Abstract
To address the limitations of single-function medical products, which often lack synchronized monitoring, storage, and analysis capabilities, this study proposes a comprehensive computerized nursing model for multi-parameter vital signs monitoring based on the Model-View-Controller (MVC) architectural pattern. The integrated system combines wearable hardware for data acquisition, a mobile application for real-time display and user interaction, and a cloud-based backend for advanced signal processing and machine learning. By applying the synthetic minority over-sampling technique (SMOTE) to balance the training data, the system’s StackingClassifier achieves 96% accuracy for arrhythmia detection on the MIT-BIH dataset. Furthermore, the model’s robustness was validated on a large-scale, non-clinical dataset from 250 participants, achieving an excellent Area Under the Curve (AUC) of 0.974 in real-world arrhythmia detection. This experimental validation also confirms high precision in measuring heart rate, body temperature, and step count when compared to commercial devices. This work demonstrates a scalable and maintainable solution for reliable self-health management and cardiovascular disease prevention.
Introduction
Continuous monitoring of vital signs, particularly the electrocardiogram (ECG), is critical for the early detection of cardiovascular diseases. However, current consumer medical devices often suffer from being single-function, lacking synchronized data storage and advanced analysis, or having poor cross-device compatibility. Furthermore, many machine learning models for ECG analysis perform well on clean, balanced clinical datasets but fail to generalize to the noisy, imbalanced data generated during daily activities. This paper presents a holistic "computer nursing model"—an end-to-end IoT system encompassing wearable hardware, a user-friendly mobile app, and a robust cloud backend for intelligent diagnosis and health education.
System Architecture
The system is built on the Model-View-Controller (MVC) architectural pattern to cleanly separate data acquisition, the user interface, and backend processing, ensuring scalability and data security.
As illustrated in the architecture diagram, the "Backstage" handles complex diagnostic tasks using Python and Matlab, while the "APP" layer manages user interaction and real-time measurement. Data is synchronized across SQLite and MySQL archives to maintain historical records for longitudinal analysis.
Hardware and Data Acquisition
The hardware layer consists of a wearable smart garment designed for comfort and continuous use. It employs flexible conductive fabric electrodes to capture high-quality ECG signals without the skin irritation common with traditional gel electrodes.
The system features a detachable electronics module containing an ADAS1000 analog front-end for signal conditioning and an ADuCM361 microcontroller for A/D conversion. Data is transmitted to the mobile application via a low-energy Bluetooth connection, ensuring long battery life and reliable real-time display.
Signal Processing and Machine Learning
To handle the noise inherent in daily activities, the cloud backend processes raw ECG data using an improved semi-soft thresholding method with Sym8 and DB2 wavelets. This removes baseline drift and motion noise while preserving the critical steep variations of the QRS complex.
To tackle the inherent imbalance in medical data, the authors apply the Synthetic Minority Over-sampling Technique (SMOTE). For classification, they deploy a StackingClassifier—an ensemble method that trains a meta-classifier on the predictions of several base models (like Random Forest, SVM, and XGBoost) to maximize predictive performance and avoid overfitting. This ensemble approach was found to outperform modern deep learning models (CNN, LSTM) on the SMOTE-augmented dataset, highlighting that well-engineered wavelet features combined with traditional ensemble ML can be more effective when inference speed and computational overhead are factors.
Experimental Results and Validation
The system was evaluated using both clinical benchmarks and real-world data:
- Clinical Benchmark (MIT-BIH): The StackingClassifier achieved 96% accuracy, a 95.6% recall, and an F1-score of 0.951, significantly outperforming baseline models like Random Forest and deep learning models like CNNs.
- Real-World Performance: On a proprietary dataset of 1.8 million heartbeats gathered from 250 volunteers, the model achieved an excellent AUC of 0.974. It maintained a high specificity of 98.9% and a sensitivity of 78.4%, translating to a remarkably low false alarm rate of just 1.1% on normal heartbeats.
- Hardware Accuracy: The wearable garment's sensors were validated against commercial medical devices, showing a heart rate error of only 1-2 bpm and a temperature deviation of < 1 °C.
Discussion and Conclusion
The study demonstrates that SMOTE is essential to boost recall and F1-scores to clinically acceptable levels, ensuring dangerous arrhythmias aren't missed. By testing on a massive proprietary dataset, the authors proved their model successfully balances high sensitivity with a low false-alarm rate in noisy, uncontrolled environments.
Future work will focus on optimizing hardware power consumption for longer battery life and developing more adaptive denoising algorithms to handle severe motion artifacts. Additionally, stricter data governance and validation on specific clinical cohorts are required before full-scale deployment. This integrated solution represents a significant step toward reliable self-health management and the prevention of cardiovascular diseases.
Summary updated