Sleep apnea is a common sleep disorder that affects many people around the world. There is need for developing new screening method to improve the accessibility and affordability of apnea screening. In this paper, we developed ensemble classifiers with overnight blood oxygen saturation signals (SpO2) for the purpose of accurate classification between people with sleep apnea and healthy people. The ensemble classifiers (ECLF) are built on top of 5 base classifiers, including logistic regression (LR), random forest (RF), support vector machine (SVM), linear discrimination analysis (LDA), and light gradient boosting machine (LGMB). The output of the ECLF is weighted voting of each classifier. Performance evaluation analysis showed that when heavier weights were assigned to the LR and SVM classifiers, the ECLF achieved a better balance between sensitivity (0.81 ± 0.02) and specificity (0.80 ± 0.02) while maintaining the overall performance as measured by AUC (0.81 ± 0.01). RF and LDA achieved high sensitivity (> 0.95) at the sacrifice of specificity, while LR and SVM achieved high specificity (> 0.80) at the sacrifice of sensitivity. LGMB demonstrated mediocre performance on both sensitivity and specificity

The swift advancement in wearable sensor technology has made continuous health monitoring accessible, opening doors for myriad health-related applications. Nevertheless, managing and integrating real-time data from multiple wearable sensors remains a notable challenge. This research offers an innovative solution to this issue - a unified cross-platform application capable of integrating real-time time-series data from various wearable sensors, such as the Apple Watch and Empatica E4.

The application, developed using the Flutter framework, streamlines the process of collecting, managing, and analyzing sensor data, thereby significantly easing the task for health professionals and researchers. The application can simultaneously capture and integrate various physiological signals, such as heart rate, acceleration, and skin temperature. Our application ensures compatibility across iOS and Android platforms, extending its accessibility to a broader user base.

In order to competently handle the surge of substantial time-series data, we utilize InfluxDB, a robust time-series database, to serve as the data storage infrastructure. Each recording session's data is stored in a uniquely identified InfluxDB bucket, providing efficient data management and retrieval, even when handling substantial data volumes from multiple sources. Moreover, it enables real-time analytics derived directly from sensor data.

As a significant contribution to the mHealth research field, this research's main output is a robust, scalable, and user-friendly application capable of seamlessly integrating a wide range of sensor data. It leverages the capabilities of a time-series database for effective data management and serves as a practical tool for health researchers and practitioners working with wearable sensor technology.

This paper presents a detailed design of an inexpensive, simple, and scalable Distributed Sensor Network (DSN). Each sensor’s hardware consists of a Raspberry Pi 3 Model B microprocessor, a specifically designed and factory-made Printed Circuit Board (PCB), an Uninterruptible Power Supply (UPS) Hat based on a High Capacity Lithium Polymer battery (LiPo), Power over Ethernet Splitter allowing for 5V power via microUSB (from a PoE injector), a GPS receiver, a LoRaWAN module, control push buttons, OLED screen, a Four-wire I2C connector, a 40-pin Raspberry Pi connector, a status LED, and five GPIO connectors configured for Signal, Voltage, Ground (SVG). Each sensor is built to capture GPS, Wi-Fi, and Bluetooth signals and sends this information to a network controller implementing a LoRaWAN gateway. Each sensor’s software is developed so all applications run on top of a Linux operating system. The layer above it includes system daemon applications, such as Air-mon, HCI tools, GPSd, and networking support. An SQLite database sits on top of the daemon applications and records the captured information. Information from each sensor forming the DSN that is received by the LoRaWAN gateway is communicated to the user via a web-server. The web-server is implemented by a custom Python script. A series of tests were carried out to ensure the DSN was functioning as expected. Based on the successful outcome of these tests, the DSN was deployed in a research project. The total cost to build a sensor, including the cost of the Raspberry Pi, was less than $160 (Australian). The inclusion of a LoRaWAN module, control push buttons, OLED screen, and status LED made accessing and controlling the DSN remotely a simple process. The scalability of the sensor's PCB, which has a Four-wire I2C connector, makes the device capable of performing numerous types of sensing and control functions.

The Amazon rainforest corresponds to one-third of the world's tropical forest area. This makes it indispensable for maintaining global biodiversity. However, the increasing occurrences of wildfires and deforestation in the region are notorious. In this sense, it is essential to protect forests to ensure the quality of life for future generations and prevent damages that affect the entire planet. In this work, a real-time monitoring device is proposed to identify attempts of deforestation through audio signals from tractors and chainsaws, using embedded artificial intelligence. Additionally, it is capable of communicating with a base station, reaching distances close to 1km in dense forest, through LORA communication. A user interface has also been developed, providing daily alerts such as attack identification, occurrence times, device locations, and battery status. The system has an average power consumption of around 300nA, employing power management methods defined as ultra-low power mode, sleep mode, prediction mode, and transmission mode. Hence, the device has the potential to promote the sustainable preservation of the Amazon rainforest, helping to prevent large-scale illegal deforestation.

In the rapidly evolving landscape of human-robot collaboration (HRC), a challenge lies in endowing robotic systems with the capability to adapt seamlessly to users' internal states, such as stress or relaxation. Current research in the field demonstrates significant progress in accurately classifying stress through the integration of diverse sensors that monitor vital signs. However, there is limited investigation into predicting future states. By extending our understanding beyond the instantaneous recognition of emotional states to anticipatory modeling, robotic systems can proactively tailor their interactions, fostering a deeper, more intuitive, and ultimately more productive collaboration with human users.

Our research investigates an approach to forecasting human vital signs by formulating the problem as a sequence-to-sequence (seq2seq) task, utilizing bidirectional long short-term memory models (BiLSTM). The study aims to compare the forecasting accuracy of uni- and multivariate modeling strategies while investigating their performance over different forecasting horizons ranging from 1 second to 10 seconds.

The dataset used in this research comprises sensor data collected during a lab study. Thirteen participants engaged in a collaborative assembly scenario with a collaborative robot, YuMi by ABB Robotics which is primarily used in industrial HRC assembly applications. The dataset includes diverse sensor modalities capturing heart rate (HR), pupil diameter (PD), and electrodermal activity (EDA). Prior to analysis, the data undergoes state-of-the-art preprocessing techniques. To evaluate forecasting accuracy, the Symmetric Mean Absolute Percentage Error (sMAPE) metric is utilized, maintaining consistent look-back and forecasting window lengths.

Our results show that univariate models outperform multivariate ones in terms of forecasting accuracy, offering valuable insights into accurate forecasting of human physiological parameters, with potential implications for personalized medical monitoring, diagnostics, and healthcare applications.

Poly(3,4-ethylenedioxythiophene) (PEDOT) is known as an interesting material for optoelectronic applications. Since the solubility of PEDOT compound is a critical element for the application, a lot of synthetic procedures such as the encapsulation of PEDOT backbone into the macrocyclic cavities via noncovalent interactions have been applied.

Herein, we continue to provide extensive insights into the excited state dynamics of absorption and emission (fluorescence and phosphorescence) in water and acetonitrile (ACN) solutions of two PEDOT∙TMe-βCD and PEDOT∙TMe-γCD polyrotaxanes. The transient absorption map revealed ground state bleaching bands (GSB) in the range 270- 315 nm, whereas an absorption band in excited state (ESA) occurs at shorter wavelengths from 210 to 250 nm and more than one excited state (S_n > 1). At longer wavelength, from 390 to 455 nm negative bands appeared which can be assigned to the stimulated emissions (SE). Also, the quantum yields measurements and time resolved fluorescence experiments were performed in both water and ACN in solutions. The quantum yields with absolute values from 0.0005 to 0.76 were found depending on the solvent nature, and lifetimes from 0.4 ns to 29 ns were obtained for fluorescence and phosphorescence lifetime with values ​​in the range 0.8 to 9 μs.

ACKNOWLEDGMENTS

This presentation was supported by a grant from the Romanian Ministry of Research, Innovation, and Digitization, CNCS-UEFISCDI, project number PN-III-P4-PCE-2021-0906

Gait analysis plays a vital role in medicine as it can help diagnose illness, monitor recovery, and measure physical performance. Related work in gait analysis has primarily utilized laboratory data due to its inherently low noise and ease of preprocessing. Daily data, gathered through wearables sensors, can also significantly impact medical care. Nonetheless, working with such data poses numerous challenges.

This article proposes an algorithm to solve the problems associated with gait segmentation in daily data obtained by inertial measurement units (IMUs) on wearable devices. The proposed algorithm can handle time-series data collected by wearable IMU sensors, including noise and different gaits. It also remains effective even when the placement of wearable sensors is in a non-standard manner, making it well-suited for use in various settings.

Data was collected using Xsens wearable device, which primarily employ inertial measurement units (IMUs) for capturing precise movement data. Principal component analysis (PCA) was used to reduce dimensionality. The proposed algorithm within this article can identify the start and end points of each gait segment within the time series, and the same type of gait will be grouped together.

Based on comparison with data marked manually, our algorithm achieved high performance for real life gait segmentation, while also demonstrating strong ability to distinguish between different gait patterns. Moreover, the versatility of the algorithm shows promising applications in fields such as rehabilitation, disease evaluation, and sports performance optimization.

The time-frequency analysis has been gaining a significant amount of attention for research purposes as its applications in the study of non-stationary signals offers meaningful information that is usually suppressed by the conventional analysis in the time or frequency domains. In this context, the smoothed pseudo Wigner-Ville distribution (WVD) for analytic signals arises as a reliable time-frequency tool that is used in the study of various different signal data. Due to the extensive computational operations involved in generating the WVD, the objective of this study is to explore approaches that reduce the computational cost associated with analyzing large datasets using the mentioned tool. The dataset used comprised a 9000-sample acoustic signal obtained from a milling machine run, with a sampling rate of 100 kHz captured by a sensor and filtered in its 1kHz to 8kHz band by a pass-band butterworth filter of 5th order. Two approaches were, then, pursued, the first involved calculating the average WVD of three time-frequency transformations obtained from equidistant time windows of the signal. The second involved reducing the sampling rate of the analyzed signal by a factor of k by creating an array where each nth element corresponded to the k*nth element of the original signal. The mean WVD method distorted the signal’s time-frequency diagram by adding middle range frequencies throughout its entire time axis regardless of the time window taken, while the second approach showed an identical WVD to the one of the original signal even for big k factors, decreasing the analysis time greatly. This leads to the conclusion that diminishing the sample frequency of the signal is a viable way to overcome the computational cost of the WVD calculation for the study of the behavior of the signal’s low-frequency components.

Federated learning (FL) is a field in distributed optimization. Therein, data collection and neural network (NN) training are decentralized, which means these tasks take place in many clients with limited communication and computation capabilities. In FL, the client NNs are trained with locally available data. Next, the client NNs are aggregated in order to update a global NN.

In this work, we apply FL on a small live gesture sensing NN for low-power 60GHz frequency modulated continuous wave (FMCW) radar from Infineon Technologies. Furthermore, we investigate the effects of client data heterogeneity on the gesture recognition accuracy. The presented FL algorithms are evaluated on a diverse dataset including approximately 26k gesture recordings. Each recording contains a specific gesture execution sequence, which is labeled accordingly as that distinct gesture. The rest of the recording, which does not contain any specific gesture, is simply labeled as the background. To study FL with varying data heterogeneity, the data among different clients is partitioned in two different ways. The first partition, called independently and identically distributed (iid) partition, involves shuffling and equally distributing the dataset among clients using the same number of local epochs during training. The second partition is non-iid, where each client is assigned a different number of gesture recordings and local epochs. It is shown that, FL converges in the iid partition to an accuracy higher than 92.4%. However, the increasing data heterogeneity degrades the accuracy to 78.3%. To address accuracy degradation resulting from client heterogeneity, we propose dynamically weighting the recordings during training based on the varying ratio of distinct gesture sequences and background in each client. Moreover, regularization terms are included in the loss function to prevent client drift and overconfidence in the local NN predictions. Finally, it is shown that the proposed adaptions reduce the accuracy degradation, such that 96.4% label accuracy, with the highest degree of data heterogeneity (one gesture per client), is obtained.

In this presentation transient absorption (TA) and steady state, as well as time-resolved fluorescence spectroscopy were used to investigate the mechanisms of fluorescence quenching in order to obtain new sensors for water detection in solutions of methanol, dimethylformamide and dioxane. The new synthesized naphthalimide compounds present a good fluorescence, high quantum yield, stability and sensitivity. We chose to approach the investigation of these naphthalimide derivatives for the theoretical information, and also some applications resulting from this study, as sensors. For fundamental study, the dynamic and static quenching theories as well as the combined dynamic and static method were applied. The emission lifetime in excited state, and the quantum yield were estimated. In the presence or absence of water, ground state bleaching bands (GSB) are present after 230 nm in the TA maps. An absorption band in excited state (ESA) occurs at shorter wavelengths after 210 nm, and at longer wavelength, after 430 nm negative bands appeared, which can be assigned to the stimulated emissions (SE).

The obtained results suggest that these naphthalimide-based derivatives can act as potential sensors detecting low amounts of water.