Search

Early detection of rapidly progressive kidney disease is key to improving the renal outcome and reducing complications in adult patients with type 2 diabetes mellitus (T2DM). We aimed to construct a 6-month machine learning (ML) predictive model for the risk of rapidly progressive kidney disease and the need for nephrology referral in adult patients with T2DM and an initial estimated glomerular filtration rate (eGFR) ≥ 60 mL/min/1.73 m2. We extracted patients and medical features from the electronic medical records (EMR), and the cohort was divided into a training/validation and testing data set to develop and validate the models on the basis of three algorithms: logistic regression (LR), random forest (RF), and extreme gradient boosting (XGBoost). We also applied an ensemble approach using soft voting classifier to classify the referral group. We used the area under the receiver operating characteristic curve (AUROC), precision, recall, and accuracy as the metrics to evaluate the performance. Shapley additive explanations (SHAP) values were used to evaluate the feature importance. The XGB model had higher accuracy and relatively higher precision in the referral group as compared with the LR and RF models, but LR and RF models had higher recall in the referral group. In general, the ensemble voting classifier had relatively higher accuracy, higher AUROC, and higher recall in the referral group as compared with the other three models. In addition, we found a more specific definition of the target improved the model performance in our study. In conclusion, we built a 6-month ML predictive model for the risk of rapidly progressive kidney disease. Early detection and then nephrology referral may facilitate appropriate management.

The wood-based furniture manufacturing industries prioritize quality of production to meet higher market demands. Identifying various types of edge-glued wooden panel defects are a challenge for a human worker or a camera. Several studies have shown that the detection of edge-glued defects with low, high, normal, overlong, short is identified but detection of residue and bluntness is highly challenging. Thus, the present model identifies defects of low, high, normal, overlong, short by computer vision and/or deep learning, whereas defects of residue and bluntness by deep learning based decide by pass for having better performance. The goal of this paper is to provide an improved defect detection solution for wood-based furniture manufacturing industries by process automation. Therefore, a system was designed that takes defect input images from a camera as raw image and laser-aligned image for defect detection of the edge-glued wooden panel. The process automation then performs computer vision-based image features extraction with deep learning for defect detection. The aim of this paper is to solve edge-glued defect detection problems by using design and implementation of edge-glued wooden defect detection, that can be stated as edge-glued wooden panel defect detection using deep learning (WDD-DL) for process automation by artificial intelligence and Automated Optical Inspection (AOI) consolidation. Possibly there exist several types of defects on the edges while edge-banding on the wooden panel in furniture manufacturing. Therefore, the scope is to achieve higher accuracy by raw image and laser-aligned image feature extraction using deep learning algorithms for final result defect classification in WDD-DL by AOI. The WDD-DL system uses Gabor, Harris corner, morphology, structured light detection and curvature calculation for pre-processing and InceptionResnetV2 Convolutional Neural Network algorithm to attain the best results. The applications of this work can be found in quality control of the furniture manufacturing industry for an edge, corner, joint defect detection of the wooden panels. The WDD-DL achieves best results as the precision, recall and F1 score are 0.97, 0.90 and 0.92, respectively. The experiments demonstrate higher accuracy achievement as compared to other methods with overkill and escape rate analysis. Ultimately, the discussion section provides an interesting experience sharing about the necessary factors for implementing the WDD-DL in real-time industrial operations.

A distinct security protocol is necessary for the exponential growth in intelligent edge devices. In particular, the autonomous devices need to address significant security concern to function smoothly in the high market demand. Nevertheless, exponential increase in the connected devices has made cloud networks more complex and suffer from information processing delay. Therefore, the goal of this work is to design a novel server-less mutual authentication protocol for the edge networks. The aim is to demonstrate an autonomous mutual authentication amongst the connected smart devices within the edge networks. The solution addresses applications of autonomous cars, smart things, and Internet of Things (IoT) devices in the edge or wireless sensor networks (WSN), etc. In this paper, the design proposes use of a public-key system, octet-based balanced-tree transitions, challenge–response mechanism, device unique ID (UID), pseudo-random number generator (PRNG), time-stamps, and event specific session keys. Ultimately, server-less design requires less infrastructure and avoids several types of network-based communication attacks, e.g., impersonating, Man in the middle (MITM), IoT-DDOS, etc. Additionally, the system overhead is eliminated by no secret key requirements. The results provide sufficient evidence about the protocol market competitiveness and demonstrate better benchmark comparison results.

In recent years, word embedding models receive tremendous research attentions due to their capability of capturing textual semantics. This study investigates the issue of employing word embedding models into storage-constrained edge devices for personalized item-of-interest recommendation in a smart store. The challenge lies in that the existing embedding models are often too large to fit into a storage-constrained edge device. One naive idea is to reside the word embedding model in a secondary storage and process recommendation with that storage. However, this idea suffers from the burden of additional traffics. To this end, we propose a framework called Word Embedding Quantization (WEQ) which constructs an index upon a given word embedding model and stores the index in the primary storage to enable the use of the word embedding model by edge devices. One challenge for using the index is that the exact user profile is no longer ensured. However, we find that there are opportunities for computing the correct recommendation results by knowing only an inexact user profile. In this paper, we propose a series of techniques that leverage the opportunities for computing candidates with the goal of minimizing the accessing cost to a secondary storage in edge devices. Experiments are performed to verify the efficiency of the proposed techniques, demonstrating the feasibility of the proposed framework.

Early prediction of clinical deterioration such as adverse events (AEs), improves patient safety. National Early Warning Score (NEWS) is widely used to predict AEs based on the aggregation of 6 physiological parameters. We took the same parameters as the features for AE prediction using deep learning algorithms (AEP-DLA) among hospitalized adult patients. The aim of this study is to get better performance than traditional naïve mathematical calculations by introducing novel vital sign data preprocessing schemes. We retrospectively collected the data from our electronic medical record data warehouse (2007 ~ 2017). AE rate of all 99,861 admissions was 6.2%. The dataset was divided into training and testing datasets from 2007–2015 and 2016–2017 respectively. In real-life clinical care, physiological parameters were not recorded every hour and missed frequently, for example, Glasgow Coma Scale (GCS). The expert domain suggested that missed GCS was rated as 15. We took two strategies (stack series records and align by hour) in the data preprocessing and tripling the values of negative samples for class balancing (CB). We used the last 28 hours’ serial data to predict AEs 3 hours later with Random Forest, XGBoost, Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN). It is shown that CNN with CB and align by hour got the best results comparing to the other methods. The precision, recall and area under curve were 0.841, 0.928 and 0.995 respectively. The performance of the model is also better than those proposed in the published literatures.

Sheet metal-based manufacturing industries operate on several varieties of sheet metal parts. Previously QR codes stickers were put on sheet metal parts for identification by manual workers as per their respective shape and size features, thus ensuring synchronized raw material flow for the manufacturing process. However, identifying a particular type of sheet metal part based on its different features is still a challenge for a trained manual operator. Currently, in the market, there exist some automation solutions for solving such kind of problem but are incapable of achieving better performance and possess accuracy issues. So our goal is to provide process automation by overcoming manual work-based dependency and limitations. Therefore, a system is required that can take input a high definition camera captured sheet metal part image and provide an accurately identified type as output by utilizing a deep learning classification model in computer vision. The automation of sheet metal part identification by using ERP, CAD files and scheduling among them would make a smooth workflow by IDS-DLA. This paper aims to solve the identification problem by using the design and implementation of a sheet metal part identification, given as sheet metal part IDentification System for process automation using Deep Learning Algorithms (IDS-DLA). Considering the sheet metal parts there exists a large volume of types but fewer quantities. IDS-DLA performs high accuracy sheet metal part identification from the CAD model database by using the Geometric and CNN triplet filter. The IDS-DLA also evaluates the Hu moment ranking to choose the top 5 rank predictions as final ranking results. The applications can be given as manufacturing process automation for industry, 2D CAD search, 2D measuring solution, closet formation, etc. Ultimately, from the experiments, it can be observed that better accuracy is obtained as compared with the previous benchmarks. This multi-filtering approach using a deep feature extraction algorithm concludes to be the better approach and achieves higher performance.

In this paper, we proposed a method for accelerating the transmission of a series of streaming files for multi-session file delivery scenarios. When streaming files are transmitted, the throughput of data transmission can be increased and the out of order problem of file reception (i.e., file reception delay) can also be reduced at the same time. This is achieved by using a repair file generation mechanism based on AL-FEC, which is triggered while an empty file transmission session appears but there are still uncompleted files waiting to be delivered. Performance evaluation results shows that the file transmission throughput of the proposed method are 28% to 44% higher than the throughput of a traditional file transmission method that is designed for reducing file reception delay. However, the file reception delay is the same as the above-mentioned traditional method.

This study aimed to develop an early prediction model for identifying patients with bloodstream infections. The data resource was taken from 2015 to 2019 at Taichung Veterans General Hospital, and a total of 1647 bloodstream infection episodes and 3552 non-bloodstream infection episodes in the intensive care unit (ICU) were included in the model development and evaluation. During the data analysis, 30 clinical variables were selected, including patients’ basic characteristics, vital signs, laboratory data, and clinical information. Five machine learning algorithms were applied to examine the prediction model performance. The findings indicated that the area under the receiver operating characteristic curve (AUROC) of the prediction performance of the XGBoost model was 0.825 for the validation dataset and 0.821 for the testing dataset. The random forest model also presented higher values for the AUROC on the validation dataset and testing dataset, which were 0.855 and 0.851, respectively. The tree-based ensemble learning model enabled high detection ability for patients with bloodstream infections in the ICU. Additionally, the analysis of importance of features revealed that alkaline phosphatase (ALKP) and the period of the central venous catheter are the most important predictors for bloodstream infections. We further explored the relationship between features and the risk of bloodstream infection by using the Shapley Additive exPlanations (SHAP) visualized method. The results showed that a higher prothrombin time is more prominent in a bloodstream infection. Additionally, the impact of a lower platelet count and albumin was more prominent in a bloodstream infection. Our results provide additional clinical information for cut-off laboratory values to assist clinical decision-making in bloodstream infection diagnostics.

Sheet metal-based products serve as a major portion of the furniture market and maintain higher quality standards by being competitive. During industrial processes, while converting a sheet metal to an end product, new defects are observed and thus need to be identified carefully. Recent studies have shown scratches, bumps, and pollution/dust are identified, but orange peel defects present overall a new challenge. So our model identifies scratches, bumps, and dust by using computer vision algorithms, whereas orange peel defect detection with deep learning have a better performance. The goal of this paper was to resolve artificial intelligence (AI) as an AI landing challenge faced in identifying various kinds of sheet metal-based product defects by ALDB-DL process automation. Therefore, our system model consists of multiple cameras from two different angles to capture the defects of the sheet metal-based drawer box. The aim of this paper was to solve multiple defects detection as design and implementation of Industrial process integration with AI by Automated Optical Inspection (AOI) for sheet metal-based drawer box defect detection, stated as AI Landing for sheet metal-based Drawer Box defect detection using Deep Learning (ALDB-DL). Therefore, the scope was given as achieving higher accuracy using multi-camera-based image feature extraction using computer vision and deep learning algorithm for defect classification in AOI. We used SHapley Additive exPlanations (SHAP) values for pre-processing, LeNet with a (1 × 1) convolution filter, and a Global Average Pooling (GAP) Convolutional Neural Network (CNN) algorithm to achieve the best results. It has applications for sheet metal-based product industries with improvised quality control for edge and surface detection. The results were competitive as the precision, recall, and area under the curve were 1.00, 0.99, and 0.98, respectively. Successively, the discussion section presents a detailed insight view about the industrial functioning with ALDB-DL experience sharing.

Background: Antinuclear antibody pattern recognition is vital for autoimmune disease diagnosis but labor-intensive for manual interpretation. To develop an automated pattern recognition system, we established machine learning models based on the International Consensus on Antinuclear Antibody Patterns (ICAP) at a competent level, mixed patterns recognition, and evaluated their consistency with human reading. Methods: 51,694 human epithelial cells (HEp-2) cell images with patterns assigned by experienced medical technologists collected in a medical center were used to train six machine learning algorithms and were compared by their performance. Next, we choose the best performing model to test the consistency with five experienced readers and two beginners. Results: The mean F1 score in each classification of the best performing model was 0.86 evaluated by Testing Data 1. For the inter-observer agreement test on Testing Data 2, the average agreement was 0.849 (κ) among five experienced readers, 0.844 between the best performing model and experienced readers, 0.528 between experienced readers and beginners. The results indicate that the proposed model outperformed beginners and achieved an excellent agreement with experienced readers. Conclusions: This study demonstrated that the developed model could reach an excellent agreement with experienced human readers using machine learning methods.

Video surveillance systems are deployed at many places such as airports, train stations, and malls for security and monitoring purposes. However, it is laborious to search for and retrieve persons in multicamera surveillance systems, especially with cluttered backgrounds and appearance variations among multiple cameras. To solve these problems, this paper proposes a person retrieval method that extracts the attributes of a masked image using an instance segmentation module for each object of interest. It uses attributes such as color and type of clothes to describe a person. The proposed person retrieval system involves four steps: (1) using the YOLACT++ model to perform pixelwise person segmentation, (2) conducting appearance-based attribute feature extraction using a multiple convolutional neural network classifier, (3) employing a search engine with a fundamental attribute matching approach, and (4) implementing a video summarization technique to produce a temporal abstraction of retrieved objects. Experimental results show that the proposed retrieval system can achieve effective retrieval performance and provide a quick overview of retrieved content for multicamera surveillance systems.

Surveillance systems are everywhere deployed for criminal prevention and suspect tracking for specific security events. The recognition of the target person cross multiple cameras remains a challenge because of the large spatio-temporal uncertainty, which means the entry and the exit of persons in cameras are unpredictable. Therefore, this study design and implement a software system that help users to continuous and stable track of the same target person across multiple cameras. Firstly, the system leverages YOLO and Correlation filter to track the suspicious in single camera. The features of the target suspect are also collected in this stage. Once the target person leaves the scope of a camera, the camera network topology is used to predict the candidate entrance cameras based on the geographical information and tracking path in the first stage. In the second stage, the person re-identification algorithm is used to calculate the similarity possibility and help to identify the person automatically between cameras. Experimental results on public datasets show our system can accurately recognize the target suspect across cameras with acceptable system performance.

Acute kidney injury (AKI) refers to rapid decline of kidney function and is manifested by decreasing urine output or abnormal blood test (elevated serum creatinine). Electronic health records (EHRs) is fundamental for clinicians and machine learning algorithms to predict the clinical outcome of patients in the Intensive Care Unit (ICU). Early prediction of AKI could automatically warn the clinicians to review the possible risk factors and act in advance to prevent it. However, the enormous amount of patient data usually consists of a relatively incomplete data set and is very challenging for supervised machine learning process. In this paper, we propose an entropy-based feature engineering framework for vital signs based on their frequency of records. In particular, we address the missing at random (MAR) and missing not at random (MNAR) types of missing data according to different clinical scenarios. Regarding its applicability, we applied it to establish a prediction model for future AKI in ICU patients using 4278 ICU admissions from a tertiary hospital. Our result shows that the proposed entropy-based features are feasible to be used in the AKI prediction model and its performance improves as the data availability increases. In addition, we study the performance of AKI prediction model by comparing different time gaps and feature windows with the proposed vital sign entropy features. This work could be used as a guidance for feature windows selection and missing data processing during the development of a prediction model in ICU.

Sheet metal part numbers are used to retrieve recipes of a working stage, and greatly help the scheduling performance between working stages. This study proposed to integrate Hu moment and CNN algorithm to match CAD models and real sheet metal images. The experiment shows that the top-5 accuracy of predictions reaches 90% for 28K CAD models and 98% for 4K CAD model database for each.

There is strong demand for real-time suspicious tracking across multiple cameras in intelligent video surveillance for public areas, such as universities, airports and factories. Most criminal events show that the nature of suspicious behavior are carried out by un-known people who try to hide themselves as much as possible. Previous learning-based studies collected a large volume data set to train a learning model to detect humans across multiple cameras but failed to recognize newcomers. There are also several feature-based studies aimed to identify humans within-camera tracking. It would be very difficult for those methods to get necessary feature information in multi-camera scenarios and scenes. It is the purpose of this study to design and implement a suspicious tracking mechanism across multiple cameras based on correlation filters, called suspicious tracking across multiple cameras based on correlation filters (STAM-CCF). By leveraging the geographical information of cameras and YOLO object detection framework, STAM-CCF adjusts human identification and prevents errors caused by information loss in case of object occlusion and overlapping for within-camera tracking cases. STAM-CCF also introduces a camera correlation model and a two-stage gait recognition strategy to deal with problems of re-identification across multiple cameras. Experimental results show that the proposed method performs well with highly acceptable accuracy. The evidences also show that the proposed STAM-CCF method can continuously recognize suspicious behavior within-camera tracking and re-identify it successfully across multiple cameras.

A defect inspection of resin films involves processes of detecting defects, size measuring, type classification and reflective action planning. It is not only a process requiring heavy investment in workforce, but also a tension between quality assurance with a 50-micrometer tolerance and visibility of the naked eye. To solve the difficulties of the workforce and time consumption processes of defect inspection, an apparatus is designed to collect high-quality images in one shot by leveraging a large field-of-view microscope at 2K resolution. Based on the image dataset, a two-step method is used to first locate possible defects and predict their types by a defect-shape-based deep learning model using the LeNet-5-adjusted network. The experimental results show that the proposed method can precisely locate the position and accurately inspect the fine-grained defects of resin films.

Street lighting is a fundamental aspect of security systems in homes, industrial facilities, and public places. To detect parking lot occupancy in outdoor environments, street light control plays a crucial role in smart surveillance applications that can perform robustly in extreme surveillance environments. However, traditional parking occupancy systems are mostly implemented for outdoor environments using costly sensor-based techniques. This study uses the Jetson TX2 to develop a method that can accurately identify street parking occupancy and control streetlights to assist occupancy detection, thereby reducing costs, and can adapt to various weather conditions. The proposed method adopts You Only Look Once version 3 (YOLO v3, Seattle, WA, USA) based on MobileNet version 2 (MobileNet v2, Salt Lake City, UT, USA), which is area-based and uses voting to stably recognize occupancy status. This solution was verified using the CNRPark + EXT dataset, a simulated model, and real scenes photographed with a camera. Our experiments revealed that the proposed framework can achieve stable parking occupancy detection in streets.

In recent years, word embedding models receive tremendous research attentions due to their capability of capturing textual semantics. This study investigates the issue of employing word embedding models into resource-limited smartphones for personalized item recommendation. The challenge lies in that the existing embedding models are often too large to fit into a resource-limited smartphones. One naive idea is to incorporate a secondary storage by residing the model in the secondary storage and processing recommendation with the secondary storage. However, this idea suffers from the burden of additional traffics. To this end, we propose a framework called Word Embedding Quantization (WEQ) that constructs an index upon a given word embedding model and stores the index on the primary storage to enable the use of the word embedding model on smartphones. One challenge for using the index is that the exact user profile is no longer ensured. However, we find that there are opportunities for computing the correct recommendation results by knowing only inexact user profile. In this paper, we propose a series of techniques that leverage the opportunities for computing candidates with the goal of minimizing the accessing cost to a secondary storage. Experiments are made to verify the efficiency of the proposed techniques, which demonstrates the feasibility of the proposed framework.

Recommender systems have been proven useful in numerous contemporary applications and helping users effectively identify items of interest within massive and potentially overwhelming collections. Among the recommender system techniques, the collaborative filtering mechanism is the most successful; it leverages the similar tastes of similar users, which can serve as references for recommendation. However, a major weakness for the collaborative filtering mechanism is its performance in computing the pairwise similarity of users. Thus, the MapReduce framework was examined as a potential means to address this performance problem. This paper details the development and employment of the MapReduce framework, examining whether it improves the performance of a personal ontology based recommender system in a digital library. The results of this extensive performance study show that the proposed algorithm can scale recommender systems for all-pairs similarity searching.

This paper focuses on flood-region detection using monitoring images. However, adverse weather affects the outcome of image segmentation methods. In this paper, we present an experimental comparison of an outdoor visual sensing system using region-growing methods with two different growing rules—namely, GrowCut and RegGro. For each growing rule, several tests on adverse weather and lens-stained scenes were performed, taking into account and analyzing different weather conditions with the outdoor visual sensing system. The influence of several weather conditions was analyzed, highlighting their effect on the outdoor visual sensing system with different growing rules. Furthermore, experimental errors and uncertainties obtained with the growing rules were compared. The segmentation accuracy of flood regions yielded by the GrowCut, RegGro, and hybrid methods was 75%, 85%, and 87.7%, respectively.

Flood hazard prevention and mitigation is an emergent environmental problem. Traditional flood monitoring devices operated under adverse environmental conditions are typically influenced by changes in weather conditions such as haze, fog and rain. Consequently, the video images obtained from such devices are often blurred or damaged, increasing the possibility of erroneous assessments in hazard mitigation processes. To ensure the efficient use of image analysis technology to improve degraded images captured under hazy weather conditions, this study proposes an automated single image dehazing method for flood monitoring. This method is based on the dark channel prior for the removal of haze from a single image. The concept of the dark channel prior is that most local patches in haze-free outdoor images contain some pixels having extremely low intensities in at least one colour channel. When this dark channel prior is used, the thickness of haze in the image can be directly estimated and a high-quality haze-free image can be obtained. The proposed method can be used to accurately improve flood detection and monitoring results. The ability to detect and remove haze from a single image is a crucial function when applying automated computer vision to disaster-monitoring applications. The experimental results show that the proposed method can efficiently alleviate the degradation of surveillance images and effectively identify flooded regions in particular areas.

Severe weather conditions greatly impair the performance of outdoor imaging. In this study, two region-based image segmentation methods, Grow Cut and Region Growing (RegGro), were applied to rain scenes. This study demonstrates that segmentation accuracy depends on fog and rain stains. In severe rainfall periods, heavy rain and fog reduced the overall image quality, and both methods yielded segmentation failure. The results show that both region-based methods are effective for segmenting objects in images captured under poor weather conditions. Both methods have unique advantages and disadvantages for fog and stain conditions. The segmentation accuracy yielded by the Grow Cut and RegGrow methods was 75% and 85%, respectively.

Recommender system can help users to effectively identify interested items from a potentially overwhelming huge collection of items, and it has been shown to be very useful in many e-commerce applications. Collaborative filtering (CF), which assumes that similar users may have similar tastes, is one of the most widely used Recommender system techniques. However, one of the major weaknesses for the CF mechanism is the computational cost in computing pairwise similarity of users. This paper attempts to tackle the computational problem of all pairs similarity using the MapReduce technique in the Hadoop framework. We give an overview of our development on using a parallel filtering algorithm to improve the performance of a personal ontology based recommender system for digital library. The experimental results show that the proposed algorithm can indeed scale out the recommender systems for all pairs search.