AFib ECG Classifier

A classical machine learning pipeline classifies ECG recordings as Normal, AFib, Other rhythm or Noisy. The demo shows held-out examples, detected beats and predictions. There is no clinical validation.

The Problem

This project asks whether classical machine learning can classify single-lead ECG recordings from the PhysioNet/CinC Challenge 2017 dataset as Normal, AFib, Other rhythm or Noisy. Beat timing is part of the signal: AFib often appears as irregularly spaced heartbeats.

What I Built

A Python pipeline that reads ECG recordings, detects candidate beats, extracts features and trains a gradient boosting classifier. It compares the model with a majority-class baseline and logistic regression based on beat intervals.

A React demo shows curated held-out recordings with detected peaks, the true label, the model prediction and probabilities. It includes mistakes. The selection is for inspection and is not an evaluation set.

How It Works

1. Read the signalConvert each recording to millivolts using its WFDB header. Read the sampling rate from the header instead of assuming it.
2. Detect candidate beatsApply a band-pass filter, an energy envelope and SciPy peak detection. Markers follow absolute filtered extrema, so they can land on an S wave.
3. Extract featuresSummarize the intervals between beats, signal quality, waveform shape and recording duration.
4. Select the modelTrain a class-balanced scikit-learn HistGradientBoostingClassifier. Choose settings through cross-validation on the training split only. There is no deep learning.
5. Evaluate and inspectEvaluate on held-out recordings, then show selected examples and errors in the demo.

The design and evaluation rules were frozen before fitting. The test set was never used for tuning. Training writes the results to artifacts/results.json, which the README identifies as the source of its reported numbers.

Results

The gradient boosting model scored above the logistic regression and majority-class baselines on the held-out test set. The reported challenge metric averages F1 for Normal, AFib and Other rhythm; Noisy predictions still count in errors. These are local held-out results, not official leaderboard results, because the challenge’s private test set was not used.

Other rhythm and Noisy are the weakest classes. The README publishes per-class precision, recall and F1 alongside a confusion matrix, so the overall score can be read with the errors behind it.

Limitations

• The split is by recording, not patient. The inspected headers and labels have no documented patient IDs. Recordings from the same patient may appear in both splits, which can inflate scores. An exact duplicate crossed the split; its test copy was excluded before fitting any model.
• The data covers a limited setting. The dataset is small, with short recordings from a single lead and device. Labels are noisy, Other combines different rhythms, and Noisy has few test examples.
• Beat detection has limited validation. It was checked on synthetic signals only. The README claims no accuracy against expert beat annotations.
• Probabilities are uncalibrated. The curated demo includes errors and cannot establish model performance.
• Saved artifacts have caveats. The README points to a release review covering stale final-split count metadata and limits of the saved model hashes.
• No clinical validation. This is not a medical device and is not for diagnosis.