measurement strategies on the diagnosis of sarcopenia and its prognostic predictions. Fourthly, gait speed was assessed over a 2.5-meter walking distinct acceleration and deceleration phases, which may result in an overestimation of physical dysfunction. Fifthly, although the CHARLS database utilized in this study possesses strong national representativeness, its primary focus on community-dwelling older adults with sarcopenia in China inherently limits the generalizability of conclusions. Furthermore, to circumvent the issues of model overfitting and subsequent interpretive bias, prediction models stratified by sarcopenia severity were not developed. Subsequent research, particularly those with access to larger sample sizes, would be well-positioned to pursue severity-stratified subgroup analyses or construct corresponding predictive models. Such work could ultimately support early, precise identification and tailored intervention strategies for patients at different stages of the condition.

Conclusions

In summary, this study established and validated a machine learning-driven fall risk prediction model tailored for Chinese older adults with sarcopenia, identifying multifactorial contributors across functional domains, particularly ADL, IADL, and social participation as critical determinants of fall susceptibility. The random forest model exhibited superior prediction accuracy and classification consistency, rendering it well-suited for clinical risk stratification and the formulation of targeted intervention strategies. The findings offer a robust scientific basis for personalized fall prevention and public health policy formulation, as well as a foundation for intelligent health monitoring and community health screening. Future research should further assess the real-world effectiveness of intervention strategies derived from this model, to achieve more precise fall risk management.

Abbreviations

CHARLS

China Health and Retirement Longitudinal Study

AUC

Area under the receiver operating characteristic curve

ADL

Activities of daily living

IADL

Instrumental activities of daily living

AWGS

the Asian Working Group on Sarcopenia

ASM

Skeletal muscle mass

DXA

Dual-energy X‐ray absorptiometry

CST

Chair stand test

SPPB

Short Physical Performance Battery

CESD

the Center for Epidemiologic Studies Depression Scale

MoCA

the Montreal Cognitive Assessment

MMSE

the Mini-Mental State Examination

MCMC

the Markov Chain Monte Carlo

SMOTE

the Synthetic Minority Over-sampling Technique

BPNN

Back Propagation Neural Network

ELM

Extreme Learning Machine

RF

Random Forest

DT

Decision Tree

SVM

Support Vector Machine

LR

Logistic Regression

TP

True Positives

FP

False Positives

FN

False Negatives

TN

True Negatives

SP

Specificity

Se

Sensitivity

Acc

Accuracy

Kappa1, Kappa2

Kappa coefficient and its variants

AUCRF

the Area under the curve of the random forest model

Authors' contributions

Ruihan Wan: Conceptualization, study design, data curation, project administration, formal analysis, writing-original draft. Danting Long: Study design, investigation, methodology, writing-original draft preparation. Kangle Wang and Kaifeng Xu: Writing-original draft, writing-review and editing. Yuxuan Sun and Xiuling Sun: Formal analysis. Weidong He: Conceptualization, supervision, writing-review and editing. Zhizhen Liu: Conceptualization, supervision, writing-review and editing.

Funding

This study was supported by grants from the National Natural Science Foundation of China (No. 82575191), the Key Research and Development project funded by the Ministry of Science and Technology of the People’s Republic of China (No. 2023YFC3503701), Scientific Research Foundation for the Top Youth Talents of Fujian University of Traditional Chinese Medicine (No. XQC2023005), and Fujian University of Traditional Chinese Medicine Research Fund (No. XJB2022007).

Data availability

Data can be downloaded from the https://charls.pku.edu.cn/.

Declarations

Ethics approval and consent to participate

The data, examination physical examination, and survey results are acquired from the CHARLS, the original CHARLS was approved by the Ethical Review Committee of Peking University (IRB00001052 − 11015). And informed consent was obtained from all participants.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.