EMG Exercise Form Classification

Turns raw two-channel surface EMG from a dumbbell press into a Good / Fair / Bad form label, using the relationship between the deltoid and pectoral channels rather than either channel alone.

Research code behind a peer-reviewed conference paper.

Status: research notebooks, released as-is — and the accuracy figures in them do not measure what the title suggests. The label is a threshold on a quantity that is also handed to the model as an input column, so the task the network is set is to recover a rule it can already see. A three-line if on that one column reproduces every label exactly. Read What these numbers are not before quoting anything from here.

---

The problem

Judging lifting form from EMG looks like a modelling problem and is mostly a signal problem. A dumbbell press recruits the anterior deltoid and the pectoralis together, and bad form usually means the wrong one is doing the work. Neither channel alone says much: a high deltoid reading is fine during part of the movement and wrong during another. The informative quantity is the balance between them over a window.

So the pipeline spends most of its effort on windowed feature extraction, and the model on top is deliberately simple.

---

Method

Signal. Two channels, deltoid and pectoral, sampled at 256 Hz over 123,072 samples.

Windowed features, computed per channel over a sliding window:

• root mean square, as an activation-magnitude proxy
• integrated area under the rectified signal, as total work in the window
• skewness and kurtosis, to capture burst shape rather than just amplitude
• Wilcoxon amplitude statistics between channels

Label derivation. diff = pec_area - del_area is min-max scaled, and the scaled difference is thresholded into Good, Fair, and Bad. This is a heuristic label, not an expert annotation.

And the quantity it is thresholded from stays in the feature matrix. X = areas_df.iloc[:, :-1] takes every column except the label, which means scaled_difference — and diff, and the two areas that determine both — are all inputs. The target is a deterministic function of the input. This is target leakage, and it is the single most important thing to know about every number below.

Models. Keras Sequential stacks of LSTM and SimpleRNN layers with batch normalization and dropout, into a 3-way softmax. Trained for 10 to 30 epochs at batch size 32 with a 0.2 validation split.

Note that these are recurrent and dense models, not convolutional. There is no Conv1D or Conv2D anywhere in either notebook.

---

What these numbers are not

Taken from the committed cell outputs in EMG.ipynb, not from memory:

Metric	Value
Highest validation accuracy recorded	0.9250
Lowest validation accuracy recorded	0.4500
Distinct val_accuracy values in the committed outputs	10
Approximate validation set size	~80 samples

Three reasons not to read 0.9250 as a result.

A three-line rule beats it. Because scaled_difference is an input column and the label is a threshold on scaled_difference, the label can be recovered exactly without any model:

rule = np.where((s > 0) & (s < 0.4), 'Bad',
       np.where((s >= 0.3) & (s < 0.5), 'Fair', 'Good'))   # s = X['scaled_difference']
# reproduces the labels: 100.00%  (586/586)

A network reported at 0.9250 is scoring below a rule it could have learned from one column.

0.9250 is the maximum over ten recorded runs, whose values run down to 0.4500. Reporting the best of a spread that wide, on a validation set this small, is a selection effect rather than a measurement.

Train and test are not independent. The sliding window is 6000 samples with a 200-sample hop, so consecutive windows overlap by 96.7%. train_test_split(..., test_size=0.2, random_state=42) splits those overlapping windows at random, putting near-identical windows on both sides. Even with the leakage removed, the held-out score would not measure generalization. A segment-wise or subject-wise split is the fix.

The validation accuracies quantize on 1/80, which is how the validation set size was inferred. Running the same windowing over the committed EMG_DATASET.csv yields 586 windows in total, which is consistent with an inner validation split of that size.

---

Files

File	Contents
EMG.ipynb	Main pipeline: load, window, extract features, label, train, evaluate
EMG_TEST.ipynb	Shorter end-to-end pass from raw CSV through cleaning and visual inspection
EMG_DATASET.csv	Recorded signal, 123,072 rows
PRANAV_256.txt	The same recording in the acquisition tool's raw text format

---

Technology stack

Python, TensorFlow / Keras (LSTM, SimpleRNN, BatchNormalization, Dropout), pandas, NumPy, SciPy (skew, kurtosis, wilcoxon), scikit-learn (MinMaxScaler, train_test_split), Matplotlib, Seaborn.

---

Setup

Requires Python 3.9 or newer.

git clone https://github.com/patsypppe/EMG-Dumbbell-press.git
cd EMG-Dumbbell-press
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
jupyter notebook EMG.ipynb

EMG.ipynb cell 0 reads preprocessed.csv, which is not committed and is not written by any cell in either notebook — so a fresh clone cannot run the pipeline as-is. EMG_DATASET.csv holds the same recording (time, deltoid, pectoral) and can stand in for it; the windowing in cell 0 then produces 586 windows.

---

Current limitations

• Target leakage. The column the label is thresholded from is one of the model's inputs. Nothing reported here survives that, and it is not fixable by dropping one column: diff and the two area features determine the label just as completely.
• There is no independent ground truth. No one annotated these windows as good or bad form. The labels are a rule applied to the signal, so even a leakage-free model would only be learning the rule. Establishing whether EMG balance predicts form needs labels from outside the signal.
• Overlapping windows, random split. 96.7% overlap between consecutive windows, split at random.
• Small validation set. Roughly 80 samples. One sample moves the figure by 1.25 points.
• preprocessed.csv is not in the repository. EMG.ipynb cell 0 reads it and no cell in either notebook writes it, so a fresh clone cannot run the pipeline top to bottom as committed.
• The threshold rule has overlapping branches. In the labelling cell, the first condition 0 < i < 0.4 absorbs everything below 0.4, so the Fair branch 0.3 <= i < 0.5 can only ever fire between 0.4 and 0.5. The intended boundaries and the effective ones differ.
• Single subject, single session. The split is not subject-independent, so the numbers say nothing about generalization to a new person.
• The two data files are the same recording in two formats, roughly 8 MB committed twice.
• Notebooks carry embedded output images, which is why EMG.ipynb is around 780 KB.
• A stray load_iris import is left over from earlier scratch work.
• The models are LSTM and SimpleRNN, so any description of this work as CNN-based is incorrect.

---

Citation

This work was published at a peer-reviewed IEEE conference. Citation details and DOI are available on request or via LinkedIn.

---

License

MIT. See LICENSE.