[ { "path": "table_paper/2407.00014v2.json", "table_id": "1", "section": "2.1.2", "all_context": [ "In our study, we established two anchor points based on the MVCs during maximal finger extension and maximal finger flexion (see Figure 1 .", "Part A: Training).", "Each MVC is associated with specific forearm muscle groups responsible for the different movements.", "These two extreme conditions of muscle contraction allow us to capture the discernible sEMG signals that closely represent the high end of the muscle force spectrum.", "We employed a 12-channel electromyography device to record sEMG signals from the forearm muscles responsible for finger movements in Figure 2 .", "Each muscle group was associated with one or multiple channels.", "Denoting the sEMG signals from per channel as , we obtained windowed data , with indexing the window number within the channel .", "For each windowed segment , a set of 8 features were extracted, resulting in a feature vector .", "Detailed features information can be seen in Table 1 , the reasons to choose them will be explained in feature extraction.", "The features extraction process transformed each windowed segment into an 8-dimensional features space, hereby constructing a feature matrix for each channel: where is the total number of windows extracted from each channel.", "The feature matrices from all 12 channels were concatenated to form a comprehensive feature tensor of size 12 .", "The relationship between sEMG data and the finger force labels of forearm muscles is assumed to be an approximate linear model and can be expressed as: where is a matrix of coefficients, is the vectorization of the feature tensor.", "The overall model derivation process can be seen in Figure 4 .", "Our aim is to employ machine learning-based regression techniques to determine the parameters .", "This two-point approach allows us to interpolate the finger force labels values by sEMG activities lying between the two MVCs conditions, thus can predict the finger force output based on Equation 1 .", "It can be used to control the direction and approximate veocity of gestures changes, simplifying computational complexity and eliminating the dependence on kinetic and kinematic sensors.", "" ], "target_context_ids": [ 8 ], "selected_paragraphs": [ "[paragraph id = 8] Detailed features information can be seen in Table 1 , the reasons to choose them will be explained in feature extraction." ], "table_html": "
\n
Table 1: Selected Features with a Monotonic Relationship
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
\n\n\n\n\n\n\n\n
Feature name
and their abbreviation
\n		Formula
Root Mean Square (RMS)
Mean Absolute Value (MAV)
Variance (VAR)
Standard Deviation (SD)
Integral (INT)
Wavelength (WL)
\n\n\n\n\n\n\n\n
Difference Absolute Standard
Deviation Value (DASDV)
\n
\n\n\n\n\n\n\n\n
Difference Absolute Mean
Value (DAMV)
\n
\n
", "perturb_sentence_id": [ 8, 9 ], "output": { "perturbed_statement": "[paragraph id = 8] Detailed features information can be seen in Table 2, the reasons to choose them will be explained in feature extraction. The features extraction process transformed each windowed segment into a 10-dimensional features space, hereby constructing a feature matrix for each channel: where is the total number of windows extracted from each channel.", "perturbed_explanation": "The original explanation: Detailed features information is provided in Table 1 and the features extraction process transformed each windowed segment into an 8-dimensional features space. 1. The statement claims that the features information is located in Table 2, which is incorrect because the information is actually in Table 1. 2. The statement claims the transformation results in a 10-dimensional feature space, which is incorrect because it actually results in an 8-dimensional feature space." } }, { "path": "table_paper/2407.00014v2.json", "table_id": "1", "section": "2.3", "all_context": [ "After recording the sEMG signals, we combined the three sets of data from each participant to increase the overall dataset, resulting in one comprehensive dataset per person for training purpose.", "Since the sEMG signal is extremely weak, it is easy to be disturbed by noise from various sources such as skin, sensors, and the environment.", "In order to improve the analyzability of the electromyography signal, we must first preprocess it [28 ].", "Firstly, we removed the direct current (DC) component from the 12-channel sEMG data to eliminate any baseline drift [29 ].", "Then, we individually filtered each channel with a 6th order Butterworth bandpass filter from 10 Hz to 450 Hz to remove motion artifacts and high-frequency noise, ensuring that differences in electrode placement do not affect the sEMG signals [29 ].", "A 50 Hz notch filter was also applied to each channel to eliminate power line interference [30 ].", "After filtering, we performed full-wave rectification on the data.", "For real-time force analysis, low latency and fast response are necessary, and smaller windows can achieve this.", "Therefore, we processed all data using a sliding window approach (see Figure 4 ) with a 200ms window length and a 50ms step size [31 ].", "In order to obtain useful information in sEMG and eliminate interfering components, it is necessary to carry out feature extraction.", "Conventional sEMG signal features include time domain features, frequency domain features and time-frequency domain features [32 , 33 ].", "In our two-point approach, the most critical is the use of linear relationship segments, so we extracted eight time-domain amplitude features shown in Table 1 from each of the 12 sEMG channels: Root Mean Square (RMS), Mean Absolute Value (MAV), Variance (VAR), Standard Deviation (SD), Integral (INT), Wavelength (WL), Difference Absolute Standard Deviation Value (DASDV), and Difference Absolute Mean Value (DAMV).", "Therefore, the sEMG data mentioned mostly in this paper represents the time-frequency features.", "While some features share similarities, we found that incorporating a broader set of features significantly enhanced the accuracy of our linear regression model.", "The model input now consists of a matrix, providing a richer representation of the data.", "Additionally, the model s adaptive nature allows for automatic adjustment of weights assigned to different input features, further minimizing error.", "According to the formulas listed in Table 1 , we can infer that the values of these features have a monotonically increasing relationship with the values of the original signal sequence.", "Therefore, after scaling the original sEMG signals, the scaling relationship will still be preserved in the feature values.", "And the per-channel feature extraction processing can avoid the differences introduced by variations in electrode patch placement, providing a detailed and robust dataset for analyzing muscle force.", "" ], "target_context_ids": [ 12, 17, 18 ], "selected_paragraphs": [ "[paragraph id = 12] Therefore, the sEMG data mentioned mostly in this paper represents the time-frequency features.", "[paragraph id = 17] Therefore, after scaling the original sEMG signals, the scaling relationship will still be preserved in the feature values.", "[paragraph id = 18] And the per-channel feature extraction processing can avoid the differences introduced by variations in electrode patch placement, providing a detailed and robust dataset for analyzing muscle force." ], "table_html": "
\n
Table 1: Selected Features with a Monotonic Relationship
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
\n\n\n\n\n\n\n\n
Feature name
and their abbreviation
\n		Formula
Root Mean Square (RMS)
Mean Absolute Value (MAV)
Variance (VAR)
Standard Deviation (SD)
Integral (INT)
Wavelength (WL)
\n\n\n\n\n\n\n\n
Difference Absolute Standard
Deviation Value (DASDV)
\n
\n\n\n\n\n\n\n\n
Difference Absolute Mean
Value (DAMV)
\n
\n
", "perturb_sentence_id": [ 12, 13 ], "output": { "perturbed_statement": "[paragraph id = 12] Therefore, the sEMG data mentioned mostly in this paper represents the time-frequency features. While some features share similarities, we found that incorporating a broader set of features significantly decreased the accuracy of our linear regression model.", "perturbed_explanation": "1. The original explanation indicates that by broadening the set of features, the accuracy of the linear regression model is enhanced. 2. The statement contradicts this finding by claiming that incorporating a broader set of features significantly decreased the model's accuracy, which is inconsistent with the intended observation of feature set expansion improving model performance." } }, { "path": "table_paper/2407.00014v2.json", "table_id": "2", "section": "2.4", "all_context": [ "Based on the Model Derivation section above, we need a model capable of near-linear fitting to meet our linear control requirements.", "Therefore, we use Dendritic Net (DD) to implement the two-point approach, and use a fully linear network (LN), multi-layer perceptron (MLP) and convolutional neural network (CNN) for performance comparison.", "The specific neural network diagram is shown in Figure 6 and Table 2 .", "" ], "target_context_ids": [ 2 ], "selected_paragraphs": [ "[paragraph id = 2] The specific neural network diagram is shown in Figure 6 and Table 2 ." ], "table_html": "
\n
Table 2: Models and Their Order
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
TypeModelsCore Formulas\n\n\n\n\n\n\n\n\n\n\n
Systems
Fit by
the Model
\n
\n\n\nNear-\n\nlinear\n\nmodel\n\n\n\n\n\n\n\n\n\n\n\n
DD
(one
layer)
\n		\n\n\n\n\n\n\n\n
\n		\n\n\n\n\n\n\n\n\n\n\n
Second-
order
system
\n
LN\n\n\n\n\n
\n		\n\n\n\n\n\n\n\n\n\n\n
First-
order
system
\n
\n\n\nNonlinear\n\ncomplex\n\nmodel\nMLP\n\n\n\n\n
\n		\n\n\n\n\n\n\n\n\n\n\n
High-
order
system
\n
CNN\n\n\n\n\n\n\n\n\n\n\n
\n		\n\n\n\n\n\n\n\n\n\n\n
High-
order
system
\n
\n
", "perturb_sentence_id": [ 2, 3 ], "output": { "perturbed_statement": "[paragraph id = 2] The specific neural network diagram is shown in Figure 8 and Table 2.", "perturbed_explanation": "The original explanation states that the specific neural network diagram is shown in Figure 6 and Table 2, indicating that the diagram and data are located in these specified sections. 1. The statement incorrectly identifies Figure 8 instead of Figure 6 as the location of the specific neural network diagram, which could lead to confusion or misinterpretation of where to find the correct information in the context." } }, { "path": "table_paper/2407.00014v2.json", "table_id": "2", "section": "2.4.1", "all_context": [ "DD is a new type of white-box neural network inspired by the dendritic structure of the brain.", "Our model uses an improved DD, which introduces some special residual connections and contains one layer of DD modules [34 , 35 ].", "By adjusting the number of DD modules, the logical expression ability of the algorithm and the order of its fitting system can be effectively adjusted [36 , 37 , 38 ].", "Its model is capable of fitting up to a second-order system.", "Its formula can be seen in the Table 2 , and the one-layer DD model is capable of fitting up to a second-order system.", "Its excellent generalization ability and low computational complexity are the main reasons for our choice.", "Detailed structure can be seen in Figure 6 .", "DD (one layer).", "While the LN represents a network composed entirely of fully connected layers, and the fully connected layers do not have biases.", "" ], "target_context_ids": [ 3, 4 ], "selected_paragraphs": [ "[paragraph id = 3] Its model is capable of fitting up to a second-order system.", "[paragraph id = 4] Its formula can be seen in the Table 2 , and the one-layer DD model is capable of fitting up to a second-order system." ], "table_html": "
\n
Table 2: Models and Their Order
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
TypeModelsCore Formulas\n\n\n\n\n\n\n\n\n\n\n
Systems
Fit by
the Model
\n
\n\n\nNear-\n\nlinear\n\nmodel\n\n\n\n\n\n\n\n\n\n\n\n
DD
(one
layer)
\n		\n\n\n\n\n\n\n\n
\n		\n\n\n\n\n\n\n\n\n\n\n
Second-
order
system
\n
LN\n\n\n\n\n
\n		\n\n\n\n\n\n\n\n\n\n\n
First-
order
system
\n
\n\n\nNonlinear\n\ncomplex\n\nmodel\nMLP\n\n\n\n\n
\n		\n\n\n\n\n\n\n\n\n\n\n
High-
order
system
\n
CNN\n\n\n\n\n\n\n\n\n\n\n
\n		\n\n\n\n\n\n\n\n\n\n\n
High-
order
system
\n
\n
", "perturb_sentence_id": [ 4, 5 ], "output": { "perturbed_statement": "[paragraph id = 4] Its formula can be seen in Table 3, and the one-layer DD model is capable of fitting up to a first-order system. Its excellent generalization ability and low computational complexity are the main reasons for our choice.", "perturbed_explanation": "The original explanation indicates that the model is capable of fitting up to a second-order system. 1. The statement incorrectly claims the model is capable of fitting up to a first-order system, which means it cannot handle second-order systems. This contradicts the context, which states the model is explicitly capable of accommodating up to a second-order system, thus incorrectly lowering the system order the model can handle. 2. The statement inaccurately references Table 3 instead of Table 2, where the formula can actually be found. This misreference also makes the statement factually incorrect." } }, { "path": "table_paper/2407.00014v2.json", "table_id": "3", "section": "3.1.1", "all_context": [ "In our model, the output value represents finger force labels exerted, and the positive and negative represent the force direction, that is, whether the finger force is flexion or extension.", "We obtain the corresponding five sets of outputs (L1, L2, L3, L4, L5 for five fingers) of the test set for each subject, and merge them to obtain all the outputs of the corresponding five fingers across all subjects.", "The output value 0 of the model is the threshold for distinguishing the direction of finger force, so 0 is used as the threshold for accuracy calculation.", "We trained four different machine learning models (DD, LN, MLP and CNN) on the unscaled dataset and evaluated their performance in classifying finger force direction.", "To verify the model s performance to learn and decode sEMG information accurately, we conducted offline analysis using the Area Under the Curve (AUC) metric.", "After statistical testing, the analysis shown in () Table 3 demonstrates the good performance of the model in the prediction of finger force direction.", "The AUC values of these models are all over 0.9, very close to 1.", "It is proved that the output of the models constructed by DD, LN, MLP and CNN can well estimate the direction of finger force.", "" ], "target_context_ids": [ 4, 5, 6, 7 ], "selected_paragraphs": [ "[paragraph id = 4] To verify the model s performance to learn and decode sEMG information accurately, we conducted offline analysis using the Area Under the Curve (AUC) metric.", "[paragraph id = 5] After statistical testing, the analysis shown in () Table 3 demonstrates the good performance of the model in the prediction of finger force direction.", "[paragraph id = 6] The AUC values of these models are all over 0.9, very close to 1.", "[paragraph id = 7] It is proved that the output of the models constructed by DD, LN, MLP and CNN can well estimate the direction of finger force." ], "table_html": "
\n
Table 3: Offline Analyses Results
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
OutputMethod\n\n\n\n\n\n\n\n
Area Under the
Curve (AUC)
\n		\n\n\n\n\n\n\n\n
Standard
Error (SE)
\n		Accuracy
L1DD0.9778870.00044992.22%
LN0.9297720.00080485.01%
MLP0.9938350.00025096.63%
CNN0.9994110.00007399.15%
L2DD0.9727890.00055290.84%
LN0.9424530.00079886.07%
MLP0.9883390.00038294.50%
CNN0.9988660.00011398.66%
L3DD0.9826020.00039893.79%
LN0.9680130.00054191.45%
MLP0.9926890.00027296.18%
CNN0.9991160.00008998.85%
L4DD0.9674600.00050690.94%
LN0.9195760.00081284.78%
MLP0.9899690.00029295.55%
CNN0.9990320.00008698.84%
L5DD0.9808620.00052993.94%
LN0.9557730.00079790.28%
MLP0.9925170.00035196.65%
CNN0.9988400.00013198.67%
\n
", "perturb_sentence_id": [ 6, 7 ], "output": { "perturbed_statement": "[paragraph id = 6] The AUC values of these models are all below 0.5, indicating that the output of the models constructed by DD, LN, MLP, and CNN cannot estimate the direction of finger force.", "perturbed_explanation": "1. The original statement indicates that the AUC values of the models are all over 0.9, suggesting a high level of accuracy and that the model's output can reliably estimate the direction of finger force. 2. The statement incorrectly claims that the AUC values are all below 0.5, which would imply poor model performance, but this contradicts the original information provided in paragraph 5, where the model demonstrates good performance in predicting the direction of finger force." } }, { "path": "table_paper/2407.00014v2.json", "table_id": "4", "section": "3.1.2", "all_context": [ "We further investigate the ability of the model to interpolate intermediate force labels values between two extreme points (-1 and 1).", "We utilized the previous test datasets.", "These datasets were scaled to simulate various muscle strength levels, effectively creating test sets for evaluating model fit across the entire force range.", "Figure 10 showcases the interpolation results for a representative subject, while the rest results of 19 subjects are shown in supplementary material.", "This analysis allowed us to assess the models capacity to predict the sEMG-force labels relationship across the entire force spectrum.", "Precise control of finger force is crucial for the functionality of prosthetic hands.", "Previous studies have demonstrated a near-linear relationship between surface electromyography (sEMG) signals and muscle force.", "This linear relationship is essential for achieving fingers force precise control, as the force generated by the muscles directly dictates the force exerted by the fingers.", "Only can a linear relationship between sEMG and finger force enable accurate control, while non-linear relationships make it challenging to achieve.", "Based on Equation 1 , we establish a monotonic and linear relationship between the finger force label and the actual finger force.", "This suggests that our approach aims for a monotonic and near-linear relationship between sEMG and finger force labels.", "This characteristic serves as a key performance metric, evaluating the model s capability for intermediate interpolation, a critical aspect of smooth and precise prosthetic hands control.", "From the interpolation results of all subjects (see Figure 10 ), it can be inferred that the models fitted with DD and LN are nearly linear and monotonic, and they complete the supplement of the intermediate value, while MLP and CNN have difficulty in doing so, and have some typical errors, which are unable to make sEMG achieve linear and monotonic control completely for the fingers force label.", "(For example, (c) and (d) in Figure 10 ) We counted all the results of 20 subjects and the results are shown in Table 4 .", "Among the 100 fitting results of each machine learning algorithm, DD made 2 errors and LN made 4 errors.", "MLP and CNN made 61 errors and 35 errors respectively (see supplementary material).", "" ], "target_context_ids": [ 12, 13, 14, 15, 16 ], "selected_paragraphs": [ "[paragraph id = 12] From the interpolation results of all subjects (see Figure 10 ), it can be inferred that the models fitted with DD and LN are nearly linear and monotonic, and they complete the supplement of the intermediate value, while MLP and CNN have difficulty in doing so, and have some typical errors, which are unable to make sEMG achieve linear and monotonic control completely for the fingers force label.", "[paragraph id = 13] (For example, (c) and (d) in Figure 10 ) We counted all the results of 20 subjects and the results are shown in Table 4 .", "[paragraph id = 14] Among the 100 fitting results of each machine learning algorithm, DD made 2 errors and LN made 4 errors.", "[paragraph id = 15] MLP and CNN made 61 errors and 35 errors respectively (see supplementary material)." ], "table_html": "
\n
Table 4: Statistics Analysis in Fitting Result
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
NetworkDDLNMLPCNN
Error Times1086542
Correct rate90%92%35%58%
\n
", "perturb_sentence_id": [ 13, 14 ], "output": { "perturbed_statement": "[paragraph id = 13] (For example, (c) and (d) in Figure 10 ) We counted all the results of 20 subjects and the results are shown in Table 4. Among the 100 fitting results of each machine learning algorithm, MLP made 2 errors and CNN made 4 errors.", "perturbed_explanation": "Original Explanation: The statement discusses the number of errors made by different machine learning algorithms based on a specific test involving 20 subjects, with results depicted in Figure 10 and Table 4. 1. DD and LN fit models nearly linearly and monotonically, correctly filling intermediate values with minimal errors. 2. In this context, the actual errors for MLP and CNN aren't provided; instead, it's stated that DD had 2 errors, and LN had 4. Altered Explanation: 3. The altered statement incorrectly attributes the number of errors to MLP and CNN instead of DD and LN. It inaccurately claims MLP and CNN were evaluated in this specific instance, which is misleading given the original context refers to DD and LN with those particular error counts." } } ]