OPTIMIZATION OF DEEP NEURAL NETWORKS FOR PPG-BASED BLOOD PRESSURE ESTIMATION ON EDGE DEVICES

Continuous and non invasive blood pressure monitoring is crucial for hypertension diagnosis and cardiovascular diseases prevention. Photoplethysmography (PPG) sensors offer a promising solution to solve this challenge, but current estimation methods lack the precision needed to meet medical standards and do diagnosis. The most promising approach is based on deep learning models that have achieved remarkable accuracy in controlled environments; on the other hand, their deployment on wearable devices faces fundamental constraints. The massive computational requirements and memory footprint of these neural networks make them unfit for edge devices, which must operate within strict power and resource limitations. Hence, the current challenge resides in maintaining high estimation performance while using more lightweight Deep Neural Network (DNN) models that can fit the constraints of ultra low power edge devices, such as smartwatches. To cope with this challenge, this thesis proposes a fully automated DNN pipeline encompassing HW-aware Neural Architecture Search (NAS), Pruning and Quantization to generate models deployable on an ultra-low-power multicore System-on-Chip (SoC), GAP8. This pipeline leverages the gradient-based DNN optimization algorithms available in the PLiNIO library: SuperNet for coarse differential NAS, Pruning-In-Time (PIT) for architecture refinement and Mixed-Precision-Search for Quantization-Aware-Training. This thesis go throughout three main contributions for lightweight PPG-based blood-pressure estimations: i) first, we did a preliminary investigation on the relation between the PPG and Arterial Blood Pressure signals and on the impact of commonly used techniques like regularization or data augmentation to adapt the training of the new automatically-searched models; ii) then, we selected four open-source benchmarking datasets and two "seed" models, i.e., state-of-the-art deep learning models to be used as starting points for our optimization pipeline; in particular, we selected models for two different approaches: either a direct signal-to-label regression or the reconstruction of the whole Arterial Blood Pressure signal from PPG, followed by a peak detection. We utilized the best SoA models: a UNet for signal-to-signal PPG-to-ABP reconstruction and a ResNet for direct systolic and diastolic blood pressure regression. iii) finally, we applied the full pipeline to these models. The first phase of the pipeline obtained optimized architecture by selecting from different layer alternatives, achieving up to 4.99% lower error or a 73.36% parameter reduction at iso-error. By applying quantization at this stage, we showed that all models found can fit in GAP8 memory without loss in accuracy, while SoA networks are too large to fit the limited 512 kB on-chip memory. During the second step, we further refine the models by using the PIT NAS improving the Pareto front on all datasets and reaching a new accuracy record on the biggest three of them. PIT achieved up to 8.4% lower MAE or a 97.5% parameter reduction at iso-error.