🫀 AI-Augmented Heart Monitoring using ECG, PCG, and PPG Signals

This project provides a complete, AI-based system for non-invasive detection of cardiovascular diseases (CVD) by analyzing a combination of physiological signals — namely Electrocardiogram (ECG), Phonocardiogram (PCG), and Photoplethysmogram (PPG). By integrating signal processing, feature engineering, and machine learning (ML), this framework enables early diagnosis of conditions like Coronary Artery Disease (CAD), arrhythmias, valvular defects, and vascular dysfunctions.

---

🌐 Motivation

Cardiovascular diseases remain the leading cause of death globally. However, traditional diagnostic tools such as echocardiograms, angiograms, and stress tests are:

• Costly
• Not portable
• Often inaccessible in rural/remote areas

This project aims to address this gap by developing a low-cost, scalable, AI-driven diagnostic tool that can work with wearable sensors, digital stethoscopes, or simple optical devices to collect biosignals and analyze them using smart algorithms.

---

🧠 System Overview

📥 1. Signal Acquisition

The system works on three physiological signals:

• ECG (Electrical): Captures electrical impulses of the heart via electrodes placed on the skin (Lead II).
• PCG (Acoustic): Captures heart sounds through a digital stethoscope, usually stored as .wav files.
• PPG (Optical): Measures changes in blood volume via light absorption using infrared LEDs (smartwatches, pulse oximeters).

These signals are either recorded in real-time using biomedical hardware or taken from publicly available datasets.

---

🧹 2. Preprocessing

All raw signals are cleaned and filtered to ensure quality before analysis.

ECG:

• DC offset removal
• Bandpass filtering (0.5–100 Hz)
• High-pass & low-pass filtering
• Normalization

PCG:

• Bandpass filtering (20–950 Hz)
• Spike noise removal
• Amplitude normalization

PPG:

• Baseline wander correction
• Motion artifact filtering
• Moving average smoothing

✅ Why this is important: Noise can obscure important clinical information like murmurs or QRS complexes. Preprocessing improves both human and AI interpretability.

---

🧩 3. Segmentation & Fiducial Point Detection

ECG:

• Pan-Tompkins algorithm detects QRS complexes
• Additional logic used to locate P and T waves
• RR intervals and waveform intervals extracted

PCG:

• Uses Springer Hidden Semi-Markov Model (HSMM) to segment into S1 → Systole → S2 → Diastole
• Clear isolation of murmurs and valve sounds

PPG:

• Peaks, Dicrotic Notch, and Diastolic end points detected
• Used for Pulse Rate Variability (PRV) analysis (analogous to HRV from ECG)

🧠 These fiducial points are crucial for extracting clinical-grade features.

---

📊 4. Feature Extraction

The extracted features fall into three domains:

🔹 Time-Domain Features:

• RR Interval (ECG)
• Heart Rate (PPG)
• Systolic and Diastolic duration (PCG)
• Pulse Transit Time (ECG + PPG combined)

🔹 Frequency-Domain Features:

• Power Spectral Density (PSD)
• Band energy in specific ranges
• Heart sound energy distribution (PCG)

🔹 Cepstral/Statistical Features:

• Linear Frequency Cepstral Coefficients (LFCC)
• Skewness, Kurtosis, Variance
• Slope and Area under the curve

All features are normalized and compiled into feature matrices for model training.

---

🤖 5. Machine Learning & Deep Learning Models

✅ Classical ML Models:

• Support Vector Machine (SVM)
• Decision Trees / Random Forest
• K-Nearest Neighbors (KNN)

These models work well with engineered features like PSD or RR intervals.

🚀 Deep Learning:

• 1D CNNs: Learn from raw waveforms directly (especially ECG/PPG)
• 2D CNNs: Trained on spectrogram images (for PCG/PPG)
• LSTM/GRU (optional): For time series classification (sequential waveforms)

🔁 Hybrid Approach:

• Combines ECG + PPG or ECG + PCG for improved diagnostic accuracy
• Feature fusion or late-stage ensemble learning

---

📏 6. Model Evaluation Metrics

Each classifier is evaluated on:

• Accuracy
• Sensitivity (Recall) – true positive rate
• Specificity – true negative rate
• Precision and F1-Score
• ROC Curve and AUC

This ensures models are clinically reliable and not just technically accurate.

---

🎯 Clinical Use Cases & Applications

Application	Signal(s) Used	Model Type	Outcome
CAD Detection	PCG + ECG	SVM / CNN	75–85% Accuracy
Murmur Identification	PCG	LFCC + CNN	80%+ Sensitivity
Arrhythmia Detection	ECG	Pan-Tompkins + CNN	Real-time QRS classification
Pulse Monitoring / PRV	PPG	Time-domain ML	Wellness analytics
Wearable Health Alerts	ECG + PPG	Ensemble	Real-time abnormal alert

---

🔬 Technical Stack

• MATLAB (Signal processing, ML training, visualization)
• .wav / .txt / .mat file support
• Custom plotting for ECG/PCG/PPG annotation
• Adaptive filters (NLMS, Wiener)
• Springer segmentation model for PCG

---

📁 Repository Structure

├── ECG/            # ECG signal processing: Pan-Tompkins, plots, QRS detection
├── PCG/            # PCG segmentation and LFCC-based feature extraction
├── PPG/            # PPG peak detection, notch detection, PRV extraction
├── ML_models/      # Trained SVM, CNN scripts
├── Output_images/  # Visual plots of each signal and step
├── README.md       # Full documentation

---

🛠️ Signal Processing Workflow

① ECG Signal (Electrocardiogram)

• Pan-Tompkins algorithm used for QRS detection
• Filtering (low-pass, high-pass, derivative)
• Squaring & MWI (Moving Window Integration)
• Final output includes annotated P, QRS, T waves

Files: /ECG/ECG01.m, /Output_images/final_output_withPTQRS.png

---

② PCG Signal (Heart Sounds)

• Springer HSMM for segmentation into S1–S2 cycle
• Mel Spectrogram generation for audio signals
• LFCC features for murmur classification

Files: /PCG/*.wav, /PCG/PCGSegment.m, /PCG/runSpringerSegmentationAlgorithm.m

---

③ PPG Signal (Optical)

• Bandpass filtering
• Peak, notch, and end diastolic point detection
• Features: HR, PRV, dicrotic notch timing

Files: /PPG/PPG01.m, /PPG/Output_images/*.png

---

🤖 AI/ML Model Pipeline

CNN Model (for PCG Mel Spectrograms)

model = Sequential([
    Conv1D(32, 5, activation='relu'),
    MaxPooling1D(2),
    Conv1D(64, 5, activation='relu'),
    MaxPooling1D(2),
    Conv1D(128, 3, activation='relu'),
    GlobalAveragePooling1D(),
    Dense(128, activation='relu'),
    Dropout(0.4),
    Dense(5, activation='softmax')
])

---

• Input shape: (24, 64)
• Classes: artifact, aunlabelledtest, extrahls, murmur, normal
• Training Samples: 140
• Testing Samples: 36

---

🧪 Evaluation Summary

• Test Accuracy: 61.11%
• Class Imbalance Affects Performance

Confusion Matrix:

    [[5, 3, 0, 0, 0],
    [2, 6, 1, 0, 2],
    [0, 1, 0, 0, 3],
    [0, 1, 0, 5, 1],
    [0, 0, 0, 0, 6]]

---

Classification Report:

                    precision   recall   f1-score    support

       artifact       0.71      0.62      0.67         8
aunlabelledtest       0.55      0.55      0.55        11
       extrahls       0.00      0.00      0.00         4
         murmur       1.00      0.71      0.83         7
         normal       0.50      1.00      0.67         6

        accuracy                           0.61        36
       macro avg       0.55      0.58      0.54        36
    weighted avg       0.60      0.61      0.59        36

---

🔍 Murmur Detection & Segmentation

• PCG signal is segmented in 0.2 second energy windows
• High-energy regions are treated as potential murmurs
• Visualized with ellipse overlays to highlight:
  • a → b: duration of murmur
  • a → c: delay to peak
  • c → d: murmur amplitude (baseline to peak)
• Output visualization: murmur_analysis_graph.jpg

---

🧬 Signal Type Comparison

Signal	Captures	Sensor Type	Use Case
ECG	Electrical activity	Electrodes	QRS detection, HRV, PRV analysis
PCG	Heart sounds	Microphone/Steth	Murmur classification, S1/S2
PPG	Blood volume changes	IR-based Optical	Pulse rate, vascular health

---

🏥 Clinical Applications

• Early screening for:
  • Coronary Artery Disease (CAD)
  • Aortic Sclerosis
  • Hypertrophic Obstructive Cardiomyopathy (HOCM)
• Wearable health systems: smartwatches, stethoscope dongles
• Suitable for remote and low-resource healthcare environments
• Enables telemedicine and home-based diagnostics