app.py

import torchaudio
import torch
import gradio as gr
import keras
import pandas as pd

from transformers import pipeline


def get_murmur_from_recordings(audio):
    pipe = pipeline("audio-classification",
                    model="cogniveon/eeem069_heart_murmur_classification")

    sampling_rate, data = audio
    waveform = torch.tensor(data).float()

    # Resample the audio to 16 kHz (if necessary)
    if sampling_rate != 16000:
        resampler = torchaudio.transforms.Resample(sampling_rate, 16000)
        waveform = resampler(waveform)

    results = pipe(waveform.numpy())

    sorted_results = sorted(results, key=lambda x: x['score'], reverse=True)
    label_scores = {item['label']: item['score'] for item in sorted_results}
    return label_scores


def get_patient_outcome(age, sex, height, weight, is_pregnant, murmur):
    model = keras.models.load_model('patient_outcome_classifier.keras')

    is_pregnant = 1 if is_pregnant else 0
    sex2int = {'Male': 0, 'Female': 1}
    sex = sex2int[sex]

    age2int = {'Neonate': 0, 'Infant': 1, 'Child': 2, 'Adolescent': 3}
    age = age2int[age]

    murmur = 0 if murmur == 'Absent' else (1 if murmur == 'Present' else 2)

    data = pd.DataFrame({
        'Age': float(age),
        'Sex': float(sex),
        'Height': float(height),
        'Weight': float(weight),
        'Pregnancy status': float(is_pregnant),
        'Murmur': float(murmur),
    }, index=[0])

    output = model.predict(data)[0]
    # 0 - Normal, 1 - Abnormal
    results = {
        'Normal': output[0],
        'Abnormal': output[1]
    }

    return results


def predict(audio, age, sex, height, weight, is_pregnant):
    assert audio is not None, 'Audio cannot be None'

    murmur_scores = get_murmur_from_recordings(audio)
    murmur = "Unknown"
    if murmur_scores['Present'] > 0.70:
        murmur = "Present"
    if murmur_scores['Absent'] > 0.80:
        murmur = "Absent"

    outcome = get_patient_outcome(
        age, sex, height, weight, is_pregnant, murmur)

    return outcome


demo = gr.Interface(
    fn=predict,
    inputs=[
        "audio",
        gr.Radio(label="Age", choices=[
                 "Neonate", "Infant", "Child", "Adolescent"],
                 value="Adolescent"),
        gr.Radio(label="Sex", choices=["Male", "Female"], value="Male"),
        gr.Number(label="Height", value="98.0"),
        gr.Number(label="Weight", value="38.1"),
        gr.Checkbox(label="Pregnant", value=False)
    ],
    outputs="label"
)
demo.launch()