研究目的
To assess the clinical usability of fingertip Photoplethysmogram (PPG) waveform by developing a sensor-agnostic method for discriminating between clean and noisy signals using novel Signal Quality Indices (SQIs) and a random forest classifier.
研究成果
The proposed sensor-agnostic method effectively classifies clean and noisy PPG signals with high sensitivity and specificity across different sensors (pulse oximeter and smartphone). It demonstrates robustness in real-time applications and improves the accuracy of disease classification, such as in Coronary Artery Disease analysis, by ensuring only high-quality signals are used. The novel SQIs contribute to reliable performance, and the approach shows potential for extension to other physiological signals like ECG or PCG.
研究不足
The annotation process is subjective and relies on expert majority voting, which may introduce bias. The method assumes a minimum window length of 8 seconds for analysis, which might not be optimal for all real-time applications. The datasets are specific to Indian populations and may not generalize globally. The signal sufficiency check criteria (e.g., heart rate range) are based on heuristics and could exclude valid physiological variations.
1：Experimental Design and Method Selection:
The study uses a hierarchical decision-making approach combining heuristics and machine learning. It involves signal sufficiency checks, template creation from cardiac cycles, extraction of novel and existing SQIs, and classification using a random forest classifier to distinguish clean and noisy PPG signals.
2：Sample Selection and Data Sources:
Datasets include PPG signals from 85 subjects (35 using a pulse oximeter and 50 using a smartphone camera), collected from an ICU in an urban hospital and a rural healthcare unit in India. Signals are segmented into 8-second windows, resulting in 1553 samples from the oximeter and 243 from the smartphone.
3：List of Experimental Equipment and Materials:
Pulse oximeter (Contec CMS 50D+), smartphone with in-house camera app, Matlab GUIDE software for annotation, and bandpass Butterworth filter for signal processing.
4：Experimental Procedures and Operational Workflow:
Data collection from the right index finger, filtering signals (
5：5-6 Hz cutoff), manual annotation by four experts using a GUI, signal sufficiency checks for clipping/flat signals and heart rate range (40-150 bpm), template creation from median cardiac cycles, extraction of 10 SQIs (including 4 novel ones), feature selection using mRMR, and classification with random forest. Data Analysis Methods:
Performance evaluation includes sensitivity, specificity, Welch's t-test for SQI significance, 5-fold cross-validation, SMOTE for class imbalance, and comparison with prior methods and classifiers like SVM, ANN, and KNN.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
