LSTM-only Model for Low-complexity HR Estimation from Wrist PPG.

Continuous and non-invasive cardiovascular monitoring has gained attention due to the miniaturization of wearable devices. Particularly, wrist-worn photoplethysmography (PPG) sensors present an alternative to electrocardiogram recording for heart rate (HR) monitoring as it is cheaper and non-intrusive for daily activities. Yet, the accuracy of PPG measurements is heavily affected by motion artifacts which are inherent to ambulatory environments. In this paper, we propose a low-complexity LSTM-only neural network for HR estimation from a single PPG channel during intense physical activity. This work explored the trade-off between model complexity and accuracy by exploring different model dataflows, number of layers, and number of training epochs to capture the intrinsic time-dependency between PPG samples. The best model achieves a mean absolute error of 4.47 ± 3.68 bpm when evaluated on 12 IEEE SPC subjects.Clinical relevance- This work aims to improve the quality of HR inference from PPG signals using neural network, enabling continuous vital signal monitoring with little interference in daily activities from embedded monitoring devices.

Fixed-Point NLMS and IPNLMS VLSI Architectures for Accurate FECG and FHR Processing.

Capturing signals without noise and interference while monitoring the maternal abdomen's fetal electrocardiogram (FECG) is a challenging task. This method can provide fetal monitoring for long hours, not harming the pregnant woman or the fetus. Such non-invasive FECG raw signal suffers from various interference sources as the bio-electric maternal potentials include her ECG component. Therefore, a critical step in the non-invasive FECG is to design the filtering of components derived from the maternal ECG. There is an increasing demand for portable devices to extract a pure FECG signal and to detect fetal heart rate (FHR) with precision. Dedicated CMOS architectures enable higher energy efficiency in portable devices. This paper proposes VLSI architectures dedicated to FECG extraction and FHR processing. Fixed-point architectures for the FECG detection exploring the NLMS (normalized least mean square), IPNLMS (improved proportional NLMS), and three different division VLSI CMOS architectures are designed herein. An architecture based on the Pan-Tompkins algorithm that processes the FECG for extracting the FHR, extending the functionally of the system, is also proposed. The results show that the NLMS and IPNLMS based architectures effectively detect the R-peaks of FECG with a detection accuracy of 92.86% and 93.75%, respectively. The synthesis results shows that our NLMS architecture proposal saves 13.3 % energy, due to a reduction of 279 clock cycles, compared to the state of the art. On the other hand, the IPNLMS algorithm results in +0.89% detection accuracy at the price of 42% additional energy consumption w.r.t NLMS.

Low-energy motion estimation memory system with dynamic management.

The digital video coding process imposes severe pressure on memory traffic, leading to considerable power consumption related to frequent DRAM accesses. External off-chip memory demand needs to be minimized by clever architecture/algorithm co-design, thus saving energy and extending battery lifetime during video encoding. To exploit temporal redundancies among neighboring frames, the motion estimation (ME) algorithm searches for good matching between the current block and blocks within reference frames stored in external memory. To save energy during ME, this work performs memory accesses distribution analysis of the test zone search (TZS) ME algorithm and, based on this analysis, proposes both a multi-sector scratchpad memory design and dynamic management for the TZS memory access. Our dynamic memory management, called neighbor management, reduces both static consumption-by employing sector-level power gating-and dynamic consumption-by reducing the number of accesses for ME execution. Additionally, our dynamic management was integrated with two previously proposed solutions: a hardware reference frame compressor and the Level C data reuse scheme (using a scratchpad memory). This system achieves a memory energy consumption savings of 99.8 % and, when compared to the baseline solution composed of a reference frame compressor and data reuse scheme, the memory energy consumption was reduced by 44.1 % at a cost of just 0.35 % loss in coding efficiency, on average. When compared with related works, our system presents better memory bandwidth/energy savings and coding efficiency results.

Analysis of ultrasonic vocalizations from mice using computer vision and machine learning.

Mice emit ultrasonic vocalizations (USVs) that communicate socially relevant information. To detect and classify these USVs, here we describe VocalMat. VocalMat is a software that uses image-processing and differential geometry approaches to detect USVs in audio files, eliminating the need for user-defined parameters. VocalMat also uses computational vision and machine learning methods to classify USVs into distinct categories. In a data set of >4000 USVs emitted by mice, VocalMat detected over 98% of manually labeled USVs and accurately classified ≈86% of the USVs out of 11 USV categories. We then used dimensionality reduction tools to analyze the probability distribution of USV classification among different experimental groups, providing a robust method to quantify and qualify the vocal repertoire of mice. Thus, VocalMat makes it possible to perform automated, accurate, and quantitative analysis of USVs without the need for user inputs, opening the opportunity for detailed and high-throughput analysis of this behavior.

Binary CorNET: Accelerator for HR Estimation From Wrist-PPG.

Research on heart rate (HR) estimation using wrist-worn photoplethysmography (PPG) sensors have progressed rapidly owing to the prominence of commercial sensing modules, used widely for lifestyle monitoring. Reported methodologies have been fairly successful in mitigating the effect of motion artifacts (MA) in ambulatory environment for HR estimation. Recently, a learning framework, CorNET, employing two-layer convolution neural networks (CNN) and two-layer long short-term network (LSTM) was successfully reported for estimating HR from MA-induced PPG signals. However, such a network topology with large number of parameters presents a challenge, towards low-complexity hardware implementation aimed at on-node processing. In this paper, we demonstrate a fully binarized network (bCorNET) topology and its corresponding algorithm-to-architecture mapping and energy-efficient implementation for HR estimation. The proposed framework achieves a MAE of 6.67 ± 5.49 bpm when evaluated on 22 IEEE SPC subjects. The design, synthesized with ST65 nm technology library achieving 3 GOPS @ 1 MHz, consumes 56.1 µJ per window with occupied 1634K NAND2 equivalent cell area and had a latency of 32 ms when estimating HR every 2 s from PPG signals.