Continuous Stream Research Articles

Golden parachutes under the threat of accidents

AbstractThis paper addresses a continuous‐time contracting model that extends Sannikov's problem. In our model, a principal hires a risk‐averse agent to carry out a project. Specifically, the agent can perform two different tasks, namely to increase the instantaneous growth rate of the project's value, and to reduce the likelihood of accidents occurring. In order to compensate for these costly actions, the principal offers a continuous stream of payments throughout the entire duration of a contract, which concludes at a random time, potentially resulting in a lump‐sum payment. We examine the consequences stemming from the introduction of accidents, modeled by a compound Poisson process that negatively impact the project's value. Furthermore, we investigate whether certain economic scenarii are still characterized by a golden parachute as in Sannikov's model. A golden parachute refers to a situation where the agent stops working and subsequently receives a compensation, which may be either a lump‐sum payment leading to termination of the contract or a continuous stream of payments, thereby corresponding to a pension.

Customer Churn Prediction Approach Based on LLM Embeddings and Logistic Regression

Nowadays, predicting customer churn is essential for the success of any company. Loyal customers generate continuous revenue streams, resulting in long-term success and growth. Moreover, companies are increasingly prioritizing the retention of existing customers due to the higher costs associated with attracting new ones. Consequently, there has been a growing demand for advanced methods aimed at enhancing customer loyalty and satisfaction, as well as predicting churners. In our work, we focused on building a robust churn prediction model for the telecommunications industry based on large embeddings from large language models and logistic regression to accurately identify churners. We conducted extensive experiments using a range of embedding techniques, including OpenAI Text-embedding, Google Gemini Text Embedding, bidirectional encoder representations from transformers (BERT), Sentence-Transformers, Sent2vec, and Doc2vec, to extract meaningful features. Additionally, we tested various classifiers, including logistic regression, support vector machine, random forest, K-nearest neighbors, multilayer perceptron, naive Bayes, decision tree, and zero-shot classification, to build a robust model capable of making accurate predictions. The best-performing model in our experiments is the logistic regression classifier, which we trained using the extracted feature from the OpenAI Text-embedding-ada-002 model, achieving an accuracy of 89%. The proposed model demonstrates a high discriminative ability between churning and loyal customers.

Efficient designs for three-sequence stepped wedge trials with continuous recruitment.

The standard approach to designing stepped wedge trials that recruit participants in a continuous stream is to divide time into periods of equal length. But the choice of design in such cases is infinitely more flexible: each cluster could cross from the control to the intervention at any point on the continuous time-scale. We consider the case of a stepped wedge design with clusters randomised to just three sequences (designs with small numbers of sequences may be preferred for their simplicity and practicality) and investigate the choice of design that minimises the variance of the treatment effect estimator under different assumptions about the intra-cluster correlation. We make some simplifying assumptions in order to calculate the variance: in particular that we recruit the same number of participants, , from each cluster over the course of the trial, and that participants present at regularly spaced intervals. We consider an intra-cluster correlation that decays exponentially with separation in time between the presentation of two individuals from the same cluster, from a value of for two individuals who present at the same time, to a value of for individuals presenting at the start and end of the trial recruitment interval. We restrict attention to three-sequence designs with centrosymmetry - the property that if we reverse time and swap the intervention and control conditions then the design looks the same. We obtain an expression for the variance of the treatment effect estimator adjusted for effects of time, using methods for generalised least squares estimation, and we evaluate this expression numerically for different designs, and for different parameter values. There is a two-dimensional space of possible three-sequence, centrosymmetric stepped wedge designs with continuous recruitment. The variance of the treatment effect estimator for given and can be plotted as a contour map over this space. The shape of this variance surface depends on and on the parameter , but typically indicates a broad, flat region of close-to-optimal designs. The 'standard' design with equally spaced periods and 1:1:1 allocation rarely performs well, however. In many different settings, a relatively simple design can be found (e.g. one based on simple fractions) that offers close-to-optimal efficiency in that setting. There may also be designs that are robustly efficient over a wide range of settings. Contour maps of the kind we illustrate can help guide this choice. If efficiency is offered as one of the justifications for using a stepped wedge design, then it is worth designing with optimal efficiency in mind.

군인가정 자녀에 대한 국내 연구 동향 분석

The purpose of this study is to analyze the trends in research on military children in Korea and to suggest directions for promoting and supporting related research. To understand the overall trends in this field, the study addresses the following research question: “What are the trends in military children research in Korea?” A systematic literature review was conducted, focusing on previous studies concerning military children in Korea. Through this process, 56 articles were selected for in-depth review and analyzed according to research period, research field, methodology, subject population, research region, and variables or factors examined. The results of the analysis are as follows. The research on military children in Korea began in 1982, with a continuous stream of publications appearing from 2008 to 2022. The majority of studies were conducted within the social sciences, with a strong preference for quantitative research methods. Elementary school students were the most frequently studied population. Geographically, the research was concentrated in the Chungcheongbuk-do region. Commonly studied variables included academic, emotional, behavioral, social, family, and environmental factors, with family and environmental factors appearing most frequently. The 179 studies reviewed addressed variables primarily related to home environment, adjustment, school transfers, and parenting. This study is significant as it provides academic and policy implications by systematically reviewing research trends in the context of limited trend analyses on military children in Korea, and it highlights the need for a more diverse range of studies to support the healthy development of military children.

Diversity-Representativeness Replay and Knowledge Alignment for Lifelong Vehicle Re-identification

Lifelong vehicle re-identification (LVReID) aims to match a target vehicle across multiple cameras, considering non-stationary and continuous data streams, which fits the needs of the practical application better than traditional vehicle re-identification. Nonetheless, this area has received relatively little attention. Recently, methods for lifelong person re-identification (LPReID) have been emerging, with replay-based methods achieving the best results by storing a small number of instances from previous tasks for retraining, thus effectively reducing catastrophic forgetting. However, these methods cannot be directly applied to LVReID because they fail to simultaneously consider the diversity and representativeness of replayed data, resulting in biases between the subset stored in the memory buffer and the original data. They randomly sample classes, which may not adequately represent the distribution of the original data. Additionally, these methods fail to consider the rich variation in instances of the same vehicle class due to factors such as vehicle orientation and lighting conditions. Therefore, preserving more informative classes and instances for replay helps maintain information from previous tasks and may mitigate the model’s forgetting of old knowledge. In view of this, we propose a novel Diversity-Representativeness Dual-Stage Sampling Replay (DDSR) strategy for LVReID that constructs an effective memory buffer through two stages, i.e. , Cluster-Centric Class Selection and Diverse Instance Mining. Specifically, we first perform class-level sampling based on density in the clustered class-centered feature space and then further mine the diverse, high-quality instances within the selected classes. In addition, we introduce Maximum Mean Discrepancy loss to align the feature distribution between replay data and the new arrivals and apply L2 regularisation in the parameter space to facilitate knowledge transfer, thus enhancing the model’s generalization ability to new tasks. Extensive experiments demonstrate effective improvements of our method compared to current state-of-the-art lifelong ReID methods on the VeRi-776, VehicleID, and VERI-Wild datasets.

Recall as a Window into Hippocampally Defined Events.

We experience the present as a continuous stream of information, but often experience the past in parcels of unique events or episodes. Decades of research have helped to articulate how we perform this event segmentation in the moment, as well as how events and their boundaries influence what we later remember. More recently, neuroscientific research has suggested that the hippocampus plays a role at critical moments during event formation alongside its established role in enabling subsequent recall. Here, we review and explore the relationship between event processing and recall with the perspective that it can be uniquely characterized by the contributions of the hippocampus and its interactions with the rest of the brain. Specifically, we highlight a growing number of empirical studies suggesting that the hippocampus is important for processing events that have just ended, bridging the gap between the prior and current event, and influencing the contents and trajectories of recalled information. We also catalogue and summarize the multifaceted sets of findings concerning how recall is influenced by event structure. Lastly, we discuss several exciting directions for future research and how our understanding of events might be enriched by characterizing them in terms of the operations of different regions of the brain.

Sensitivity analysis for multi-measurement points based SHM in the mooring lines of floating offshore wind turbines

Abstract Structural health monitoring in floating offshore wind turbines’ mooring lines is vital for detecting early faults and preventing disruptions. Currently, sporadic and expensive monitoring is conducted via remote-operating vehicles. Efficient, automated, economical monitoring methods based on a continuous stream of data acquired via multiple measurement points in the wind turbine, are required. Such methods based on vibration signals have been investigated limitedly for steel chains. Recently, faulty synthetic mooring lines of a simulated 10MW semi-submersible wind turbine have been detected under varying environmental conditions and via the Functional Model Based Method (FMBM) equipped with functional models. The uncertainty due to the conditions’ stochasticity has been addressed by training the models using ten signal realizations per condition (wind) under the healthy wind turbine and from each of two measurement points. The current study presents a preliminary sensitivity analysis of the FMBM’s capability in detecting the simulated wind turbine’s faulty mooring lines when the functional models are trained using one signal realization per varying condition (wind) from each measurement point. The method’s effectiveness is evaluated with acceleration signals from 11 healthy/ 66 faulty (reduced stifness in one mooring line) cases. The FMBM detects all cases even when trained using one signal realization.

Initial Validation of the NOL Nociception Level Index® Monitoring System in Black and Multiracial People.

The NOL-Nociception Level Index® is a multiparameter index, based on artificial intelligence for the monitoring of nociception (physiological pain response) during anesthesia that has recently been authorized by the FDA. The monitor utilizes continuous streams of information from a finger probe comprising four sensors, including photoplethysmography, to provide a personalized nociception score on a scale of 0-100. Recent studies have suggested racial bias in pulse oximeter measurements due to the effect of melanin levels on photoplethysmography measurements. Therefore, there is a need to clinically validate new monitoring technologies in patients of all skin tones. The performance of the NOL scale in 8 patients that self-identified as Black or multiracial was compared to a database of 447 consented patients, assessing the response to surgical events at different levels of intensity. The descriptive, pilot data suggest that NOL performance in Black and multiracial patients is not different from the performance shown for the large database. Larger studies utilizing recognized skin tone scales to ensure accurate assessment of skin pigmentation are planned for the future.

Molecular epidemiological study on tick-borne pathogens in Qinghai Province, northwestern China

Recently, there has been a continuous stream of reports on emerging tick-borne pathogens affecting humans. Qinghai Province, located in the northeastern region, is one of China’s major pastoral areas, providing a suitable environment for the survival of ticks and the transmission of tick-borne pathogens. Here, we collected 560 free-living and parasitic ticks from 11 locations in Qinghai Province using the flag-drag method or tweezers, identifying them as belonging to 4 species of ticks. The overall positivity rate for tick-borne pathogens was 51.61 %, comprising Rickettsia (34.64 %), Anaplasma (5.00 %), Ehrlichia (2.14 %), Borrelia burgdorferi sensu lato (BBSL) (7.50 %), Babesia (0.18 %), and Theileria (5.89 %). Sequencing revealed the presence of 7 species of Rickettsia, 4 species of Anaplasma, 2 species of Ehrlichia, 2 species of BBSL, 1 species of Babesia, and 3 species of Theileria. Among the ticks, 6.43 % were found to be co-infected with 2 pathogens, while 0.36 % exhibited co-infection with 3 pathogens. Significant correlations (P < 0.05) were observed between the prevalence of tick-borne pathogens and factors including tick species, sex, developmental stages, parasitic status, and blood-feeding status. The results highlight the diverse distribution of tick-borne pathogens in the Qinghai Province, posing a significant threat to both local animal husbandry and human health. It underscores the need to enhance systematic monitoring of tick-borne pathogens in the local population and livestock.

Optimizing Open Radio Access Network Systems with LLAMA V2 for Enhanced Mobile Broadband, Ultra-Reliable Low-Latency Communications, and Massive Machine-Type Communications: A Framework for Efficient Network Slicing and Real-Time Resource Allocation.

This study presents an advanced framework integrating LLAMA_V2, a large language model, into Open Radio Access Network (O-RAN) systems. The focus is on efficient network slicing for various services. Sensors in IoT devices generate continuous data streams, enabling resource allocation through O-RAN's dynamic slicing and LLAMA_V2's optimization. LLAMA_V2 was selected for its superior ability to capture complex network dynamics, surpassing traditional AI/ML models. The proposed method combines sophisticated mathematical models with optimization and interfacing techniques to address challenges in resource allocation and slicing. LLAMA_V2 enhances decision making by offering explanations for policy decisions within the O-RAN framework and forecasting future network conditions using a lightweight LSTM model. It outperforms baseline models in key metrics such as latency reduction, throughput improvement, and packet loss mitigation, making it a significant solution for 5G network applications in advanced industries.

A real scale application of a novel set of spatial and similarity features for detection and classification of natural seismic sources from distributed acoustic sensing data

SUMMARY Distributed acoustic sensing (DAS) turns a fibre optic into a very dense network of equally distributed seismic sensors. We focused on the high-density sampling of the seismic wavefield, expressed in strain rates, measured by DAS. Classical approaches used to identify seismic signals rely on the recorded features at one station, but it is difficult to include spatial information in case of dense seismic station networks. This work aims at introducing new spatial and similarity features for seismic event classification suitable to analyse DAS observations. We propose a processing chain based on the XGBoost algorithm and the use of specifically designed spatiotemporal and similarity features for the event classification, and Markov random field for the spatial clustering. The methodology is designated to be applied on a continuous stream of DAS observations. We tested our processing chain to detect earthquakes and quarry blasts recorded in the region by permanent seismic networks and included in the RENASS catalogue. These events are part of a strain-rate seismic survey carried out during a 3 weeks campaign of DAS measurements along à 91 km fibre optic cable deployed in the central Pyrenees mountains (France). Despite the high anthropogenic activities along the fibre optic path, the proposed method succeeded in detecting earthquakes of magnitude &amp;gt;0.4 and quarry blasts of magnitude &amp;gt;1.0 while limiting the number of false alarms. This performance is particularly noteworthy for low-magnitude events, where detection is accomplished despite a lower signal-to-noise ratio compared to traditional seismometers. The methodology opens the door to real time detection and classification of seismic events measured with long-distance fibre optic systems.

Dynamics of real-time forecasting failure and recovery due to data gaps: A study using EnKF-based assimilation with the Lorenz model

Data assimilation-based real-time forecasting is widely used in meteorological and hydrological applications, where continuous data streams are employed to update forecasts and maintain accuracy. However, the reliability of the data source can be compromised due to sensor and communication failures or physical or cyber-attacks, and the impact of data stream failures on the accuracy of the forecasting system is not well understood. This study aims to systematically investigate the process of data stream failure and recovery for the first time. To achieve this, data gaps with varying lengths and timings are introduced to EnKF-based data assimilation system on the Lorenz model operating in both chaotic and periodic modes. Results show that the forecasting error grows exponentially in the chaotic mode but was limited in the periodic mode from the start of the data gap. For chaotic mode, the recovery of the system depends on the length of the data gap if the model error is not saturated; after saturation, the timing of the data stream recovery is important. Moreover, even long after restarting the data assimilation in the chaotic mode, the forecasting system cannot fully restore the original accuracy, while the periodic mode is generally resilient to disruption. This research introduces new metrics for quantifying system resilience and provides crucial insights into the long-term implications of data gaps, advancing our understanding of forecasting system behavior and reliability.

A standardised open science framework for sharing and re-analysing neural data acquired to continuous stimuli

Neurophysiology research has demonstrated that it is possible and valuable to investigate sensory processing in scenarios involving continuous sensory streams, such as speech and music. Over the past 10 years or so, novel analytic frameworks combined with the growing participation in data sharing has led to a surge of publicly available datasets involving continuous sensory experiments. However, open science efforts in this domain of research remain scattered, lacking a cohesive set of guidelines. This paper presents an end-to-end open science framework for the storage, analysis, sharing, and re-analysis of neural data recorded during continuous sensory experiments. We propose a data structure that builds on existing custom structures (Continuous-event Neural Data), providing a precise naming convention and data types, as well as providing a workflow for storing and loading data in the general-purpose BIDS structure. The framework has been designed to interface easily with existing toolboxes, such as EelBrain, NapLib, MNE, and the mTRF-Toolbox. We present guidelines by taking both the user view (how to rapidly re-analyse existing data) and the experimenter view (how to store, analyse, and share), making the process as straightforward and accessible. Additionally, we introduce a web-based data browser that enables the effortless replication of published results and data re-analysis.

Protecting Infinite Data Streams from Wearable Devices with Local Differential Privacy Techniques

The real-time data collected by wearable devices enables personalized health management and supports public health monitoring. However, sharing these data with third-party organizations introduces significant privacy risks. As a result, protecting and securely sharing wearable device data has become a critical concern. This paper proposes a local differential privacy-preserving algorithm designed for continuous data streams generated by wearable devices. Initially, the data stream is sampled at key points to avoid prematurely exhausting the privacy budget. Then, an adaptive allocation of the privacy budget at these points enhances privacy protection for sensitive data. Additionally, the optimized square wave (SW) mechanism introduces perturbations to the sampled points. Afterward, the Kalman filter algorithm is applied to maintain data flow patterns and reduce prediction errors. Experimental validation using two real datasets demonstrates that, under comparable conditions, this approach provides higher data availability than existing privacy protection methods for continuous data streams.

Sustainable enhancement of biogas production from a cold-region municipal wastewater anaerobic digestion process using optimized sludge-derived and commercial biochar additives

Anaerobic digestion (AD) of municipal wastewater sludges produces valuable solid digestate and biogas. Biogas is a source of clean energy and enhancement of its production has been of recent interest for increased electricity generation, among other products. The objective of this study was the development of a novel municipal sludge-derived biochar and its application in a municipal wastewater AD system to increase the biogas production rate. Thickened waste-activated sludge (TWAS) samples were collected from the cold-region municipal wastewater treatment plant and used to synthesize biochar applied in the simulation of AD processes using laboratory-scale reactors. The TWAS-derived biochar was synthesized using the commonly used furnace pyrolysis (sludge-based biochar, SBC), and more novel microwave pyrolysis including phosphoric acid as a microwave activator (activated sludge-based biochar, ASBC). The microwave pyrolysis conditions were optimized using a computational fluid dynamics (CFD) technique. In addition, various commercially available carbon-based additives were assessed for their impacts on the AD process including activated carbon, wood-derived biochar, and forest residue-derived biochar. Results showed that the ASBC increased the cumulative methane production by 50% (333 mL/g VS) versus the control sample (221 mL/g VS) after 30 d. The ASBC showed higher surface area, electrical conductivity, and metal contents versus the other biochars which boosted the AD microbial community growth leading to higher organic matter conversion into biogas. In the ASBC-amended digesters, the bacterial phylum Bacteroidota, which contains a major genus of the dgA-11-gut-group, exhibited a synergy between organic substrate fermentation and volatile fatty acid production, resulting in enhanced biogas production. The TWAS biochar demonstrated promising performance in enhancing the AD process fostering energy and resource self-sufficiency at municipal wastewater treatment plants. This smart sludge management aligns well with sustainable waste management practices and clean energy production strategies, especially considering that the biochar was sourced from a readily available continuous waste-product stream.

Learning incremental audio–visual representation for continual multimodal understanding

Deep learning methods have demonstrated remarkable success in processing static datasets for various video tasks. However, when confronted with continuous data streams, these approaches often encounter the challenge of catastrophic forgetting. This phenomenon leads to a significant decline in overall performance when learning new classes incrementally. Moreover, existing methods tend to overlook the correlation between audio and visual modalities in video incremental learning, despite their joint significance in scene comprehension. How to continuously learn from new classes while maintaining the knowledge of old videos with limited storage and computing resources is becoming imperative in the field of multimodal learning. In this paper, we introduce CavRL, a pioneering benchmark for audio–visual representation learning under class incremental scenarios. To mitigate catastrophic forgetting, we propose a rehearsal-based training approach that leverages a small exemplar set from previous classes. Our approach constrains the memory buffer within strict storage limits, optimizing exemplar selection by learning correlative audio–visual representations. Additionally, we employ a distillation method to mitigate forgetting in a self-supervised manner. Evaluations on two prevalent multimodal tasks: audio–visual event classification and audio–visual speaker recognition, which demonstrate that CavRL outperforms existing state-of-the-art incremental learning methods across various settings. We anticipate that CavRL will significantly advance research in continual multimodal learning.

Consonant lengthening marks the beginning of words across a diverse sample of languages

Speech consists of a continuous stream of acoustic signals, yet humans can segment words and other constituents from each other with astonishing precision. The acoustic properties that support this process are not well understood and remain understudied for the vast majority of the world’s languages, in particular regarding their potential variation. Here we report cross-linguistic evidence for the lengthening of word-initial consonants across a typologically diverse sample of 51 languages. Using Bayesian multilevel regression, we find that on average, word-initial consonants are about 13 ms longer than word-medial consonants. The cross-linguistic distribution of the effect indicates that despite individual differences in the phonology of the sampled languages, the lengthening of word-initial consonants is a widespread strategy to mark the onset of words in the continuous acoustic signal of human speech. These findings may be crucial for a better understanding of the incremental processing of speech and speech segmentation.

Floor-Level Occupancy Estimation of a Multi-Story Building Using Coarse Wi-Fi Data

In recent years, there has been an extraordinary increase in wireless capable devices and network infrastructure, which spawned a corresponding rise in data produced from the interactions of these technologies. Mobile devices constantly roam, leading to a perpetual dialog between a mobile device and wireless access points. This dialogue generates a continuous stream of device-specific data, including but not limited to a device’s media access control address, time of access, and received signal strength. Given the knowledge of the access point’s location and received signal strength, it is possible to infer the position of user devices and estimate their mobility and occupancy. This article presents two methods for accurately measuring floor-level occupancy in a multi-story building at Texas State University using coarse Wi-Fi log data. The first method employs a static filter, while the second incorporates user-role data and user location to create a dynamic filter. Quantitative methods are used to evaluate these filters against field-collected reference data and existing internal people-counting sensors. Our results demonstrate that the dynamic filter, leveraging variable thresholds, provides a more accurate estimation of occupancy compared to the fixed 5-min static filter which consistently overestimated occupancy. This research sheds light on the potential of dynamic filters derived from user-role data for precise floor-level occupancy estimations, with implications for various applications.

Stream salinity prediction in data-scarce regions: Application of transfer learning and uncertainty quantification

Scarcity of stream salinity data poses a challenge to understanding salinity dynamics and its implications for water supply management in water-scarce salt-prone regions around the world. This paper introduces a framework for generating continuous daily stream salinity estimates using instance-based transfer learning (TL) and assessing the reliability of the synthetic salinity data through uncertainty quantification via prediction intervals (PIs). The framework was developed using two temporally distinct specific conductance (SC) datasets from the Upper Red River Basin (URRB) located in southwestern Oklahoma and Texas Panhandle, United States. The instance-based TL approach was implemented by calibrating Feedforward Neural Networks (FFNNs) on a source SC dataset of around 1200 instantaneous grab samples collected by United States Geological Survey (USGS) from 1959 to 1993. The trained FFNNs were subsequently tested on a target dataset (1998-present) of 220 instantaneous grab samples collected by the Oklahoma Water Resources Board (OWRB). The framework's generalizability was assessed in the data-rich Bird Creek watershed in Oklahoma by manipulating continuous SC data to simulate data-scarce conditions for training the models and using the complete Bird Creek dataset for model evaluation. The Lower Upper Bound Estimation (LUBE) method was used with FFNNs to estimate PIs for uncertainty quantification. Autoregressive SC prediction methods via FFNN were found to be reliable with Nash Sutcliffe Efficiency (NSE) values of 0.65 and 0.45 on in-sample and out-of-sample test data, respectively. The same modeling scenario resulted in an NSE of 0.54 for the Bird Creek data using a similar missing data ratio, whereas a higher ratio of observed data increased the accuracy (NSE = 0.84). The relatively narrow estimated PIs for the North Fork Red River in the URRB indicated satisfactory stream salinity predictions, showing an average width equivalent to 25 % of the observed range and a confidence level of 70 %.

Adaptive Detection in Real-Time Gait Analysis through the Dynamic Gait Event Identifier.

The Dynamic Gait Event Identifier (DGEI) introduces a pioneering approach for real-time gait event detection that seamlessly aligns with the needs of embedded system design and optimization. DGEI creates a new standard for gait analysis by combining software and hardware co-design with real-time data analysis, using a combination of first-order difference functions and sliding window techniques. The method is specifically designed to accurately separate and analyze key gait events such as heel strike (HS), toe-off (TO), walking start (WS), and walking pause (WP) from a continuous stream of inertial measurement unit (IMU) signals. The core innovation of DGEI is the application of its dynamic feature extraction strategies, including first-order differential integration with positive/negative windows, weighted sleep time analysis, and adaptive thresholding, which together improve its accuracy in gait segmentation. The experimental results show that the accuracy rate of HS event detection is 97.82%, and the accuracy rate of TO event detection is 99.03%, which is suitable for embedded systems. Validation on a comprehensive dataset of 1550 gait instances shows that DGEI achieves near-perfect alignment with human annotations, with a difference of less than one frame in pulse onset times in 99.2% of the cases.