Frailty Identification using a Sensor-based Upper- extremity Function Test: A Deep Learning Approach

Frailty Identification using a Sensor-based Upper- extremity Function Test: A Deep Learning Approach | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Frailty Identification using a Sensor-based Upper- extremity Function Test: A Deep Learning Approach Mehran Asghari, Hossein Ehsani, Nima Toosizadeh This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-4458153/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 22 Apr, 2025 Read the published version in Scientific Reports → Version 1 posted 10 You are reading this latest preprint version Abstract The global increase in the older adult population highlights the need for effective frailty assessment, a condition linked to adverse health outcomes such as hospitalization and mortality. Existing frailty assessment tools, like the Fried phenotype and Rockwood score, have practical limitations, necessitating a more efficient approach. This study aims to enhance frailty prediction accuracy in older adults using a combined biomechanical and deep learning approach. We recruited 312 participants (126 non-frail, 145 pre-frail, 41 frail) and assessed frailty using the Fried index, upper-extremity function (UEF) test, and muscle force calculations. Machine learning (ML) models, including logistic regression and support vector machine (SVM), were employed alongside deep learning with long short-term memory (LSTM) networks. Results showed that incorporating muscle model parameters significantly improved frailty prediction. The LSTM model achieved the highest accuracy (74%), outperforming SVM (67%) and regression (66%), with precision and F1 scores of 81% and 75%, respectively. Notably, muscle co-contraction emerged as a critical predictor, with frail individuals exhibiting substantially higher levels. Our findings demonstrate that integrating UEF tasks with deep learning models provides superior frailty prediction, potentially offering a robust, efficient clinical tool. However, further validation with larger, more diverse populations is needed to confirm the generalizability of our results. This study underscores the potential of advanced computational techniques to improve the identification and management of frailty in older adults. Biological sciences/Biological techniques Biological sciences/Computational biology and bioinformatics Health sciences/Diseases Physical sciences/Engineering Physical sciences/Mathematics and computing Frailty assessment Long short-term memory (LSTM) Deep learning Muscle co-contraction Figures Figure 1 Introduction The escalating proportion of the older adults worldwide underscores the urgency of addressing aging-related syndromes, notably frailty, which is both prevalent and clinically burdensome [ 1 – 3 ]. Frailty is a clinical condition characterized by an excessive vulnerability of an individual to endogenous and exogenous stressors [ 4 ]. Frailty generates a high risk of developing negative health-related events including hospitalization, cognitive decline, institutionalization, delirium, and mortality [ 5 – 7 ]. Identifying frailty has the potential to improve the clinical care of vulnerable older adults in the hospital setting and select the best treatment procedure [ 8 ]. Frailty affects approximately 11% of community-dwelling older adults and up to 30–70% of older surgical patients [ 9 , 10 ]. Furthermore, 20% inappropriate medications have reported in older adults due to miss-assessment of frailty [ 11 ]. Given the heterogeneous nature of older adult populations, and adverse consequences of frailty miss-assessment, identifying frailty becomes imperative for healthcare providers to effectively tailor treatment decisions. Numerous multidimensional assessment tools have been previously devised for frailty assessment [ 12 – 14 ]. The most well-known and widely used phenotype was defined by Fried et al. [ 15 , 16 ], which identifies frailty as the presence of three or more of five criteria, including weight loss, exhaustion, weak grip strength, slow walking speed, and low physical activity [ 15 ]. In additional to the Fried phenotype, the Rockwood score, being among the most widely utilized frailty measures, and have undergone validation within sizable cohorts [ 16 ]. Nonetheless, practical constraints often prevent their adoption in clinical contexts [ 13 ]. These constraints encompass spatial limitations, safety concerns, and the physical capabilities necessary for gait assessment within the Fried phenotype, as well as the demand for an array of clinical data points with the Rockwood score, coupled with the time-intensive nature of assessment for both measures [ 13 ]. Despite these obstacles, the importance of these frailty assessment tools cannot be overstated, as they offer valuable insights into an individual's vulnerability to adverse health outcomes. Nevertheless, an alternative objective, fast, feasible, and easy to perform approach is needed to measure frailty as a standard clinical tool. We previously developed an upper-extremity function (UEF) test to identify frailty in older adult [ 17 ]. The approach demonstrated convergent validity with established frailty measures including the Fried index and the Rockwood score [ 18 ]. This method, integrating low-cost sensors, provided an objective and rapid assessment, which is applicable to both ambulatory and non-ambulatory older adults. UEF employs physical frailty features including slowness, weakness, exhaustion, and flexibility of upper-extremity movements based on kinematics data [ 19 ]. Additionally, in a more recent study we showed that assessing muscle performance, implementing a subject-specific arm muscle model, may provide additional information regarding frailty-related neuromuscular deficiencies compared to kinematics data [ 20 ]. In our initial frailty assessment model, we utilized logistic regression to predict frailty using kinematics features in general community dwelling older adults [ 19 ]. Further, we conducted a separate study examining the impact of employing various machine learning (ML) approaches, including K-nearest neighbors (KNN), support vector machine (SVM), and logistic regression, to predict adverse outcomes using features extracted from UEF, in older adults with chronic obstructive pulmonary disease (COPD) [ 21 ]. Although there were only slight differences in accuracy among these methods, our findings indicated that SVM yields superior accuracy compared to other methods. Although, in previous research, we explored the extraction of kinematic, muscle model, and nonlinear features for frailty assessment [ 20 – 22 ], the utilization of deep learning approaches to directly learn features from the original signals, has yet to be studied. In prior investigations, deep learning methodologies have been utilized to explore the connection between motor performance such as functional mobility (walking, sit and stand from chair maintaining balance), balance control, and gait speed with frailty [ 23 – 25 ]. Various ML models, such as convolutional neural networks (CNN), and artificial neural networks (ANN), were employed to achieve high accuracy rates in identifying frail individuals, using gait data [ 24 ]. Furthermore, other studies highlighted the effectiveness of deep learning techniques, particularly long short-term memory (LSTM) networks, in discerning frail individuals using gait time-series from wearable sensors [ 23 ]. In another study, our team developed an LSTM model to differentiate between pre-frail/frail and non-frail older adults based on heart rate responses to gait [ 26 ]. The findings revealed that deep learning approaches offer superior accuracy compared to traditional ML methods for predicting frailty [ 26 ]. Our current study aimed to evaluate the frailty status of older adults by employing a combined biomechanical and deep learning approach. Given the strong performance of deep learning algorithms in pattern recognition and classification of frailty stages [ 27 – 29 ], particularly the LSTM method [ 30 ], in the current study, we developed an LSTM model using muscle force signals to improve frailty prediction accuracy compared with traditional ML approaches. Our first hypothesis was that augmenting kinematic with muscle model features would enhance frailty prediction accuracy in older adults compared to kinematic features alone. To test this, we computed additional muscle model features highlighting key indicators of muscle weakness identified in previous research [ 20 ]. Our second hypothesis was that utilizing a deep learning approach, rather than feature extraction, would enable the model to autonomously identify outlier features, thereby improving frailty prediction. Method Participants Older adults (≥ 65 years) residing in the community or attending outpatient services were recruited from the Banner University Medical Center in Tucson and the Banner Sun Health Research Institute in Sun City, Arizona. Inclusion criteria were older adults over 65 years old. Exclusion criteria encompassed known disorders linked with severe motor deficits (such as stroke, diagnosed Parkinson’s disease, and amputation), major mobility impairments (i.e., inability to walk approximately 10 meters), and upper-extremity disorders (including severe bilateral shoulder or elbow rheumatoid or osteoarthritis). Participants were permitted to use assistive devices regularly in their daily activities and during walking test, including canes and walkers, with the exception of wheelchairs. The study received approval from the Institutional Review Boards of the University of Arizona and the Banner Sun Health Research Institute. Prior to participation, all subjects provided written informed consent in accordance with the principles outlined in the Declaration of Helsinki [ 31 ]. Frailty and cognitive evaluation The Fried index, incorporating unintentional weight loss, self-reported exhaustion, low physical activity, grip strength, and walking speed, served as the gold standard [ 15 ]. Grip strength was evaluated three times for both right and left hands, and the average for both sides was recorded. Walking speed was determined as the time required to cover 4.6 meters (15 feet). Individuals were categorized as frail if they exhibited three or more positive Fried criteria, as pre-frail if they met one or two criteria, and as non-frail if they met none. UEF assessment To conduct the UEF task, participants positioned themselves on a chair to ensure unrestricted arm movement. They were instructed to rapidly flex and extend their elbow for a duration of 20 seconds. This timeframe adhered to established UEF testing protocols aimed at minimizing the influence of fatigue [ 18 , 19 ]. Prior to starting the task with their dominant arm, participants practiced the movement with their non-dominant arm without data recording. If a severe injury or vital sign attachments precluded the use of the dominant arm, the task was performed with the non-dominant arm. Previous studies among older adults have yielded comparable results from both arms [ 17 ]. During data recording, participants maintained their elbow fully extended for at least two seconds as the starting position before each UEF task. Upper-arm and forearm angular velocity were measured using commercialized wireless wearable motion sensors (tri-axial wearable gyroscope sensor, Bio-Sensics LLC, Cambridge, MA). Each sensor captured tri-axial gyroscope data with a sampling frequency of 100Hz. Two sensors were employed: one placed on the wrist and the other on the upper arm to monitor forearm and upper-arm angular velocity, facilitating the estimation of elbow joint rotations. Gyroscope data underwent initial noise and drift removal through a first-order high-pass Butterworth filter with a cutoff of 2.5Hz. Subsequently, integration and differentiation with respect to time were performed to compute joint angles and angular acceleration, respectively. Muscle force calculation We previously developed and verified a two degree of freedom arm model to characterize muscle forces during arm flexion-extension within the UEF task [ 20 ]. This model consisted of the elbow and shoulder joints in the sagittal plane and seven muscles. We used optimization approach along with the entropy assisted cost function to predict muscle forces. The advantage of entropy-assisted cost function is that it predicts muscle co-contraction (i.e., duration that both biceps and triceps are active) using kinematics data. We extracted UEF kinematics and kinetics features using anthropometric data and the musculoskeletal arm model (see Table 1 for parameter definitions). Machine Learning Approach Drawing from our prior research, we employed two ML approaches well-suited for nonlinear problems and one-dimensional time series: logistic regression and SVM [ 32 ]. The logistic regression model utilizes the natural logarithm of odds as a regression function for predictors [ 33 ]. SVM, on the other hand, is a supervised learning method designed for classification, particularly effective in high-dimensional spaces, and widely utilized in handling complex and nonlinear relationships that regression methods may struggle with [ 34 ]. SVM identifies a hyperplane in the feature space that maximizes the margin between classes (e.g., frailty categories) [ 34 ]. Based on our previous finding, SVM demonstrated superior accuracy compared to other models based on UEF signals [ 20 ]. On the other hand, regression was employed to facilitate comparison with the initial frailty model we previously developed [ 17 ]. In the ML approaches, eleven extracted features were utilized for frailty prediction, detailed in Table 1 . We employed L1 and L2 regularization techniques to select the optimal feature set for frailty prediction, with parameter definitions provided in Table 1 . L1 regularization adds a penalty to the loss function of the model based on the absolute values of coefficients, encouraging sparsity and effectively selecting relevant features. In contrast, L2 regularization penalizes large coefficient values by adding a penalty based on the squared magnitudes of coefficients, helping prevent overfitting without forcing coefficients to zero, thus retaining more features [ 35 ]. Table 1 kinematics and kinetics features during 20 second of elbow flexion/extension Kinematics UEF features Speed Mean value of the elbow angular velocity range (maximum minus minimum speed) Power Mean value of the product of the angular acceleration range and the range of angular velocity Rise time Mean value of the time required to reach the maximum angular velocity Moment Mean value of the maximum moment on elbow within each flexion/extension; estimated from the moment of inertia of the forearm and the hand, and elbow motion Flexibility Mean value of the elbow flexion range Speed Variability Coefficient of variation (standard deviation divided by the mean) of the angular velocity range Speed Reduction Difference in the angular velocity range between the last and the first five seconds of elbow flexion as a percentage of initial angular velocity range Flexion number Number of flexion/extensions during 20 seconds Flexion Muscle Force Max of muscle force mean value across flexion cycles Flexion Activation Duration Average of percentage of muscle activation duration (force > 0) during flexion across cycles Co-contraction Index Overlapping activity duration of flexor and extensor muscles during the cycle period \(T\) divided by the cycle duration, averaged across all cycles To evaluate the generalizability of the classification algorithm, we employed a stratified 5-fold cross-validation procedure. The dataset, including all three classes of frailty groups, was randomly divided into 5 folds, with four folds allocated for training and one-fold for testing. To ensure consistency, an equal number of participants within each frailty category were randomly assigned to all folds. This process was iterated 5 times, with each iteration utilizing a different fold for testing. Given the random nature of data partitioning, the 5-fold cross-validation process was repeated 1000 times to ascertain the potential range of classifier performance. To compare classifiers, we computed metrics including accuracy, F1 score, and precision. Deep learning approach In this study, we employed the LSTM model, a type of recurrent neural network (RNN) algorithm, specifically designed for analyzing time series data [ 36 ]. The LSTM algorithm was chosen for its ability to retain and recall information over extended periods, particularly suited for classifying physiological time series data. We utilized muscle force signals calculated by the muscle model as inputs for our deep learning model because based on our previous findings we found that muscle model data provides better information regarding neuromuscular deficiencies compared to kinematics data [ 37 ]. Hyperparameters and the LSTM structure were determined through previous research and optimization efforts [ 26 ]. Sigmoid activation function was applied for the output layer classification, while rectified linear unit (ReLU) and hyperbolic tangent (Tanh) were utilized for activation functions in other hidden layers. The Adam optimizer updated weights during back-propagation. The LSTM network comprised five LSTM layers and two fully connected layers, with node numbers set at 55, 50, 40, 30, 20 for the LSTM layers and 10 and 5 for the fully connected layers, respectively (Fig. 1 ). Three classes of frail, pre-frail, and non-frail labels were considered for the model. Data augmentation was employed, and the LSTM algorithm was used for frailty category prediction using muscle force data, with performance evaluated via 5-fold cross-validation. Similar to other ML evaluations, each fold underwent a 4:1 training-testing split iteratively until all folds were used for testing. Performance metrics were averaged across each training set to determine the overall model performance (Fig. 1 ). Data Augmentation We incorporated data augmentation to expand the dataset for training the model, addressing potential issues stemming from insufficient data that may impede the LSTM model convergence. Additionally, an imbalanced dataset, characterized by a smaller sample of frail individuals compared to pre-frail and non-frail cases, which was the case for the current sample, can introduce bias into classification outcomes [ 35 ]. To rectify this, we performed data augmentation to increase the training sample size and achieve balanced dataset. Our approach involved applying a combination of scaling and jittering methods [ 36 ] to generate new muscle force training data. Scaling entailed multiplying a random scalar value with the muscle force time series vector, with the random value drawn from a normal distribution with a mean of one and standard deviations (SDs) of 0.1 or 0.2 [ 37 ], [ 14 ]. Subsequently, Gaussian noise was introduced to the scaled dataset, incorporating a vector from a normal distribution with a zero mean and SDs of 0.05 or 0.1 [ 37 ], [ 14 ]. Statistical Analysis Univariate analysis of variance (ANOVA) models were utilized to examine differences in demographic information among frailty groups. The Chi-square (χ 2 ) test was employed to assess differences in sex distributions across groups. ANOVA models were deployed to explore variations in muscle model and UEF kinematics parameters among the three frailty groups. Results are summarized as mean (standard deviation-SD). All statistical analyses were conducted using MATLAB, with statistical significance determined at p < 0.05. Results Participants We recruited three hundred and twelve participants, comprising 126 (40%) non-frail, 145 (46%) pre-frail, and 41 (13%) frail older adults, as defined by the Fried index. Significant differences in age and height were observed among frailty groups ( p < 0.01). On average, frail individuals were ~ 3 years older and 1% shorter compared to pre-frail counterparts, while pre-frail participants were ~ 5 years older and 2% shorter compared to non-frail individuals. Differences in demographic parameters are detailed in Table 2 . Table 2 Demographic information for the three frailty categories defined using the Fried index. Variable Frail Pre-frail Non-frail p -value (effect size) Number, n (%of the total) 41 (13%) 145 (46%) 126 (40%) - Male, n (%of the total) 10 (25%) 45 (31%) 48 (38%) 0.22 Age, year (SD) 82.64 (7.23) 79.76 (9.27) 77.24 (7.92) 0.01* (0.35) Weight, kg (SD) 74.14 (17.52) 73.66 (18.27) 74.63 (16.92) 0.95 (0.01) Height, cm (SD) 161.69 (9.29) 163.56 (9.29) 166.11 (10.92) 0.01* (0.14) BMI, kg/m 2 (SD) 28.12 (6.14) 27.43 (6.57) 27.01 (4.59) 0.21 (0.09) The asterisk symbols represent a significant difference between frailty groups. Significant differences were noted in all kinetics and muscle model parameters among frailty groups ( p < 0.01). The highest effect size was obtained for muscle model parameters, especially muscle co-contraction (effect size = 1.72, Table 3 ). Average effect size achieved by kinematics parameters was 0.87, while the average effect size obtained by muscle model parameters was 1.25. Frail individuals exhibited notably worse performance across all variables compared to pre-frail and non-frail counterparts, suggesting diminished function, which was characterized by slower movements, reduced muscle power (higher percentage and duration of muscle activity), and increased variability and muscle co-contraction. The optimum feature selection obtained by L1, L2 regularization approach were consisted of muscle co-contraction, flexion activation duration, and speed. Table 3 Differences in UEF kinematics and muscle model parameters for the three frailty categories defined using the Fried index. Variable Frail Pre frail Non-frail p -value (effect size) Speed, degrees/s (SD) 452 (23) 789 (185) 1184 (262) < 0.01* (1.12) Power, degrees 2 /s 3 *100000 (SD) 22.7 (14.1) 75.2 (37.8) 207.5 (114.5) < 0.01* (1.08) Rise time, msec (SD) 42.1 (17,7) 30.4 (7.4) 26.8 (4.2) < 0.01* (0.78) Moment, Nm (SD) 15.1 (7.6) 41.9 (22.5) 59.9 (25.8) < 0.01* (0.68) Flexibility, degrees (SD) 85 (26) 111 (21) 138 (25) < 0.01* (0.67) Speed Variability, % (SD) 31.2 (37.1) 12.5 (5.7) 9.2 (3.1) < 0.01* (0.88) Speed Reduction, % (SD) 23.5 (24.9) 8.1 (8.3) 1.9 (5.6) < 0.01* (0.71) Flexion number, n (SD) 11.9 (4.1) 19.7 (4.7) 24.8 (5.9) < 0.01* (0.83) Flexion Muscle Force, % (SD) 61.7 (21.1) 48.5 (23.7) 32.3 (24.5) < 0.01* (1.08) Flexion Activation Duration, % (SD) 88.5 (27.8) 70.4 (29.8) 45.3 (25.8) < 0.01* (0.94) Co-contraction Index, % (SD) 81.7 (23.6) 56.7 (29.2) 33.7 (26.6) < 0.01* (1.72) The asterisk symbols represent a significant difference between frailty groups. With the data augmentation, the dataset size expanded from 312 samples (prior to augmentation) to 1284 samples (post-augmentation). Results of the 5-fold cross-validation using LSTM, regression, and SVM models indicated that the LSTM model achieved the highest accuracy of 74%, outperforming both SVM and regression models. With F1 score and precision values of 75% and 81%, respectively, the LSTM model demonstrated superior capability in accurately identifying frail and pre-frail individuals while maintaining a high precision rate. Table 4 Average of performance metrics in 5-Fold cross validation using LSTM, regression, and SVM models for predicting non-frail, pre-frail, and frail individuals. Model Accuracy (%) F1 score Precision (%) SVM 67 70 75 Regression 66 67 73 LSTM 74 75 81 Discussion Kinematics and muscle model features As hypothesized, our study revealed that incorporating muscle model parameters significantly enhances the accuracy of frailty prediction. Among features, two key parameters from the muscle model, namely flexion activation duration and co-contraction, emerged as influential predicting variables. This underscores the potential value of assessing muscle force characteristics as an alternative and efficient approach for studying frailty. Numerous studies have established the association between frailty and sarcopenia/dynapenia, characterized by muscle loss and function [ 38 – 40 ]. The observed muscle weakness in frail patients may stem from various factors, including alterations in the structure, function, and metabolism of peripheral skeletal muscles [ 41 ]. Additionally, social frailty status has been linked with muscle weakness in previous research [ 42 ]. In this context, social frailty encompasses factors such as social isolation, loneliness, and limited social support, indicating vulnerability in social domains beyond physical health. Consistent with these findings, our study revealed that muscle model parameters exhibited the highest effect size among all included parameters (average effect size = 1.25). Specifically, co-contraction emerged as the parameter with the highest effect size (effect size = 1.72). Co-contraction occurs when agonist and antagonist muscles contract simultaneously, resulting in prolonged muscle activation, increased resistance to movement, and heightened muscle activity. While co-contraction typically enhances joint stability in healthy individuals [ 43 ], abnormal increases in co-contraction have been observed in certain conditions [ 44 ]. Our previous research on COPD patients indicated elevated levels of co-contraction compared to healthy controls [ 20 ]. Other studies have documented age-related increases in muscle co-contraction [ 45 ] as well as associations with factors such as fatigue. Moreover, investigations focusing on lower limb co-contraction have demonstrated its potential in estimating fall risk during walking, with individuals at higher risk of falls exhibiting elevated levels of muscle co-contraction. In agreement with these studies, our investigation revealed higher levels of co-contraction within the frail group compared to both pre-frail (31% less) and non-frail (59% less) participants. These findings suggest a potential link between increased co-contraction and the risk of frailty in older adults. Frailty prediction using machine learning model In this study, we extended upon prior investigations by introducing a novel methodology for identifying frailty through the analysis of UEF combined with deep learning techniques. We employed SVM and regression methodologies and compared their performance with that of LSTM networks. Among the ML approaches utilized, SVM exhibited a slightly superior performance by 1%. This aligns with previous research in which SVM outperformed other methods, such as k-nearest neighbors (KNN) and logistic regression, in predicting adverse outcomes in older adults [ 21 ]. Similarly, in a separate study comparing SVM, decision trees, and KNN for frailty prediction using functional mobility, balance control, and gait speed data, SVM consistently demonstrated the highest accuracy [ 46 ]. These consistent findings demonstrate the efficacy of SVM in developing models for predicting adverse outcomes and frailty based on motion data. By leveraging computational musculoskeletal arm models and deep learning algorithms, we enhanced the accuracy of frailty prediction in older adults compared to traditional ML approaches. According to Table 4 , the LSTM model exhibited a 12% increase in accuracy, 8% improvement in precision, and enhancements of 9% in F1 score. Our study showcases the effectiveness of deep learning methodologies, specifically the LSTM model, in predicting frailty based on muscle force signals. Similar findings, across 16 public time series data (e.g., sleep monitoring, gait analysis, daily activity, and etc.), suggest that the LSTM model performs superior than conventional ML techniques like logistic regression and SVM [ 47 ]. This suggests the potential of deep learning algorithms to autonomously discern patterns and features directly from raw signals, eliminating the need for manual feature extraction and selection. The observed enhancement in model prediction may stem from the fact that shallow learning approaches are adept at handling simpler and more linear classification tasks [ 48 ], whereas LSTM models excel at learning complex, nonlinear functions and can effectively capture long-range dependencies [ 49 ]. The association observed between the UEF task and frailty suggests that integrating the UEF task with muscle model and LSTM methodologies could serve as a robust measure of frailty. Limitation and Future Direction In this study, 312 participants were included, among which less than less than 14% were frail older adults. Although the small number of participants and unbalanced frailty group were, to some extent, remedied with augmentation, the sample size limitation still exists. The relatively small sample size and specific cohort of older adults recruited from outpatient services may limit the generalizability of our findings. Additionally, while our deep learning model showed promising results, further validation in larger and more diverse populations, especially those that are hospitalized at the time of assessment, is warranted to confirm UEF efficacy and reliability. Conclusion In conclusion, our study contributes to the growing body of research on frailty assessment by introducing a novel approach that integrates computational modeling and deep learning techniques. By leveraging muscle force signals and advanced ML algorithms, we provide a comprehensive framework for frailty prediction that is objective, efficient, and clinically relevant. Future research should focus on validating our findings in larger cohorts and exploring the integration of additional physiological signals, such as heart rate and brain function, to enhance the accuracy and predictive power of frailty models. Declarations Author Contribution Mehran Asghari contributed to the conceptualization, data analysis, and post-processing of the study. He also led the writing of the manuscript. Dr. Toosizadeh collected the data in his lab and provided critical review and guidance throughout the research process. Hossein Ehsani, proficient in machine learning, contributed by making significant improvements to the analysis and manuscript. All authors reviewed and approved the final version of the paper. Data Availability The database used for the current study is available upon reasonable request. Please contact the corresponding author, Dr. Nima Toosizadeh, at [email protected] for more information. References Buckinx, F., et al., Burden of frailty in the elderly population: perspectives for a public health challenge . Archives of public health, 2015. 73(1): p. 1–7. Horak, F.B., C.L. Shupert, and A. Mirka, Components of postural dyscontrol in the elderly: a review . Neurobiology of aging, 1989. 10(6): p. 727–738. Mitnitski, A., S.E. Howlett, and K. Rockwood, Heterogeneity of human aging and its assessment . Journals of Gerontology Series A: Biomedical Sciences and Medical Sciences, 2017. 72(7): p. 877–884. Proietti, M. and M. Cesari, Frailty: what is it? Frailty and Cardiovascular Diseases: Research into an Elderly Population, 2020: p. 1–7. Boyd, C.M., et al., Frailty, hospitalization, and progression of disability in a cohort of disabled older women . The American journal of medicine, 2005. 118(11): p. 1225–1231. Dent, E., P. Kowal, and E.O. Hoogendijk, Frailty measurement in research and clinical practice: a review . European journal of internal medicine, 2016. 31: p. 3–10. Sacha, J., et al., Is it time to begin a public campaign concerning frailty and pre-frailty? A review article . Frontiers in physiology, 2017. 8: p. 484. McMillan, G. and R. Hubbard, Frailty in older inpatients: what physicians need to know . QJM: An International Journal of Medicine, 2012. 105(11): p. 1059–1065. McRae, P.J., et al., Frailty and geriatric syndromes in vascular surgical ward patients . Annals of vascular surgery, 2016. 35: p. 9–18. Walters, S., et al., The prevalence of frailty in patients admitted to hospital with vertebral fragility fractures . Current Rheumatology Reviews, 2016. 12(3): p. 244–247. Willcox, S.M., D.U. Himmelstein, and S. Woolhandler, Inappropriate drug prescribing for the community-dwelling elderly . Jama, 1994. 272(4): p. 292–296. De Vries, N., et al., Outcome instruments to measure frailty: a systematic review . Ageing research reviews, 2011. 10(1): p. 104–114. Goldstein, J.P., M.K. Andrew, and A. Travers, Frailty in older adults using pre-hospital care and the emergency department: a narrative review . Canadian geriatrics journal: CGJ, 2012. 15(1): p. 16. Mohler, M.J., et al., The Frailty syndrome: clinical measurements and basic underpinnings in humans and animals . Experimental gerontology, 2014. 54: p. 6–13. Fried, L.P., et al., Frailty in older adults: evidence for a phenotype. The Journals of Gerontology Series A: Biological Sciences and Medical Sciences, 2001. 56(3): p. M146-M157. Rockwood, K., M. Andrew, and A. Mitnitski, A comparison of two approaches to measuring frailty in elderly people . The Journals of Gerontology Series A: Biological Sciences and Medical Sciences, 2007. 62(7): p. 738–743. Toosizadeh, N., J. Mohler, and B. Najafi, Assessing upper extremity motion: an innovative method to identify frailty . Journal of the American Geriatrics Society, 2015. 63(6): p. 1181–1186. Toosizadeh, N., et al., Assessing upper-extremity motion: An innovative method to quantify functional capacity in patients with chronic obstructive pulmonary disease . PloS one, 2017. 12(2): p. e0172766. Toosizadeh, N., et al., Frailty assessment in older adults using upper-extremity function: index development . BMC geriatrics, 2017. 17: p. 1–7. Asghari, M., et al., A computational musculoskeletal arm model for assessing muscle dysfunction in chronic obstructive pulmonary disease . Medical & Biological Engineering & Computing, 2023: p. 1–14. Asghari, M., A Novel Sensor-Based Approach to Predict Adverse Outcomes in Chronic Obstructive Pulmonary Disease . 2023, The University of Arizona. Asghari, M., et al., Nonlinear analysis of the movement variability structure can detect aging-related differences among cognitively healthy individuals . Human movement science, 2021. 78: p. 102807. Aponte-Hao, S., et al., Machine learning for identification of frailty in Canadian primary care practices . International journal of population data science, 2021. 6(1). Arshad, M.Z., et al. Gait-based frailty assessment using image representation of IMU signals and deep CNN . in 2021 43rd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC) . 2021. IEEE. Friedrich, B., et al. A deep learning approach for TUG and SPPB score prediction of (pre-) frail older adults on real-life IMU data . in Healthcare . 2021. MDPI. Eskandari, M., et al., Frailty identification using heart rate dynamics: A deep learning approach . IEEE Journal of Biomedical and Health Informatics, 2022. 26(7): p. 3409–3417. Idris, S. and N. Badruddin. Classification of cognitive frailty in elderly people from blood samples using machine learning . in 2021 IEEE EMBS International Conference on Biomedical and Health Informatics (BHI) . 2021. IEEE. Radha, M., et al., Sleep stage classification from heart-rate variability using long short-term memory neural networks . Scientific reports, 2019. 9(1): p. 14149. Wang, L. and X. Zhou, Detection of congestive heart failure based on LSTM-based deep network via short-term RR intervals . Sensors, 2019. 19(7): p. 1502. Thirugnanam, M. and M.S. Pasupuleti, Cardiomyopathy-induced arrhythmia classification and pre-fall alert generation using convolutional neural network and long short-term memory model . Evolutionary Intelligence, 2021. 14: p. 789–799. Shrestha, B. and L. Dunn, The declaration of Helsinki on medical research involving human subjects: a review of seventh revision . Journal of Nepal Health Research Council, 2019. 17(4): p. 548–552. Austria, Y.D., et al., Comparison of machine learning algorithms in breast cancer prediction using the coimbra dataset . cancer, 2019. 7(10): p. 23–1. Dayton, C.M., Logistic regression analysis . Stat, 1992. 474: p. 574. Sen, P.C., M. Hajra, and M. Ghosh. Supervised classification algorithms in machine learning: A survey and review . in Emerging Technology in Modelling and Graphics: Proceedings of IEM Graph 2018 . 2020. Springer. Ng, A.Y. Feature selection, L 1 vs. L 2 regularization, and rotational invariance . in Proceedings of the twenty-first international conference on Machine learning . 2004. DiPietro, R. and G.D. Hager, Deep learning: RNNs and LSTM , in Handbook of medical image computing and computer assisted intervention . 2020, Elsevier. p. 503–519. Asghari, M., et al., A computational musculoskeletal arm model for assessing muscle dysfunction in chronic obstructive pulmonary disease . Medical & Biological Engineering & Computing, 2023. 61(9): p. 2241–2254. Kim, J.C., K. Kalantar-Zadeh, and J.D. Kopple, Frailty and protein-energy wasting in elderly patients with end stage kidney disease . Journal of the American Society of Nephrology, 2013. 24(3): p. 337–351. Pourhassan, M., et al., The impact of malnutrition on acute muscle wasting in frail older hospitalized patients . Nutrients, 2020. 12(5): p. 1387. VanItallie, T.B., Frailty in the elderly: contributions of sarcopenia and visceral protein depletion . Metabolism, 2003. 52: p. 22–26. Wüst, R.C. and H. Degens, Factors contributing to muscle wasting and dysfunction in COPD patients . International journal of chronic obstructive pulmonary disease, 2007. 2(3): p. 289. Makizako, H., et al., Associations of social frailty with loss of muscle mass and muscle weakness among community-dwelling older adults . Geriatrics & gerontology international, 2019. 19(1): p. 76–80. Hirokawa, S., et al., Muscular co-contraction and control of knee stability . Journal of Electromyography and Kinesiology, 1991. 1(3): p. 199–208. Busse, M., C.M. Wiles, and R.W.M. Van Deursen, Muscle co-activation in neurological conditions . Physical therapy reviews, 2005. 10(4): p. 247–253. Potvin, J. and P. O'brien, Trunk muscle co-contraction increases during fatiguing, isometric, lateral bend exertions: possible implications for spine stability . Spine, 1998. 23(7): p. 774–780. Ambagtsheer, R., et al., The application of artificial intelligence (AI) techniques to identify frailty within a residential aged care administrative data set . International journal of medical informatics, 2020. 136: p. 104094. Abbasimehr, H. and R. Paki, Improving time series forecasting using LSTM and attention models . Journal of Ambient Intelligence and Humanized Computing, 2022: p. 1–19. Allam, A., et al., Neural networks versus Logistic regression for 30 days all-cause readmission prediction . Scientific reports, 2019. 9(1): p. 9277. Güney, S. and Ç.B. Erdaş. A deep LSTM approach for activity recognition . in 2019 42nd International Conference on Telecommunications and Signal Processing (TSP) . 2019. IEEE. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Published Journal Publication published 22 Apr, 2025 Read the published version in Scientific Reports → Version 1 posted Editorial decision: Revision requested 16 Aug, 2024 Reviews received at journal 12 Aug, 2024 Reviewers agreed at journal 02 Aug, 2024 Reviews received at journal 27 Jun, 2024 Reviewers agreed at journal 17 Jun, 2024 Reviewers invited by journal 17 Jun, 2024 Editor assigned by journal 17 Jun, 2024 Editor invited by journal 04 Jun, 2024 Submission checks completed at journal 31 May, 2024 First submitted to journal 22 May, 2024 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-4458153","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":313260676,"identity":"6bb87592-4be8-445e-84b0-f051e3da39e0","order_by":0,"name":"Mehran Asghari","email":"","orcid":"","institution":"Rutgers University","correspondingAuthor":false,"prefix":"","firstName":"Mehran","middleName":"","lastName":"Asghari","suffix":""},{"id":313260677,"identity":"ea8f92c2-0cbb-46ee-a1e3-51797af6e73f","order_by":1,"name":"Hossein Ehsani","email":"","orcid":"","institution":"Rutgers University","correspondingAuthor":false,"prefix":"","firstName":"Hossein","middleName":"","lastName":"Ehsani","suffix":""},{"id":313260678,"identity":"123247db-6302-4cdf-bb3e-5ffcb4b1a642","order_by":2,"name":"Nima Toosizadeh","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA9UlEQVRIiWNgGAWjYLCCBAYGxgYJ5gMIER48qnkQWtgSwCzitDCAtfAYEKfFnr358IcHf+xk+6V7vn3m/VErZ95/gPHB2zY8tvAcS5NI4Ek2njnn7ObZPAnHjWVuJDAbzsWnRSLHjCFBgjlxw43czcw8CccSZ0gwsEnz4tdi/CHBoD5x/42cxyAt9TP4D7D/JqDFQCIh4XDiBokcZqCWmgQJhgQ2ZrxazoD8cuC48YwbacaMc9IOGM6QSGyWnHMOtxb29ubDH3/8qZbtn5H8mOGNTZ28BP/hgx/elOHWgg4OM4DiiHj1QFBHkupRMApGwSgYGQAAgolO06KpanIAAAAASUVORK5CYII=","orcid":"","institution":"Rutgers University","correspondingAuthor":true,"prefix":"","firstName":"Nima","middleName":"","lastName":"Toosizadeh","suffix":""}],"badges":[],"createdAt":"2024-05-22 04:36:15","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-4458153/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-4458153/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1038/s41598-024-73854-2","type":"published","date":"2025-04-22T15:58:06+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":58595160,"identity":"6fc3aa0a-275c-4d8d-a408-9b3be3173cd3","added_by":"auto","created_at":"2024-06-18 16:37:05","extension":"jpeg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":415447,"visible":true,"origin":"","legend":"\u003cp\u003eSchematic representation of the study workflow and deep learning model\u003c/p\u003e","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-4458153/v1/65a85391da62dd9b3f5a9056.jpeg"},{"id":81569876,"identity":"54179094-f3c6-47b9-8dc7-40d33acfec09","added_by":"auto","created_at":"2025-04-28 16:12:14","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1089403,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-4458153/v1/a66df8e9-59fb-46ef-b685-d91ae4793fea.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Frailty Identification using a Sensor-based Upper- extremity Function Test: A Deep Learning Approach","fulltext":[{"header":"Introduction","content":"\u003cp\u003eThe escalating proportion of the older adults worldwide underscores the urgency of addressing aging-related syndromes, notably frailty, which is both prevalent and clinically burdensome [\u003cspan additionalcitationids=\"CR2\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Frailty is a clinical condition characterized by an excessive vulnerability of an individual to endogenous and exogenous stressors [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Frailty generates a high risk of developing negative health-related events including hospitalization, cognitive decline, institutionalization, delirium, and mortality [\u003cspan additionalcitationids=\"CR6\" citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Identifying frailty has the potential to improve the clinical care of vulnerable older adults in the hospital setting and select the best treatment procedure [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Frailty affects approximately 11% of community-dwelling older adults and up to 30\u0026ndash;70% of older surgical patients [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Furthermore, 20% inappropriate medications have reported in older adults due to miss-assessment of frailty [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Given the heterogeneous nature of older adult populations, and adverse consequences of frailty miss-assessment, identifying frailty becomes imperative for healthcare providers to effectively tailor treatment decisions.\u003c/p\u003e \u003cp\u003eNumerous multidimensional assessment tools have been previously devised for frailty assessment [\u003cspan additionalcitationids=\"CR13\" citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. The most well-known and widely used phenotype was defined by Fried et al. [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e], which identifies frailty as the presence of three or more of five criteria, including weight loss, exhaustion, weak grip strength, slow walking speed, and low physical activity [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. In additional to the Fried phenotype, the Rockwood score, being among the most widely utilized frailty measures, and have undergone validation within sizable cohorts [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. Nonetheless, practical constraints often prevent their adoption in clinical contexts [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. These constraints encompass spatial limitations, safety concerns, and the physical capabilities necessary for gait assessment within the Fried phenotype, as well as the demand for an array of clinical data points with the Rockwood score, coupled with the time-intensive nature of assessment for both measures [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. Despite these obstacles, the importance of these frailty assessment tools cannot be overstated, as they offer valuable insights into an individual's vulnerability to adverse health outcomes. Nevertheless, an alternative objective, fast, feasible, and easy to perform approach is needed to measure frailty as a standard clinical tool.\u003c/p\u003e \u003cp\u003eWe previously developed an upper-extremity function (UEF) test to identify frailty in older adult [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. The approach demonstrated convergent validity with established frailty measures including the Fried index and the Rockwood score [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. This method, integrating low-cost sensors, provided an objective and rapid assessment, which is applicable to both ambulatory and non-ambulatory older adults. UEF employs physical frailty features including slowness, weakness, exhaustion, and flexibility of upper-extremity movements based on kinematics data [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. Additionally, in a more recent study we showed that assessing muscle performance, implementing a subject-specific arm muscle model, may provide additional information regarding frailty-related neuromuscular deficiencies compared to kinematics data [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. In our initial frailty assessment model, we utilized logistic regression to predict frailty using kinematics features in general community dwelling older adults [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. Further, we conducted a separate study examining the impact of employing various machine learning (ML) approaches, including K-nearest neighbors (KNN), support vector machine (SVM), and logistic regression, to predict adverse outcomes using features extracted from UEF, in older adults with chronic obstructive pulmonary disease (COPD) [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. Although there were only slight differences in accuracy among these methods, our findings indicated that SVM yields superior accuracy compared to other methods. Although, in previous research, we explored the extraction of kinematic, muscle model, and nonlinear features for frailty assessment [\u003cspan additionalcitationids=\"CR21\" citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e], the utilization of deep learning approaches to directly learn features from the original signals, has yet to be studied.\u003c/p\u003e \u003cp\u003eIn prior investigations, deep learning methodologies have been utilized to explore the connection between motor performance such as functional mobility (walking, sit and stand from chair maintaining balance), balance control, and gait speed with frailty [\u003cspan additionalcitationids=\"CR24\" citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. Various ML models, such as convolutional neural networks (CNN), and artificial neural networks (ANN), were employed to achieve high accuracy rates in identifying frail individuals, using gait data [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. Furthermore, other studies highlighted the effectiveness of deep learning techniques, particularly long short-term memory (LSTM) networks, in discerning frail individuals using gait time-series from wearable sensors [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]. In another study, our team developed an LSTM model to differentiate between pre-frail/frail and non-frail older adults based on heart rate responses to gait [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. The findings revealed that deep learning approaches offer superior accuracy compared to traditional ML methods for predicting frailty [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eOur current study aimed to evaluate the frailty status of older adults by employing a combined biomechanical and deep learning approach. Given the strong performance of deep learning algorithms in pattern recognition and classification of frailty stages [\u003cspan additionalcitationids=\"CR28\" citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e], particularly the LSTM method [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e], in the current study, we developed an LSTM model using muscle force signals to improve frailty prediction accuracy compared with traditional ML approaches. Our first hypothesis was that augmenting kinematic with muscle model features would enhance frailty prediction accuracy in older adults compared to kinematic features alone. To test this, we computed additional muscle model features highlighting key indicators of muscle weakness identified in previous research [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. Our second hypothesis was that utilizing a deep learning approach, rather than feature extraction, would enable the model to autonomously identify outlier features, thereby improving frailty prediction.\u003c/p\u003e"},{"header":"Method","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eParticipants\u003c/h2\u003e \u003cp\u003eOlder adults (\u0026ge;\u0026thinsp;65 years) residing in the community or attending outpatient services were recruited from the Banner University Medical Center in Tucson and the Banner Sun Health Research Institute in Sun City, Arizona. Inclusion criteria were older adults over 65 years old. Exclusion criteria encompassed known disorders linked with severe motor deficits (such as stroke, diagnosed Parkinson\u0026rsquo;s disease, and amputation), major mobility impairments (i.e., inability to walk approximately 10 meters), and upper-extremity disorders (including severe bilateral shoulder or elbow rheumatoid or osteoarthritis). Participants were permitted to use assistive devices regularly in their daily activities and during walking test, including canes and walkers, with the exception of wheelchairs. The study received approval from the Institutional Review Boards of the University of Arizona and the Banner Sun Health Research Institute. Prior to participation, all subjects provided written informed consent in accordance with the principles outlined in the Declaration of Helsinki [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003eFrailty and cognitive evaluation\u003c/h2\u003e \u003cp\u003eThe Fried index, incorporating unintentional weight loss, self-reported exhaustion, low physical activity, grip strength, and walking speed, served as the gold standard [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. Grip strength was evaluated three times for both right and left hands, and the average for both sides was recorded. Walking speed was determined as the time required to cover 4.6 meters (15 feet). Individuals were categorized as frail if they exhibited three or more positive Fried criteria, as pre-frail if they met one or two criteria, and as non-frail if they met none.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003eUEF assessment\u003c/h2\u003e \u003cp\u003eTo conduct the UEF task, participants positioned themselves on a chair to ensure unrestricted arm movement. They were instructed to rapidly flex and extend their elbow for a duration of 20 seconds. This timeframe adhered to established UEF testing protocols aimed at minimizing the influence of fatigue [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. Prior to starting the task with their dominant arm, participants practiced the movement with their non-dominant arm without data recording. If a severe injury or vital sign attachments precluded the use of the dominant arm, the task was performed with the non-dominant arm. Previous studies among older adults have yielded comparable results from both arms [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. During data recording, participants maintained their elbow fully extended for at least two seconds as the starting position before each UEF task.\u003c/p\u003e \u003cp\u003eUpper-arm and forearm angular velocity were measured using commercialized wireless wearable motion sensors (tri-axial wearable gyroscope sensor, Bio-Sensics LLC, Cambridge, MA). Each sensor captured tri-axial gyroscope data with a sampling frequency of 100Hz. Two sensors were employed: one placed on the wrist and the other on the upper arm to monitor forearm and upper-arm angular velocity, facilitating the estimation of elbow joint rotations. Gyroscope data underwent initial noise and drift removal through a first-order high-pass Butterworth filter with a cutoff of 2.5Hz. Subsequently, integration and differentiation with respect to time were performed to compute joint angles and angular acceleration, respectively.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003eMuscle force calculation\u003c/h2\u003e \u003cp\u003eWe previously developed and verified a two degree of freedom arm model to characterize muscle forces during arm flexion-extension within the UEF task [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. This model consisted of the elbow and shoulder joints in the sagittal plane and seven muscles. We used optimization approach along with the entropy assisted cost function to predict muscle forces. The advantage of entropy-assisted cost function is that it predicts muscle co-contraction (i.e., duration that both biceps and triceps are active) using kinematics data. We extracted UEF kinematics and kinetics features using anthropometric data and the musculoskeletal arm model (see Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e for parameter definitions).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003eMachine Learning Approach\u003c/h2\u003e \u003cp\u003eDrawing from our prior research, we employed two ML approaches well-suited for nonlinear problems and one-dimensional time series: logistic regression and SVM [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]. The logistic regression model utilizes the natural logarithm of odds as a regression function for predictors [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]. SVM, on the other hand, is a supervised learning method designed for classification, particularly effective in high-dimensional spaces, and widely utilized in handling complex and nonlinear relationships that regression methods may struggle with [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]. SVM identifies a hyperplane in the feature space that maximizes the margin between classes (e.g., frailty categories) [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]. Based on our previous finding, SVM demonstrated superior accuracy compared to other models based on UEF signals [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. On the other hand, regression was employed to facilitate comparison with the initial frailty model we previously developed [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. In the ML approaches, eleven extracted features were utilized for frailty prediction, detailed in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. We employed L1 and L2 regularization techniques to select the optimal feature set for frailty prediction, with parameter definitions provided in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. L1 regularization adds a penalty to the loss function of the model based on the absolute values of coefficients, encouraging sparsity and effectively selecting relevant features. In contrast, L2 regularization penalizes large coefficient values by adding a penalty based on the squared magnitudes of coefficients, helping prevent overfitting without forcing coefficients to zero, thus retaining more features [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e].\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ekinematics and kinetics features during 20 second of elbow flexion/extension\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003eKinematics UEF features\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSpeed\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMean value of the elbow angular velocity range (maximum minus minimum speed)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePower\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMean value of the product of the angular acceleration range and the range of angular velocity\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRise time\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMean value of the time required to reach the maximum angular velocity\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMoment\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMean value of the maximum moment on elbow within each flexion/extension; estimated from the moment of inertia of the forearm and the hand, and elbow motion\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFlexibility\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMean value of the elbow flexion range\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSpeed Variability\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCoefficient of variation (standard deviation divided by the mean) of the angular velocity range\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSpeed Reduction\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDifference in the angular velocity range between the last and the first five seconds of elbow flexion as a percentage of initial angular velocity range\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFlexion number\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNumber of flexion/extensions during 20 seconds\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFlexion Muscle Force\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMax of muscle force mean value across flexion cycles\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFlexion Activation Duration\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAverage of percentage of muscle activation duration (force\u0026thinsp;\u0026gt;\u0026thinsp;0) during flexion across cycles\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCo-contraction Index\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eOverlapping activity duration of flexor and extensor muscles during the cycle period \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(T\\)\u003c/span\u003e\u003c/span\u003e divided by the cycle duration, averaged across all cycles\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eTo evaluate the generalizability of the classification algorithm, we employed a stratified 5-fold cross-validation procedure. The dataset, including all three classes of frailty groups, was randomly divided into 5 folds, with four folds allocated for training and one-fold for testing. To ensure consistency, an equal number of participants within each frailty category were randomly assigned to all folds. This process was iterated 5 times, with each iteration utilizing a different fold for testing. Given the random nature of data partitioning, the 5-fold cross-validation process was repeated 1000 times to ascertain the potential range of classifier performance. To compare classifiers, we computed metrics including accuracy, F1 score, and precision.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eDeep learning approach\u003c/h2\u003e \u003cp\u003eIn this study, we employed the LSTM model, a type of recurrent neural network (RNN) algorithm, specifically designed for analyzing time series data [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e]. The LSTM algorithm was chosen for its ability to retain and recall information over extended periods, particularly suited for classifying physiological time series data. We utilized muscle force signals calculated by the muscle model as inputs for our deep learning model because based on our previous findings we found that muscle model data provides better information regarding neuromuscular deficiencies compared to kinematics data [\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e]. Hyperparameters and the LSTM structure were determined through previous research and optimization efforts [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. Sigmoid activation function was applied for the output layer classification, while rectified linear unit (ReLU) and hyperbolic tangent (Tanh) were utilized for activation functions in other hidden layers. The Adam optimizer updated weights during back-propagation. The LSTM network comprised five LSTM layers and two fully connected layers, with node numbers set at 55, 50, 40, 30, 20 for the LSTM layers and 10 and 5 for the fully connected layers, respectively (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Three classes of frail, pre-frail, and non-frail labels were considered for the model. Data augmentation was employed, and the LSTM algorithm was used for frailty category prediction using muscle force data, with performance evaluated via 5-fold cross-validation. Similar to other ML evaluations, each fold underwent a 4:1 training-testing split iteratively until all folds were used for testing. Performance metrics were averaged across each training set to determine the overall model performance (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003eData Augmentation\u003c/h2\u003e \u003cp\u003eWe incorporated data augmentation to expand the dataset for training the model, addressing potential issues stemming from insufficient data that may impede the LSTM model convergence. Additionally, an imbalanced dataset, characterized by a smaller sample of frail individuals compared to pre-frail and non-frail cases, which was the case for the current sample, can introduce bias into classification outcomes [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e]. To rectify this, we performed data augmentation to increase the training sample size and achieve balanced dataset. Our approach involved applying a combination of scaling and jittering methods [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e] to generate new muscle force training data. Scaling entailed multiplying a random scalar value with the muscle force time series vector, with the random value drawn from a normal distribution with a mean of one and standard deviations (SDs) of 0.1 or 0.2 [\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e], [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. Subsequently, Gaussian noise was introduced to the scaled dataset, incorporating a vector from a normal distribution with a zero mean and SDs of 0.05 or 0.1 [\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e], [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e].\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003eStatistical Analysis\u003c/h2\u003e \u003cp\u003eUnivariate analysis of variance (ANOVA) models were utilized to examine differences in demographic information among frailty groups. The Chi-square (χ\u003csup\u003e2\u003c/sup\u003e) test was employed to assess differences in sex distributions across groups. ANOVA models were deployed to explore variations in muscle model and UEF kinematics parameters among the three frailty groups. Results are summarized as mean (standard deviation-SD). All statistical analyses were conducted using MATLAB, with statistical significance determined at \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eParticipants\u003c/h2\u003e \u003cp\u003eWe recruited three hundred and twelve participants, comprising 126 (40%) non-frail, 145 (46%) pre-frail, and 41 (13%) frail older adults, as defined by the Fried index. Significant differences in age and height were observed among frailty groups (\u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.01). On average, frail individuals were ~\u0026thinsp;3 years older and 1% shorter compared to pre-frail counterparts, while pre-frail participants were ~\u0026thinsp;5 years older and 2% shorter compared to non-frail individuals. Differences in demographic parameters are detailed in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eDemographic information for the three frailty categories defined using the Fried index.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVariable\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFrail\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePre-frail\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eNon-frail\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cem\u003ep\u003c/em\u003e-value (effect size)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNumber, n (%of the total)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e41 (13%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e145 (46%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e126 (40%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMale, n (%of the total)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e10 (25%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e45 (31%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e48 (38%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.22\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAge, year (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e82.64 (7.23)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e79.76 (9.27)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e77.24 (7.92)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.01* (0.35)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWeight, kg (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e74.14 (17.52)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e73.66 (18.27)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e74.63 (16.92)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.95 (0.01)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHeight, cm (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e161.69 (9.29)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e163.56 (9.29)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e166.11 (10.92)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.01* (0.14)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBMI, kg/m\u003csup\u003e2\u003c/sup\u003e (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e28.12 (6.14)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e27.43 (6.57)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e27.01 (4.59)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.21 (0.09)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe asterisk symbols represent a significant difference between frailty groups.\u003c/p\u003e \u003cp\u003eSignificant differences were noted in all kinetics and muscle model parameters among frailty groups (\u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.01). The highest effect size was obtained for muscle model parameters, especially muscle co-contraction (effect size\u0026thinsp;=\u0026thinsp;1.72, Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). Average effect size achieved by kinematics parameters was 0.87, while the average effect size obtained by muscle model parameters was 1.25. Frail individuals exhibited notably worse performance across all variables compared to pre-frail and non-frail counterparts, suggesting diminished function, which was characterized by slower movements, reduced muscle power (higher percentage and duration of muscle activity), and increased variability and muscle co-contraction. The optimum feature selection obtained by L1, L2 regularization approach were consisted of muscle co-contraction, flexion activation duration, and speed.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eDifferences in UEF kinematics and muscle model parameters for the three frailty categories defined using the Fried index.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVariable\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFrail\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePre frail\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eNon-frail\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cem\u003ep\u003c/em\u003e-value (effect size)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSpeed, degrees/s (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e452 (23)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e789 (185)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1184 (262)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.01* (1.12)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePower, degrees\u003csup\u003e2\u003c/sup\u003e/s\u003csup\u003e3\u003c/sup\u003e*100000 (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e22.7 (14.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e75.2 (37.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e207.5 (114.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.01* (1.08)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRise time, msec (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e42.1 (17,7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e30.4 (7.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e26.8 (4.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.01* (0.78)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMoment, Nm (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e15.1 (7.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e41.9 (22.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e59.9 (25.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.01* (0.68)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFlexibility, degrees (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e85 (26)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e111 (21)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e138 (25)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.01* (0.67)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSpeed Variability, % (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e31.2 (37.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e12.5 (5.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e9.2 (3.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.01* (0.88)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSpeed Reduction, % (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e23.5 (24.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e8.1 (8.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.9 (5.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.01* (0.71)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFlexion number, n (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e11.9 (4.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e19.7 (4.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e24.8 (5.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.01* (0.83)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFlexion Muscle Force, % (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e61.7 (21.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e48.5 (23.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e32.3 (24.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.01* (1.08)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFlexion Activation Duration, % (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e88.5 (27.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e70.4 (29.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e45.3 (25.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.01* (0.94)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCo-contraction Index, % (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e81.7 (23.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e56.7 (29.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e33.7 (26.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.01* (1.72)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe asterisk symbols represent a significant difference between frailty groups.\u003c/p\u003e \u003cp\u003eWith the data augmentation, the dataset size expanded from 312 samples (prior to augmentation) to 1284 samples (post-augmentation). Results of the 5-fold cross-validation using LSTM, regression, and SVM models indicated that the LSTM model achieved the highest accuracy of 74%, outperforming both SVM and regression models. With F1 score and precision values of 75% and 81%, respectively, the LSTM model demonstrated superior capability in accurately identifying frail and pre-frail individuals while maintaining a high precision rate.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eAverage of performance metrics in 5-Fold cross validation using LSTM, regression, and SVM models for predicting non-frail, pre-frail, and frail individuals.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAccuracy (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eF1 score\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePrecision (%)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSVM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e75\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRegression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e73\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLSTM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e74\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e75\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e81\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eKinematics and muscle model features\u003c/h2\u003e \u003cp\u003eAs hypothesized, our study revealed that incorporating muscle model parameters significantly enhances the accuracy of frailty prediction. Among features, two key parameters from the muscle model, namely flexion activation duration and co-contraction, emerged as influential predicting variables. This underscores the potential value of assessing muscle force characteristics as an alternative and efficient approach for studying frailty. Numerous studies have established the association between frailty and sarcopenia/dynapenia, characterized by muscle loss and function [\u003cspan additionalcitationids=\"CR39\" citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e]. The observed muscle weakness in frail patients may stem from various factors, including alterations in the structure, function, and metabolism of peripheral skeletal muscles [\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e]. Additionally, social frailty status has been linked with muscle weakness in previous research [\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e]. In this context, social frailty encompasses factors such as social isolation, loneliness, and limited social support, indicating vulnerability in social domains beyond physical health. Consistent with these findings, our study revealed that muscle model parameters exhibited the highest effect size among all included parameters (average effect size\u0026thinsp;=\u0026thinsp;1.25). Specifically, co-contraction emerged as the parameter with the highest effect size (effect size\u0026thinsp;=\u0026thinsp;1.72).\u003c/p\u003e \u003cp\u003eCo-contraction occurs when agonist and antagonist muscles contract simultaneously, resulting in prolonged muscle activation, increased resistance to movement, and heightened muscle activity. While co-contraction typically enhances joint stability in healthy individuals [\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e], abnormal increases in co-contraction have been observed in certain conditions [\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e]. Our previous research on COPD patients indicated elevated levels of co-contraction compared to healthy controls [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. Other studies have documented age-related increases in muscle co-contraction [\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e] as well as associations with factors such as fatigue. Moreover, investigations focusing on lower limb co-contraction have demonstrated its potential in estimating fall risk during walking, with individuals at higher risk of falls exhibiting elevated levels of muscle co-contraction. In agreement with these studies, our investigation revealed higher levels of co-contraction within the frail group compared to both pre-frail (31% less) and non-frail (59% less) participants. These findings suggest a potential link between increased co-contraction and the risk of frailty in older adults.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eFrailty prediction using machine learning model\u003c/h2\u003e \u003cp\u003eIn this study, we extended upon prior investigations by introducing a novel methodology for identifying frailty through the analysis of UEF combined with deep learning techniques. We employed SVM and regression methodologies and compared their performance with that of LSTM networks. Among the ML approaches utilized, SVM exhibited a slightly superior performance by 1%. This aligns with previous research in which SVM outperformed other methods, such as k-nearest neighbors (KNN) and logistic regression, in predicting adverse outcomes in older adults [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. Similarly, in a separate study comparing SVM, decision trees, and KNN for frailty prediction using functional mobility, balance control, and gait speed data, SVM consistently demonstrated the highest accuracy [\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e]. These consistent findings demonstrate the efficacy of SVM in developing models for predicting adverse outcomes and frailty based on motion data.\u003c/p\u003e \u003cp\u003eBy leveraging computational musculoskeletal arm models and deep learning algorithms, we enhanced the accuracy of frailty prediction in older adults compared to traditional ML approaches. According to Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, the LSTM model exhibited a 12% increase in accuracy, 8% improvement in precision, and enhancements of 9% in F1 score. Our study showcases the effectiveness of deep learning methodologies, specifically the LSTM model, in predicting frailty based on muscle force signals. Similar findings, across 16 public time series data (e.g., sleep monitoring, gait analysis, daily activity, and etc.), suggest that the LSTM model performs superior than conventional ML techniques like logistic regression and SVM [\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e]. This suggests the potential of deep learning algorithms to autonomously discern patterns and features directly from raw signals, eliminating the need for manual feature extraction and selection. The observed enhancement in model prediction may stem from the fact that shallow learning approaches are adept at handling simpler and more linear classification tasks [\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e], whereas LSTM models excel at learning complex, nonlinear functions and can effectively capture long-range dependencies [\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e]. The association observed between the UEF task and frailty suggests that integrating the UEF task with muscle model and LSTM methodologies could serve as a robust measure of frailty.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eLimitation and Future Direction\u003c/h2\u003e \u003cp\u003eIn this study, 312 participants were included, among which less than less than 14% were frail older adults. Although the small number of participants and unbalanced frailty group were, to some extent, remedied with augmentation, the sample size limitation still exists. The relatively small sample size and specific cohort of older adults recruited from outpatient services may limit the generalizability of our findings. Additionally, while our deep learning model showed promising results, further validation in larger and more diverse populations, especially those that are hospitalized at the time of assessment, is warranted to confirm UEF efficacy and reliability.\u003c/p\u003e \u003c/div\u003e"},{"header":"Conclusion","content":"\u003cp\u003eIn conclusion, our study contributes to the growing body of research on frailty assessment by introducing a novel approach that integrates computational modeling and deep learning techniques. By leveraging muscle force signals and advanced ML algorithms, we provide a comprehensive framework for frailty prediction that is objective, efficient, and clinically relevant. Future research should focus on validating our findings in larger cohorts and exploring the integration of additional physiological signals, such as heart rate and brain function, to enhance the accuracy and predictive power of frailty models.\u003c/p\u003e "},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eMehran Asghari contributed to the conceptualization, data analysis, and post-processing of the study. He also led the writing of the manuscript. Dr. Toosizadeh collected the data in his lab and provided critical review and guidance throughout the research process. Hossein Ehsani, proficient in machine learning, contributed by making significant improvements to the analysis and manuscript. All authors reviewed and approved the final version of the paper.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe database used for the current study is available upon reasonable request. Please contact the corresponding author, Dr. Nima Toosizadeh, at [email protected] for more information.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eBuckinx, F., et al., \u003cem\u003eBurden of frailty in the elderly population: perspectives for a public health challenge\u003c/em\u003e. Archives of public health, 2015. 73(1): p. 1\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHorak, F.B., C.L. Shupert, and A. Mirka, \u003cem\u003eComponents of postural dyscontrol in the elderly: a review\u003c/em\u003e. Neurobiology of aging, 1989. 10(6): p. 727\u0026ndash;738.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMitnitski, A., S.E. Howlett, and K. Rockwood, \u003cem\u003eHeterogeneity of human aging and its assessment\u003c/em\u003e. Journals of Gerontology Series A: Biomedical Sciences and Medical Sciences, 2017. 72(7): p. 877\u0026ndash;884.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eProietti, M. and M. Cesari, \u003cem\u003eFrailty: what is it?\u003c/em\u003e Frailty and Cardiovascular Diseases: Research into an Elderly Population, 2020: p. 1\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBoyd, C.M., et al., \u003cem\u003eFrailty, hospitalization, and progression of disability in a cohort of disabled older women\u003c/em\u003e. The American journal of medicine, 2005. 118(11): p. 1225\u0026ndash;1231.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDent, E., P. Kowal, and E.O. Hoogendijk, \u003cem\u003eFrailty measurement in research and clinical practice: a review\u003c/em\u003e. European journal of internal medicine, 2016. 31: p. 3\u0026ndash;10.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSacha, J., et al., \u003cem\u003eIs it time to begin a public campaign concerning frailty and pre-frailty? A review article\u003c/em\u003e. Frontiers in physiology, 2017. 8: p. 484.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMcMillan, G. and R. Hubbard, \u003cem\u003eFrailty in older inpatients: what physicians need to know\u003c/em\u003e. QJM: An International Journal of Medicine, 2012. 105(11): p. 1059\u0026ndash;1065.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMcRae, P.J., et al., \u003cem\u003eFrailty and geriatric syndromes in vascular surgical ward patients\u003c/em\u003e. Annals of vascular surgery, 2016. 35: p. 9\u0026ndash;18.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWalters, S., et al., \u003cem\u003eThe prevalence of frailty in patients admitted to hospital with vertebral fragility fractures\u003c/em\u003e. Current Rheumatology Reviews, 2016. 12(3): p. 244\u0026ndash;247.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWillcox, S.M., D.U. Himmelstein, and S. Woolhandler, \u003cem\u003eInappropriate drug prescribing for the community-dwelling elderly\u003c/em\u003e. Jama, 1994. 272(4): p. 292\u0026ndash;296.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDe Vries, N., et al., \u003cem\u003eOutcome instruments to measure frailty: a systematic review\u003c/em\u003e. Ageing research reviews, 2011. 10(1): p. 104\u0026ndash;114.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGoldstein, J.P., M.K. Andrew, and A. Travers, \u003cem\u003eFrailty in older adults using pre-hospital care and the emergency department: a narrative review\u003c/em\u003e. Canadian geriatrics journal: CGJ, 2012. 15(1): p. 16.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMohler, M.J., et al., \u003cem\u003eThe Frailty syndrome: clinical measurements and basic underpinnings in humans and animals\u003c/em\u003e. Experimental gerontology, 2014. 54: p. 6\u0026ndash;13.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFried, L.P., et al., \u003cem\u003eFrailty in older adults: evidence for a phenotype.\u003c/em\u003e The Journals of Gerontology Series A: Biological Sciences and Medical Sciences, 2001. 56(3): p. M146-M157.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRockwood, K., M. Andrew, and A. Mitnitski, \u003cem\u003eA comparison of two approaches to measuring frailty in elderly people\u003c/em\u003e. The Journals of Gerontology Series A: Biological Sciences and Medical Sciences, 2007. 62(7): p. 738\u0026ndash;743.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eToosizadeh, N., J. Mohler, and B. Najafi, \u003cem\u003eAssessing upper extremity motion: an innovative method to identify frailty\u003c/em\u003e. Journal of the American Geriatrics Society, 2015. 63(6): p. 1181\u0026ndash;1186.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eToosizadeh, N., et al., \u003cem\u003eAssessing upper-extremity motion: An innovative method to quantify functional capacity in patients with chronic obstructive pulmonary disease\u003c/em\u003e. PloS one, 2017. 12(2): p. e0172766.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eToosizadeh, N., et al., \u003cem\u003eFrailty assessment in older adults using upper-extremity function: index development\u003c/em\u003e. BMC geriatrics, 2017. 17: p. 1\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAsghari, M., et al., \u003cem\u003eA computational musculoskeletal arm model for assessing muscle dysfunction in chronic obstructive pulmonary disease\u003c/em\u003e. Medical \u0026amp; Biological Engineering \u0026amp; Computing, 2023: p. 1\u0026ndash;14.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAsghari, M., \u003cem\u003eA Novel Sensor-Based Approach to Predict Adverse Outcomes in Chronic Obstructive Pulmonary Disease\u003c/em\u003e. 2023, The University of Arizona.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAsghari, M., et al., \u003cem\u003eNonlinear analysis of the movement variability structure can detect aging-related differences among cognitively healthy individuals\u003c/em\u003e. Human movement science, 2021. 78: p. 102807.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAponte-Hao, S., et al., \u003cem\u003eMachine learning for identification of frailty in Canadian primary care practices\u003c/em\u003e. International journal of population data science, 2021. 6(1).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eArshad, M.Z., et al. \u003cem\u003eGait-based frailty assessment using image representation of IMU signals and deep CNN\u003c/em\u003e. in 2021 \u003cem\u003e43rd Annual International Conference of the IEEE Engineering in Medicine \u0026amp; Biology Society (EMBC)\u003c/em\u003e. 2021. IEEE.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFriedrich, B., et al. \u003cem\u003eA deep learning approach for TUG and SPPB score prediction of (pre-) frail older adults on real-life IMU data\u003c/em\u003e. in \u003cem\u003eHealthcare\u003c/em\u003e. 2021. MDPI.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEskandari, M., et al., \u003cem\u003eFrailty identification using heart rate dynamics: A deep learning approach\u003c/em\u003e. IEEE Journal of Biomedical and Health Informatics, 2022. 26(7): p. 3409\u0026ndash;3417.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIdris, S. and N. Badruddin. \u003cem\u003eClassification of cognitive frailty in elderly people from blood samples using machine learning\u003c/em\u003e. in 2021 \u003cem\u003eIEEE EMBS International Conference on Biomedical and Health Informatics (BHI)\u003c/em\u003e. 2021. IEEE.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRadha, M., et al., \u003cem\u003eSleep stage classification from heart-rate variability using long short-term memory neural networks\u003c/em\u003e. Scientific reports, 2019. 9(1): p. 14149.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang, L. and X. Zhou, \u003cem\u003eDetection of congestive heart failure based on LSTM-based deep network via short-term RR intervals\u003c/em\u003e. Sensors, 2019. 19(7): p. 1502.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eThirugnanam, M. and M.S. Pasupuleti, \u003cem\u003eCardiomyopathy-induced arrhythmia classification and pre-fall alert generation using convolutional neural network and long short-term memory model\u003c/em\u003e. Evolutionary Intelligence, 2021. 14: p. 789\u0026ndash;799.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShrestha, B. and L. Dunn, \u003cem\u003eThe declaration of Helsinki on medical research involving human subjects: a review of seventh revision\u003c/em\u003e. Journal of Nepal Health Research Council, 2019. 17(4): p. 548\u0026ndash;552.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAustria, Y.D., et al., \u003cem\u003eComparison of machine learning algorithms in breast cancer prediction using the coimbra dataset\u003c/em\u003e. cancer, 2019. 7(10): p. 23\u0026ndash;1.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDayton, C.M., \u003cem\u003eLogistic regression analysis\u003c/em\u003e. Stat, 1992. 474: p. 574.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSen, P.C., M. Hajra, and M. Ghosh. \u003cem\u003eSupervised classification algorithms in machine learning: A survey and review\u003c/em\u003e. in \u003cem\u003eEmerging Technology in Modelling and Graphics: Proceedings of IEM Graph 2018\u003c/em\u003e. 2020. Springer.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNg, A.Y. \u003cem\u003eFeature selection, L 1 vs. L 2 regularization, and rotational invariance\u003c/em\u003e. in \u003cem\u003eProceedings of the twenty-first international conference on Machine learning\u003c/em\u003e. 2004.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDiPietro, R. and G.D. Hager, \u003cem\u003eDeep learning: RNNs and LSTM\u003c/em\u003e, in \u003cem\u003eHandbook of medical image computing and computer assisted intervention\u003c/em\u003e. 2020, Elsevier. p. 503\u0026ndash;519.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAsghari, M., et al., \u003cem\u003eA computational musculoskeletal arm model for assessing muscle dysfunction in chronic obstructive pulmonary disease\u003c/em\u003e. Medical \u0026amp; Biological Engineering \u0026amp; Computing, 2023. 61(9): p. 2241\u0026ndash;2254.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKim, J.C., K. Kalantar-Zadeh, and J.D. Kopple, \u003cem\u003eFrailty and protein-energy wasting in elderly patients with end stage kidney disease\u003c/em\u003e. Journal of the American Society of Nephrology, 2013. 24(3): p. 337\u0026ndash;351.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePourhassan, M., et al., \u003cem\u003eThe impact of malnutrition on acute muscle wasting in frail older hospitalized patients\u003c/em\u003e. Nutrients, 2020. 12(5): p. 1387.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVanItallie, T.B., \u003cem\u003eFrailty in the elderly: contributions of sarcopenia and visceral protein depletion\u003c/em\u003e. Metabolism, 2003. 52: p. 22\u0026ndash;26.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eW\u0026uuml;st, R.C. and H. Degens, \u003cem\u003eFactors contributing to muscle wasting and dysfunction in COPD patients\u003c/em\u003e. International journal of chronic obstructive pulmonary disease, 2007. 2(3): p. 289.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMakizako, H., et al., \u003cem\u003eAssociations of social frailty with loss of muscle mass and muscle weakness among community-dwelling older adults\u003c/em\u003e. Geriatrics \u0026amp; gerontology international, 2019. 19(1): p. 76\u0026ndash;80.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHirokawa, S., et al., \u003cem\u003eMuscular co-contraction and control of knee stability\u003c/em\u003e. Journal of Electromyography and Kinesiology, 1991. 1(3): p. 199\u0026ndash;208.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBusse, M., C.M. Wiles, and R.W.M. Van Deursen, \u003cem\u003eMuscle co-activation in neurological conditions\u003c/em\u003e. Physical therapy reviews, 2005. 10(4): p. 247\u0026ndash;253.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePotvin, J. and P. O'brien, \u003cem\u003eTrunk muscle co-contraction increases during fatiguing, isometric, lateral bend exertions: possible implications for spine stability\u003c/em\u003e. Spine, 1998. 23(7): p. 774\u0026ndash;780.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAmbagtsheer, R., et al., \u003cem\u003eThe application of artificial intelligence (AI) techniques to identify frailty within a residential aged care administrative data set\u003c/em\u003e. International journal of medical informatics, 2020. 136: p. 104094.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAbbasimehr, H. and R. Paki, \u003cem\u003eImproving time series forecasting using LSTM and attention models\u003c/em\u003e. Journal of Ambient Intelligence and Humanized Computing, 2022: p. 1\u0026ndash;19.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAllam, A., et al., \u003cem\u003eNeural networks versus Logistic regression for 30 days all-cause readmission prediction\u003c/em\u003e. Scientific reports, 2019. 9(1): p. 9277.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eG\u0026uuml;ney, S. and \u0026Ccedil;.B. Erdaş. \u003cem\u003eA deep LSTM approach for activity recognition\u003c/em\u003e. in 2019 \u003cem\u003e42nd International Conference on Telecommunications and Signal Processing (TSP)\u003c/em\u003e. 2019. IEEE.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Frailty assessment, Long short-term memory (LSTM), Deep learning, Muscle co-contraction","lastPublishedDoi":"10.21203/rs.3.rs-4458153/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-4458153/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe global increase in the older adult population highlights the need for effective frailty assessment, a condition linked to adverse health outcomes such as hospitalization and mortality. Existing frailty assessment tools, like the Fried phenotype and Rockwood score, have practical limitations, necessitating a more efficient approach. This study aims to enhance frailty prediction accuracy in older adults using a combined biomechanical and deep learning approach. We recruited 312 participants (126 non-frail, 145 pre-frail, 41 frail) and assessed frailty using the Fried index, upper-extremity function (UEF) test, and muscle force calculations.\u003c/p\u003e \u003cp\u003eMachine learning (ML) models, including logistic regression and support vector machine (SVM), were employed alongside deep learning with long short-term memory (LSTM) networks. Results showed that incorporating muscle model parameters significantly improved frailty prediction. The LSTM model achieved the highest accuracy (74%), outperforming SVM (67%) and regression (66%), with precision and F1 scores of 81% and 75%, respectively. Notably, muscle co-contraction emerged as a critical predictor, with frail individuals exhibiting substantially higher levels.\u003c/p\u003e \u003cp\u003eOur findings demonstrate that integrating UEF tasks with deep learning models provides superior frailty prediction, potentially offering a robust, efficient clinical tool. However, further validation with larger, more diverse populations is needed to confirm the generalizability of our results. This study underscores the potential of advanced computational techniques to improve the identification and management of frailty in older adults.\u003c/p\u003e","manuscriptTitle":"Frailty Identification using a Sensor-based Upper- extremity Function Test: A Deep Learning Approach","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-06-18 16:37:00","doi":"10.21203/rs.3.rs-4458153/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2024-08-16T06:35:37+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2024-08-12T09:40:54+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"257140274355574116203580767719533199011","date":"2024-08-02T19:28:22+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2024-06-27T15:01:35+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"58347910876976144076762976725637220913","date":"2024-06-17T12:51:58+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2024-06-17T12:49:03+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2024-06-17T12:45:01+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2024-06-04T11:04:59+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2024-05-31T10:29:43+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2024-05-22T04:33:45+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"ffdf3f29-6850-46cf-acb5-c73b747d45c4","owner":[],"postedDate":"June 18th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[{"id":33120548,"name":"Biological sciences/Biological techniques"},{"id":33120549,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":33120550,"name":"Health sciences/Diseases"},{"id":33120551,"name":"Physical sciences/Engineering"},{"id":33120552,"name":"Physical sciences/Mathematics and computing"}],"tags":[],"updatedAt":"2025-04-28T16:06:49+00:00","versionOfRecord":{"articleIdentity":"rs-4458153","link":"https://doi.org/10.1038/s41598-024-73854-2","journal":{"identity":"scientific-reports","isVorOnly":false,"title":"Scientific Reports"},"publishedOn":"2025-04-22 15:58:06","publishedOnDateReadable":"April 22nd, 2025"},"versionCreatedAt":"2024-06-18 16:37:00","video":"","vorDoi":"10.1038/s41598-024-73854-2","vorDoiUrl":"https://doi.org/10.1038/s41598-024-73854-2","workflowStages":[]},"version":"v1","identity":"rs-4458153","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-4458153","identity":"rs-4458153","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}