Similar Performance Research Articles

Feature selection and its combination with data over-sampling for multi-class imbalanced datasets

Feature selection aims at filtering out some unrepresentative features from a given dataset in order to construct more effective learning models. Furthermore, ensemble feature selection by combining multiple feature selection methods has shown its outperformance over single feature selection. However, the performances of different (ensemble) feature selection methods have not been fully examined over multi-class imbalanced datasets. On the other hand, for class imbalanced datasets, one widely considered solution is to re-balance the datasets by data over-sampling, which generates some synthetic examples for the minority classes. However, the effect of performing (ensemble) feature selection on over-sampling multi-class imbalanced datasets has not been investigated. Therefore, the first research objective is to examine the performances of single and ensemble feature selection methods by fifteen well-known filter, wrapper, and embedded algorithms in terms of classification accuracy. For the second research objective, two orders of combining the feature selection and over-sampling steps are compared in order to find out the best combination procedure as well as the best combined algorithms. The experimental results based on ten different domain datasets containing low to very high feature dimensions show that ensemble feature selection methods slightly perform better than single ones. However, their performance differences are not big. To combine with the Synthetic Minority Oversampling Technique (SMOTE) over-sampling algorithm, performing feature selection first and over-sampling second outperforms the other procedure. Although the best combined algorithms are based on ensemble feature selection, eXtreme Gradient Boosting (XGBoost), as the single best feature selection algorithm, combined with SMOTE provides very similar classification performance to the best combined algorithms. To consider the issues of classification performance and compactional cost, the optimal solution is based on the combined XGBoost and SMOTE.

The Analysis of Physical Performance During Official Competitions in Professional English Football: Do Positions, Game Locations, and Results Influence Players' Game Demands?

Beato, M, Youngs, A, and Costin, AJ. The analysis of physical performance during official competitions in professional English football: Do positions, game locations, and results influence players' game demands? J Strength Cond Res 38(5): e226-e234, 2024-This study aimed, first, to verify if physical parameters were different between players' positions during official matches in English professional football, and second, if the game location (H and A) or results (W, L, D) affected players' physical performance. Twenty-five male professional football players of the same team were included in this data analysis (age = 27 ± 9 years) during the 2022-23 season. Players were divided into positions: center backs, wide backs, center midfielders, attacking midfielders, and strikers. The external load parameters were distance covered, high-speed running (HSR), sprinting distance, accelerations, decelerations, and high metabolic load distance (HMLD, meters > 25.5 w·kg -1 ) that were monitored using global navigation satellite systems Apex (STATSports). Linear mixed models' analysis for positions reported a significant difference in total distance ( p = 0.011), HSR ( p < 0.001), sprinting distance ( p < 0.001), accelerations ( p = 0.003), decelerations ( p = 0.002), and HMLD ( p < 0.001). Centre backs showed the lowest physical performance in the metrics analyzed, whereas players in the other positions frequently displayed a similar physical performance. Regarding locations and results, differences were only found between locations for decelerations ( p = 0.041) and between results for HMLD ( p = 0.010). In conclusion, physical performance was influenced by players' positions, whereas game location or results seem to not affect physical performance during official competitions. Consequently, practitioners can physically prepare their players independently from the location of the match or of the possible game results, whereas specific positional training is requested to optimally prepare their players.

Enhancing Smart Building Surveillance Systems in Thin Walls: An Efficient Barrier Design.

This paper introduces an efficient barrier model for enhancing smart building surveillance in harsh environment with thin walls and structures. After the main research problem of minimizing the total number of wall-recognition surveillance barriers, we propose two distinct algorithms, Centralized Node Deployment and Adaptation Node Deployment, which are designed to address the challenge by strategic placement of surveillance nodes within the smart building. The Centralized Node Deployment aligns nodes along the thin walls, ensuring consistent communication coverage and effectively countering potential disruptions. Conversely, the Adaptation Node Deployment begins with random node placement, which adapts over time to ensure efficient communication across the building. The novelty of this work is in designing a novel barrier system to achieve energy efficiency and reinforced surveillance in a thin-wall environment. Instead of a real environment, we use an ad hoc server for simulations with various scenarios and parameters. Then, two different algorithms are executed through those simulation environments and settings. Also, with detailed discussions, we provide the performance analysis, which shows that both algorithms deliver similar performance metrics over extended periods, indicating their suitability for long-term operation in smart infrastructure.

Transferring a deep learning model from healthy subjects to stroke patients in a motor imagery brain–computer interface

Objective. Motor imagery (MI) brain–computer interfaces (BCIs) based on electroencephalogram (EEG) have been developed primarily for stroke rehabilitation, however, due to limited stroke data, current deep learning methods for cross-subject classification rely on healthy data. This study aims to assess the feasibility of applying MI-BCI models pre-trained using data from healthy individuals to detect MI in stroke patients. Approach. We introduce a new transfer learning approach where features from two-class MI data of healthy individuals are used to detect MI in stroke patients. We compare the results of the proposed method with those obtained from analyses within stroke data. Experiments were conducted using Deep ConvNet and state-of-the-art subject-specific machine learning MI classifiers, evaluated on OpenBMI two-class MI-EEG data from healthy subjects and two-class MI versus rest data from stroke patients. Main results. Results of our study indicate that through domain adaptation of a model pre-trained using healthy subjects’ data, an average MI detection accuracy of 71.15% ( ±12.46% ) can be achieved across 71 stroke patients. We demonstrate that the accuracy of the pre-trained model increased by 18.15% after transfer learning (p < 0.001). Additionally, the proposed transfer learning method outperforms the subject-specific results achieved by Deep ConvNet and FBCSP, with significant enhancements of 7.64% (p < 0.001) and 5.55% (p < 0.001) in performance, respectively. Notably, the healthy-to-stroke transfer learning approach achieved similar performance to stroke-to-stroke transfer learning, with no significant difference (p > 0.05). Explainable AI analyses using transfer models determined channel relevance patterns that indicate contributions from the bilateral motor, frontal, and parietal regions of the cortex towards MI detection in stroke patients. Significance. Transfer learning from healthy to stroke can enhance the clinical use of BCI algorithms by overcoming the challenge of insufficient clinical data for optimal training.

Comparison of home detection algorithms using smartphone GPS data

Estimation of people’s home locations using location-based services data from smartphones is a common task in human mobility assessment. However, commonly used home detection algorithms (HDAs) are often arbitrary and unexamined. In this study, we review existing HDAs and examine five HDAs using eight high-quality mobile phone geolocation datasets. These include four commonly used HDAs as well as an HDA proposed in this work. To make quantitative comparisons, we propose three novel metrics to assess the quality of detected home locations and test them on eight datasets across four U.S. cities. We find that all three metrics show a consistent rank of HDAs’ performances, with the proposed HDA outperforming the others. We infer that the temporal and spatial continuity of the geolocation data points matters more than the overall size of the data for accurate home detection. We also find that HDAs with high (and similar) performance metrics tend to create results with better consistency and closer to common expectations. Further, the performance deteriorates with decreasing data quality of the devices, though the patterns of relative performance persist. Finally, we show how the differences in home detection can lead to substantial differences in subsequent inferences using two case studies—(i) hurricane evacuation estimation, and (ii) correlation of mobility patterns with socioeconomic status. Our work contributes to improving the transparency of large-scale human mobility assessment applications.

High-fidelity and polarization-insensitive universal photonic processors fabricated by femtosecond laser writing

Abstract Universal photonic processors (UPPs) are fully programmable photonic integrated circuits that are key components in quantum photonics. With this work, we present a novel platform for the realization of low-loss, low-power, and high-fidelity UPPs based on femtosecond laser writing (FLW) and compatible with a large wavelength spectrum. In fact, we demonstrate different UPPs, tailored for operation at 785 nm and 1550 nm, providing similar high-level performances. Moreover, we show that standard calibration techniques applied to FLW-UPPs result in Haar random polarization-insensitive photonic transformations implemented with average amplitude fidelity as high as 0.9979 at 785 nm (0.9970 at 1550 nm), with the possibility of increasing the fidelity over 0.9990 thanks to novel optimization algorithms. Besides being the first demonstrations of polarization-insensitive UPPs, these devices show the highest level of control and reconfigurability ever reported for a FLW circuit. These qualities will be greatly beneficial to applications in quantum information processing.

Dry ice sublimation performance as affected by binding agent, density, and age

AbstractDry ice is one of the world's most in‐demand commodities for cold chain distribution of temperature sensitive products in insulated shippers. A commonly used rule of thumb is that dry ice sublimates about 8% every 24 h, without consideration for dry ice type, size, or geometry. This study explored sublimation rates of dry ice composition (binding agent and density), age, and geometry (size and shape). Dry ice block manufacturing often involves dosing propylene glycol‐water “binder” solution to liquid carbon dioxide. Tests performed in this study revealed that the binding agent did not influence the dry ice sublimation rate. Dry ice density was found to impact sublimation rate. Fresh and aged blocks of similar dimensions and mass were found to have similar performance outcomes. Mass loss curves for pellets and various sized dry ice blocks showed a dependence on surface area and volume. Chunked blocks and pellets with and without binding agent sublimated at similar rates of 2.84 and 2.53% h−1 (w/w), respectively, whereas block varieties with significantly lower surface area to volume ratios had lower sublimation rates of 0.98–1.60% h−1.

Establishment of reaction engineering approach for saline droplet and its application in spray evaporation simulation

The statistical study of droplets in spray evaporation tower has guiding significance for understanding spray evaporation and designing drying tower. The key to spray evaporation simulation is establishing a suitable single droplet evaporation model through experiment. In this paper, a visualized single droplet evaporation experiment platform was established. The mass, diameter, and temperature during the evaporation of 10 % NaCl solution droplets were obtained. The Reaction Engineering Approach model was developed and validated to describe the single droplet evolution process in saline water spray. A comparison calculation between the spray evaporation of pure and saline water was executed, and the distributions of residence time of droplets were obtained. It reveals that droplets with initial diameters less than 117 μm can fully evaporate, and for larger droplets, pure water droplets represent a higher evaporation ratio. About 50 % of the fully evaporated saline droplets stay in the backflow area. Contrast numerical experiments of saline water spray evaporation were performed, concluded that similar performance can be attained with the same gas-liquid ratio, and excessive inlet air will reduce the residence time of droplets. The influence of other operating parameters on the performance was also discussed, providing a comprehensive understanding of the regularity of spray drying.

Reconnection of cysteine in reduced hair with alkylene dimaleates via thiol-Michael click chemistry.

Conventional hair permanent waving (PW) and permanent straightening processes typically involve two steps: reduction, for breaking -S-S- bond in cystine into cysteine and oxidation for -S-S- bond reconnection. However, it is known that the hair incurs damage during the oxidation step. In this work, we proposed a novel strategy to reconnect reduced disulfide bonds in hair via the thiol-Michael click reaction, by using a symmetric Michael reagent. Virgin black Chinese hair was reduced using 8% wt thioglycolic acid and employed as model hair containing a high content of broken disulfide bonds. The reduced hair was treated with 1,4-n-butylene dimaleate. Raman spectroscopy and Fourier transform infrared spectroscopy (FT-IR) were used to verify the chemical changes occurred in untreated and treated hair fibre. Single-fibre mechanical properties and thermal properties of the hair were evaluated using tensile testing and differential scanning calorimetry (DSC), respectively. The 1,4-n-butylene dimaleate could reconnect free thiol groups generated by disulfide bond reduction via thiol-Michael click reaction and significantly improve the mechanical strength of hair compared to that of the reduced hair. Secondary conformational resolution analysis of FT-IR results revealed that the content of α-helix structure could be restored after treatment with 1,4-n-butylene dimaleate. The intermolecular forces established by the newly generated C-S bonds compensate the broken disulfide bonds and enhance the fracture strength of the hair compared to that of reduced hair. Michael reagents of similar structure also showed similar performance in restoring the mechanical properties of reduced hair. Our data suggest that 1,4-n-butylene dimaleate can restore the mechanical properties of reduced hair by reconnecting reduced disulfide bonds and restoring the secondary conformation of hair keratin.

Rejuvenating oxidized paving asphalts – When softeners meet asphaltene dispersants

The search and development of various asphalt rejuvenators have been following a trial-and-error methodology in essence; the imperative need of improved efficiency and versatility yearns for a general principle/framework of formulating rejuvenators. In this study, we propose one such formulation concept that rejuvenators can be designed by combining asphaltene dispersants (interacting with and dispersing asphaltenes) with softeners (diluting asphaltenes). This concept was preliminarily validated in an experimental program involving three rejuvenation schemes, namely using an aromatic softener alone (as the reference) and using it together with an ionic liquid or a phenolic compound (both acting as dispersants). Immediate and long-term rejuvenating effectiveness was evaluated based on viscoelastic properties, fatigue resistance, and moisture susceptibility. The results indicated that at respective optimum dosages, the three schemes all yielded satisfactory immediate performance of rejuvenation. After long-term aging, the rejuvenated asphalts demonstrated more desired viscoelastic characteristics in resisting high-temperature deformation and low-temperature cracking, and were more fatigue resistant, as compared to the virgin binder. Rejuvenation in all cases also recovered the moisture resistance nearly to the virgin level. Infrared spectrum analysis indicated that the performance restoration was accomplished by the agents blending physically with the aged binder, which decreased the concentrations of oxidation functional groups (carbonyl and sulfoxide). The incorporation of dispersants improved the aging resistance (in terms of high temperature viscoelasticity), and substantially reduced the need of aromatic softener while providing similar or better performance compared to using the softener alone, which would lead to pronounced environmental and economic benefits. The much higher dosage efficiency was ascribed to the capability of the dispersants in peptizing and dissociating the asphaltenes, as investigated by molecular dynamics simulations. The different dispersing behaviors of the additives were examined with reference to their chemical compositions and molecular structures.

Improved accuracy and efficiency of primary care fall risk screening of older adults using a machine learning approach.

While many falls are preventable, they remain a leading cause of injury and death in older adults. Primary care clinics largely rely on screening questionnaires to identify people at risk of falls. Limitations of standard fall risk screening questionnaires include suboptimal accuracy, missing data, and non-standard formats, which hinder early identification of risk and prevention of fall injury. We used machine learning methods to develop and evaluate electronic health record (EHR)-based tools to identify older adults at risk of fall-related injuries in a primary care population and compared this approach to standard fall screening questionnaires. Using patient-level clinical data from an integrated healthcare system consisting of 16-member institutions, we conducted a case-control study to develop and evaluate prediction models for fall-related injuries in older adults. Questionnaire-derived prediction with three questions from a commonly used fall risk screening tool was evaluated. We then developed four temporal machine learning models using routinely available longitudinal EHR data to predict the future risk of fall injury. We also developed a fall injury-prevention clinical decision support (CDS) implementation prototype to link preventative interventions to patient-specific fall injury risk factors. Questionnaire-based risk screening achieved area under the receiver operating characteristic curve (AUC) up to 0.59 with 23% to 33% similarity for each pair of three fall injury screening questions. EHR-based machine learning risk screening showed significantly improved performance (best AUROC = 0.76), with similar prediction performance between 6-month and one-year prediction models. The current method of questionnaire-based fall risk screening of older adults is suboptimal with redundant items, inadequate precision, and no linkage to prevention. A machine learning fall injury prediction method can accurately predict risk with superior sensitivity while freeing up clinical time for initiating personalized fall prevention interventions. The developed algorithm and data science pipeline can impact routine primary care fall prevention practice.

Exploring optimization algorithms for establishing patient-based real-time quality control models

BackgroundPatient-based real-time quality control (PBRTQC) models must be optimized for use in different clinical laboratories, but the grid search (GS) algorithm explored in recent studies for this purpose is inefficient. Thus, finding an efficient optimization algorithm is critical for future research and implementation of the PBRTQC. MethodsWe compared the efficiency and performance of five commonly used optimization algorithms, including GS, simulated annealing (SA), genetic algorithms (GA), differential evolution (DE), and particle swarm optimization (PSO), to optimize conventional PBRTQC and regression-adjusted real-time quality control (RARTQC) models for serum alanine aminotransferase and sodium. ResultsThe GS, GA, DE, and PSO provided models with similar performances. However, GA and DE required significantly less computation time than GS. The results also demonstrate a general tradeoff between the optimization method's chance of discovering the optimum and the computation time required. ConclusionMore efficient optimization methods should be adopted when establishing PBRTQC or RARTQC models to save time and computing power that will enable the development of more complex models and increase the scalability of extensive PBRTQC applications.

Effects of applied current modes on the migration selectivity of peptides from a porcine cruor hydrolysate during electrodialysis with ultrafiltration membrane

Pepsic hydrolysis of porcine hemoglobin produces various antimicrobial peptides, however, within such complex hydrolysates, certain peptide bioactivities remain unexpressed. Electrodialysis with ultrafiltration membrane (EDUF) addresses this by separating peptides from hydrolysates based on peptide charge and molecular mass, resulting in fractions with increased bioactivity. This study aims to assess the impacts of different current modes (continuous current (CC), pulsed electric field (PEF) and polarity reversal (PR)) on peptide migration selectivity, which has never been reported. Different combinations of pulse and pause or reversed pulse were tested: 1 s/1 s and 10 s/1 s. This investigation focused on the 31 most migrating peptides and their 6 main physicochemical characteristics. Three groups of current conditions with similar performances were observed: (1) CC, PEF 10 s/1 s and PR 10 s/1 s (Averaged final migration rate of 5.0 g/m2h and energy consumption of 48.7 Wh/g of peptide); (2) PEF 1 s/1 s (9.0 g/m2h and 61.0 Wh/g of peptide), and (3) PR 1 s/1 s (3.3 g/m2h and 95.6 Wh/g of peptide). CC, PEF, and PR 10 s/1 s facilitated the migration of mainly cationic peptides with low molecular masses and were the most selective conditions. PR 1 s/1 s allowed the migration of some anionic peptides with lower molecular masses, due to the short polarity reversal. This study demonstrated for the first time the effects of the electrical current mode and condition associated on migration selectivity of peptides during EDUF and their performances. Modifying the electrical current mode offers a new way to optimize separation selectivity and generate specific peptide populations in the recovery compartment with potentially different bioactivities.

The Development of a Composite Thin Film Barrier of Tungsten Fe3O4-rGO (FerGO) for the Radiation Shielding of Medical Personnel.

Tungsten is the most effective eco-friendly material used for radiation shielding in hospitals. However, despite its commendable density and shielding performance, tungsten faces challenges in miscibility with other materials because of its elevated melting point and strength. In this study, to protect medical personnel against scattered rays, which are indirect X-rays, a lightweight material was prepared by mixing graphite oxide material, considering its thinness and flexibility. Tungsten particles were evenly dispersed in the polymer, and nanofibers were prepared using this blended polymer solution via electrospinning. Concurrently, the process technology was explored to craft a thin film sheet and obtain a lead-like shielding effect. A spinning solution was prepared by mixing Fe3O4-rGO (FerGO) and tungsten. At 60 kVp, 0.1 mm was measured as 0.097 mmPb, at 80 kVp, 0.2 mm was measured as 0.196 mmPb, and at 100 kVp, 0.3 mm was measured as 0.279 mmPb, showing similar shielding performance to lead. As density directly affects the shielding effect, graphene oxide played an important role in increasing the density of the material from 1.941 g/cm3 to 2.302 g/cm3. Thus, this study provides an effective process for producing thin film sheets equivalent to lead.

Accuracy of an AI-based automated plate reading mobile application for the identification of clinical mastitis-causing pathogens in chromogenic culture media

Using on-farm microbiological culture (OFC), based on chromogenic culture media, enables the identification of mastitis causing pathogens in about 24 h, allows rapid decision making on selective treatment and control management measures of cows with clinical mastitis (CM). However, accurate interpretation of OFC results requires trained and experienced operators, which could be a limitation for the use of OFC in dairy farms. Our hypothesis was that AI-based automated plate reading mobile application can analyze images of microorganisms’ colonies in chromogenic culture media with similar diagnostic performance as a trained specialist evaluator. Therefore, the aim of the present study was to evaluate the diagnostic accuracy of an AI-based application (Rumi; OnFarm, Piracicaba, São Paulo, Brazil) for interpreting images of mastitis causing microorganism colonies grown in chromogenic culture media. For this study two trials were organized to compare the results obtained using an AI-based application Rumi with the interpretation of: (1) a trained specialist, using MALDI-TOF MS as the gold standard; (2) farm personnel users (FPU). In trial 1, a total of 476 CM milk samples, from 11 farms located in São Paulo (n = 7) and Minas Gerais (n = 4), southeast Brazil, were inoculated in chromogenic culture media plates (Smartcolor 2, OnFarm, Piracicaba, São Paulo, Brazil) by specialists under lab conditions, and digital images were recorded 24 h after incubation at 37 °C. After that, all the 476 digital images were analyzed by the Rumi and by another specialist (who only had access to the digital images) and the diagnostic accuracy indicators sensitivity (Se) and specificity (Sp) were calculated using MALDI-TOF MS microbiological identification of the isolates as the reference. In Trial 2, a total of 208 CM milk samples, from 150 farms from Brazil, were inoculated in chromogenic culture media plates by FPU, and the results of microbiological growth were visually interpreted by FPU under on-farm conditions. After visual interpretation, results were recorded using an OnFarmApp application (herd manage application for mastitis by OnFarm, Piracicaba, São Paulo, Brazil), and the images of the chromogenic culture plates were captured by the OnFarmApp to be evaluated by Rumi and Bayesian Latent Class Models were performed to compare Rumi and the FPU. In Trial 1, Rumi presented high and intermediate accuracy results, with the only exception of the low Enterococcus spp.’s Se. In comparison with the specialist, Rumi performed similarly in Se and Sp for most groups of pathogens, with the only exception of non-aureus staphylococci where Se results were lower. Both Rumi and the specialist achieved Sp results > 0.96. In Trial 2, Rumi had similar results as the FPU in the Bayesian Latent Class Model analysis. In conclusion, the use of the AI-based automated plate reading mobile application can be an alternative for visual interpretation of OFC results, simplifying the procedures for selective treatment decisions for CM based on OFC.

AI/ML-as-a-Service for optical network automation: use cases and challenges [Invited

In recent years, artificial intelligence/machine learning (AI/ML) has played a significant role in automating optical networks. Despite this, the methods for creating, deploying, and monitoring AI/ML models still rely heavily on human intervention and trial-and-error. AI/ML-as-a-Service aims at automating the processes associated with AI/ML models, reducing the need for human intervention and thus facilitating the widespread adoption of AI/ML models. In this paper, we introduce the concept of AI/ML-as-a-Service in the context of optical network automation and propose an architecture for realizing this concept. We provide details of a reference implementation that focuses on the model creation stage. The reference implementation is tested using two use cases related to the quality-of-transmission (QoT) estimation of optical channels. We demonstrate that models created through AI/ML-as-a-Service are able to achieve similar performance as manually tuned models while drastically reducing the need for human involvement. Finally, we discuss future challenges and opportunities for applying AI/ML-as-a-Service in optical network automation.

Unified two-way shear model for steel and FRP-RC slabs: Evaluation and reliability calibration

This study introduces a unified two-way (punching) shear design model for concrete reinforced with slabs steel- and FRP bars. The unified model modifies the ACI 318-19 punching shear model to account for the axial stiffness of the tension reinforcement by providing a new modification term (ρ Ereft./Ec)0.26, representing the reinforcing ratio multiplied by the elastic modular ratio of longitudinal bars to concrete. The unified punching shear model has been evaluated over two experimental datasets: FRP-RC slabs (145 specimens) and steel-RC slabs (390 specimens). A comparative assessment between the ACI 440.11-22, CSA S806, JSCE, and the proposed models to test their generalization abilities for slabs reinforced with both types of reinforcement. The assessment indicated a similar performance of the proposed model to the CSA S806 and the JSCE models in terms of statistical measures. The ACI 440.11–22 showed a high average strength ratio (Vexp./Vpred.) of 1.83. However, the ACI and the CSA models showed reduction in the conservatism at high reinforcement axial stiffness. In addition, reliability analysis using Monte Carlo simulation indicated that the unified punching shear model provides a satisfactory safety level with a reliability index between 3.5 and 4.0 for both steel and FRP-RC slabs.

Operator learning for urban water clarification hydrodynamics and particulate matter transport with physics-informed neural networks

Computational fluid dynamics (CFD) can be a powerful tool for higher-fidelity water infrastructure planning and design. Despite decades of development and demonstration over a wide range of water systems such as clarification basins, activated sludge processes, ozone contactors, etc., CFD remains primarily used in academic research, with limited application in civil and environmental engineering practice. This limitation is contributed by its higher computational cost and demand for specialized user skills. This, however, need not be the case, if a robust and efficient surrogate model can be developed from CFD simulations and independently deployed for engineering purposes. Leveraging the emerging scientific machine learning (ML) techniques of physics-informed ML and operator learning, this study develops a composite neural network (CPNN) for learning the flow hydrodynamics and particulate matter (PM) transport and fate in clarification systems. The CPNN consists of a deep operator network (DeepONet) as an encoder and a physics-informed neural network (PINN) as a decoder. In contrast to common “black box” and lumped ML approaches, the developed CPNN directly incorporates physics principles into its architecture. Furthermore, the CPNN is designed for process-resolved and operator learning, enabling it to predict spatial hydrodynamics and PM concentration distribution (i.e., contours) across different basin geometrics and loading conditions. Compared to CFD simulation, the developed CPNN model has significantly higher computational efficiency (∼ milliseconds) while demonstrating robust predictive capability. For predicting basin hydrodynamics across 10,000 test cases, the trained CPNN model achieves an R2 above 0.8 for 66.4% of cases and an R2 above 0.4 for 89.2% of cases. A similar performance is also demonstrated by the CPNN in predicting basin PM concentration. Further investigation reveals that basin geometrics that trigger bi-modal flow solutions can be particularly challenging for ML. Additionally, this study visualizes the dependency of basin hydrodynamics and PM concentration on basin geometrics and loading conditions, providing valuable insights for optimizing basin configuration. Lastly, the potentials and benefits of web-based applications, e.g., DeepXtorm, as a user-friendly interface for the developed CPNN model is discussed. This study represents the initial step toward achieving real-time higher-fidelity water infrastructure planning, design, optimization, and regulation.

Efficient Video Transformers via Spatial-temporal Token Merging for Action Recognition

Transformer has exhibited promising performance in various video recognition tasks but brings a huge computational cost in modeling spatial-temporal cues. This work aims to boost the efficiency of existing video transformers for action recognition through eliminating redundancies in their tokens and efficiently learning motion cues of moving objects. We propose a lightweight and plug-and-play module, namely Spatial-temporal Token Merger (STTM), to merge the tokens belonging to the same object into a more compact representation. STTM first adaptively identifies crucial object clues underlying the video as meta tokens. Similarity scores between input tokens and meta tokens are hence computed and used to guide the fusion of similar tokens in both spatial and temporal domains, respectively. To compensate for motion cues lost in the merging procedure, we compute the linear aggregation of spatial-temporal positions of tokens as motion features. STTM hence outputs a compact set of tokens fusing both appearance and motion features of moving objects. This procedure substantially decreases the number of tokens that need to be processed by each Transformer block and boosts the efficiency. As a general module, STTM can be applied to different layers of various video Transformers. Extensive experiments on the action recognition datasets Kinectics-400 and SSv2 demonstrate its promising performance. For example, it reduces the computation complexity of ViT by 38% while maintaining a similar performance on Kinectics-400. It also brings 1.7% gains of top-1 accuracy on SSv2 under the same computational cost.

The performance of a deep learning system in assisting junior ophthalmologists in diagnosing 13 major fundus diseases: a prospective multi-center clinical trial

Artificial intelligence (AI)-based diagnostic systems have been reported to improve fundus disease screening in previous studies. This multicenter prospective self-controlled clinical trial aims to evaluate the diagnostic performance of a deep learning system (DLS) in assisting junior ophthalmologists in detecting 13 major fundus diseases. A total of 1493 fundus images from 748 patients were prospectively collected from five tertiary hospitals in China. Nine junior ophthalmologists were trained and annotated the images with or without the suggestions proposed by the DLS. The diagnostic performance was evaluated among three groups: DLS-assisted junior ophthalmologist group (test group), junior ophthalmologist group (control group) and DLS group. The diagnostic consistency was 84.9% (95%CI, 83.0% ~ 86.9%), 72.9% (95%CI, 70.3% ~ 75.6%) and 85.5% (95%CI, 83.5% ~ 87.4%) in the test group, control group and DLS group, respectively. With the help of the proposed DLS, the diagnostic consistency of junior ophthalmologists improved by approximately 12% (95% CI, 9.1% ~ 14.9%) with statistical significance (P < 0.001). For the detection of 13 diseases, the test group achieved significant higher sensitivities (72.2% ~ 100.0%) and comparable specificities (90.8% ~ 98.7%) comparing with the control group (sensitivities, 50% ~ 100%; specificities 96.7 ~ 99.8%). The DLS group presented similar performance to the test group in the detection of any fundus abnormality (sensitivity, 95.7%; specificity, 87.2%) and each of the 13 diseases (sensitivity, 83.3% ~ 100.0%; specificity, 89.0 ~ 98.0%). The proposed DLS provided a novel approach for the automatic detection of 13 major fundus diseases with high diagnostic consistency and assisted to improve the performance of junior ophthalmologists, resulting especially in reducing the risk of missed diagnoses. ClinicalTrials.gov NCT04723160