🫀 CardioSense ResNet-18 — Phonocardiogram (PCG) Murmur Classifier

CardioSense is a deep residual convolutional neural network trained to detect pathological heart murmurs and valvular defects from phonocardiogram (PCG) stethoscope recordings. It achieves ROC-AUC 0.8041 and 82.78% sensitivity on the PhysioNet CinC 2016 benchmark.

---

Model Summary

Property	Value
Architecture	CardioResNet-18 (Deep 2D Residual CNN)
Task	Binary Audio Classification (Normal vs. Abnormal)
Input	Log-Mel Spectrogram (1, 128, time_steps)
Output	Softmax probabilities [P_normal, P_abnormal]
Decision Threshold	tau* = 0.4900 (Youden's J optimal)
Dataset	PhysioNet CinC 2016 PCG Challenge
ROC-AUC	0.8041
Sensitivity	82.78% (murmur recall)
Specificity	63.33%
Parameters	~11.2M
Framework	PyTorch 2.x

---

DSP & Preprocessing Pipeline

Raw Audio (.wav/.mp3)
   -> 1. Resample to 2,000 Hz
   -> 2. Downmix to Mono
   -> 3. 4th-Order Butterworth Bandpass (20 Hz to 200 Hz), zero-phase filtfilt
   -> 4. Peak Amplitude Normalization to [-1.0, 1.0]
   -> 5. Sliding Window Segmentation (4.0s window, 2.0s hop, 50% overlap)
   -> 6. 128-Band Log-Mel Spectrogram (n_fft=512, hop=128, fmin=20, fmax=400 Hz)
   -> 7. Tensor: (1, 128, time_steps)

---

Usage

Install

pip install torch librosa scipy soundfile numpy

Inference

import torch, torch.nn.functional as F
import numpy as np, librosa, scipy.signal, soundfile as sf
from cardio_resnet import CardioResNet18

model = CardioResNet18(num_classes=2)
model.load_state_dict(torch.load("cardiosense_resnet18.pt", map_location="cpu"))
model.eval()

def preprocess(path, sr=2000):
    y, orig_sr = sf.read(path)
    if y.ndim > 1: y = y.mean(axis=1)
    if orig_sr != sr: y = librosa.resample(y, orig_sr=orig_sr, target_sr=sr)
    nyq = sr / 2
    b, a = scipy.signal.butter(4, [20/nyq, 200/nyq], btype="bandpass")
    y = scipy.signal.filtfilt(b, a, y).astype(np.float32)
    return y / (np.max(np.abs(y)) + 1e-6), sr

def mel(seg, sr=2000):
    m = librosa.feature.melspectrogram(y=seg, sr=sr, n_fft=512, hop_length=128,
                                        n_mels=128, fmin=20, fmax=400, power=2)
    m = librosa.power_to_db(m, ref=np.max)
    mn, mx = m.min(), m.max()
    return ((2*(m-mn)/(mx-mn)-1) if mx-mn>1e-6 else np.zeros_like(m)).astype(np.float32)

audio, sr = preprocess("heart_sound.wav")
TAU, W, H = 0.4900, 4*sr, 2*sr
probs = []
with torch.no_grad():
    for i in range(0, len(audio), H):
        seg = audio[i:i+W]
        if len(seg) < W: seg = np.pad(seg, (0, W-len(seg)))
        x = torch.from_numpy(mel(seg)).unsqueeze(0).unsqueeze(0)
        probs.append(F.softmax(model(x), dim=1)[0,1].item())

p = float(np.mean(probs))
print(f"{'Abnormal' if p >= TAU else 'Normal'}  |  Abnormal Prob: {p:.3f}  |  Conf: {max(p,1-p)*100:.1f}%")

---

Evaluation (PhysioNet Test Set, n=301)

Metric	tau*=0.4900	tau=0.50
ROC-AUC	0.8041	0.8041
Sensitivity	82.78%	62.91%
Specificity	63.33%	78.67%
True Positives	125/151	95/151
False Negatives	26	56

---

Architecture

Input (B, 1, 128, T)
  Stem: Conv7x7(64) -> BN -> ReLU -> MaxPool
  Stage1: 2x ResBlock(64->64,  s=1)
  Stage2: 2x ResBlock(64->128, s=2)
  Stage3: 2x ResBlock(128->256,s=2)
  Stage4: 2x ResBlock(256->512,s=2)
  AdaptiveAvgPool2d(1,1)
  Dropout(0.35)
  Linear(512->2)

---

Clinical Disclaimer

CardioSense is an investigational acoustic screening tool. It does NOT provide a definitive medical diagnosis. All findings must be correlated with clinical history, stethoscope auscultation, and echocardiography by a licensed healthcare professional.