4 : Pinch! Test Of Learning
Mean absolute error (AE), constant error (CE), mean difference, intraclass correlation coefficient (ICC) with 95% confidence interval (CI), standard error of the measurement (SEM), and 95% limits of agreement (LOA) between the test and retest of the tip pinch at three force levels
4 : Pinch! Test of Learning
Mean absolute error (AE), constant error (CE), mean difference, intraclass correlation coefficient (ICC) with 95% confidence interval (CI), standard error of the measurement (SEM), and 95% limits of agreement (LOA) between the test and retest of the key pinch at three force levels
Mean absolute error (AE), constant error (CE), mean difference, intraclass correlation coefficient (ICC) with 95% confidence interval (CI), standard error of the measurement (SEM), and 95% limits of agreement (LOA) between the test and retest of the palmar pinch at three force levels
To test for the presence of rigidity, we need to passively manipulate the limbs of the patient. However, If the disease is in its early stage or the symptoms are well controlled with medications, we may not be able to see rigidity. We will need to use some activation maneuvers, that basically consist in performing repetitive movements with the limb contralateral to the one that is being tested.
Action or kinetic tremor is a type of tremor that is uncovered only when the patient is carrying out a movement. To test for kinetic tremor we can use the finger to nose test. In performing this test, the patients are instructed to alternatively touch their nose and our finger. In doing so, the patients should stretch their arm completely and should not move too fast. In this way we have more chances of triggering the tremor.
There is a different variant of the finger to nose test, in which the finger of the examiner changes position every time that the patient tries to reach it. This variant of the test is used to test for ataxia, rather than for tremor.
In this test the examiner stands behind the patients and by pulling on their shoulder tries to make them fall backwards. If the patients are able to correct their center of gravity in just one or two steps, the test is negative for a balance abnormalities. The test is instead considered positive if the patients catch their balance in more than two steps or if they do not stabilize at all and tend to fall to the ground.
Hill-type muscle models are widely employed in simulations of human movement. Yet, the parameters underlying these models are difficult or impossible to measure in vivo. Prior studies demonstrate that Hill-type muscle parameters are encoded within dynamometric data. But, a generalizable approach for estimating these parameters from dynamometric data has not been realized. We aimed to leverage musculoskeletal models and artificial neural networks to classify one Hill-type muscle parameter (maximum isometric force) from easily measurable dynamometric data (simulated lateral pinch force). We tested two neural networks (feedforward and long short-term memory) to identify if accounting for dynamic behavior improved accuracy.
We generated four datasets via forward dynamics, each with increasing complexity from adjustments to more muscles. Simulations were grouped and evaluated to show how varying the maximum isometric force of thumb muscles affects lateral pinch force. Both neural networks classified these groups from lateral pinch force alone.
Here, we examine to what extent artificial neural networks can be used to predict underlying muscle parameters from easily measurable datasets. Artificial neural networks are a machine learning method and have multiple advantages making them attractive for complex biomechanical analyses. First, artificial neural networks can approximate complex nonlinear mappings, which are common to the musculoskeletal system. One example of this nonlinearity is displayed by the work of Pearlman et al. [20], which identified thumb-tip force to be a nonlinear function of muscle force in cadaveric specimens. Second, artificial neural networks can infer unseen relationships, such as the mapping between Hill-type muscle parameters and the joint movements they influence. Lastly, after training, artificial neural networks are computationally efficient and rapidly perform time-series classification [21], which can enhance analysis of dynamic data (e.g., joint angle trajectories, joint torques, and/or external forces versus time).
To examine the impact of varying the maximum isometric force of thumb muscles, we generated four datasets of lateral pinch. Dynamic lateral pinch simulations were produced via forward dynamics in OpenSim v. 4.1 [28]. Feedforward and LSTM neural network models were tested to predict the maximum isometric force of the muscle actuators varied using only dynamic thumb-tip force. The performance of each neural network model was quantified as test losses and accuracies using a 5-fold cross-validation process. We evaluated overall performance in predicting maximum isometric force as well as relative performance between the two neural network models.
(a) Lateral pinch model used to generate dynamic thumb-tip force datasets [29]. (b) Values 1 through 4 correspond to Datasets 1 through 4, which varied the thumb muscle actuators shown. Datasets 1 through 4 contained 120, 1024, 2197, and 4096 simulations, respectively, representing different combinations of maximum isometric force values. For example, Dataset 1 included 120 variations to the FPL, while Dataset 4 included 8 variations to each the FPL, APL, ADPt, and ADPo.
To enable isolated variation of the maximum isometric force of thumb muscles, we used forward dynamics to simulate lateral pinch (Fig 2a). Muscle activations calculated via computed muscle control served as inputs to the forward dynamic simulations. These activations were calculated to produce increasing thumb-tip force from 0 to 36.4 N magnitude (35 N in the palmar direction and 10 N in the ulnar direction) across 1.5 seconds while maintaining a target thumb posture (-15 carpometacarpal flexion, -20 carpometacarpal abduction, 20 metacarpophalangeal flexion, 40 interphalangeal flexion). The described target forces are both sufficient for many activities of daily living [22] and activated the muscles of interest. These activations were held constant across all simulations; thus, observed changes in output represent isolated changes in maximum isometric muscle force. The output of each forward dynamic simulation was a time-varying, three-component vector describing dynamic changes in thumb-tip force.
(a) Workflow for producing simulated lateral pinch data. Muscle activations were attained via computed muscle control. (b) Workflow for training and testing artificial neural networks on simulated thumb-tip forces and time.
We generated four datasets (Fig 1b) to examine the impact of varying the maximum isometric force input of thumb muscles. Each dataset increased in complexity by including changes to the maximum isometric force for an increasing number of muscles. We first altered the maximum isometric force of the FPL (Dataset 1) and then added variations in APL (Dataset 2), ADPt (Dataset 3), and ADPo (Dataset 4) in order. This means that Dataset 1 varied only one muscle, whereas Dataset 4 included variations across all four. The FPL and APL were varied first as these muscles contribute the most to lateral pinch thumb-tip force [20] and importantly contribute to different directional components of thumb-tip force (the FPL flexes, the APL abducts). Within each dataset, the maximum isometric force values used for each muscle were uniformly sampled within the approximated range (Table 1). Dataset 1 included 120 simulations corresponding to uniform sampling across the full range of maximum isometric force values of the FPL. Datasets 2, 3, and 4 included 1024, 2197, and 4096 simulations, respectively. Importantly, these datasets scale substantially in size as needed to adequately train both neural network models for use in increasingly complex classifications. With the exception of Dataset 1, where the size was selected to not over-sample the isometric force space, dataset sizes were selected to be approximately double for each sequential dataset, while still uniformly sampling the maximum isometric forces for each muscle. For example, Dataset 4 incorporated 4096 simulations corresponding to all combinations of 8 maximum isometric force values for each of the four muscles (i.e., all combinations of maximum isometric force values = 84 or 4096 simulations).
We also analyzed the distribution of test accuracies and losses associated with the final neural network models as evaluated via 5-fold cross-validation. For a well-performing model, accuracies should generally increase and losses should decrease across epochs until stabilizing at their final values. Training was terminated after 100 epochs, as the losses and accuracies were stable, indicating solution convergence. We calculated 95% confidence intervals for the test accuracies and losses for each neural network model across epochs. To test whether accounting for the dynamic behavior of the simulation increases neural network accuracy, we performed two-sample t-tests (p
Both neural network models achieved comparably high accuracies relative to random guess (Fig 5). The LSTM model trained from Dataset 4 achieved the lowest accuracies, but still reached 27.9% accuracy as compared to 6.25% for a random guess. Comparing the models via two-sample t-tests revealed the analysis of Dataset 2 and Dataset 3 to produce significantly higher peak accuracies (p
Leveraging artificial neural networks and musculoskeletal models, we predicted changes in difficult-to-measure muscle parameters from minimal, measurable data. We specifically quantified feedforward and LSTM neural network performance in predicting the maximum isometric force of thumb muscles from simulated lateral pinch data. We report two key findings: 1. Under relatively simple conditions, neural network models can predict high and low maximum isometric force from lateral pinch data and 2. Accounting for long-term temporal dependencies in simulated lateral pinch data did not significantly improve neural network model performance. Although we did not predict explicit values of maximum isometric force in this study, accurately predicting changes in this parameter is a critical step that demonstrates the feasibility of using artificial neural networks to perform this task. With considerations for task complexity, dataset size, and Hill-type parameter of interest, the foundational work developed here could be expanded to define specific values of Hill-type parameters through categorization or regression. 041b061a72