Five gestures from three forearm channels
The paper's target is a myoelectric prosthetic hand: surface electrodes read the small electrical signals produced when forearm muscles contract, and a classifier maps those signals to a gesture that then drives the hand's motors. The design brief was deliberately narrow — five gesture classes, three input channels, and hardware cheap enough for amputees in developing countries, where the paper states only 5% of an estimated 40 million amputees have access to any prosthesis.
The five classes were hand open, hand close (grasp), clockwise wrist rotation (extension), counter-clockwise wrist rotation (flexion), and rest. Signals came from three electrodes placed arbitrarily on the upper forearm, following general EMG recording guidelines. Each gesture was performed for 10 seconds per trial, and the recordings intentionally included the transition from rest into a gesture as well as sustained holds, because the authors argue a classifier trained only on steady contractions degrades during the transitions that dominate real-time use.
Time and frequency features fed the same network shape
Both analyses used a 2000 samples/second sampling rate and a 128-sample moving window, so the comparison held sampling rate and window size constant. In the frequency domain, a Fast Fourier Transform converted each window into 64 bins of about 15.6 Hz each; to fit the Arduino's memory, only eight bins per channel — those with the highest power — became network inputs, giving 24 input neurons. In the time domain, variance, mean absolute deviation, and waveform length were computed per channel, giving nine input neurons.
Both versions used a shallow feedforward network with one 10-neuron hidden layer using a tansig transfer function and a five-neuron softmax output layer, so the output was a probability per gesture and the highest probability became the classified gesture. Training ran in MATLAB on an external computer; only classification ran on the Arduino Mega 2560, a split the authors made because training is computationally expensive and they wanted to keep system cost low.
Time-domain features won on accuracy and on speed
On pre-stored training and testing datasets, the time-domain network reached an average classification accuracy of 94.8% and the frequency-domain network 93.2%. The gap is small, but the accompanying table shows a much larger difference in cost: algorithm complexity was O(n) for time domain versus O(n log n) for frequency domain, feature extraction took 26 ms per window versus 198 ms, and training took 1 second versus 33 seconds. Training epochs ran the other way, 99 for time domain versus 70 for frequency domain.
Those numbers explain the adoption decision. The authors chose the time-domain pipeline as the prosthetic's classification method because it gave both better average accuracy and a faster response, which matters when the hand must react as a user moves between gestures. The accuracy figures are averages over the paper's stored test data, not per-gesture guarantees, and the paper reports them alongside confusion matrices for each domain rather than as a single headline number.
Two gestures drive a wrist rotation mechanism
The base hand uses two servo motors for two degrees of freedom. The authors added a wrist rotation mechanism controlled by two of the five gestures, one gesture per rotation direction, which raised the overall design to three degrees of freedom. This is a design and testing result rather than a classification-accuracy result: the mechanism extends what the same five-class output can command, without adding new gesture classes to the classifier.
That distinction matters for reading the paper's claims. The 94.8% figure describes gesture classification, not the reliability of wrist rotation or grasp force. The wrist mechanism's contribution is functional scope — the same classified output now drives more of the hand's movement.
Force sensors and vibration motors add touch feedback
The paper also built a tactile feedback path: force-sensitive resistors measure the force on the hand, and coin vibration motors convey it to the user. The hand was divided into three regions, each mapped to one vibration motor, and each motor operates between 0 and 2.5 V, with 2.5 V as maximum vibration. Higher force on a sensor lowers its resistance; if force exceeds a maximum threshold, the motors stay at maximum power.
The authors describe this feedback as a significant improvement to the prosthetic as a whole, framing it as part of sensory integration with the user. The evidence packet does not report a user study measuring whether wearers could interpret the vibration, so the feedback claim rests on the system being built and described, not on measured sensory performance.
The cost target is a parts list, not a clinical result
The paper's cost table itemizes the build: an Arduino Mega 2560 at 38.50 USD, three AD620 amplifiers at 12.60 USD each, three AD8618 units at 6.12 USD each, three GWS MICRO 2BBMG servos at 15.85 USD each, 3D-printed PLA parts at 80 USD, a steel ball bearing at 2.00 USD, five FSR 400 force sensors at 6.95 USD each, and three coin vibration motors at 0.30 USD each. These are component prices, not a total system price or a comparison against commercial prostheses.
The affordability argument therefore rests on the bill of materials and on design choices that keep computation light — three channels, eight frequency bins per channel, training offloaded to MATLAB, classification on a 256 KB-flash Arduino. The paper calls the accuracy and complexity results a proof of concept for a low-cost myoelectric hand, which is the appropriate scope: it demonstrates a working build, not field performance with amputees.
The accuracy claim is bounded by its test setup
The 94.8% and 93.2% figures come from pre-stored training and testing datasets collected under the protocol described, with the same sampling rate and window size for both analyses. The paper does not report a held-out user study, a count of participants, or performance across different people wearing the hand over time, so the numbers should be read as classification accuracy on this dataset rather than as expected accuracy for a new user.
The paper's own comparison section reinforces why context matters: it cites prior work where reducing testing data relative to training data inflated accuracy, and where dropping from six channels to three lowered accuracy from 94.1% to 85.25%. The three-channel design here is a deliberate cost and complexity trade-off, and the reported accuracy belongs to that configuration and that recording protocol.
Speed, not the accuracy gap, drove the design choice
The most decision-relevant comparison in this paper is not 94.8% versus 93.2%. It is 26 ms versus 198 ms per window for feature extraction and O(n) versus O(n log n) complexity, on hardware limited to an Arduino Mega 2560. A 1.6-point accuracy difference would rarely justify roughly seven times the per-window feature-extraction time in a hand that must respond as the user transitions between gestures.
So the paper's practical priority is clear: for a low-cost, three-channel, five-gesture prosthetic controlled on a microcontroller, time-domain features with a shallow network are the configuration the authors chose, and the wrist rotation mechanism and vibration feedback are additions layered on that same five-class output. The open question the evidence leaves is how that configuration performs with amputees and across sessions, which the reported dataset does not measure.
Sources used in this answer
Development of a Low-Cost Prosthetic Hand Using Electromyography and Machine Learning
A shallow neural network using time-domain EMG features from three forearm channels classified five gestures at 94.8% average accuracy, versus 93.2% for frequency-domain features, with faster feature extraction (26 ms vs 198 ms per window) and lower complexity, so the time-domain pipeline was adopted for the prosthetic's control.