inference.py

"""
CardioScreen AI — Lightweight Inference Engine
No PyTorch. No transformers. Just signal processing.

Detects murmurs using spectral analysis of heart sounds:
- Heart rate via Hilbert envelope peak detection
- Murmur screening via frequency analysis between S1/S2 beats
  (murmurs produce abnormal energy in 100–600 Hz between heartbeats)
"""
import io
import os
import numpy as np

# Numpy 2.0 compatibility
if not hasattr(np, 'trapz'):
    np.trapz = np.trapezoid
if not hasattr(np, 'in1d'):
    np.in1d = np.isin

import librosa
import scipy.signal

# Lazy-loaded model dependencies
_cnn_model = None
_cnn_available = None
_finetuned_model = None
_finetuned_available = None
_resnet_model = None
_resnet_available = None
_gru_model = None
_gru_available = None

TARGET_SR = 16000
GRU_SR = 4000  # Bi-GRU uses 4kHz (McDonald et al.)

# 4-class murmur timing labels
CLASS_NAMES = ["Normal", "Systolic Murmur", "Diastolic Murmur", "Continuous Murmur"]
NUM_CLASSES  = 4

# Brief clinical notes per murmur type (shown in UI + PDF)
MURMUR_TYPE_NOTES = {
    "Normal":           "No murmur detected. Heart sounds are within normal limits.",
    "Systolic Murmur":  "Systolic murmur (S1→S2). Common causes: mitral insufficiency, "
                        "pulmonic or aortic stenosis, VSD. Recommend echocardiography.",
    "Diastolic Murmur": "Diastolic murmur (S2→S1). Uncommon in dogs — often indicates "
                        "aortic insufficiency. Specialist evaluation strongly advised.",
    "Continuous Murmur":"Continuous (machinery) murmur throughout the cardiac cycle. "
                        "Classic finding in patent ductus arteriosus (PDA). Urgent referral advised.",
}

print("CardioScreen AI engine loaded (lightweight mode)", flush=True)


# ─── Noise Reduction ─────────────────────────────────────────────────────────

def reduce_noise(y, sr, noise_percentile=10, smooth_ms=25):
    """
    Spectral gating noise reduction.
    Estimates a noise floor from the quietest frames, then subtracts it.
    """
    n_fft = 2048
    hop = n_fft // 4
    S = np.abs(librosa.stft(y, n_fft=n_fft, hop_length=hop))
    phase = np.angle(librosa.stft(y, n_fft=n_fft, hop_length=hop))

    # Estimate noise from the quietest frames
    frame_energy = np.mean(S ** 2, axis=0)
    threshold_idx = max(1, int(len(frame_energy) * noise_percentile / 100))
    quietest = np.argsort(frame_energy)[:threshold_idx]
    noise_profile = np.mean(S[:, quietest], axis=1, keepdims=True)

    # Subtract noise floor with soft masking (avoid artifacts)
    gain = np.maximum(S - noise_profile * 1.5, 0) / (S + 1e-10)

    # Smooth the gain to prevent musical noise
    smooth_frames = max(1, int((smooth_ms / 1000) * sr / hop))
    if smooth_frames > 1:
        kernel = np.ones(smooth_frames) / smooth_frames
        for i in range(gain.shape[0]):
            gain[i] = np.convolve(gain[i], kernel, mode='same')

    S_clean = S * gain
    y_clean = librosa.istft(S_clean * np.exp(1j * phase), hop_length=hop, length=len(y))
    return y_clean


def load_audio(audio_bytes: bytes):
    """Decode audio bytes → noise-reduce → bandpass filter → normalize."""
    import soundfile as sf
    y, sr = sf.read(io.BytesIO(audio_bytes))

    if len(y.shape) > 1:
        y = np.mean(y, axis=1)

    if sr != TARGET_SR:
        y = librosa.resample(y, orig_sr=sr, target_sr=TARGET_SR)

    # Step 1: Noise reduction (spectral gating)
    y = reduce_noise(y, TARGET_SR)

    # Step 2: Cardiac bandpass filter (25–600 Hz)
    nyq = 0.5 * TARGET_SR
    b, a = scipy.signal.butter(4, [25.0 / nyq, 600.0 / nyq], btype='band')
    y_filtered = scipy.signal.filtfilt(b, a, y)

    return librosa.util.normalize(y_filtered)


def calculate_bpm(y, sr):
    """
    Extract BPM and cardiac cycle count from a PCG recording.

    Uses TWO complementary methods:
    1. Envelope peak detection — reliable for normal/slow rates, provides peak positions
    2. Autocorrelation — robust at ALL heart rates, used to validate/correct BPM

    Key clinical considerations:
    - Each cardiac cycle produces TWO sounds: S1 (lub) and S2 (dup)
    - Peak detection with refractory window detects only S1 peaks
    - Autocorrelation finds the dominant cycle period (S1-to-S1) naturally
    - Valid canine heart rate range: 40–250 BPM
    """
    try:
        # ── 1. RMS Noise Gate ──────────────────────────────────────────────────
        rms = np.sqrt(np.mean(y ** 2))
        if rms < 0.01:
            return 0, 0, np.array([])

        # ── 2. Multi-scale Hilbert Envelope ───────────────────────────────────
        envelope = np.abs(scipy.signal.hilbert(y))

        fine_len = int(0.05 * sr) | 1        # 50 ms, must be odd
        coarse_len = int(0.12 * sr) | 1      # 120 ms, must be odd

        fine_env   = scipy.signal.savgol_filter(envelope, fine_len,   polyorder=2)
        coarse_env = scipy.signal.savgol_filter(fine_env, coarse_len, polyorder=2)
        coarse_env = np.clip(coarse_env, 0, None)

        # ── 3. Adaptive Height Threshold ──────────────────────────────────────
        v90    = np.percentile(coarse_env, 90)
        height = v90 * 0.40

        # ── 4. Peak Detection (proven parameters for normal/slow rates) ───────
        min_dist_sec  = 0.50           # 500 ms → max 120 BPM via peaks
        min_dist_samp = int(min_dist_sec * sr)
        prominence = v90 * 0.30

        peaks, props = scipy.signal.find_peaks(
            coarse_env,
            distance=min_dist_samp,
            height=height,
            prominence=prominence,
        )

        # Fallback for quiet recordings
        if len(peaks) < 2:
            peaks, _ = scipy.signal.find_peaks(
                coarse_env,
                distance=min_dist_samp,
                height=v90 * 0.15,
            )
            if len(peaks) < 2:
                return 0, 0, np.array([])

        # Post-detection refractory check
        refractory = int(0.40 * sr)
        clean_peaks = [peaks[0]]
        for pk in peaks[1:]:
            if pk - clean_peaks[-1] >= refractory:
                clean_peaks.append(pk)

        peaks = np.array(clean_peaks)
        if len(peaks) < 2:
            return 0, 0, peaks

        # Peak-based BPM
        intervals = np.diff(peaks)
        median_interval = np.median(intervals)
        peak_bpm = (60.0 * sr) / median_interval

        # ── 5. Autocorrelation BPM (robust at all heart rates) ────────────────
        # Autocorrelation finds the dominant periodicity in the envelope.
        # The full cardiac cycle (S1+S2+pause) repeats, so the autocorrelation
        # peak corresponds to the S1-to-S1 interval — regardless of heart rate.
        acorr_bpm = 0
        try:
            # Normalize envelope for autocorrelation
            env_norm = coarse_env - np.mean(coarse_env)
            autocorr = np.correlate(env_norm, env_norm, mode='full')
            autocorr = autocorr[len(autocorr) // 2:]  # keep positive lags only
            autocorr = autocorr / autocorr[0]  # normalize

            # Search for dominant peak in physiological range:
            # 40 BPM → 1.5s period, 250 BPM → 0.24s period
            min_lag = int(0.24 * sr)   # 250 BPM
            max_lag = int(1.5 * sr)    # 40 BPM

            if max_lag <= len(autocorr):
                search_region = autocorr[min_lag:max_lag]
                if len(search_region) > 0:
                    acorr_peaks, _ = scipy.signal.find_peaks(search_region, prominence=0.1)
                    if len(acorr_peaks) > 0:
                        # First prominent peak = fundamental cardiac cycle period
                        best_lag = acorr_peaks[0] + min_lag
                        acorr_bpm = (60.0 * sr) / best_lag
        except Exception:
            pass

        # ── 6. Choose best BPM estimate ───────────────────────────────────────
        # Peak detection is reliable for normal rates but caps at ~120 BPM.
        # Autocorrelation works at all rates. Use autocorrelation when it
        # detects a meaningfully faster rate (suggesting tachycardia that
        # peak detection is missing).
        if acorr_bpm > 0 and acorr_bpm > peak_bpm * 1.3:
            # Autocorrelation found significantly faster rate → tachycardia
            bpm = acorr_bpm
        else:
            bpm = peak_bpm

        # Clamp to physiological canine range (40–250 BPM)
        bpm = int(max(40, min(250, bpm)))
        return bpm, len(peaks), peaks

    except Exception as e:
        print(f"BPM Error: {e}", flush=True)
        return 0, 0, np.array([])



def detect_murmur(y, sr, peaks):
    """
    Murmur detection via spectral analysis of inter-beat intervals.
    
    Normal heart sounds (S1, S2) are brief, low-frequency thuds.
    Murmurs are prolonged, higher-frequency sounds BETWEEN the beats.
    
    We analyze the spectral content between detected heartbeats:
    - High energy ratio in 100-600Hz between beats → murmur likely
    - Low spectral entropy → normal clean silence between beats  
    - High spectral entropy → turbulent flow (murmur indicator)
    """
    if len(peaks) < 3:
        return {
            "label": "Insufficient Data",
            "confidence": 0.0,
            "is_disease": False,
            "details": "Need at least 3 heartbeats for analysis",
            "all_classes": [
                {"label": "Insufficient Data", "probability": 1.0},
            ]
        }

    # Analyze the intervals BETWEEN heartbeats
    inter_beat_energies = []
    inter_beat_entropies = []
    beat_energies = []

    for i in range(len(peaks) - 1):
        # Region around the beat itself (±50ms)
        beat_start = max(0, peaks[i] - int(0.05 * sr))
        beat_end = min(len(y), peaks[i] + int(0.05 * sr))
        beat_segment = y[beat_start:beat_end]

        # Region between beats (the "gap" where murmurs live)
        gap_start = peaks[i] + int(0.08 * sr)  # skip 80ms after beat
        gap_end = peaks[i + 1] - int(0.08 * sr)  # stop 80ms before next beat

        if gap_end <= gap_start:
            continue

        gap_segment = y[gap_start:gap_end]

        # RMS energy of the beat vs the gap
        beat_rms = np.sqrt(np.mean(beat_segment ** 2)) if len(beat_segment) > 0 else 0
        gap_rms = np.sqrt(np.mean(gap_segment ** 2)) if len(gap_segment) > 0 else 0

        beat_energies.append(beat_rms)
        inter_beat_energies.append(gap_rms)

        # Spectral entropy of the gap (high entropy = turbulent flow = murmur)
        if len(gap_segment) > 256:
            freqs = np.abs(np.fft.rfft(gap_segment))
            freqs = freqs / (np.sum(freqs) + 1e-12)
            entropy = -np.sum(freqs * np.log2(freqs + 1e-12))
            inter_beat_entropies.append(entropy)

    if not inter_beat_energies:
        return {
            "label": "Insufficient Data",
            "confidence": 0.0,
            "is_disease": False,
            "details": "Could not isolate inter-beat intervals",
            "all_classes": [
                {"label": "Insufficient Data", "probability": 1.0},
            ]
        }

    # Key metrics
    avg_beat_energy = np.mean(beat_energies)
    avg_gap_energy = np.mean(inter_beat_energies)
    energy_ratio = avg_gap_energy / (avg_beat_energy + 1e-12)
    avg_entropy = np.mean(inter_beat_entropies) if inter_beat_entropies else 0

    # Inter-beat energy consistency (murmurs are consistent; noise is random)
    gap_energy_std = np.std(inter_beat_energies) / (avg_gap_energy + 1e-12)
    consistency = 1.0 - min(1.0, gap_energy_std)  # High = consistent inter-beat energy

    # High-frequency energy ratio (murmurs have more energy in 200-600Hz band)
    # Analyze frequency distribution in the gaps
    hf_ratios = []
    for i in range(len(peaks) - 1):
        gap_start = peaks[i] + int(0.08 * sr)
        gap_end = peaks[i + 1] - int(0.08 * sr)
        if gap_end <= gap_start:
            continue
        gap_segment = y[gap_start:gap_end]
        if len(gap_segment) > 512:
            fft_mag = np.abs(np.fft.rfft(gap_segment))
            freqs_hz = np.fft.rfftfreq(len(gap_segment), 1.0 / sr)
            # Energy in murmur band (150-500Hz) vs total
            murmur_band = np.sum(fft_mag[(freqs_hz >= 150) & (freqs_hz <= 500)])
            total = np.sum(fft_mag) + 1e-12
            hf_ratios.append(murmur_band / total)
    
    hf_ratio = np.mean(hf_ratios) if hf_ratios else 0.0

    # Also extract MFCCs for overall spectral characterization
    mfccs = librosa.feature.mfcc(y=y, sr=sr, n_mfcc=13)
    mfcc_var = np.mean(np.var(mfccs, axis=1))

    # ─── Trained classifier (logistic regression on 21 canine recordings) ───
    # Features: [energy_ratio, consistency, hf_ratio, entropy, mfcc_var]
    # Trained on VetCPD (5) + Hannover Examples (9) + Hannover Grading (7)
    # Results: 95% accuracy, 94% sensitivity, 100% specificity
    #
    # Model weights (scikit-learn logistic regression):
    SCALER_MEAN = [0.4315, 0.7709, 0.2588, 7.1566, 220.9825]
    SCALER_STD  = [0.1573, 0.0888, 0.1294, 0.7728, 124.5063]
    WEIGHTS     = [1.2507, 0.3728, -0.4740, -0.3317, 1.1285]
    INTERCEPT   = 0.8248
    SCREENING_THRESHOLD = 0.40  # Optimized for screening sensitivity

    # Scale features
    raw_features = [energy_ratio, consistency, hf_ratio, avg_entropy, mfcc_var]
    scaled = [(f - m) / (s + 1e-12) for f, m, s in zip(raw_features, SCALER_MEAN, SCALER_STD)]

    # Logistic regression: P(murmur) = sigmoid(w·x + b)
    logit = sum(w * x for w, x in zip(WEIGHTS, scaled)) + INTERCEPT
    murmur_prob = float(1.0 / (1.0 + np.exp(-logit)))
    normal_prob = float(1.0 - murmur_prob)

    is_murmur = bool(murmur_prob >= SCREENING_THRESHOLD)

    return {
        "label": "Murmur" if is_murmur else "Normal",
        "confidence": round(murmur_prob if is_murmur else normal_prob, 4),
        "is_disease": is_murmur,
        "details": f"Energy ratio: {energy_ratio:.3f}, HF ratio: {hf_ratio:.3f}, Consistency: {consistency:.3f}, Entropy: {avg_entropy:.1f}, MFCC var: {mfcc_var:.1f}",
        "all_classes": [
            {"label": "Normal", "probability": round(normal_prob, 4)},
            {"label": "Murmur", "probability": round(murmur_prob, 4)},
        ]
    }


# ─── CNN Inference ────────────────────────────────────────────────────────────
def _load_cnn_model():
    """Lazy-load the trained CNN model (only once)."""
    global _cnn_model, _cnn_available
    if _cnn_available is not None:
        return _cnn_available

    try:
        import torch
        import torch.nn as nn

        class HeartSoundCNN(nn.Module):
            def __init__(self, num_classes=NUM_CLASSES):
                super().__init__()
                self.features = nn.Sequential(
                    nn.Conv2d(1, 32, 3, padding=1), nn.BatchNorm2d(32), nn.ReLU(), nn.MaxPool2d(2),
                    nn.Conv2d(32, 64, 3, padding=1), nn.BatchNorm2d(64), nn.ReLU(), nn.MaxPool2d(2),
                    nn.Conv2d(64, 128, 3, padding=1), nn.BatchNorm2d(128), nn.ReLU(),
                    nn.AdaptiveAvgPool2d((1, 1)),
                )
                self.classifier = nn.Sequential(nn.Dropout(0.3), nn.Linear(128, num_classes))

            def forward(self, x):
                x = self.features(x)
                return self.classifier(x.view(x.size(0), -1))

        weights_path = os.path.join(os.path.dirname(os.path.abspath(__file__)),
                                    "weights", "cnn_heart_classifier.pt")

        # If weights not found locally, try downloading from HF repo
        if not os.path.exists(weights_path):
            try:
                from huggingface_hub import hf_hub_download
                print("Downloading CNN weights from HF...", flush=True)
                os.makedirs(os.path.dirname(weights_path), exist_ok=True)
                hf_hub_download(
                    repo_id="mahmoud611/cardioscreen-api",
                    filename="weights/cnn_heart_classifier.pt",
                    repo_type="space",
                    local_dir=os.path.dirname(os.path.dirname(weights_path)),
                )
                print("CNN weights downloaded ✓", flush=True)
            except Exception as dl_err:
                print(f"CNN weights not found and download failed: {dl_err}", flush=True)
                _cnn_available = False
                return False

        model = HeartSoundCNN()
        model.load_state_dict(torch.load(weights_path, map_location="cpu", weights_only=True))
        model.eval()
        _cnn_model = model
        _cnn_available = True
        print("CNN model loaded ✓", flush=True)
        return True

    except ImportError:
        print("PyTorch not installed — CNN disabled", flush=True)
        _cnn_available = False
        return False
    except Exception as e:
        print(f"CNN load error: {e}", flush=True)
        _cnn_available = False
        return False


def predict_cnn(y, sr):
    """
    Classify audio using the trained Mel-spectrogram CNN (4-class).
    Returns Normal / Systolic Murmur / Diastolic Murmur / Continuous Murmur.
    """
    if not _load_cnn_model():
        return None

    import torch

    # Config must match training
    N_MELS, N_FFT, HOP = 64, 1024, 512
    CLIP_SEC   = 5
    target_len = sr * CLIP_SEC

    # Split into 5-sec clips (with timestamps)
    clips = []
    clip_starts = []   # start sample index for each clip
    if len(y) >= target_len:
        for s in range(0, len(y) - target_len + 1, target_len):
            clips.append(y[s:s + target_len])
            clip_starts.append(s)
    else:
        clips.append(np.pad(y, (0, target_len - len(y))))
        clip_starts.append(0)

    # Classify each clip
    target_t = int(np.ceil(CLIP_SEC * sr / HOP))
    probs = []
    for clip in clips:
        S    = librosa.feature.melspectrogram(y=clip, sr=sr, n_mels=N_MELS, n_fft=N_FFT, hop_length=HOP)
        S_db = librosa.power_to_db(S, ref=np.max)
        S_db = (S_db - S_db.min()) / (S_db.max() - S_db.min() + 1e-8)

        if S_db.shape[1] < target_t:
            S_db = np.pad(S_db, ((0, 0), (0, target_t - S_db.shape[1])))
        else:
            S_db = S_db[:, :target_t]

        tensor = torch.FloatTensor(S_db).unsqueeze(0).unsqueeze(0)  # (1,1,64,T)
        with torch.no_grad():
            logits = _cnn_model(tensor)
            p = torch.softmax(logits, dim=1)[0]  # shape: (NUM_CLASSES,)
            probs.append(p.numpy())

    # Build per-segment results (for timeline + table in the UI)
    MURMUR_THRESHOLD_SEG = 0.30
    segments = []
    for i, (p, start_samp) in enumerate(zip(probs, clip_starts)):
        seg_normal_p = float(p[0])
        seg_murmur_p = float(1.0 - seg_normal_p)
        seg_pred_idx = int(np.argmax(p))
        is_seg_murmur = seg_murmur_p > MURMUR_THRESHOLD_SEG
        if is_seg_murmur and seg_pred_idx == 0:
            seg_pred_idx = int(np.argmax(p[1:])) + 1
        segments.append({
            "segment_idx":  i,
            "start_sec":    round(start_samp / sr, 2),
            "end_sec":      round((start_samp + target_len) / sr, 2),
            "top_label":    CLASS_NAMES[seg_pred_idx] if is_seg_murmur else "Normal",
            "is_murmur":    is_seg_murmur,
            "probs":        {CLASS_NAMES[j]: round(float(p[j]), 4) for j in range(NUM_CLASSES)},
            "murmur_prob":  round(seg_murmur_p, 4),
        })

    # Average probabilities across clips
    avg_prob = np.mean(probs, axis=0)  # (NUM_CLASSES,)

    # --- Murmur detection threshold (binary: Normal vs. any murmur type) ---
    # P(any murmur) = 1 - P(Normal).  Threshold 0.30 keeps high sensitivity.
    normal_p   = float(avg_prob[0])
    murmur_p   = float(1.0 - normal_p)   # P(any murmur type)
    MURMUR_THRESHOLD = 0.30
    is_murmur  = murmur_p > MURMUR_THRESHOLD

    # --- Murmur type: argmax over 4 classes ---
    predicted_class = int(np.argmax(avg_prob))
    # If we detect a murmur but the model's top class is Normal (border case),
    # fall back to the highest-probability murmur subclass.
    if is_murmur and predicted_class == 0:
        predicted_class = int(np.argmax(avg_prob[1:])) + 1

    murmur_type       = CLASS_NAMES[predicted_class]
    type_confidence   = float(avg_prob[predicted_class])
    overall_label     = murmur_type if is_murmur else "Normal"
    overall_conf      = round(murmur_p if is_murmur else normal_p, 4)

    return {
        "label":       overall_label,
        "confidence":  overall_conf,
        "is_disease":  bool(is_murmur),
        "murmur_type": murmur_type,
        "murmur_type_confidence": round(type_confidence, 4),
        "murmur_type_note": MURMUR_TYPE_NOTES.get(murmur_type, ""),
        "method":      "CNN (Mel-Spectrogram, 4-class)",
        "clips_analyzed": len(clips),
        "segments":    segments,   # per-5s-window breakdown for UI timeline
        "all_classes": [
            {"label": CLASS_NAMES[i], "probability": round(float(avg_prob[i]), 4)}
            for i in range(NUM_CLASSES)
        ],
    }


# ─── Fine-tuned CNN Inference ─────────────────────────────────────────────────
def _load_finetuned_model():
    """Lazy-load the fine-tuned CNN model."""
    global _finetuned_model, _finetuned_available
    if _finetuned_available is not None:
        return _finetuned_available
    try:
        import torch
        import torch.nn as nn

        class HeartSoundCNN(nn.Module):
            def __init__(self, num_classes=NUM_CLASSES):
                super().__init__()
                self.features = nn.Sequential(
                    nn.Conv2d(1, 32, 3, padding=1), nn.BatchNorm2d(32), nn.ReLU(), nn.MaxPool2d(2),
                    nn.Conv2d(32, 64, 3, padding=1), nn.BatchNorm2d(64), nn.ReLU(), nn.MaxPool2d(2),
                    nn.Conv2d(64, 128, 3, padding=1), nn.BatchNorm2d(128), nn.ReLU(),
                    nn.AdaptiveAvgPool2d((1, 1)),
                )
                self.classifier = nn.Sequential(nn.Dropout(0.5), nn.Linear(128, num_classes))
            def forward(self, x):
                x = self.features(x)
                return self.classifier(x.view(x.size(0), -1))

        weights_path = os.path.join(os.path.dirname(os.path.abspath(__file__)),
                                    "weights", "cnn_finetuned.pt")
        if not os.path.exists(weights_path):
            print("Fine-tuned CNN weights not found", flush=True)
            _finetuned_available = False
            return False

        model = HeartSoundCNN()
        model.load_state_dict(torch.load(weights_path, map_location="cpu", weights_only=True))
        model.eval()
        _finetuned_model = model
        _finetuned_available = True
        print("Fine-tuned CNN loaded ✓", flush=True)
        return True
    except Exception as e:
        print(f"Fine-tuned CNN load error: {e}", flush=True)
        _finetuned_available = False
        return False


def predict_finetuned(y, sr):
    """Classify using the fine-tuned CNN (2-step transfer learning)."""
    if not _load_finetuned_model():
        return None
    import torch
    N_MELS, N_FFT, HOP, CLIP_SEC = 64, 1024, 512, 5
    target_len = sr * CLIP_SEC
    clips = [y[s:s+target_len] for s in range(0, len(y)-target_len+1, target_len)] if len(y) >= target_len else [np.pad(y, (0, target_len-len(y)))]
    probs = []
    for clip in clips:
        S = librosa.feature.melspectrogram(y=clip, sr=sr, n_mels=N_MELS, n_fft=N_FFT, hop_length=HOP)
        S_db = librosa.power_to_db(S, ref=np.max)
        S_db = (S_db - S_db.mean()) / (S_db.std() + 1e-8)
        tensor = torch.FloatTensor(S_db).unsqueeze(0).unsqueeze(0)
        with torch.no_grad():
            probs.append(torch.softmax(_finetuned_model(tensor), 1)[0].numpy())
    avg = np.mean(probs, 0)
    pred = int(np.argmax(avg))
    return {
        "label": CLASS_NAMES[pred], "confidence": round(float(avg[pred]), 4),
        "is_disease": pred != 0, "method": "Fine-tuned CNN (2-step Transfer Learning)",
        "all_classes": [{"label": CLASS_NAMES[i], "probability": round(float(avg[i]), 4)} for i in range(NUM_CLASSES)],
    }


# ─── ResNet-18 Inference ─────────────────────────────────────────────────────
def _load_resnet_model():
    """Lazy-load the ImageNet-pretrained ResNet-18 model."""
    global _resnet_model, _resnet_available
    if _resnet_available is not None:
        return _resnet_available
    try:
        import torch
        import torch.nn as nn
        from torchvision.models import resnet18

        weights_path = os.path.join(os.path.dirname(os.path.abspath(__file__)),
                                    "weights", "cnn_resnet_classifier.pt")
        if not os.path.exists(weights_path):
            print("ResNet-18 weights not found", flush=True)
            _resnet_available = False
            return False

        model = resnet18(weights=None)
        model.fc = nn.Sequential(nn.Dropout(0.3), nn.Linear(512, NUM_CLASSES))
        model.load_state_dict(torch.load(weights_path, map_location="cpu", weights_only=True))
        model.eval()
        _resnet_model = model
        _resnet_available = True
        print("ResNet-18 loaded ✓", flush=True)
        return True
    except Exception as e:
        print(f"ResNet-18 load error: {e}", flush=True)
        _resnet_available = False
        return False


def predict_resnet(y, sr):
    """Classify using ImageNet-pretrained ResNet-18 (frozen backbone)."""
    if not _load_resnet_model():
        return None
    import torch
    import torch.nn.functional as F
    N_MELS, N_FFT, HOP, CLIP_SEC = 64, 1024, 512, 5
    # Apply bandpass 50-500 Hz (Bisgin et al.)
    nyq = sr / 2
    b, a = scipy.signal.butter(4, [50/nyq, 500/nyq], btype='band')
    y_bp = scipy.signal.filtfilt(b, a, y).astype(np.float32)
    target_len = sr * CLIP_SEC
    clips = [y_bp[s:s+target_len] for s in range(0, len(y_bp)-target_len+1, target_len)] if len(y_bp) >= target_len else [np.pad(y_bp, (0, target_len-len(y_bp)))]
    probs = []
    for clip in clips:
        S = librosa.feature.melspectrogram(y=clip, sr=sr, n_mels=N_MELS, n_fft=N_FFT, hop_length=HOP)
        S_db = librosa.power_to_db(S, ref=np.max)
        S_db = (S_db - S_db.mean()) / (S_db.std() + 1e-8)
        t = torch.FloatTensor(S_db).unsqueeze(0)
        t = F.interpolate(t.unsqueeze(0), (224, 224), mode='bilinear', align_corners=False).squeeze(0)
        t = t.expand(3, -1, -1).unsqueeze(0)
        with torch.no_grad():
            probs.append(torch.softmax(_resnet_model(t), 1)[0].numpy())
    avg = np.mean(probs, 0)
    pred = int(np.argmax(avg))
    return {
        "label": CLASS_NAMES[pred], "confidence": round(float(avg[pred]), 4),
        "is_disease": pred != 0, "method": "ResNet-18 (ImageNet Pretrained)",
        "all_classes": [{"label": CLASS_NAMES[i], "probability": round(float(avg[i]), 4)} for i in range(NUM_CLASSES)],
    }


# ─── Bi-GRU Inference (McDonald et al., Cambridge 2024) ──────────────────────
def _load_gru_model():
    """Lazy-load the Bi-GRU (McDonald et al.) model."""
    global _gru_model, _gru_available
    if _gru_available is not None:
        return _gru_available
    try:
        import torch
        import torch.nn as nn

        class HeartSoundGRU(nn.Module):
            def __init__(self, input_dim=129, hidden_dim=64, num_layers=2,
                         num_classes=2, dropout=0.4):
                super().__init__()
                self.input_norm = nn.LayerNorm(input_dim)
                self.gru = nn.GRU(input_dim, hidden_dim, num_layers,
                                  batch_first=True, bidirectional=True,
                                  dropout=dropout if num_layers > 1 else 0)
                self.attention = nn.Sequential(
                    nn.Linear(hidden_dim * 2, 64), nn.Tanh(), nn.Linear(64, 1))
                self.classifier = nn.Sequential(
                    nn.Dropout(dropout), nn.Linear(hidden_dim * 2, 32),
                    nn.ReLU(), nn.Dropout(dropout * 0.5), nn.Linear(32, num_classes))
            def forward(self, x):
                x = self.input_norm(x)
                gru_out, _ = self.gru(x)
                attn = torch.softmax(self.attention(gru_out), dim=1)
                ctx = (gru_out * attn).sum(dim=1)
                return self.classifier(ctx)

        weights_path = os.path.join(os.path.dirname(os.path.abspath(__file__)),
                                    "weights", "gru_canine_finetuned.pt")
        if not os.path.exists(weights_path):
            print("Bi-GRU weights not found", flush=True)
            _gru_available = False
            return False

        model = HeartSoundGRU()
        model.load_state_dict(torch.load(weights_path, map_location="cpu", weights_only=True))
        model.eval()
        _gru_model = model
        _gru_available = True
        print("Bi-GRU (McDonald) loaded ✓", flush=True)
        return True
    except Exception as e:
        print(f"Bi-GRU load error: {e}", flush=True)
        _gru_available = False
        return False


def predict_gru(y, sr):
    """
    Classify using Bi-GRU with log-spectrogram (McDonald et al., 2024).

    Uses 5-second windows with 2.5-second stride (50% overlap), matching the
    AryanGit720 reference implementation for clinical segment-level analysis.
    Windows: 0-5s, 2.5-7.5s, 5-10s, 7.5-12.5s, ...
    """
    if not _load_gru_model():
        return None
    import torch

    # Resample to 4 kHz (GRU training SR)
    y_4k = librosa.resample(y, orig_sr=sr, target_sr=GRU_SR)

    N_FFT_G  = 256
    HOP_G    = 64
    CLIP_SEC = 5
    STEP_SEC = 2.5          # 50% overlap stride

    target_len = int(GRU_SR * CLIP_SEC)   # 20 000 samples @ 4 kHz
    step_len   = int(GRU_SR * STEP_SEC)   # 10 000 samples

    GRU_BINARY_NAMES = ["Normal", "Murmur"]
    MURMUR_THRESHOLD = 0.50   # standard 50/50 threshold for binary GRU

    # ── Build overlapping windows ──────────────────────────────────────────
    starts = []
    if len(y_4k) >= target_len:
        s = 0
        while s + target_len <= len(y_4k):
            starts.append(s)
            s += step_len
    else:
        starts = [0]   # short recording: single padded clip

    probs      = []   # (N_windows, 2)
    raw_starts = []   # sample start in y_4k for each window

    for s in starts:
        clip = y_4k[s: s + target_len]
        if len(clip) < target_len:
            clip = np.pad(clip, (0, target_len - len(clip)))

        S     = np.abs(librosa.stft(clip, n_fft=N_FFT_G, hop_length=HOP_G)) ** 2
        log_S = np.log1p(S)
        log_S = (log_S - log_S.mean()) / (log_S.std() + 1e-8)
        spec  = log_S.T.astype(np.float32)   # (time_frames, freq_bins)

        t = torch.FloatTensor(spec).unsqueeze(0)
        with torch.no_grad():
            p = torch.softmax(_gru_model(t), 1)[0].numpy()
        probs.append(p)
        raw_starts.append(s)

    # ── Per-segment results (for timeline + table in UI) ──────────────────
    segments = []
    for i, (p, s_samp) in enumerate(zip(probs, raw_starts)):
        murmur_p   = float(p[1])
        is_seg_mur = murmur_p >= MURMUR_THRESHOLD
        start_sec  = round(s_samp / GRU_SR, 2)
        end_sec    = round((s_samp + target_len) / GRU_SR, 2)
        segments.append({
            "segment_idx": i,
            "start_sec":   start_sec,
            "end_sec":     end_sec,
            "top_label":   "Murmur" if is_seg_mur else "Normal",
            "is_murmur":   is_seg_mur,
            "murmur_prob": round(murmur_p, 4),
            "probs": {
                "Normal": round(float(p[0]), 4),
                "Murmur": round(murmur_p, 4),
            },
        })

    # ── Record-level aggregate (average across all windows) ───────────────
    avg  = np.mean(probs, axis=0)
    pred = int(np.argmax(avg))
    is_murmur = bool(avg[1] >= MURMUR_THRESHOLD)
    label = "Murmur" if is_murmur else "Normal"

    return {
        "label":      label,
        "confidence": round(float(avg[1] if is_murmur else avg[0]), 4),
        "is_disease": is_murmur,
        "method":     "Bi-GRU Binary (McDonald et al., Cambridge 2024)",
        "clips_analyzed": len(probs),
        "segments":   segments,   # per-2.5s-step window breakdown for UI
        "all_classes": [
            {"label": GRU_BINARY_NAMES[i], "probability": round(float(avg[i]), 4)}
            for i in range(2)
        ],
    }


# ─── Signal Quality Scoring ──────────────────────────────────────────────────

def score_quality(y, sr, peaks):
    """
    Rate recording quality 0-100 based on:
    - SNR (signal vs noise floor)
    - Peak regularity (consistent inter-beat intervals)
    - Clipping detection (microphone overload)
    - Duration adequacy (minimum useful length)
    """
    duration = len(y) / sr
    warnings = []
    score = 100

    # ── 1. Duration Check ────────────────────────────────────────────────
    if duration < 2.0:
        score -= 40
        warnings.append("Recording too short (< 2s) — please record for at least 5 seconds")
    elif duration < 5.0:
        score -= 15
        warnings.append("Short recording — 5+ seconds recommended for accuracy")

    # ── 2. SNR (Signal-to-Noise Ratio) ───────────────────────────────────
    rms = np.sqrt(np.mean(y ** 2))
    if rms < 0.005:
        score -= 35
        warnings.append("Very low signal — ensure microphone is close to the chest")
    elif rms < 0.02:
        score -= 15
        warnings.append("Weak signal — try moving the microphone closer")

    # Estimate SNR from peak vs inter-peak energy
    if len(peaks) >= 2:
        peak_window = int(0.08 * sr)  # 80ms around each peak
        peak_energy = 0
        noise_energy = 0
        noise_mask = np.ones(len(y), dtype=bool)

        for pk in peaks:
            start = max(0, pk - peak_window)
            end = min(len(y), pk + peak_window)
            peak_energy += np.sum(y[start:end] ** 2)
            noise_mask[start:end] = False

        noise_samples = y[noise_mask]
        if len(noise_samples) > 0:
            noise_energy = np.sum(noise_samples ** 2)
            if noise_energy > 0:
                snr_db = 10 * np.log10(peak_energy / noise_energy + 1e-10)
            else:
                snr_db = 30.0  # Very clean
        else:
            snr_db = 15.0

        if snr_db < 3:
            score -= 25
            warnings.append("High background noise detected")
        elif snr_db < 8:
            score -= 10
            warnings.append("Moderate background noise")
    else:
        snr_db = 0.0
        score -= 20
        warnings.append("Unable to detect heartbeat peaks — poor signal quality")

    # ── 3. Peak Regularity ───────────────────────────────────────────────
    if len(peaks) >= 3:
        intervals = np.diff(peaks) / sr  # in seconds
        cv = np.std(intervals) / (np.mean(intervals) + 1e-10)  # coefficient of variation
        if cv > 0.5:
            score -= 15
            warnings.append("Irregular heartbeat intervals — possible motion artifact")
        elif cv > 0.3:
            score -= 5
            warnings.append("Slightly irregular intervals")
    
    # ── 4. Clipping Detection ────────────────────────────────────────────
    clip_ratio = np.mean(np.abs(y) > 0.98)
    if clip_ratio > 0.05:
        score -= 20
        warnings.append("Audio clipping detected — reduce microphone sensitivity")
    elif clip_ratio > 0.01:
        score -= 8
        warnings.append("Minor audio clipping")

    # Clamp score
    score = max(0, min(100, score))

    # Grade
    if score >= 70:
        grade = "Good"
    elif score >= 40:
        grade = "Fair"
    else:
        grade = "Poor"

    return {
        "score": score,
        "grade": grade,
        "warnings": warnings,
        "metrics": {
            "snr_db": round(snr_db, 1) if len(peaks) >= 2 else None,
            "duration_s": round(duration, 1),
            "peak_count": int(len(peaks)),
            "clip_ratio": round(float(clip_ratio) * 100, 1),
        }
    }


# ─── Rhythm Analysis ──────────────────────────────────────────────────

def analyze_rhythm(peaks, sr):
    """
    Classify cardiac rhythm pattern from detected S1 peak intervals.

    Uses three metrics:
    1. Coefficient of Variation (CV) — overall regularity
    2. Poincaré plot analysis (SD1/SD2 ratio) — short vs long-term variability
    3. Alternating pattern detection — Regularly Irregular (AV block) vs random (AF)

    Rhythm labels:
    - Regular Sinus Rhythm   : CV < 0.10, consistent intervals
    - Sinus Arrhythmia       : CV 0.10–0.20, gradual variation (physiologic in dogs)
    - Regularly Irregular    : alternating short-long pattern — 2nd degree AV block
    - Irregularly Irregular  : random variation — Atrial Fibrillation pattern
    """
    if len(peaks) < 3:
        return {
            "label": "Insufficient Data",
            "short_label": "N/A",
            "confidence": 0.0,
            "cv": None,
            "note": "Need ≥3 beats for rhythm analysis",
            "color": "#94a3b8",
        }

    intervals = np.diff(peaks).astype(float) / sr  # RR intervals in seconds
    mean_rr = np.mean(intervals)
    std_rr  = np.std(intervals)
    cv      = std_rr / (mean_rr + 1e-10)

    # ── Poincaré plot: SD1 (beat-to-beat) vs SD2 (trend) ────────────
    # SD1 measures short-term variability (perpendicular to identity line)
    # SD2 measures long-term variability (along identity line)
    # AF: SD1 ≈ SD2 (random); Sinus: SD1 << SD2 (structured)
    if len(intervals) >= 3:
        rr_n  = intervals[:-1]   # RR_n
        rr_n1 = intervals[1:]    # RR_n+1
        diff  = rr_n1 - rr_n
        sd1   = np.std(diff) / np.sqrt(2)
        sd2   = np.sqrt(2 * std_rr**2 - sd1**2 + 1e-12)
        sd1_sd2_ratio = sd1 / (sd2 + 1e-10)  # >0.8 suggests AF
    else:
        sd1_sd2_ratio = 0.0

    # ── Alternating pattern: Regularly Irregular ─────────────────
    # In 2nd degree AV block, every other interval is longer (dropped beat)
    # Detect by checking if even and odd intervals are consistently different
    alternating = False
    if len(intervals) >= 4:
        even_mean = np.mean(intervals[0::2])
        odd_mean  = np.mean(intervals[1::2])
        ratio = max(even_mean, odd_mean) / (min(even_mean, odd_mean) + 1e-10)
        even_std = np.std(intervals[0::2])
        odd_std  = np.std(intervals[1::2])
        # Alternating if: groups are consistently different (ratio>1.3)
        # AND each group is internally consistent (low internal CV)
        internal_cv = (even_std + odd_std) / (2 * mean_rr + 1e-10)
        alternating = (ratio > 1.3) and (internal_cv < 0.10) and (cv > 0.10)

    # ── Classification logic ───────────────────────────────────
    if len(peaks) < 4:
        # Too few beats for confident rhythm classification
        label       = "Regular Sinus Rhythm"
        short_label = "Regular"
        note        = "Appears regular — record longer for rhythm classification"
        color       = "#10b981"
        confidence  = 0.60

    elif cv < 0.08:
        # Very consistent intervals — normal sinus rhythm
        label       = "Regular Sinus Rhythm"
        short_label = "Regular"
        note        = "Consistent R-R intervals. Normal rhythm."
        color       = "#10b981"  # green
        confidence  = round(1.0 - cv, 2)

    elif cv < 0.22 and not alternating:
        # Mildly variable but structured — sinus arrhythmia
        # This is NORMAL in dogs (respiratory sinus arrhythmia)
        label       = "Sinus Arrhythmia"
        short_label = "Sinus Arrhythmia"
        note        = "Mild rate variation. Normal finding in dogs — often respiratory in origin."
        color       = "#06b6d4"  # cyan
        confidence  = round(0.9 - cv, 2)

    elif alternating:
        # Alternating long-short pattern — suggests 2nd degree AV block
        label       = "Regularly Irregular"
        short_label = "Regularly Irregular"
        note        = "Alternating long-short R-R pattern. Consider 2nd degree AV block — evaluate with ECG."
        color       = "#f59e0b"  # amber
        confidence  = round(min(0.85, (cv - 0.08) * 3), 2)

    elif cv >= 0.22 and sd1_sd2_ratio > 0.75:
        # High variability + SD1≈SD2 (random scatter) — AF pattern
        label       = "Irregularly Irregular"
        short_label = "Irregularly Irregular"
        note        = "Random R-R variation without pattern. Consistent with Atrial Fibrillation — ECG required for diagnosis."
        color       = "#ef4444"  # red
        confidence  = round(min(0.85, sd1_sd2_ratio * 0.9), 2)

    else:
        # High variability but not clearly AF or alternating — may be artifact
        label       = "Irregular — Possible Artifact"
        short_label = "Irregular"
        note        = "Irregular intervals detected. Could be motion artifact or true arrhythmia — re-record recommended."
        color       = "#f59e0b"  # amber
        confidence  = round(0.50 + cv * 0.2, 2)

    return {
        "label":       label,
        "short_label": short_label,
        "confidence":  min(1.0, max(0.0, confidence)),
        "color":       color,
        "note":        note,
        "metrics": {
            "cv":           round(float(cv), 4),
            "mean_rr_ms":   round(float(mean_rr * 1000), 1),
            "sd1_ms":       round(float(sd1 * 1000), 2) if len(intervals) >= 3 else None,
            "sd2_ms":       round(float(sd2 * 1000), 2) if len(intervals) >= 3 else None,
            "sd1_sd2_ratio":round(float(sd1_sd2_ratio), 3) if len(intervals) >= 3 else None,
            "beat_count":   int(len(peaks)),
        },
    }


# ─── Heart Score ─────────────────────────────────────────────────────────────

def calculate_heart_score(dsp_result, cnn_result, quality):
    """
    Compute a composite Heart Score (1-10) for clinical communication.
    
    CNN is the primary predictor (70% weight) — validated at 96.3% sensitivity.
    DSP provides a secondary signal (30% weight).
    Quality dampens confidence when recording is poor.
    
    Score: 10 = healthy heart, 1 = high murmur risk.
    """
    # ── 1. Get murmur probabilities from each system ─────────────────────
    # CNN: use the raw murmur probability (higher = more likely murmur)
    if cnn_result and "probabilities" in cnn_result:
        cnn_murmur_prob = cnn_result["probabilities"].get("Murmur", 0.5)
    elif cnn_result:
        cnn_murmur_prob = cnn_result["confidence"] if cnn_result["is_disease"] else (1 - cnn_result["confidence"])
    else:
        cnn_murmur_prob = 0.5  # no CNN available, neutral

    # DSP: convert confidence to murmur probability
    if dsp_result["is_disease"]:
        dsp_murmur_prob = dsp_result["confidence"]
    else:
        dsp_murmur_prob = 1 - dsp_result["confidence"]

    # ── 2. Weighted combination (CNN primary) ────────────────────────────
    cnn_weight = 0.90
    dsp_weight = 0.10
    combined_murmur_prob = (cnn_weight * cnn_murmur_prob) + (dsp_weight * dsp_murmur_prob)

    # ── 3. Map to 1-10 scale (10 = healthy, 1 = high risk) ──────────────
    raw_score = 10 - (combined_murmur_prob * 9)  # 0% murmur → 10, 100% → 1

    # ── 4. Quality dampening — reduce confidence if recording is poor ────
    quality_factor = quality["score"] / 100.0  # 0.0 to 1.0
    # Pull score toward 5 (uncertain) if quality is poor
    dampened_score = 5 + (raw_score - 5) * quality_factor

    # ── 5. Clamp and round ───────────────────────────────────────────────
    heart_score = max(1, min(10, round(dampened_score)))

    # ── 6. Clinical interpretation ───────────────────────────────────────
    if heart_score >= 8:
        interpretation = "Normal — no significant findings"
        risk_level = "low"
    elif heart_score >= 6:
        interpretation = "Borderline — consider monitoring"
        risk_level = "moderate"
    elif heart_score >= 4:
        interpretation = "Suspicious — further evaluation recommended"
        risk_level = "elevated"
    else:
        interpretation = "Abnormal — recommend echocardiography"
        risk_level = "high"

    return {
        "score": heart_score,
        "max_score": 10,
        "interpretation": interpretation,
        "risk_level": risk_level,
        "breakdown": {
            "cnn_murmur_prob": round(cnn_murmur_prob, 3),
            "dsp_murmur_prob": round(dsp_murmur_prob, 3),
            "combined_prob": round(combined_murmur_prob, 3),
            "quality_factor": round(quality_factor, 2),
        }
    }


def predict_audio(audio_bytes: bytes):
    """Main inference — returns DSP, CNN results, signal quality, and Heart Score."""
    try:
        waveform = load_audio(audio_bytes)
        duration = len(waveform) / TARGET_SR
        print(f"Audio: {len(waveform)} samples, {duration:.1f}s", flush=True)

        bpm, heartbeat_count, peaks = calculate_bpm(waveform, TARGET_SR)
        print(f"BPM: {bpm}, Beats: {heartbeat_count}", flush=True)

        # Signal quality scoring
        quality = score_quality(waveform, TARGET_SR, peaks)
        print(f"Quality: {quality['grade']} ({quality['score']}/100)", flush=True)

        # Rhythm analysis
        rhythm = analyze_rhythm(peaks, TARGET_SR)
        print(f"Rhythm: {rhythm['label']} (CV={rhythm['metrics'].get('cv', 'N/A')})", flush=True)

        # DSP-based classification
        dsp_result = detect_murmur(waveform, TARGET_SR, peaks)

        # Quality-gated DSP dampening: reduce DSP confidence when noise is present
        # (noise creates spectral features that DSP misinterprets as murmur)
        noise_warnings = [w for w in quality.get("warnings", []) if "noise" in w.lower()]
        if noise_warnings and dsp_result["is_disease"]:
            # Dampening: pull murmur_prob toward 0.5 based on quality + noise severity
            # Quality score: 100 → no dampening, 0 → fully neutral
            # Noise penalty: noise in the murmur frequency band directly corrupts DSP
            #   features, so we apply an extra 0.5x penalty per noise warning
            quality_damp = quality["score"] / 100.0
            noise_penalty = max(0.2, 1.0 - 0.5 * len(noise_warnings))  # cap at 0.2
            damp_factor = quality_damp * noise_penalty
            raw_murmur = dsp_result["all_classes"][1]["probability"]
            dampened_murmur = 0.5 + (raw_murmur - 0.5) * damp_factor
            dampened_normal = 1.0 - dampened_murmur
            # When noise is present, require higher confidence to call Murmur
            # (0.40 is too low when we know noise is corrupting the features)
            is_murmur = dampened_murmur >= 0.65
            dsp_result = {
                **dsp_result,
                "label": "Murmur" if is_murmur else "Normal",
                "confidence": round(dampened_murmur if is_murmur else dampened_normal, 4),
                "is_disease": is_murmur,
                "all_classes": [
                    {"label": "Normal", "probability": round(dampened_normal, 4)},
                    {"label": "Murmur", "probability": round(dampened_murmur, 4)},
                ],
                "details": dsp_result["details"] + f" | Quality-dampened ({quality['score']}/100)",
            }
        print(f"DSP: {dsp_result['label']} ({dsp_result['confidence']:.1%})", flush=True)

        # CNN-based classification (Joint CNN — Run 6)
        cnn_result = predict_cnn(waveform, TARGET_SR)
        if cnn_result:
            print(f"CNN: {cnn_result['label']} ({cnn_result['confidence']:.1%})", flush=True)

        # ── Run all 4 models for comparison ──────────────────────────────────
        finetuned_result = predict_finetuned(waveform, TARGET_SR)
        if finetuned_result:
            print(f"Fine-tuned: {finetuned_result['label']} ({finetuned_result['confidence']:.1%})", flush=True)

        resnet_result = predict_resnet(waveform, TARGET_SR)
        if resnet_result:
            print(f"ResNet: {resnet_result['label']} ({resnet_result['confidence']:.1%})", flush=True)

        gru_result = predict_gru(waveform, TARGET_SR)
        if gru_result:
            print(f"Bi-GRU: {gru_result['label']} ({gru_result['confidence']:.1%})", flush=True)

        # Build model comparison array
        model_comparison = []
        if cnn_result:
            model_comparison.append({
                "name": "Joint CNN", "tag": "BASELINE",
                "color": "#8B5CF6",
                "description": "2D CNN · Mel-spectrogram · Joint training",
                "score": "5/10 canine",
                **cnn_result
            })
        if finetuned_result:
            model_comparison.append({
                "name": "Fine-tuned CNN", "tag": "TRANSFER",
                "color": "#F59E0B",
                "description": "2D CNN · 2-step transfer learning",
                "score": "5/10 canine",
                **finetuned_result
            })
        if resnet_result:
            model_comparison.append({
                "name": "ResNet-18", "tag": "IMAGENET",
                "color": "#10B981",
                "description": "ImageNet pretrained · Frozen backbone",
                "score": "8/10 canine",
                **resnet_result
            })
        if gru_result:
            model_comparison.append({
                "name": "Bi-GRU", "tag": "PRIMARY",
                "color": "#06B6D4",
                "description": "McDonald et al. · Temporal RNN · Log-spec",
                "score": "10/10 canine",
                **gru_result
            })

        # Heart Score (1-10)
        heart_score = calculate_heart_score(dsp_result, cnn_result, quality)
        print(f"Heart Score: {heart_score['score']}/10 ({heart_score['interpretation']})", flush=True)

        # Combined summary — Bi-GRU is primary, CNN is fallback
        dsp_disease = dsp_result["is_disease"]
        # Use GRU as primary decision-maker (10/10 canine accuracy)
        if gru_result:
            primary_result = gru_result
            primary_name = "Bi-GRU"
        elif cnn_result:
            primary_result = cnn_result
            primary_name = "CNN"
        else:
            primary_result = dsp_result
            primary_name = "DSP"

        is_disease = primary_result["is_disease"]

        # Murmur type from primary model
        murmur_type      = None
        murmur_type_conf = None
        murmur_type_note = None
        if is_disease:
            murmur_type      = primary_result.get("label", "Murmur")
            murmur_type_conf = primary_result.get("confidence")
            murmur_type_note = MURMUR_TYPE_NOTES.get(murmur_type, "")

        if quality["grade"] == "Poor":
            summary   = "⚠️ Poor recording quality — results may be unreliable, please re-record"
            agreement = "poor_quality"
        elif is_disease and dsp_disease:
            type_str  = f" ({murmur_type})" if murmur_type else ""
            summary   = f"⚠️ Murmur detected{type_str} — confirmed by {primary_name} and DSP analysis"
            agreement = "both_murmur"
        elif is_disease and not dsp_disease:
            type_str  = f" ({murmur_type})" if murmur_type else ""
            summary   = f"⚠️ Murmur detected{type_str} by {primary_name} — DSP analysis was inconclusive"
            agreement = "primary_only"
        elif not is_disease and dsp_disease:
            summary   = f"Normal heart sound ({primary_name}) — DSP flagged minor irregularity, likely artifact"
            agreement = "dsp_only"
        else:
            summary   = "Normal heart sound — no murmur detected"
            agreement = "both_normal"

        # Downsample waveform for frontend (~800 points)
        num_points = 800
        step = max(1, len(waveform) // num_points)
        vis_waveform = waveform[::step].tolist()
        vis_duration = len(vis_waveform)

        peak_times_sec = [round(float(p) / TARGET_SR, 3) for p in peaks]
        peak_vis_indices = [int(p // step) for p in peaks if int(p // step) < vis_duration]

        return {
            "bpm":              bpm,
            "heartbeat_count":  heartbeat_count,
            "duration_seconds": round(duration, 1),
            "rhythm":           rhythm,
            "is_disease":       is_disease,
            "murmur_type":      murmur_type,
            "murmur_type_confidence": murmur_type_conf,
            "murmur_type_note": murmur_type_note,
            "agreement":        agreement,
            "clinical_summary": summary,
            "heart_score":      heart_score,
            "ai_classification":  dsp_result,
            "dsp_classification":  dsp_result,
            "cnn_classification":  cnn_result,
            "model_comparison":   model_comparison,   # NEW: all 4 models
            "gru_classification": gru_result,         # NEW: Bi-GRU (primary)
            "signal_quality":   quality,
            "waveform":         vis_waveform,
            "peak_times_seconds": peak_times_sec,
            "peak_vis_indices": peak_vis_indices,
        }

    except Exception as e:
        import traceback
        print(f"Error:\n{traceback.format_exc()}", flush=True)
        return {"error": str(e)}