| --- |
| language: |
| - en |
| license: mit |
| library_name: tflite |
| tags: |
| - medical |
| - time-series |
| - health-monitoring |
| - cnn-bilstm |
| - attention-mechanism |
| datasets: |
| - custom-sensor-data |
| metrics: |
| - accuracy |
| - f1 |
| pipeline_tag: audio-classification |
| --- |
| |
| # Model Card for Asclepius_v1 |
| |
| Asclepius_v1 is a specialized deep learning model designed for real-time medical crisis prediction. It processes multi-modal time-series data from wearable sensors to detect and classify health emergencies. |
|
|
| ## Model Details |
|
|
| ### Model Description |
|
|
| Asclepius_v1 utilizes a hybrid architecture combining **1D-CNN**, **Bidirectional LSTM (BiLSTM)**, and an **Attention Mechanism**. This design allows the model to extract local features from high-frequency sensor noise while maintaining a long-term memory of physiological trends. |
| |
| - **Developed by:** Koussay Chaanbi |
| - **Model type:** Hybrid CNN-BiLSTM-Attention Network |
| - **Language(s):** English (Technical Documentation) |
| - **License:** MIT |
| - **Model Architecture:** - **1D-CNN:** For local feature extraction from sensor streams. |
| - **BiLSTM:** To capture temporal dependencies in both forward and backward directions. |
| - **Attention Layer:** To assign weights to critical physiological events within the data window. |
| |
| ### Model Sources |
| |
| - **Repository:** [Asclepius_v1 on Hugging Face](https://huggingface.co/Koussay/Asclepius_v1) |
| - **Hardware Target:** Raspberry Pi 4 / Raspberry Pi 5 |
| |
| ## Uses |
| |
| ### Direct Use |
| |
| The model is intended for integration into wearable health devices (such as smart shirts) to monitor: |
| 1. Heart Rate (HR) |
| 2. Oxygen Saturation (SpO2) |
| 3. Movement (Accelerometer data) |
| |
| It classifies input data into one of five specific medical crisis categories. |
| |
| ### Out-of-Scope Use |
| |
| This model is **not** a replacement for professional medical diagnosis or clinical-grade monitoring equipment. It should not be used as the sole basis for medical intervention without human-in-the-loop verification. |
| |
| ## Bias, Risks, and Limitations |
| |
| - **Technical Limitations:** The model is optimized for edge deployment (float16 quantization), which may lead to a slight drop in precision compared to full-precision desktop models. |
| - **Data Bias:** Performance is dependent on the quality and placement of the hardware sensors. |
| |
| ### Recommendations |
| |
| Users should ensure proper sensor calibration and skin contact for reliable inference. The model should be used as a proactive "early warning" system rather than a definitive diagnostic tool. |
| |
| ## How to Get Started with the Model |
| |
| Since the model is provided in `.tflite` format, use the following code to run inference: |
| |
| ```python |
| import tensorflow as tf |
| import numpy as np |
| |
| # Load the TFLite model and allocate tensors. |
| interpreter = tf.lite.Interpreter(model_path="model_quantized.tflite") |
| interpreter.allocate_tensors() |
|
|
| # Get input and output details. |
| input_details = interpreter.get_input_details() |
| output_details = interpreter.get_output_details() |
|
|
| # Prepare your sensor data (HR, SpO2, Movement) |
| # Shape should match: [1, window_size, features] |
| input_data = np.array(your_sensor_window, dtype=np.float32) |
| interpreter.set_tensor(input_details[0]['index'], input_data) |
| |
| interpreter.invoke() |
| |
| # Extract prediction |
| output_data = interpreter.get_tensor(output_details[0]['index']) |
| print(f"Crisis Classification Result: {np.argmax(output_data)}") |
