app.py

import os
import cv2
import numpy as np
import mediapipe as mp
from fastapi import FastAPI, UploadFile, File
from fastapi.middleware.cors import CORSMiddleware
from scipy import signal
from scipy.signal import find_peaks
import tempfile

print("[Init] Loading MediaPipe...", flush=True)
mp_face_mesh = mp.solutions.face_mesh
face_mesh = mp_face_mesh.FaceMesh(
    static_image_mode=False,
    max_num_faces=1,
    refine_landmarks=True,
    min_detection_confidence=0.5,
    min_tracking_confidence=0.5
)
print("[Init] MediaPipe OK", flush=True)

app = FastAPI(title="AF Detector API")

app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_methods=["*"],
    allow_headers=["*"],
)

# ── Signal Processing ─────────────────────────────────────

def extract_rppg_signal(frames, fps=30):
    """
    Extrait le signal rPPG depuis les frames vidéo.
    Utilise le canal vert de la ROI du visage (MediaPipe).
    """
    green_signal = []
    valid_frames = 0

    for frame in frames:
        rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
        result = face_mesh.process(rgb)

        if result.multi_face_landmarks:
            lm = result.multi_face_landmarks[0].landmark
            h, w = frame.shape[:2]

            # ROI = joues + front (zones riches en vaisseaux)
            roi_points = [
                # Front
                (int(lm[10].x * w),  int(lm[10].y * h)),
                (int(lm[151].x * w), int(lm[151].y * h)),
                # Joue gauche
                (int(lm[234].x * w), int(lm[234].y * h)),
                (int(lm[93].x * w),  int(lm[93].y * h)),
                # Joue droite
                (int(lm[454].x * w), int(lm[454].y * h)),
                (int(lm[323].x * w), int(lm[323].y * h)),
            ]

            # Bounding box de la ROI
            xs = [p[0] for p in roi_points]
            ys = [p[1] for p in roi_points]
            x1, x2 = max(0, min(xs)), min(w, max(xs))
            y1, y2 = max(0, min(ys)), min(h, max(ys))

            if x2 > x1 and y2 > y1:
                roi = frame[y1:y2, x1:x2]
                # Canal vert (le plus sensible aux pulsations)
                g_mean = np.mean(roi[:, :, 1])
                green_signal.append(g_mean)
                valid_frames += 1
        else:
            # Pas de visage — utiliser frame entière comme fallback
            g_mean = np.mean(frame[:, :, 1])
            green_signal.append(g_mean)

    face_ratio = valid_frames / len(frames) if frames else 0
    return np.array(green_signal), face_ratio


def bandpass_filter(signal_data, fs, lowcut=0.7, highcut=4.0, order=4):
    """
    Filtre passe-bande Butterworth.
    0.7 Hz = 42 BPM (min)
    4.0 Hz = 240 BPM (max)
    """
    nyq = fs / 2.0
    low = lowcut / nyq
    high = highcut / nyq
    b, a = signal.butter(order, [low, high], btype='band')
    return signal.filtfilt(b, a, signal_data)


def detect_peaks_rr(filtered_signal, fps):
    """
    Détecte les pics du signal cardiaque → intervalles RR.
    """
    min_distance = int(fps * 0.35)  # min 350ms entre pics (max ~170 BPM)
    threshold    = np.std(filtered_signal) * 0.3

    peaks, properties = find_peaks(
        filtered_signal,
        distance=min_distance,
        height=threshold
    )
    return peaks


def compute_hrv_metrics(rr_intervals_ms):
    """
    Calcule les métriques HRV classiques utilisées pour détecter la FA.
    """
    if len(rr_intervals_ms) < 5:
        return None

    rr = np.array(rr_intervals_ms)

    # Métriques temporelles
    mean_rr  = np.mean(rr)
    sdnn     = np.std(rr)                          # variabilité globale
    rmssd    = np.sqrt(np.mean(np.diff(rr)**2))    # variabilité court terme
    pnn50    = np.sum(np.abs(np.diff(rr)) > 50) / len(rr) * 100  # % diff > 50ms

    # BPM
    bpm = round(60000 / mean_rr)

    # Coefficient de variation (CV) — clé pour FA
    cv = (sdnn / mean_rr) * 100

    # Irregularity index — entropie approchée
    diffs = np.abs(np.diff(rr))
    irr_index = round(min(100, (np.mean(diffs) / mean_rr) * 100))

    return {
        "bpm":       int(np.clip(bpm, 30, 250)),
        "mean_rr":   round(float(mean_rr), 1),
        "sdnn":      round(float(sdnn), 1),
        "rmssd":     round(float(rmssd), 1),
        "pnn50":     round(float(pnn50), 1),
        "cv":        round(float(cv), 2),
        "irr_index": irr_index,
        "rr_count":  len(rr_intervals_ms),
    }


def compute_af_score(metrics):
    """
    Score de risque FA (0-100) basé sur les métriques HRV.
    
    Critères cliniques FA :
    - Absence de onde P régulière → RR irréguliers
    - RMSSD élevé
    - CV élevé (>10%)
    - pNN50 élevé
    - Pattern d'irrégularité sans rythme
    """
    score = 0
    reasons = []

    bpm      = metrics["bpm"]
    rmssd    = metrics["rmssd"]
    cv       = metrics["cv"]
    pnn50    = metrics["pnn50"]
    irr      = metrics["irr_index"]
    sdnn     = metrics["sdnn"]

    # BPM anormal
    if bpm < 50:
        score += 15; reasons.append(f"Bradycardie ({bpm} BPM)")
    elif bpm > 100:
        score += 20; reasons.append(f"Tachycardie ({bpm} BPM)")

    # RMSSD — variabilité élevée = irrégularité
    if rmssd > 100:
        score += 30; reasons.append(f"RMSSD très élevé ({rmssd}ms)")
    elif rmssd > 60:
        score += 18; reasons.append(f"RMSSD élevé ({rmssd}ms)")
    elif rmssd > 40:
        score += 8

    # CV — coefficient de variation
    if cv > 15:
        score += 25; reasons.append(f"Variabilité RR critique (CV={cv}%)")
    elif cv > 10:
        score += 15; reasons.append(f"Variabilité RR élevée (CV={cv}%)")
    elif cv > 6:
        score += 6

    # pNN50
    if pnn50 > 40:
        score += 15; reasons.append(f"pNN50 élevé ({pnn50}%)")
    elif pnn50 > 20:
        score += 8

    # Irregularity index
    if irr > 25:
        score += 10; reasons.append(f"Irrégularité marquée ({irr}%)")

    score = int(min(100, score))

    # Classification
    if score < 25:
        result = "NORMAL"
        label  = "Normal Sinus Rhythm"
        risk   = "LOW"
    elif score < 50:
        result = "IRREGULAR"
        label  = "Irregular Pattern Detected"
        risk   = "MODERATE"
    else:
        result = "AF_SUSPECTED"
        label  = "Atrial Fibrillation Suspected"
        risk   = "HIGH"

    return {
        "af_score": score,
        "result":   result,
        "label":    label,
        "risk":     risk,
        "reasons":  reasons,
    }


# ── API Endpoints ─────────────────────────────────────────

@app.get("/health")
def health():
    return {"status": "ok", "service": "AF Detector API v1.0"}


@app.post("/analyze/")
async def analyze_video(video_file: UploadFile = File(...), fps: float = 30.0):
    """
    Analyse une vidéo de 30s pour détecter la FA via rPPG.
    Retourne les métriques HRV et le score AF.
    """
    # Sauvegarder la vidéo temporairement
    suffix = os.path.splitext(video_file.filename)[-1] or ".mp4"
    with tempfile.NamedTemporaryFile(delete=False, suffix=suffix) as tmp:
        tmp.write(await video_file.read())
        tmp_path = tmp.name

    try:
        # Lire la vidéo
        cap = cv2.VideoCapture(tmp_path)
        if not cap.isOpened():
            return {"error": "Cannot open video file"}

        real_fps = cap.get(cv2.CAP_PROP_FPS) or fps
        frames = []

        while True:
            ret, frame = cap.read()
            if not ret: break
            # Subsample si fps > 30
            if len(frames) % max(1, int(real_fps / 30)) == 0:
                frames.append(frame)

        cap.release()
        os.remove(tmp_path)

        if len(frames) < 60:
            return {"error": "Video too short. Minimum 30 seconds required.", "frames": len(frames)}

        print(f"[Analyze] Frames: {len(frames)} | FPS: {real_fps:.1f}", flush=True)

        # ── rPPG extraction ──────────────────────────────
        green_signal, face_ratio = extract_rppg_signal(frames, real_fps)
        print(f"[Analyze] Face detected: {face_ratio:.1%}", flush=True)

        if face_ratio < 0.3:
            return {"error": "Face not detected in most frames. Ensure good lighting.", "face_ratio": face_ratio}

        # ── Détrending + filtrage ────────────────────────
        # Supprimer la tendance lente (illumination)
        detrended = signal.detrend(green_signal)

        # Normaliser
        detrended = (detrended - np.mean(detrended)) / (np.std(detrended) + 1e-8)

        # Bandpass 0.7-4Hz
        filtered = bandpass_filter(detrended, real_fps)

        # ── Peak detection ───────────────────────────────
        peaks = detect_peaks_rr(filtered, real_fps)
        print(f"[Analyze] Peaks detected: {len(peaks)}", flush=True)

        if len(peaks) < 8:
            return {"error": "Signal too noisy. Stay still and ensure good lighting.", "peaks": len(peaks)}

        # ── RR intervals (ms) ────────────────────────────
        rr_intervals = [(peaks[i] - peaks[i-1]) / real_fps * 1000
                        for i in range(1, len(peaks))]

        # Filtrer les RR aberrants (< 300ms ou > 2000ms)
        rr_intervals = [rr for rr in rr_intervals if 300 < rr < 2000]

        if len(rr_intervals) < 5:
            return {"error": "Not enough valid beats detected."}

        # ── HRV metrics ──────────────────────────────────
        metrics = compute_hrv_metrics(rr_intervals)
        if not metrics:
            return {"error": "Cannot compute HRV metrics."}

        # ── AF Score ─────────────────────────────────────
        af_data = compute_af_score(metrics)

        print(f"[Analyze] BPM={metrics['bpm']} | RMSSD={metrics['rmssd']} | CV={metrics['cv']} | AF_Score={af_data['af_score']}", flush=True)

        return {
            "success":      True,
            "face_ratio":   round(face_ratio, 2),
            "frames":       len(frames),
            "fps":          round(real_fps, 1),
            "peaks_count":  len(peaks),
            "rr_intervals": [round(rr, 1) for rr in rr_intervals[-30:]],  # derniers 30
            **metrics,
            **af_data,
            "disclaimer": "Experimental AI tool. Not a medical diagnosis. Consult a cardiologist."
        }

    except Exception as e:
        if os.path.exists(tmp_path):
            os.remove(tmp_path)
        print(f"[Error] {e}", flush=True)
        return {"error": str(e)}


@app.post("/analyze_frames/")
async def analyze_frames_json(data: dict):
    """
    Alternative : reçoit le signal vert directement depuis le frontend.
    Plus rapide — pas besoin d'encoder/décoder la vidéo.
    """
    green_signal = np.array(data.get("green_signal", []))
    fps = float(data.get("fps", 30.0))

    if len(green_signal) < 60:
        return {"error": "Signal too short. Minimum 60 samples required."}

    try:
        detrended = signal.detrend(green_signal)
        detrended = (detrended - np.mean(detrended)) / (np.std(detrended) + 1e-8)
        filtered  = bandpass_filter(detrended, fps)
        peaks     = detect_peaks_rr(filtered, fps)

        if len(peaks) < 5:
            return {"error": "Signal too noisy. Stay still.", "peaks": len(peaks)}

        rr_intervals = [(peaks[i] - peaks[i-1]) / fps * 1000
                        for i in range(1, len(peaks))]
        rr_intervals = [rr for rr in rr_intervals if 300 < rr < 2000]

        if len(rr_intervals) < 4:
            return {"error": "Not enough valid beats."}

        metrics = compute_hrv_metrics(rr_intervals)
        af_data = compute_af_score(metrics)

        print(f"[Frames] BPM={metrics['bpm']} | RMSSD={metrics['rmssd']} | AF={af_data['af_score']}", flush=True)

        return {
            "success": True,
            "rr_intervals": [round(rr, 1) for rr in rr_intervals],
            **metrics,
            **af_data,
        }

    except Exception as e:
        print(f"[Error] {e}", flush=True)
        return {"error": str(e)}