HAPI-BELT: Revolutionizing Hypoglycemia Detection in Preterm Infants

Introduction

As healthcare technology progresses, it has started to impact even the most vulnerable patient populations, including preterm infants. HAPI-BELT is a groundbreaking system that continuously monitors these infants for signs of hypoglycemia, harnessing the power of smart sensors and deep learning algorithms to facilitate early detection and timely intervention.

The Core Components of HAPI-BELT

HAPI-BELT primarily comprises three vital elements:

1. Photoplethysmography (PPG) Sensor: This sensor, embedded within a smart belt, continuously captures cardiac activity and heart rate data.
2. Smart Camera: Located inside the incubator, it observes breathing patterns, skin color changes, and movement.
3. Data Processing Unit: Utilizing a Raspberry Pi Zero-2 W, the collected data undergoes processing and classification through a GRU-LSTM network.

The Role of the PPG Sensor

The PPG sensor is strategically mounted on a smart belt worn around the infant’s abdomen or chest, allowing it to acquire detailed information about heartbeats and other cardiorespiratory parameters. This sensor detects variations in blood volume as the heart beats, resulting in waveform data that can signal hypoglycemia-related symptoms such as bradycardia. The data captured is processed to gauge vital physiological parameters, including heart rate and blood oxygen saturation.

Smart Camera Integration

Additionally, a smart camera in the incubator captures real-time visual data, such as the infant’s breathing patterns, skin color changes indicative of health conditions, and movements. This multimodal monitoring system helps identify symptoms like paleness or excessive sweating that can also be associated with hypoglycemia.

Data Collection and Pre-processing

Data collection is inseparable from effective monitoring in the HAPI-BELT system. The PPG sensor and smart camera continuously gather vital data, which is subsequently transmitted to the Raspberry Pi for preprocessing.

The Importance of Pre-processing

The initial step converts raw sensor data into a digital format suitable for analysis. This preprocessing stage is critical since it prepares the data for feature extraction (FE) and classification. The system relies on proxy indicators—heart rate variability, abnormal respiratory patterns, movement, and skin color changes—rather than direct glucose measurement, which presents ethical and practical challenges.

Acts such as bradycardia, respiratory irregularities, and color changes have been linked to hypoglycemic events in infants, proving to be effective early warning signs.

Deep Learning Model: GRU-LSTM

The real-time classification of the physiological state is executed by the GRU-LSTM model, a combination that leverages the strengths of both Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM) networks. GRUs are efficient in handling simpler sequences, while LSTMs are adept at recognizing longer-term patterns—a perfect suit for the diverse data streams from HAPI-BELT.

Data Training

Training the model requires a robust dataset carefully collected from infants at the University of Science and Technology in Mazandaran, Iran. This dataset consists of various physiological signals, age, gender, and invasive blood glucose levels.

Preprocessing of PPG Data

Like any effective monitoring system, PPG data preprocessing is crucial for ensuring high-quality, standardized signals. Steps include:

1. Filtering Noise: Implementing a Butterworth bandpass filter eliminates extraneous noise outside the typical heartbeat range.
2. Eliminating Baseline Wander: High-pass filters remove low-frequency interference caused by movement or respiration.
3. Normalization: Both min-max and z-score normalization ensure that the data is standardized, facilitating more accurate analyses.

Feature Extraction

Following the preprocessing steps, the PPG signals are segmented and prepared for feature extraction. Notable features include the number of peaks in the PPG signal, heart rate variability (HRV), and power distribution across frequency components.

Preprocessing of Visual Data

Visual data preprocessing complements PPG data by enhancing image quality and extracting relevant features.

1. Uniform Frame Rates: Resampling videos to a consistent frame rate ensures uniformity in the dataset.
2. Noise Reduction and Enhancement: Gaussian blur and histogram equalization techniques improve image clarity.
3. Region of Interest (ROI) Identification: Features such as facial and chest regions are marked for monitoring skin color changes or movement patterns.

Quantification of Motion and Features

Optical flow is calculated, and methods such as Local Binary Patterns (LBP) are used to characterize changes in the visual data effectively.

Optimizing Classification: Cat Swarm Optimization

To improve classification accuracy, Cat Swarm Optimization (CSO) is employed. This optimization technique mimics the natural behavior of cats and helps identify critical features from the data collected by the HAPI-BELT system. Through feature selection, CSO enhances the performance of the classification models, making hypoglycemia detection in preterm infants more reliable.

Through these innovative technologies—the smart PPG sensor, smart camera, real-time data processing, and state-of-the-art learning algorithms—HAPI-BELT is setting a new standard in continuous monitoring for the safety of preterm infants. By harnessing the power of technology, it not only ensures timely medical intervention but also paves the way for future advancements in neonatal care.