This repository presents a signal processing and exploratory analysis study investigating the relationship between PPG-derived features and glucose levels, supported by invasively measured physiological parameters.
The work was conducted as part of a Master’s-level Signal Analysis coursework and is based on data collected using a custom-built physiological signal acquisition prototype.
Hemoglobin (Hb) directly absorbs green light and primarily affects the DC component of the PPG signal, while blood pressure (BP) and glucose influence vascular tone, thereby altering the AC component of the waveform.
This physiological relationship motivates the use of PPG morphology and derivative-based features for indirect glucose analysis, as supported by optical and cardiovascular literature.
- Data collected personally from mostly clinically healthy volunteers, including students and professionals, with a limited number of pre-diabetic individuals.
- All measurements were acquired under controlled conditions using a custom prototype.
- Invasive glucose measurements were obtained during data collection and used as the gold-standard reference.
- Additional measurements included:
- Systolic Blood Pressure (SBP)
- Diastolic Blood Pressure (DBP)
- Hemoglobin (Hb)- Invasive measurement & non inavsive from prototype
The final dataset consisted of 71 subjects after preprocessing.
The signal processing workflow includes:
- DC removal and Butterworth band-pass filtering for noise suppression
- Segmentation of clean PPG pulses based on signal quality criteria
- First and second derivative computation (APG analysis)
- Peak detection on PPG and APG signals
These steps follow established methodologies in PPG signal analysis literature.
Features were extracted from selected high-quality PPG pulses rather than short sliding windows, as glucose-related physiological changes occur gradually.
Extracted feature categories include:
- Statistical features
- Morphological PPG features
- Derivative-based (APG) features
- Physiological ratio-based features linked to vascular stiffness
Feature selection strategies were informed by prior studies (e.g., Elgendi 2012; Oh et al. 2003).
- Z-score normalization was applied to address large inter-feature scale variations.
- Spearman correlation was used to assess monotonic relationships between PPG-derived features and physiological parameters (Glucose, SBP, DBP, Hb).
- Correlation heatmaps were used to identify multicollinearity and feature redundancy.
Dimensionality reduction was applied separately using:
- Principal Component Analysis (PCA)
- Unsupervised
- Glucose excluded from PCA input
- Linear Discriminant Analysis (LDA)
- Supervised
- Glucose used as the class label
Multiple feature sets were evaluated independently.
- Glucose values were binarized using the median glucose level (101 mg/dL):
- Below median → Class 0
- Above median → Class 1
- Random Forest classifiers were trained on:
- Raw features
- PCA-reduced features
- LDA-transformed features
- Model selection used 5-fold cross-validation and grid search.
- Individual PPG features showed weak correlations with glucose.
- Combined physiological features improved discrimination.
- PCA revealed that ~95.75% of variance could be captured using 13 components.
- Outlier removal improved variance concentration and class separation.
- Best performance was achieved using LDA with all features, yielding:
- AUC ≈ 0.84 on held-out test data.
The results indicate that while individual PPG-derived features have limited predictive power, combined feature interactions, particularly involving blood pressure and hemoglobin, contribute meaningfully to glucose estimation.
This study highlights the potential of integrating physiological parameters into PPG-based glucose analysis but does not claim clinical applicability.
Future work includes validation on larger cohorts, algorithm optimization, and real-time integration into wearable systems.
- Main analysis notebook:
Effect_of_Physiological_Parameters_on_PPG_Signal_based_Glucose_Estimation.ipynb