Matrix Profile ECG Features with a Lightweight CNN-Attention Model and Severity Scoring

About the sleep apnea detection M.Tech project

Many people stop breathing for short spells during sleep and never find out. The standard lab test is costly, so most cases stay undiagnosed. This sleep apnea detection M.Tech project works with only one ECG lead, the kind a patch or a home device can record, and asks a clear research question: can richer waveform features, minute-to-minute context and a small model give a reliable per-minute label and a per-night severity score that a doctor can trust?

Research behind it

Base paper. MPCNN: A Novel Matrix Profile Approach for CNN-based Single Lead Sleep Apnea in Classification Problem, IEEE Journal of Biomedical and Health Informatics, 2024. It computes minimum, maximum and mean distance profiles from ECG subsequences and feeds them to a modified LeNet-5 and two light networks, tested on Apnea-ECG and UCDDB.

Research gap. The distance profiles are used on their own and are not joined with RR-interval and amplitude features. Each one-minute segment is judged alone, without the minutes around it. There is no explained severity score, and CPU cost for home devices is not studied with detection quality.

Proposed system and novelty. Every minute of ECG becomes two inputs: matrix profile distance profiles, and the RR-interval and R-amplitude series. A two-branch 1D CNN with channel attention encodes both, and a small temporal block looks at neighbouring minutes. Minute labels are summed into a nightly apnea index and mapped to normal, mild, moderate or severe. An int8 copy is timed on a plain CPU.

Models compared. RR features with Random Forest and SVM, a plain 1D CNN, a CNN-LSTM and the base MPCNN method rebuilt on the same subject-wise split. The ablation tests each feature branch alone, removes channel attention and minute context, and compares the float model with the int8 one.

Students who like medical signals and tabular risk models together can also look at the stacked ensemble project for diabetes, heart and kidney risk.

How it works

• An overnight record from PhysioNet Apnea-ECG is read, filtered and split into one-minute segments.
• R-peaks are detected and noisy minutes are dropped.
• Matrix profile distance features and RR-interval series are built for each minute.
• The two-branch model labels each minute as apnea or normal, using the minutes on either side.
• Minute labels give an hourly apnea count, a nightly index and a severity group.
• The doctor sees the ECG with apnea minutes shaded and downloads the report.

Users and modules

The doctor or sleep technician adds a patient, uploads a recording, studies the shaded ECG and the hourly chart, reads the index and severity group and saves a PDF report. The admin manages users, switches between the float and int8 models, sets severity cut-offs and opens the comparison page. Everything runs on a normal laptop; the CNN-LSTM baseline is trained once on a free cloud notebook.

What you get

• Phase 1 Research: Literature survey of 15-20 recent ECG apnea papers; Base paper (IEEE JBHI, 2024) with research gap and novelty; Apnea-ECG data study with signal and label EDA; Problem statement and objectives; Synopsis and Phase-1 report.
• Phase 2 Implementation: Flask web app with doctor and admin panels; ECG filtering, R-peak detection and minute segmentation; Matrix profile distance and RR-interval feature builder; Baselines: RR features with RF / SVM, 1D CNN, CNN-LSTM, base MPCNN; Proposed two-branch CNN with channel attention and minute context; Severity module, int8 model and MySQL database.
• Results: Accuracy, sensitivity, specificity, F1 and AUC per minute; Per-night apnea index correlation and severity accuracy; Ablation: feature branches, attention, context, int8; Model size and CPU time per night table; results.xlsx and all graphs in high resolution.
• Thesis and paper: Thesis written for you (about 90 pages); IEEE conference paper draft; Journal paper draft; Publication guide.
• Reviews, viva and support: PPTs for Review 1, Review 2 and final viva; Viva book with 50+ questions; A1 poster and project diary; Demo and code walkthrough videos; 1-on-1 setup, 2 revisions and support till viva.

Who can use this project

M.Tech and M.E. students of CSE and IT who want a biomedical signal dissertation with real deep learning work will find a good fit here. It also suits searches for sleep apnea detection using deep learning or a single-lead ECG classification project.