Binary Brains — Cough Detection at 750 Hz
YŪGŌ Hackathon 2026 | SSN College of Engineering
Overview
This repository presents a software-only, offline cough detection system developed for Pack B – Cough Detection at 750 Hz in the YŪGŌ Hackathon 2026. The system processes pre-recorded 750 Hz mono WAV audio files and detects cough events using MFCC-based signal features and a gradient-boosted tree classifier (XGBoost). Outputs include structured JSON cough events, publication-quality plots, and an interactive Streamlit dashboard.
⚠️ This tool is a research prototype and NOT a medical device.
Novelty & Innovation
Unlike the provided baseline model, which performs end-to-end inference on fixed representations, our approach introduces a segment-level, sliding-window detection framework that enables fine-grained temporal localization of cough events. By extracting statistically aggregated MFCC descriptors from short overlapping audio segments and classifying them using an optimized gradient-boosted ensemble, the system achieves improved discrimination while retaining interpretability and computational efficiency. The pipeline further introduces post-processing logic to aggregate segment-level predictions into clinically interpretable cough events with timestamps, confidence scores, and severity grading.
Technical Approach
1. Input: 750 Hz mono WAV files (offline)
2. Segmentation: 0.5-second windows with 0.3-second overlap
3. Feature Extraction: 13 MFCC coefficients (mean + std per segment)
4. Model: XGBoost binary classifier
5. Post-processing: Probability thresholding and temporal aggregation
6. Outputs: JSON cough events, ROC curve, confusion matrix, spectrogram, probability timeline, Streamlit dashboard
Results (Compared to Baseline)
Baseline Accuracy: ~87%
Our Accuracy: 91.67%
Baseline AUC: ~0.91
Our AUC: 0.957
Publication-Quality Plots (300 DPI)
The following plots are exported at 300 DPI (PNG):
• ROC Curve
• Confusion Matrix
• Spectrogram
• Probability Timeline
How to Run (Offline Demo)
python run_demo.py --offline
Dataset Citation
NCSU Robust-Cough Dataset (750 Hz subset), CC-BY 4.0 License.
Sharma et al., Robust Cough Detection Using Audio Signals.
https://datadryad.org/dataset/doi:10.5061/dryad.mkkwh717r
Model Citation
Chen, T. & Guestrin, C. (2016). XGBoost: A Scalable Tree Boosting System. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining.
Patent Language Stub
The system describes a software-based method for detecting and temporally localizing cough events from low-frequency audio signals using short-duration overlapping segmentation, MFCC-based statistical feature extraction, and gradient-boosted ensemble classification. The method includes post-processing logic to aggregate segment-level predictions into clinically interpretable events with confidence and severity grading, enabling scalable, non-invasive respiratory monitoring.
Disclaimer
This tool is a hackathon prototype and is NOT a medical device. The outputs are estimates intended for research and demonstration only. Do not use this tool to diagnose, treat, cure, or prevent any disease. If you have symptoms or concerns, consult a qualified clinician.
Team
Binary Brains
Akhshayaa S — SSN College of Engineering
Parvathi PR — SSN College of Engineering