Classification of Fine-ADL Using sEMG Signals Under Different
Measurement Conditions
Surya Naidu
a
, Anish Turlapaty
b
and Vidya Sagar
c
Biosignal Analysis Group, IIIT Sri City, Sri City, India
Keywords:
sEMG, Fine-ADL, Feature Extraction, Neural Networks, CNN Bi-LSTM, Class-Wise Analysis.
Abstract:
Most studies on surface electromyography (sEMG) related to finger activities have concentrated on grips,
grasps and general arm movements without any emphasis on the correlation of body postures and hand po-
sitions on the finger-centric activities. The main objective of the new dataset is to investigate activities of
daily living (ADL) needing focus on finer motor control in diverse measurement conditions. In this paper, we
present a novel dataset of finger-centric activities of daily living comprising 5-channel sEMG signals collected
under different body postures and hand positions. The dataset encompasses sEMG signals acquired from 10
subjects, performing 10 distinct fine-ADLs in various body postures and hand positions. Further, a machine
learning framework for classification of the fine-ADL is developed as follows. Each signal is segmented into
250ms windows and Time Domain (TD), Frequency Domain (FD), Wavelet Domain (WD) and Eigenvalues
features are extracted. Four classifier frameworks using the features are implemented for the analyses. The
results reveal that a hybrid CNN Bi-LSTM achieves an average test accuracy of 76.85%, followed by a 5-
layered fully connected neural network (FCNN) with 72.42%, in aggregate scenario. An average subject-wise
test accuracy of 88% is achieved by the FCNN across all body postures and hand positions combined. Most
importantly, the CNN Bi-LSTM, enhances subject-wise classification by an average test accuracy of 27.47%
than the FCNN under varying body postures. Dependencies of the test accuracy on measurement conditions
are analyzed. Stand body posture is found to be the easiest to classify in Aggregate scenario, whereas Folded
Knees was the most difficult to classify. An increase in test accuracy with an increase in training data is ob-
served body postures combinations analysis.
1 INTRODUCTION
1.1 Background
Activities of Daily Living (ADL) refer to the basic
tasks and activities that individuals perform on a daily
basis. ADL are particularly important for individuals
with disabilities or chronic illnesses who may require
assistance or accommodations to perform them. Fur-
thermore, insights gained from studying ADL can in-
form the design of assistive technologies to help indi-
viduals with disabilities perform daily tasks indepen-
dently.
The applications of ADL assessment are wide-
ranging and include geriatric care, rehabilitation, dis-
ability evaluation (Chen et al., 2022), and assistive
a
https://orcid.org/0009-0007-6390-8074
b
https://orcid.org/0000-0003-0078-3845
c
https://orcid.org/0009-0006-3464-9205
technology design. In geriatric care (Sandberg et al.,
2012), the ADL assessment can help identify func-
tional decline and enable healthcare providers to im-
plement interventions to maintain independence and
quality of life (Faria et al., 2020). In rehabilitation
care, it is important for establishing baselines, track-
ing progress (Dai et al., 2021), and developing effec-
tive treatment plans (He et al., 2021). In the design
of assistive technology, it can help ensure that devices
are tailored to the specific needs and abilities of a user
(Park et al., 2020).
We propose Fine-ADL as a class of activities of
daily living that require fine motor ability as described
in (Fauth et al., 2017) for precise control of the fingers
and wrists. Examples include ADL such as writing,
typing, and using standard mechanical tools such as
a kitchen knife. Assessment of Fine-ADL is impor-
tant in evaluation of the ADL score (Katz, 1983) and
various related applications. ADL are analyzed using
motion sensors, visual information and sEMG signals
Naidu, S., Turlapaty, A. and Sagar, V.
Classification of Fine-ADL Using sEMG Signals Under Different Measurement Conditions.
DOI: 10.5220/0012346000003657
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 17th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2024) - Volume 1, pages 591-598
ISBN: 978-989-758-688-0; ISSN: 2184-4305
Proceedings Copyright © 2024 by SCITEPRESS – Science and Technology Publications, Lda.
591

(Toledo-P
´
erez et al., 2019).
Surface electromyography (sEMG) is a non-
invasive measurement technique to record a muscle’s
electrical activity providing valuable information into
its contractions. Hence, this method finds applica-
tion in the examination of these finer activities. Pat-
tern recognition of sEMG signals is a promising ap-
proach for wearable robotic control. To realize their
widespread real-world adaptation, it is important to
develop control technologies that can perform Fine-
ADL for the users (Castellini et al., 2009). Toward
realization of this goal a fundamental requirement is
a pattern recognition framework that can reliably clas-
sify intended actions of a user based on multi-channel
sEMG signals.
One major challenge is the limited performance
of the pattern recognition algorithms under variable
measurement conditions different from the controlled
laboratory experiments. One reason for this limita-
tion is tasks performed in controlled circumstances
with limited trials and specific instructions may not
reflect the variability and complexity of Fine-ADL in
the everyday life. (Rosenburg and Seidel, 1989) ob-
served a significant correlation between sEMG sig-
nals and body postures. However, the underlying
mechanisms driving this relationship have remained
elusive. The observed variability has been attributed
to bio-mechanical factors such as lever arm of mus-
cles, force distribution across muscles, gravity and
others. Consequently, it is important to incorpo-
rate diverse measurement conditions when conduct-
ing sEMG studies.
There are a few efforts available in literature
studying the impact of measurement conditions,
specifically body and hand positions on classification
of ADL and other limb movements based on different
measurement modalities. In (Song and Kim, 2018),
a classification algorithm using a single inertial sen-
sor to categorize three fundamental gait activities was
proposed. The experiments in this study included
measurements both within a gait lab and in an out-
door walking course, allowing analysis under vary-
ing conditions. Another study (Williams et al., 2022)
explored control strategies for myoelectric prostheses
that incorporate position awareness. By considering
the positional information of the prosthetic hand, a
natural and intuitive control can be achieved. In (Yang
et al., 2015) and the references therein, the impact
of upper limb positions and dynamic movements on
classification of finger motions, which usually con-
tribute to Fine-ADL, is demonstrated. However in the
existing literature there is no study or sEMG dataset
with focus on Fine-ADL under different measurement
conditions. Furthermore, our objective is to collect
Table 1: List of Activities and activity duration.
Activity Name (Class) Action
Swiping on a Phone (C1) 5s
Zoom In on a Phone (C2) 5s
Zoom Out on a Phone (C3) 5s
Pressing Button using Thumb (C4) 5s
Flipping a Switch (C5) 5s
Cutting a fruit (knife) (C6) 7s
Eating With Spoon(C7) 7s
Flipping a Bottle cap (C8) 5s
Writing on paper with pen (C9) 7s
Using Scissors (C10) 7s
Table 2: Body Postures and Hand Positions.
Body Postures Hand Positions
Folded Knees (b1)
Folded Legs (b2) P1
Sit (b3)
Sit to Stand (b4) P2
Stand (b5)
sEMG signals during Fine-ADL from a group of In-
dian subjects and analyze the impact of measurement
conditions on ADL classification.
1.2 Contributions
• A novel sEMG dataset is developed that corre-
sponds to Fine-ADL under different body pos-
tures and hand positions.
• The impact of different body postures and arm
positions on the classification of Fine-ADL and
class-wise analysis are studied through various
experiments.
• Various classical Machine Learning (ML) frame-
works and hybrid CNN Bi-LSTM are imple-
mented for the classification of Fine-ADL.
2 METHODOLOGY
2.1 Fine-ADL sEMG Dataset
In this work, a new sEMG dataset corresponding to
Fine-ADL is presented. The data is collected from 10
subjects who have no abnormalities or impairments in
their upper limbs. The group of subjects have diverse
demographics, consisting of 7 males and 3 females,
with 2 left-handed and 8 right-handed individuals. In
terms of age distribution, there are 8 subjects in the
17 − 20 age group, 1 subject above 30, and 1 subject
above 40. The research study was approved by the
BIOSIGNALS 2024 - 17th International Conference on Bio-inspired Systems and Signal Processing
592

10 Subjects
5 Body Postures
2 Hand Positions
10 Activities
Labels
Experiments:
Aggregate Analysis
Subject -wise Analysis
Body -Postures Combinations
ML Models Training:
FCNN
RF
kNN
EMGHandNet
Classification
Analysis
Class -wise
Analysis
5-channels sEMG signals
Preprocessing:
20Hz HPF, 50Hz BSF
Time Domain – 16
Frequency Domain – 8
Wavelet Domain -3
Eigenvalues - 5
Feature Extraction
Figure 1: Block Diagram.
ethics committee at the Indian Institute of Information
Technology, Sri City (No. IIITS/EC/2022/01) and is
in accordance with the principles of the Declaration of
Helsinki. The data acquisition process is non-invasive
and prior to a measurement session, each subject gave
a written informed consent and was introduced to the
experimental protocol.
In this study, each subject performed a series of
10 activities of different durations as listed in Table
1. Further, to ensure accurate measurements, these
activities in the selected body postures and hand po-
sitions were demonstrated to the subjects. The EMG
signal from the hand is recorded using a wireless 5-
channel Noraxon Ultium sEMG sensor configuration.
The electrical contact is made with dual Ag/AgCl
self-adhesive electrodes at the densest region of the
selected forearm muscle sites (Criswell, 2010) given
in Table 3. Prior to electrode placement, the hands are
cleaned using an alcohol-based wet wipe. Each sub-
ject is asked to perform each of the fine-activities of
Figure 2: Body Postures.
Figure 3: Hand Positions.
daily living in three phases: rest, action, and release.
Subjects start with the rest phase, where they relax the
muscles and refrain from any physical activity. Dur-
ing the action phase, the activity is executed and the
release phase denotes a smooth transition from the ac-
tion state to the rest state.
During a measurement session, the subject per-
forms 10 Fine-ADL in 5 different body postures as
shown in Fig. 2 and 2 different hand positions as
shown in Fig. 3. Thus, the total number of measure-
ment conditions is 10. The signal specifications are as
follows: 1) sampling rate: 4000 samples/sec, 2) dura-
tion of a trial: 11s or 13s, and 3) the break between
two consecutive body postures: 5 minutes. Each ac-
tivity in a given posture is repeated 5 times. Hence
the total number of trials is 10 subjects × 5 postures
× 2 positions × 10 activities × 5 trials = 5000.
2.2 Methodology
The methodology for classification of the sEMG sig-
nals corresponding to Fine-ADLs consists of the fol-
lowing stages: 1) data preparation, 2) segmentation,
3) feature extraction, 4) ML model training, and 5)
testing and analysis as shown in Fig. 1.
2.2.1 Data Preparation
In this phase, all the 5-channeled sEMG signals are
processed by two digital filters. Initially, the sEMG
signals are high-pass filtered with a lower cut off fre-
quency of 20Hz to remove any motion artifacts. The
output of this filter is processed by a 50Hz notch fil-
ter to remove any electric line noise. From the filter
Classification of Fine-ADL Using sEMG Signals Under Different Measurement Conditions
593

Folded Knees Folded Legs Sit Sit to Stand Stand P1 P2 All
Poses
CNN Bi-LSTM
FCNN
Random Forest
kNN
Subjects
0.795
0.688
0.771
0.693
0.756
0.668
0.768
0.651
0.598
0.839
0.764
0.712
0.748
0.713
0.646
0.713
0.73
0.715
0.671
0.794
0.852
0.847
0.782
0.43
0.466
0.517
0.491
0.478
0.531
0.551 0.476
0.416
0.45
0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
Figure 4: Test accuracies from classification under various Conditions using different Classifiers in the aggregate scheme.
output, the rest phase and the samples from the final
one second of the release phase are discarded to retain
the action and 2 seconds of the release phase. These
processed sEMG signals are annotated with the rele-
vant class labels corresponding to the Fine-ADL cat-
egories.
2.2.2 Feature Extraction
To improve classification performance, the pre-
processed sEMG signals are segmented into win-
dows of 250ms duration. For each of these sEMG
segments, 16 time domain features (Sapsanis et al.,
2013),(Karnam et al., 2021), 8 frequency domain fea-
tures, 3 wavelet domain features and eigenvalues are
computed. The features from each of these segments
and the 5 channels of a signal are concatenated to
build the full feature vector. The specific combina-
tion of features is used after analysing various feature
combinations.
2.2.3 Classification Framework
The next stage consists of model training and testing
four classifiers: Fully Connected Neural Networks
(FCNN), Random Forests (RF), k-Nearest Neighbors
(k-NN), and the CNN Bi-LSTM (Karnam et al., 2022)
with minor modifications. Note that in the case of the
deep learning model the feature extraction is implicit
within the ML framework. The modifications to the
CNN Bi-LSTM include changing the dropout rate to
0.3, the CNN window size to 4×7, and the batch size
to 8. The FCNN consists of five dense layers with re-
spective number of neurons: 256, 128, 64, 16, and
10. The ReLU activation function is utilized in the
first four layers and the softmax function used in the
output layer.
3 IMPLEMENTATION
In this paper, three distinct experiments, 1) the aggre-
gate scheme, 2) the subject-wise scheme and, 3) im-
Table 3: Sensor placement on hand muscles.
Channel No. Muscle Name
1 Abductor digiti minimi
2 Extensor pollicis brevis
3 First dorsal interosseous
4 Abductor pollicis brevis
5 Brachioradialis
pact analysis of body postures combinations are con-
ducted. These experiments investigate the impact of
the body postures and the hand positions on classifi-
cation performance.
3.1 Aggregate Scheme
This experiment involves the utilization of the afore-
mentioned four classifiers to classify the feature data
from all of the 10 subjects. The objective of this
scheme is to investigate the general impact of differ-
ent measurement conditions on the classification of
Fine-ADL and also to analyse if the classifiers can
learn and classify across different individuals. The
classifiers are evaluated for the following three condi-
tions.
Body Postures. In this case, the trials of the 10
subjects from any set of four body postures are em-
ployed for training purposes while the trials in the left
out body posture are utilized for testing. This process
is repeated till trials from the five body postures are
tested separately(Fougner et al., 2011). This method-
ology is referred to as Leave-One-Posture-Out analy-
sis.
Hand Positions. In this case, the trials of all sub-
jects from one hand position are utilized for training
while the trials from the other hand position are used
for testing. This process is repeated for both hand po-
sitions(Fougner et al., 2011).
All Positions. In this case, 80% of all trials from
each combination of a body posture and a hand po-
sition, aggregated across all subjects are utilized for
training. The remaining 20% of the trials are tested.
BIOSIGNALS 2024 - 17th International Conference on Bio-inspired Systems and Signal Processing
594

Folded Knees Folded Legs Sit Sit to Stand Stand P1 P2 All
Poses
1
2
3
4
5
6
7
8
9
10
Average
Subjects
0.704
0.737
1
0.869
0.768
0.859
0.768
0.697
0.723
0.899
0.735
0.889
0.816
0.859
0.808
0.745
0.778
0.745
0.758
0.802
0.848
1
0.724
0.646
0.859
1
0.687
0.806
0.687
0.786
0.838
0.717
0.889
0.848
0.687
0.796
1
0.838
0.747
0.795
0.848
0.727
0.908
0.859
0.806
0.796
0.939
0.758
0.889
0.848
0.837
0.831
0.707
0.759
0.729
0.782
0.765
0.859
0.745
0.803
0.757
0.843
0.726
0.843
0.731
0.807
0.699
0.775
0.655
0.759
0.819
0.765
0.899
0.838
0.944
0.853
0.884
0.858
0.919
0.784
0.832
0.889
0.87
0.434
0.404
0.606
0.596 0.586
0.5
0.6
0.7
0.8
0.9
1
Figure 5: Test accuracies from Subject-wise average performance under various conditions using the hybrid CNN Bi-LSTM.
3.2 Subject-Wise Analysis
In the subject-wise experiment, a model’s learning ca-
pability at the subject level is evaluated. In this experi-
ment, only the results from the hybrid CNN Bi-LSTM
are reported. For the signals from each of the ten sub-
jects, the training, testing, and performance analysis
is carried out under the three sets of conditions men-
tioned previously. Hence, in the subject-wise scheme,
in a given classification analysis, the amount of data
analyzed is only one tenth of the first experiment. Fi-
nally, the results averaged across the subjects are also
reported.
3.3 Body Postures Combinations
Analysis
In the body posture combinations analysis, the impact
of diverse body postures on Fine-ADL is investigated.
The study encompasses data from each of the 10 sub-
jects. The methodology involves training the model
using different sets of body postures and testing on
the remaining body postures (Fougner et al., 2011).
The number of specific combinations when i condi-
tions are used in training is 5
C
i
. The total is
∑
4
i=1
5
C
i
resulting in the examination of 30 distinct scenarios.
The averaged results from FCNN across the 4 cate-
gories and the class-wise F
1
scores are reported.
4 RESULTS AND ANALYSIS
4.1 Aggregate Analysis
Fig. 4 illustrates the results of the aggregate scheme.
Across various body postures and hand positions, on
average, the hybrid CNN Bi-LSTM performs the best,
achieving an average test accuracy of 76.89%, fol-
lowed by the FCNN with 72.42%. In the classifica-
tion analysis versus the body postures, Fig. 4 shows
that the Stand position is the least challenging posture
for the model to understand with a test accuracy of
83.93% by hybrid CNN Bi-LSTM, while Sit posture
is the most difficult with 75.62%. An 8% difference in
accuracy is observed between the Stand and Sit pos-
tures indicating an impact of body postures on Fine-
ADL classification. However, in the hand position
analysis, the FCNN is the top-performing classifier
with an average test accuracy of 73.9%. Moreover,
reduction of the training data to 50% has little impact
on the accuracy for most classifiers. The hand posi-
tions seem to be less influential in case of aggregate
analysis as they produce similar accuracies. In the
third condition where all positions are considered, the
FCNN outperforms others with an accuracy of 85.2%,
followed by the RF with 84.7% and the hybrid CNN
Bi-LSTM with 79.4%.
4.2 Subject-Wise Analysis
Fig. 5 illustrates the subject-wise test accuracy of
the hybrid CNN Bi-LSTM across the different pos-
tures and positions. CNN Bi-LSTM outperforms the
FCNN in the body posture analysis, achieving an av-
erage test accuracy of 78.93% compared to FCNN’s
51.46%. In terms of variations across subjects, the
data from subjects 3 and 7 have the highest test ac-
curacies averaged across the different conditions at
89.8% and 88.2% respectively. The lowest average
performance is observed for data from subject 8 at
68.9%. In the body postures analysis, Folded Knees
prove to be the most challenging posture for the clas-
Classification of Fine-ADL Using sEMG Signals Under Different Measurement Conditions
595

C1 C2 C3 C4 C5 C6 C7 C8 C9 C10
Class
b1
b2
b3
b4
b5
b1, b2
b1, b3
b1, b4
b1, b5
b2, b3
b2, b4
b2, b5
b3, b4
b3, b5
b4, b5
b1, b2, b3
b1, b2, b4
b1, b2, b5
b1, b3, b4
b1, b3, b5
b1, b4, b5
b2, b3, b4
b2, b3, b5
b2, b4, b5
b3, b4, b5
b1, b2, b3, b4
b1, b2, b3, b5
b1, b2, b4, b5
b1, b3, b4, b5
b2, b3, b4, b5
P1
P2
All Conditions
Combination
0.588
0.602
0.653
0.601
0.594
0.608
0.612
0.624
0.638
0.666
0.672
0.619
0.685
0.609
0.658
0.659
0.689
0.666
0.657
0.767
0.637
0.659
0.644
0.751
0.637
0.826
0.609
0.633
0.643
0.659
0.588
0.658
0.615
0.626
0.619
0.587
0.719
0.598
0.631
0.595
0.668
0.606
0.752
0.618
0.599
0.606
0.714
0.661
0.744
0.606
0.598
0.587
0.625
0.687
0.688
0.729
0.649
0.667
0.669
0.7
0.707
0.704
0.728
0.778
0.698
0.732
0.677
0.73
0.672
0.752
0.756
0.719
0.774
0.706
0.791
0.703
0.788
0.698
0.7
0.843
0.88
0.879
0.845
0.864
0.895
0.881
0.919
0.925
0.892
0.916
0.892
0.884
0.884
0.907
0.912
0.942
0.926
0.858
0.938
0.905
0.914
0.917
0.92
0.907
0.91
0.966
0.914
0.884
0.92
0.927
0.913
0.911
0.949
0.854
0.878
0.875
0.852
0.909
0.907
0.894
0.894
0.924
0.907
0.932
0.914
0.878
0.905
0.926
0.915
0.948
0.919
0.884
0.917
0.936
0.925
0.909
0.947
0.917
0.947
0.905
0.964
0.914
0.932
0.935
0.92
0.954
0.72
0.795
0.737
0.69
0.78
0.827
0.817
0.804
0.803
0.865
0.837
0.82
0.768
0.844
0.833
0.881
0.862
0.791
0.845
0.858
0.831
0.896
0.881
0.851
0.879
0.927
0.875
0.807
0.861
0.918
0.838
0.891
0.902
0.688
0.697
0.635
0.623
0.73
0.786
0.732
0.812
0.754
0.747
0.76
0.762
0.757
0.763
0.806
0.799
0.847
0.778
0.836
0.747
0.835
0.793
0.771
0.813
0.828
0.884
0.742
0.834
0.834
0.838
0.768
0.802
0.914
0.59
0.599
0.623
0.651
0.724
0.702
0.657
0.663
0.667
0.645
0.715
0.655
0.771
0.742
0.605
0.757
0.772
0.708
0.695
0.733
0.683
0.745
0.851
0.751
0.657
0.789
0.745
0.685
0.736
0.859
0.728
0.708
0.653
0.668
0.736
0.773
0.808
0.802
0.775
0.788
0.761
0.746
0.713
0.793
0.748
0.847
0.815
0.753
0.846
0.861
0.793
0.784
0.824
0.764
0.795
0.895
0.842
0.758
0.869
0.815
0.814
0.824
0.912
0.511
0.434
0.527
0.541
0.564
0.58
0.503
0.527
0.507
0.388
0.458
0.532
0.555
0.551
0.54
0.566
0.572
0.483
0.572
0.577
0.558
0.562
0.522
0.472
0.394
0.47
0.393
0.472
0.526
0.524
0.541
0.47
0.563
0.506
0.529
0.508
0.53
0.535
0.507
0.485
0.525
0.584
0.586
0.585
0.572
0.531
0.523
0.528
0.544
0.566
0.54
0.578
0.534
0.51
0.575
0.4
0.5
0.6
0.7
0.8
0.9
Figure 6: Class-wise F1 score averaged across the test conditions as function of combinations of training conditions.
sifier, with an average test accuracy of 72.39%. In
contrast, Stand posture yields the highest accuracy of
83.78%, exhibiting an 11.39% difference. The re-
sults show that the body postures have a significant
influence among the subject level results. On aver-
age, there is a 23% gap in accuracy between the high-
est and lowest performing postures. This distinction
is particularly pronounced in subjects 9 and 6, with
a substantial 42.42% and 48.48% disparity between
Folded Knees and Stand, respectively. Interestingly,
while Stand posture boasts the highest average accu-
racy, Folded Knees and Folded Legs emerge as the top
performers in three subjects each, while Stand pre-
vails in only two subjects. These results suggest that
the impact of body postures is notably intricate on a
subject-specific basis.
In the hand positions analysis, as shown in Fig. 5,
the hybrid CNN Bi-LSTM seems to have an average
accuracy (across subjects) close to 76% for both the
hand positions. The standard deviation across sub-
jects is 7.55% and 6.3% respectively. Subjects 3, 8, 7,
and 6 exhibit considerable discrepancies in hand po-
sition accuracy, highlighting distinct subject-specific
trends. Generally, the performance closely aligns
with the aggregate analysis. Finally, in all positions
case, the hybrid CNN Bi-LSTM achieves an average
test accuracy of 86.99%, with a minimum of 78.39%
and a maximum of 94.44%.
4.3 Body Postures Combinations
Analysis
Figs. 6 and 7 illustrate class-wise F1 scores when
different combinations of body postures are used for
BIOSIGNALS 2024 - 17th International Conference on Bio-inspired Systems and Signal Processing
596

1 2 3 4 5 6 7 8 9 10
Classes
0.4
0.5
0.6
0.7
0.8
0.9
1
Test Accuracy
Overall Average Single body Posture Double body Posture Triple body Posture Quadriple body Posture Hand Postures
Figure 7: Class-wise test F1 score averaged across a group of training conditions when a number of conditions are used for
training.
training. Fig. 6, shows the test F1 scores for each of
the classes (columns) when a specific combination of
conditions is used for training (rows). Fig. 7 shows
the test F1 scores further averaged across the 5
c
i
con-
ditions for each i, indicating the average performance
when any of i conditions are used for training. From
these Figs., the F1 scores increase as the number of
body postures used for training increases. On aver-
age, using one posture yields an F1 score of 64.64%,
while employing two, three, and four postures results
in F1 scores of 72.11%, 75.87%, and 78.3% respec-
tively. The combination {b1, b2, b3, b4} exhibits the
highest F1 score, reaching 84.5%. It’s worth noting
that b5 proves to be the easiest posture to learn, even
without explicit training from the same condition. In-
terestingly, as the number of postures used for training
increases, the increase in F1 score diminishes. The
class-wise analysis reinforces the observation that C3
has the lowest F1 score, while C5 and C6 consistently
demonstrate the highest scores across all combina-
tions of body postures.
Moreover, the average for hand postures falls be-
tween that of a single combination and all other com-
binations. This suggests that while hand postures do
contribute to accuracy, they are not as influential as
combined body postures. Furthermore, it’s notable
that activities C2 (Zoom In) and C3 (Zoom Out), be-
ing very similar in nature, might be causing substan-
tial confusion for the model. Surprisingly, a dip in F1
score is observed for C9 (Writing), which could be at-
tributed to potential similarities with other activities,
leading to mis-classification.
4.4 Discussion
The hybrid CNN Bi-LSTM model consistently out-
performs other models in both Subject-wise scenario
and Body Posture analyses within the Aggregate sce-
nario. Moreover, the FCNN demonstrates superior
classification ability specifically for Body Posture
Combinations analysis. Consistent trends are ob-
served in both aggregate and subject-wise analyses.
Stand posture is always the easiest to classify on av-
erage. Additionally, as the amount of training data in-
creased, there was a noticeable improvement in test
accuracy, particularly evident in the hand position
analysis where 50% of the data was utilized, com-
pared to body postures which utilized 80% of the
data. This incremental trend was also apparent in
the analysis of body posture combinations. The con-
sistent shape of Fig. 7 indicates that the ranking of
class performance remains stable across various com-
binations. The subject-wise analysis further empha-
sized the impact of measurement conditions and the
subject-specific nature of Fine-ADL. Overall, these
findings underscore the impact of measurement con-
ditions on Fine-ADL.
5 CONCLUSION AND FUTURE
WORK
This paper presents a new sEMG dataset consisting of
10 Fine-ADL activities conducted under various mea-
Classification of Fine-ADL Using sEMG Signals Under Different Measurement Conditions
597

surement conditions. The dataset includes data cap-
tured in different body postures and hand positions.
The analysis of the dataset is carried out from two
perspectives: Aggregate and Subject-wise, consider-
ing three cases: body postures, hand positions, and all
positions in both experiments along with class-wise
analysis on impact of body postures. Among the clas-
sifiers examined, the hybrid CNN Bi-LSTM demon-
strates the best performance, successfully recognizing
Fine-ADL even in diverse measurement conditions.
The most challenging body posture for the classifier is
Folded Knees, while the least challenging is the Stand
posture. Interestingly, both the hand positions consid-
ered yield similar accuracies. Nevertheless, the cur-
rent outcomes highlight the potential of the proposed
framework for real-time Fine-ADL and also demon-
strate the impact of various measurement conditions
on Fine-ADL. In terms of future work, there is po-
tential for further enhancing the model through fine-
tuning to achieve improved results. Additionally, the
impact of the amount of training data, pertaining to a
specific measurement condition, on testing accuracy
needs to be investigated. Furthermore, feature selec-
tion analysis also requires further improvements and
the generalization ability of the model to new subjects
needs to be explored.
REFERENCES
Castellini, C., Fiorilla, A. E., and Sandini, G. (2009). Multi-
subject/daily-life activity emg-based control of me-
chanical hands. Journal of neuroengineering and re-
habilitation, 6(1):1–11.
Chen, M., Dutt, A. S., and Nair, R. (2022). Systematic re-
view of reviews on activities of daily living measures
for children with developmental disabilities. Heliyon,
8(6):e09698.
Criswell, E. (2010). Cram’s introduction to surface elec-
tromyography. Jones & Bartlett Publishers, Burling-
ton.
Dai, S., Wang, Q., and Meng, F. (2021). A telehealth frame-
work for dementia care: an adls patterns recognition
model for patients based on nilm. In 2021 Interna-
tional Joint Conference on Neural Networks (IJCNN),
pages 1–8.
Faria, A. L., Pinho, M. S., and Berm
´
udez i Badia, S. (2020).
A comparison of two personalization and adaptive
cognitive rehabilitation approaches: a randomized
controlled trial with chronic stroke patients. Journal
of NeuroEngineering and Rehabilitation, 17(1):78.
Fauth, E. B., Schaefer, S. Y., Zarit, S. H., Ernsth-Bravell,
M., and Johansson, B. (2017). Associations between
fine motor performance in activities of daily living and
cognitive ability in a nondemented sample of older
adults: Implications for geriatric physical rehabilita-
tion. Journal of Aging and Health, 29(7):1144–1159.
Fougner, A., Scheme, E., Chan, A. D. C., Englehart, K.,
and Stavdahl, Ø (2011). Resolving the limb position
effect in myoelectric pattern recognition. IEEE Trans-
actions on Neural Systems and Rehabilitation Engi-
neering, 19(6):644–651.
He, C., Xiong, C.-H., Chen, Z.-J., Fan, W., Huang, X.-L.,
and Fu, C. (2021). Preliminary assessment of a pos-
tural synergy-based exoskeleton for post-stroke upper
limb rehabilitation. IEEE Transactions on Neural Sys-
tems and Rehabilitation Engineering, 29:1795–1805.
Karnam, N. K., Dubey, S. R., Turlapaty, A. C., and
Gokaraju, B. (2022). Emghandnet: A hybrid cnn and
bi-lstm architecture for hand activity classification us-
ing surface emg signals. Biocybernetics and Biomed-
ical Engineering, 42(1):325–340.
Karnam, N. K., Turlapaty, A. C., Dubey, S. R., and
Gokaraju, B. (2021). Classification of semg signals
of hand gestures based on energy features. Biomedi-
cal Signal Processing and Control, 70:102948.
Katz, S. (1983). Assessing self-maintenance: activities of
daily living, mobility, and instrumental activities of
daily living. J Am Geriatr Soc, 31(12):721–727.
Park, J., Zahabi, M., Kaber, D., Ruiz, J., and Huang, H.
(2020). Evaluation of activities of daily living tes-
beds for assessing prosthetic device usability. In 2020
IEEE International Conference on Human-Machine
Systems (ICHMS), pages 1–4.
Rosenburg, R. and Seidel, H. (1989). Electromyography of
lumbar erector spinae muscles–influence of posture,
interelectrode distance, strength, and fatigue. Eur J
Appl Physiol Occup Physiol, 59(1-2):104–114.
Sandberg, M., Kristensson, J., Midl
¨
ov, P., Fagerstr
¨
om, C.,
and Jakobsson, U. (2012). Prevalence and predictors
of healthcare utilization among older people (60+):
Focusing on adl dependency and risk of depression.
Archives of Gerontology and Geriatrics, 54(3):e349–
e363.
Sapsanis, C., Georgoulas, G., and Tzes, A. (2013). Emg
based classification of basic hand movements based
on time-frequency features. In 21st Mediterranean
Conference on Control and Automation, pages 716–
722.
Song, M. and Kim, J. (2018). An ambulatory gait moni-
toring system with activity classification and gait pa-
rameter calculation based on a single foot inertial sen-
sor. IEEE Transactions on Biomedical Engineering,
65(4):885–893.
Toledo-P
´
erez, D. C., Rodr
´
ıguez-Res
´
endiz, J., G
´
omez-
Loenzo, R. A., and Jauregui-Correa, J. (2019). Sup-
port vector machine-based EMG signal classification
techniques: A review. Applied Sciences, 9(20):4402.
doi: 10.3390/app9204402.
Williams, H., Shehata, A. W., Dawson, M., Scheme, E.,
Hebert, J., and Pilarski, P. (2022). Recurrent convo-
lutional neural networks as an approach to Position-
Aware myoelectric prosthesis control. IEEE Trans
Biomed Eng, 69(7):2243–2255.
Yang, D., Yang, W., Huang, Q., and Liu, H. (2015). Clas-
sification of multiple finger motions during dynamic
upper limb movements. IEEE journal of biomedical
and health informatics, 21(1):134–141.
BIOSIGNALS 2024 - 17th International Conference on Bio-inspired Systems and Signal Processing
598