app.py

import gradio as gr
import numpy as np
import tensorflow as tf
from tensorflow.keras.layers import Layer
import pickle
import json

# --- Register custom layers before loading model ---
@tf.keras.utils.register_keras_serializable(package='PhysioWatch')
class ExpandDims(Layer):
    def call(self, x):
        return tf.expand_dims(x, -1)
    def compute_output_shape(self, input_shape):
        return (input_shape[0], input_shape[1], 1)
    def get_config(self):
        return super().get_config()

@tf.keras.utils.register_keras_serializable(package='PhysioWatch')
class ReduceSum(Layer):
    def call(self, x):
        return tf.reduce_sum(x, axis=1)
    def compute_output_shape(self, input_shape):
        return (input_shape[0], input_shape[2])
    def get_config(self):
        return super().get_config()

# --- Load model and scaler ---
model  = tf.keras.models.load_model('physiowatch_model_clean.keras')
with open('scaler.pkl', 'rb') as f:
    scaler = pickle.load(f)

print("✅ Model and scaler loaded successfully")

# --- JSON prediction endpoint ---
def predict_json(json_input):
    try:
        data = json.loads(json_input)
        sequences = np.array(data["sequences"])

        # Handle both single (24,5) and batch (n,24,5)
        if sequences.ndim == 2:
            sequences = sequences[np.newaxis, ...]

        n = sequences.shape[0]
        flat = sequences.reshape(-1, 5)
        normalized = scaler.transform(flat)
        X = normalized.reshape(n, 24, 5)

        probs = model.predict(X, verbose=0).flatten()

        results = []
        for prob in probs:
            prob = float(prob)
            results.append({
                "prediction": "Abnormal" if prob > 0.5 else "Normal",
                "confidence": round(max(prob, 1 - prob) * 100, 2),
                "probability_abnormal": round(prob * 100, 2),
                "probability_normal": round((1 - prob) * 100, 2)
            })

        return json.dumps({"results": results, "status": "success"}, indent=2)

    except Exception as e:
        return json.dumps({"error": str(e), "status": "failed"})

# --- Manual input endpoint ---
def predict_vitals(heart_rate_seq, systolic_bp_seq, diastolic_bp_seq,
                   respiratory_rate_seq, spo2_seq):
    try:
        hr   = [float(x) for x in heart_rate_seq.strip().split(',')]
        sbp  = [float(x) for x in systolic_bp_seq.strip().split(',')]
        dbp  = [float(x) for x in diastolic_bp_seq.strip().split(',')]
        rr   = [float(x) for x in respiratory_rate_seq.strip().split(',')]
        spo2 = [float(x) for x in spo2_seq.strip().split(',')]

        if not all(len(s) == 24 for s in [hr, sbp, dbp, rr, spo2]):
            return {"error": "Each vital sign must have exactly 24 values"}

        raw_sequence = np.array([hr, sbp, dbp, rr, spo2]).T
        normalized = scaler.transform(raw_sequence)
        sequence = normalized.reshape(1, 24, 5)

        prob = float(model.predict(sequence, verbose=0)[0][0])
        label = "Abnormal" if prob > 0.5 else "Normal"

        return {
            "prediction": label,
            "confidence": round(max(prob, 1 - prob) * 100, 2),
            "probability_abnormal": round(prob * 100, 2),
            "probability_normal": round((1 - prob) * 100, 2),
            "status": "success"
        }

    except Exception as e:
        return {"error": str(e), "status": "failed"}

# --- Gradio Interface ---
with gr.Blocks(title="PhysioWatch API") as demo:
    gr.Markdown("# PhysioWatch — Abnormal Health Pattern Detection API")
    gr.Markdown("CNN-BiLSTM-Attention model trained on MIMIC-IV ICU data")

    with gr.Tab("JSON API"):
        gr.Markdown("**Input format:** `{\"sequences\": [[24 timesteps × 5 vitals]]}`")
        gr.Markdown("**Column order:** Heart Rate, Systolic BP, Diastolic BP, Respiratory Rate, SpO2")
        json_in  = gr.Textbox(label="JSON Input", lines=10,
                              placeholder='{"sequences": [[[72, 120, 80, 16, 98], ...]]}')
        json_out = gr.Textbox(label="Result", lines=10)
        gr.Button("Run Prediction", variant="primary").click(
            predict_json, inputs=json_in, outputs=json_out
        )

    with gr.Tab("Manual Input"):
        gr.Markdown("Enter exactly **24 comma-separated values** per vital sign")
        hr_in   = gr.Textbox(label="Heart Rate (bpm)")
        sbp_in  = gr.Textbox(label="Systolic BP (mmHg)")
        dbp_in  = gr.Textbox(label="Diastolic BP (mmHg)")
        rr_in   = gr.Textbox(label="Respiratory Rate (breaths/min)")
        spo2_in = gr.Textbox(label="SpO2 (%)")
        manual_out = gr.JSON(label="Result")
        gr.Button("Predict", variant="primary").click(
            predict_vitals,
            inputs=[hr_in, sbp_in, dbp_in, rr_in, spo2_in],
            outputs=manual_out
        )

demo.launch()