End-to-end Automated Deep Neural Network Optimization for PPG-based Blood Pressure Estimation on Wearables
Authors: Francesco Carlucci, Giovanni Pollo, Xiaying Wang, Massimo Poncino, Enrico Macii, Luca Benini, Sara Vinco, Alessio Burrello, +1 more
Organizations: Interuniversity Department of Regional and Urban Studies and Planning at Politecnico di Torino, Torino, Italy · Department of Control and Computer Engineering at Politecnico di Torino, Torino, Italy · Department of Information Technology and Electrical Engineering at ETH Zurich, Zurich, Switzerland · Department of Electrical, Electronic, and Information Engineering "Guglielmo Marconi" at University of Bologna, Bologna, Italy
Photoplethysmography (PPG)-based blood pressure (BP) estimation is a challenging task, particularly on resource-constrained wearable devices. However, fully on-board processing is desirable to ensure user data confidentiality. Recent deep neural networks (DNNs) have achieved high BP estimation accuracy by reconstructing BP waveforms or directly regressing BP values, but their large memory, computation, and energy requirements hinder deployment on wearables. This work introduces a fully automated DNN design pipeline that combines hardware-aware neural architecture search (NAS), pruning, and mixed-precision search (MPS) to generate accurate yet compact BP prediction models optimized for ultra-low-power multicore systems-on-chip (SoCs). Starting from state-of-the-art baseline models on four public datasets, our optimized networks achieve up to 7.99% lower error with a 7.5x parameter reduction, or up to 83x fewer parameters with negligible accuracy loss. All models fit within 512 kB of memory on our target SoC (GreenWaves' GAP8), requiring less than 55 kB and achieving an average inference latency of 142 ms and energy consumption of 7.25 mJ. Patient-specific fine-tuning further improves accuracy by up to 64%, enabling fully autonomous, low-cost BP monitoring on wearables.
Figures & tables
Figure 1. SBP and DBP estimation from PPG (left) and ABP (right) signals.
Null data and saturated points detection, Butterworth Filter [0.5, 8] Hz, Min-Max Normalization, band-pass Butterworth filter [0.5, 45] Hz
6.57
14.39
Cheng et al. , 2021 ( Cheng et al., 2021 )
MIMIC-II
ABP-Net
Removal of short and unrealiable signals
3.73
6.41
Table 1. State-of-the-art table of methods with Custom data pre-processing
Work
Dataset
Architecture
Preprocessing
DBP MAE
SBP MAE
Gonzalez et al. , 2023 (Benchmark) ( González et al., 2023 )
Sensors
ResNet
Standard benchmark preprocessing (no fine-tuning)
8.33
17.46
UCI
8.30
16.59
BCG
7.76
12.20
PPGBP
8.61
13.62
Sensors
U-Net
7.66
15.64
UCI
7.88
16.93
Table 2. State-of-the-art table of methods with standardized pre-processing from ( González et al., 2023 ) . The bold numbers represents the most accurate models on the four datasets on both DBP and SBP.
Figure 2. Overview of the proposed automated DNN optimization flow. Starting from a dataset and a baseline model, the pipeline performs Neural Architecture Search (NAS), followed by structured pruning and mixed-precision search (quantization). Before deployment, selected candidate models can optionally undergo patient-specific fine-tuning using data from an individual subject to further improve accuracy.
Figure 3. SuperNet-based NAS. From left to right, the figure reports the initial condition of the network (a), the different layer options (b), the selection process during training (c) and finally the output model (d).
Figure 4. Pruning-in-Time (PIT) overview: starting from the initial layer sequence (a), PIT injects a trainable mask tensor θ2 to gate slices of the convolutional kernel W2 during training (b), yielding a pruned kernel W2′ and a reduced layer C2′ in the final network (c).
Figure 5. Mixed Precision Search (MPS). The figure shows the selection of the bitwidth of each convolutional layer, which leverages a SuperNet-like approach. Supported bitwidth are currently 2, 4, and 8.
Step
Description
1
Train the model on all samples from patients in Ptrain .
2
Evaluate on 20% of Ptest to obtain baseline accuracy.
3
Fine-tune the model on the remaining samples of Ptest .
4
Evaluate again on 20% of Ptest to assess the effect of fine-tuning.
Table 3. Subject-specific fine-tuning protocol
Figure 6. NAS results on all datasets on DBP (top row) and SBP prediction (bottom row).
Figure 7. Pruning results on all datasets on DBP (top row) and SBP (bottom row) prediction. Black circles are PIT seeds, red circles are the models discussed in Section 4.4 .
Figure 8. Results of MPS on all datasets on DBP (top row) and SBP (bottom row) prediction. Black circles are MPS seeds, red circles are the models discussed in Section 4.5 .
Figure 9. MAE comparison pre- and post- fine-tuning, without data shuffling, using 80% of the samples as training set and 20% test set (left), or 20% as training set and 20% as test set (right)
Figure 10. MAE comparison pre- and post- fine-tuning, with data shuffling, using 80% of the samples as training set and 20% test set (left), or 20% as training set and 20% as test set (right)
Model
SBP
DBP
MAE
Params
MAE
Params
[mmHg]
[#]
[mmHg]
[#]
PPGBP
Random Forest
13.17
20.3k
8.12
20.4k
Support Vector Regression
13.15
29.0k
8.04
16.9k
Conv-Transformer ( Lim et al., 2025 )
14.82
-
9.17
-
Table 4. Comparison with State of the Art methods from ( González et al., 2023 ) and ( Lim et al., 2025 ) .
Network
MACs [M]
Cycles
MACs/Cycles
DBP MAE
SBP MAE
Latency [ms]
Energy [mJ]
Memory [KByte]
RF (DBP + SBP)
n.a.
366.6K + 1.2M
n.a.
7.89
12.88
3.7 + 12.2
0.19 + 0.62
85.3 + 0.21
SVR (DBP + SBP)
n.a.
764.6K + 1.4M
n.a.
7.34
11.45
7.64 + 13.89
0.39 + 0.71
55.9 + 10.7
VNet
64.4
6.7M
9.55
8.01
11.42
67.49
3.44
491
Best DBP
47.09
13.2M
3.56
7.51 (2.68) †
11.31 (4.62) †
132.24
6.74
54.35
Best SBP
48.59
24.1M
2.01
8.28 (2.80) †
11.06 (4.70) †
241.33
12.31
30.26
Best Params
11.61
5.3M
2.20
9.32 (7.72) †
12.25 (11.06) †
52.87
2.70
2.15
Table 5. Deployment results on GAP8.
Figure 11. Visualization of the choices made by the pipeline during the optimization of the ’Best Params’ deployed model
Figure 12. Visualization of the choices made by the pipeline during the optimization of the ’Best SBP’ deployed model
Department of Informatics and Networked Systems University of Pittsburgh Pittsburgh, USA · Department of Biostatistics and Health Data Science University of Pittsburgh Pittsburgh, USA · Language Technologies Institute Carnegie Mellon University Pittsburgh, USA
1. Vanderbilt Institute for Surgery and Engineering, Vanderbilt University, Nashville, TN, 37232 · 6. Department of Biomedical Engineering, Vanderbilt University, Nashville, TN, 37235 · 2. Department of Electrical and Computer Engineering, Vanderbilt University, Nashville, TN, 37235 +3