Introduction
Photoplethysmography (PPG) offers a promising, non-invasive method for continuous blood pressure (BP) monitoring. However, a fundamental debate persists in the research community: should deep learning models treat BP prediction as a continuous regression task or a discrete classification problem? While regression aims for precise values (e.g., 120/80 mmHg), classification focuses on identifying clinically relevant categories like hypertension or normotension. This study investigates this trade-off, particularly in the context of model personalization.
Methodology
The researchers evaluated four Deep Neural Network (DNN) architectures: a modified 1D AlexNet and three ResNet variants (18, 34, and 50). To facilitate a direct comparison, regression outputs were mapped into discrete bins. Four distinct segmentation schemes were employed to test varying levels of granularity:
- hph: 3 bins (Hypotension, Normal, Hypertension).
- even4: 4 equally sized bins.
- dgk: 6 bins based on German Cardiac Society guidelines.
- even10: 10 narrow bins (10 mmHg wide), approximating regression granularity.
A key focus was the impact of personalization. Models were pre-trained on a large population and then fine-tuned using a small subset (10%) of individual-specific data.
Dataset and Processing
The study utilized the MIMIC-III database, extracting data from 4,000 subjects. PPG signals were processed into 5-second segments, filtered, and normalized. Ground truth Systolic Blood Pressure (SBP) was obtained from synchronized invasive arterial blood pressure (ABP) measurements. Training sets were carefully balanced across BP ranges to ensure the models learned physiological features rather than dataset averages.
Results and Discussion
The experimental results highlight several critical insights:
The Personalization Multiplier
Personalization proved to be the decisive factor in model performance. While pre-trained models showed comparable performance between regression and classification, fine-tuning on subject-specific data allowed classification models to achieve significantly higher accuracy. For the 3-bin (hph) task, personalized ResNet models reached accuracies between 80% and 90%.
Architecture and Granularity
Modern architectures like ResNet, which utilize skip connections, demonstrated a superior ability to benefit from personalization compared to AlexNet. This suggests that deeper models are more effective at extracting individual-specific PPG biomarkers. As expected, accuracy decreased as the number of bins increased, reflecting the inherent difficulty of fine-grained prediction.
Error Analysis
Confusion matrices revealed that model failures were often "near misses," with predictions falling into bins adjacent to the ground truth. This indicates that even when the exact category is missed, the models successfully capture the general hemodynamic trends.
Conclusion
For practical clinical applications, a coarse classification approach appears more robust and reliable than continuous regression. Personalization is not merely an enhancement but a fundamental requirement for achieving medically useful accuracy in PPG-based BP monitoring. Future work should focus on expanding subject diversity and addressing the challenges of real-world noise and motion artifacts in wearable settings.
Summary updated