Model Training Phase Research Articles

Behavioral Account of Attended Stream Enhances Neural Tracking.

During the past decade, several studies have identified electroencephalographic (EEG) correlates of selective auditory attention to speech. In these studies, typically, listeners are instructed to focus on one of two concurrent speech streams (the “target”), while ignoring the other (the “masker”). EEG signals are recorded while participants are performing this task, and subsequently analyzed to recover the attended stream. An assumption often made in these studies is that the participant’s attention can remain focused on the target throughout the test. To check this assumption, and assess when a participant’s attention in a concurrent speech listening task was directed toward the target, the masker, or neither, we designed a behavioral listen-then-recall task (the Long-SWoRD test). After listening to two simultaneous short stories, participants had to identify keywords from the target story, randomly interspersed among words from the masker story and words from neither story, on a computer screen. To modulate task difficulty, and hence, the likelihood of attentional switches, masker stories were originally uttered by the same talker as the target stories. The masker voice parameters were then manipulated to parametrically control the similarity of the two streams, from clearly dissimilar to almost identical. While participants listened to the stories, EEG signals were measured and subsequently, analyzed using a temporal response function (TRF) model to reconstruct the speech stimuli. Responses in the behavioral recall task were used to infer, retrospectively, when attention was directed toward the target, the masker, or neither. During the model-training phase, the results of these behavioral-data-driven inferences were used as inputs to the model in addition to the EEG signals, to determine if this additional information would improve stimulus reconstruction accuracy, relative to performance of models trained under the assumption that the listener’s attention was unwaveringly focused on the target. Results from 21 participants show that information regarding the actual – as opposed to, assumed – attentional focus can be used advantageously during model training, to enhance subsequent (test phase) accuracy of auditory stimulus-reconstruction based on EEG signals. This is the case, especially, in challenging listening situations, where the participants’ attention is less likely to remain focused entirely on the target talker. In situations where the two competing voices are clearly distinct and easily separated perceptually, the assumption that listeners are able to stay focused on the target is reasonable. The behavioral recall protocol introduced here provides experimenters with a means to behaviorally track fluctuations in auditory selective attention, including, in combined behavioral/neurophysiological studies.

FMD-Yolo: An efficient face mask detection method for COVID-19 prevention and control in public

Coronavirus disease 2019 (COVID-19) is a world-wide epidemic and efficient prevention and control of this disease has become the focus of global scientific communities. In this paper, a novel face mask detection framework FMD-Yolo is proposed to monitor whether people wear masks in a right way in public, which is an effective way to block the virus transmission. In particular, the feature extractor employs Im-Res2Net-101 which combines Res2Net module and deep residual network, where utilization of hierarchical convolutional structure, deformable convolution and non-local mechanisms enables thorough information extraction from the input. Afterwards, an enhanced path aggregation network En-PAN is applied for feature fusion, where high-level semantic information and low-level details are sufficiently merged so that the model robustness and generalization ability can be enhanced. Moreover, localization loss is designed and adopted in model training phase, and Matrix NMS method is used in the inference stage to improve the detection efficiency and accuracy. Benchmark evaluation is performed on two public databases with the results compared with other eight state-of-the-art detection algorithms. At IoU = 0.5 level, proposed FMD-Yolo has achieved the best precision AP50 of 92.0% and 88.4% on the two datasets, and AP75 at IoU = 0.75 has improved 5.5% and 3.9% respectively compared with the second one, which demonstrates the superiority of FMD-Yolo in face mask detection with both theoretical values and practical significance.

An ultra-fast bi-phase advanced network for segmenting crop plants from dense weeds

Fast and accurate perception of crop plants is critical for robotic weeding. However, due to the similar appearance, it remains challenging to precisely and efficiently segment crop plants from weeds, especially when the weed density is high. To tackle this problem, we devise an ultra-fast bi-phase advanced network (BA-Net) architecture that advances both the feature extraction phase and the model training phase. A bi-path fusion tree structure and a bi-polar distributed loss are introduced for the two phases, respectively. The bi-path fusion tree merges features from two adjacent convolutional stages at each step and efficiently models the scene through a tree-like structure. The bi-polar distributed loss strengthens the training at ambiguous areas, which can greatly enhance the distinctness at detailed structures like the crop plant boundaries. Experiments demonstrate the superiority of the proposed BA-Net to other state-of-the-art methods in terms of both accuracy and efficiency. BA-Net can accurately segment the crop plants from the weedy background, achieving an intersection over union score of 0.963, a F-measure score of 0.981, and a mean absolute error of 0.00376. Also, BA-Net is lightweight with only 0.55 M trainable parameters and achieves ultra-fast speed, i.e. , over 340 FPS on a desktop and over 26 FPS on a Jetson TX2 embedded computer. The code of BA-Net is available at https://github.com/ZhangXG001/BA-Net . • An ultra-fast deep network is proposed for segmenting crop plants from dense weeds. • The bi-path fusion tree structure can perform efficient multi-scale feature fusion. • The bi-polar distributed loss can improve the segmentation at the obscure areas. • The network achieves superior accuracy and efficiency over state-of-the-arts. • The network runs over 26 FPS on a Jetson TX2 embedded computer.

Large-scale 3D point-cloud semantic segmentation of urban and rural scenes using data volume decomposition coupled with pipeline parallelism

This study proposes a generic approach which performs a series of systematic analyses by first introducing a data volume decomposition method to generate useful data features for performing semantic segmentation analysis involving 3D point-cloud data. Pipeline parallelism protocol is then implemented to accelerate the deep learning model's training phase. Our proposed approach is verified by decomposing around 2.0 billion point-cloud data points, as extracted from an open-source Semantic3D dataset, into many 3D regular structures with defined numbers of voxels. Each derived 3D structure has imposed normality in their data distribution of the respective label classes. Using the optimal hyperparameters for model training, the resulting trained model achieves average overall accuracy (mOA) and average intersection over union (mIOU) values of 0.984 and 0.752, respectively, on a testing dataset having close to 800 million point-cloud data points. The results are comparable with that of other state-of-the-art models in the literature.

Prognostic Factors of Survival of Advanced Liver Cancer Patients Treated With Palliative Radiotherapy: A Retrospective Study.

AimsSurvival benefit of liver cancer patients who undergo palliative radiotherapy varies from person to person. The present study aims to identify indicators of survival of advanced liver cancer patients receiving palliative radiotherapy.Patients and MethodsOne hundred and fifty-nine patients treated with palliative radiotherapy for advanced liver cancer were retrospectively assessed. Of the 159 patients, 103 patients were included for prediction model construction in training phase, while other 56 patients were analyzed for external validation in validation phase. In model training phase, clinical characteristics of included patients were evaluated by Kaplan-Meier curves and log-rank test. Thereafter, multivariable Cox analysis was taken to further identify characteristics with potential for prediction. In validation phase, a separate dataset including 56 patients was used for external validation. Harrell’s C-index and calibration curve were used for model evaluation. Nomograms were plotted based on the model of multivariable Cox analysis.ResultsThirty-one characteristics of patients were investigated in model training phase. Based on the results of Kaplan-Meier plots and log-rank tests, 6 factors were considered statistically significant. On multivariable Cox regression analysis, bone metastasis (HR = 1.781, P = 0.026), portal vein tumor thrombus (HR = 2.078, P = 0.015), alpha-fetoprotein (HR = 2.098, P = 0.007), and radiation dose (HR = 0.535, P = 0.023) show significant potential to predict the survival of advanced liver cancer patients treated with palliative radiotherapy. Moreover, nomograms predicting median overall survival, 1- and 2-year survival probability were plotted. The Harrell’s C-index of the predictive model is 0.709(95%CI, 0.649-0.769) and 0.735 (95%CI, 0.666-0.804) for training model and validation model respectively. Calibration curves of the 1- and 2-year overall survival of the predictive model indicate that the predicted probabilities of OS are very close to the actual observed outcomes both in training and validation phase.ConclusionBone metastasis, portal vein tumor thrombus, alpha-fetoprotein and radiation dose are independent prognostic factors for the survival of advanced liver cancer patients treated with palliative radiotherapy.

Breast ultrasound tumour classification: A Machine Learning—Radiomics based approach

AbstractPrediction of breast tumour malignancy using ultrasound imaging, is an important step for early detection of breast cancer. An efficient prediction system can be a great help to improve the survival chances of the involved patients. In this work, a machine learning (ML)—radiomics based classification pipeline is proposed, to perform this predictive modelling task, in a much more efficient manner. Multiple different types of image features of the region of interests are considered in this work, followed by a recursive feature elimination based feature selection step. Furthermore, a synthetic minority oversampling technique based step is also included in the pipeline, to deal with the class imbalance problem, that is often present in medical imaging datasets. Various ML models are considered in the subsequent model training phase, on a publicly available breast ultrasound image dataset (BUSI). From experimental analysis it has been observed that, shape, texture and histogram oriented gradients related features are the most informative, with respect to the predictive modelling task. Furthermore, it was observed that ensemble learners such as random forest, gradient boosting and AdaBoost classifiers are able to achieve significant results with respect to multiple evaluation metrics. The proposed approach achieved the state‐of‐the‐art accuracy, area under the curve, F1‐score and Mathews correlation coefficient values of 0.974, 0.97, 0.94 and 0.959, respectively, on the BUSI dataset. Such kind of impressive results suggest that the proposed approach can have a very high practical utility, in real medical diagnostic settings.

Integrated lithology identification based on images and elemental data from rocks

In order to realize low-cost, fast and accurate lithology identification in reservoir evaluation, stratum development potential evaluation and underground engineering construction, we propose an intelligent lithology identification method, which can achieve deep fusion of images and elemental data from rocks, and a step-wise model training method. Lithological features are extracted from both images and the rock elemental data using convolution layers and fully connected layers; those features are then concatenated and fed to the fully connected layers to achieve deep fusion. In the model training phase, the image feature extraction module of the fusion model is first trained only with images, and then the entire fusion model is trained using both the images and the rock elemental data. We also conduct reliability assessment and then compare the performance of our fusion identification model to that of an identification model that relies solely on images. The maximum accuracy (94.62%), recall (94.30%), F1-score (93.98%), and Kappa coefficient (0.94) values of the fusion method far exceed that of the image identification method. The step-wise model training method noticeably reduces the amount of rock elemental data required by the fusion model. High image similarity and smaller lithology features degrade the performance of the image identification method, while the fusion method can significantly overcome these challenges and thus greatly improve the accuracy of lithology identification. The fusion identification method is more accurate and robust than the image identification method.

Stochastic modeling for tracking and prediction of gradual and transient battery performance degradation

As rechargeable battery-powered devices become a pervasive part of daily life, consumer standards for quality and reliability in these devices increase. One standard of particular interest is the ability to reliably track and predict the daily and lifetime capacity degradation of the batteries, over successive charge–discharge cycles. Tracking and predicting the degradation trend over batteries’ entire service lives can be challenging, not only due to the nonlinear degradation patterns subject to different usage conditions, but also because of the presence of transient regeneration events that can rapidly increase the battery's maximum capacity and affect degradation rates. Current methods in the literature commonly apply an exponential model in a Bayesian inference framework to track and predict global degradation trend, neglecting the transient behaviors and local fluctuations, hence leading to unsatisfactory performance. This paper presents a new customary model specifically designed to track the gradual battery degradation pattern, combined with a compound Poisson process-based model that aims to capture the transient behaviors. The integration of the two parts, forming a comprehensive degradation model, provides a more accurate description of capacity variation throughout battery's life cycle. During the model training phase, upon historical data, parameters involved in the two models are estimated through two Bayesian inference techniques: step-by-step estimation by a Particle Filter and batch estimation by a Markov Chain Monte Carlo algorithm, with their performance compared. The estimated parameters are then utilized for generating transient events and predicting future capacity degradation. NASA's lithium-ion battery data are analyzed to evaluate the effectiveness of the developed degradation model.

Mapless LiDAR Navigation Control of Wheeled Mobile Robots Based on Deep Imitation Learning

This paper addresses the problems related to the mapless navigation control of wheeled mobile robots based on deep learning technology. The traditional navigation control framework is based on a global map of the environment, and its navigation performance depends on the quality of the global map. In this paper, we proposes a mapless Light Detection and Ranging (LiDAR) navigation control method for wheeled mobile robots based on deep imitation learning. The proposed method is a data-driven control method that directly uses LiDAR sensors and relative target position for mobile robot navigation control. A deep convolutional neural network (CNN) model is proposed to predict motion control commands of the mobile robot without the requirement of the global map to achieve navigation control of the mobile robot in unknown environments. While collecting the training dataset, we manipulated the mobile robot to avoid obstacles through manual control and recorded the raw data of the LiDAR sensor, the relative target position, and the corresponding motion control commands. Next, we applied a data augmentation method on the recorded samples to increase the number of training samples in the dataset. In the network model design, the proposed CNN model consists of a LiDAR CNN module to extract LiDAR features and a motion prediction module to predict the motion behavior of the robot. In the model training phase, the proposed CNN model learns the mapping between the input sensor data and the desired motion behavior through end-to-end imitation learning. Experimental results show that the proposed mapless LiDAR navigation control method can safely navigate the mobile robot in four unseen environments with an average success rate of 75%. Therefore, the proposed mapless LiDAR navigation control system is effective for robot navigation control in an unknown environment without the global map.

VMAN: A Virtual Mainstay Alignment Network for Transductive Zero-Shot Learning.

Transductive zero-shot learning (TZSL) extends conventional ZSL by leveraging (unlabeled) unseen images for model training. A typical method for ZSL involves learning embedding weights from the feature space to the semantic space. However, the learned weights in most existing methods are dominated by seen images, and can thus not be adapted to unseen images very well. In this paper, to align the (embedding) weights for better knowledge transfer between seen/unseen classes, we propose the virtual mainstay alignment network (VMAN), which is tailored for the transductive ZSL task. Specifically, VMAN is casted as a tied encoder-decoder net, thus only one linear mapping weights need to be learned. To explicitly learn the weights in VMAN, for the first time in ZSL, we propose to generate virtual mainstay (VM) samples for each seen class, which serve as new training data and can prevent the weights from being shifted to seen images, to some extent. Moreover, a weighted reconstruction scheme is proposed and incorporated into the model training phase, in both the semantic/feature spaces. In this way, the manifold relationships of the VM samples are well preserved. To further align the weights to adapt to more unseen images, a novel instance-category matching regularization is proposed for model re-training. VMAN is thus modeled as a nested minimization problem and is solved by a Taylor approximate optimization paradigm. In comprehensive evaluations on four benchmark datasets, VMAN achieves superior performances under the (Generalized) TZSL setting.

Deep Learning Method based on Big Data for Defects Detection in Manufacturing Systems Industry 4.0

Due to the technological advancement in Today’s manufacturing systems, a large amount of data is generated in different volume, velocity, and variety of kinds. Extracting information from these data and make a real-time decision is a big challenge to the current manufacturing systems. This study presents a novel model that converts the iFactory learning facility into a fully Industry 4.0 (I4.0) manufacturing system. To achieve this purpose, we utilized the cyber physical system (CPS) components and sensors, the Internet of Things (IoT), deep learning methods, and cloud computing to fully meet the I4.0 enablers. Cloud computing is utilised in two phases: (1) during the model training phase to hold a large amount of product image data collected from the inspection station, and (2) during the execution of the model. The core learning model is based on a convolutional neural network (CNN) that is trained from the captured product images in the production line to predict the defective items in the line. The model was initialized by Resnet method and optimized to improve the learning rate and reduced loss function. The supervised learning model achieved high accuracy prediction performance up to 96.75% in the real-time decision making process. The model was able to extract the feature map of the normal non-detective product and use it to improving the accuracy and reducing the traffic between iFactory station and the cloud server. The model exploits the parallel computing big-data framework to achieve a real time decision making. The model can be applied to the current system and adopted as with all it is functionalities for the newer systems.

Machine learning approaches for estimation of compressive strength of concrete

Estimation of compressive strength of rubberized concrete is important for engineering safety. In this study, measured data (the compressive strength of rubberized concrete and its impacting factors) were collected by literature review (457 samples). In order to accurately predict the compressive strength of rubberized concrete, four machine learning models [artificial neural network (ANN), k-nearest neighbor (KNN), regression trees (RT), and random forests (RF)] were developed and compared to estimate the compressive strength of rubberized concrete, and the modeling results were compared with two traditional expressions. The model performance was evaluated using three performance indicators: the Nash–Sutcliffe efficiency coefficient (NSC), the root-mean-squared error (RMSE), and the mean absolute error (MAE). The results showed that the RT model performs the best, followed by the ANN and RF in the model training phase. In the model testing phase, the ANN model performs the best, followed by the RT, RF, and KNN. The overall results indicated that the ANN model performs the best, followed by RT and RF, and the KNN model performs the worst. The ANN and RT models outperformed the two traditional expressions. The tree-based models (RT and RF) and KNN model may not be applicative to estimate the compressive strength of rubberized concrete due to the generally poor performances in the model testing phase compared with that in the model training phase. The results showed that the traditional ANN model is sufficient for the accurate estimation of the compressive strength of rubberized concrete when the model is properly trained. The results in the present research can provide reference for the prediction of the compressive strength of rubberized concrete, which will benefit engineering management and safety as well.

Detection of Low-Frequency and Multi-Stage Attacks in Industrial Internet of Things

The increasingly sophisticated cyber attacks have become a serious challenge for Industrial Internet of Things (IIoT), which presents two new characteristics: low frequency and multi-stage. That is, hackers could gain authority to attack industrial equipment/infrastructure gradually in a long interval through lurking, lateral intrusion and privilege escalation. While, the existing Machine Learning (ML) based intrusion detection schemes all require the participation of expert knowledge, so it is difficult to adaptively select an attack interval and a retraining period of the detection model in IIoT, resulting in poor detection performance. To solve above problems, a bidirectional long and short-term memory network with multi-feature layer (B-MLSTM) is designed. Firstly, sequence and stage feature layers are introduced in the model training phase model which can learn the corresponding attack interval from historical data, so that the model can effectively detect attacks with different intervals. Then, a double-layer reverse unit is introduced to update the detection model. By collecting information from test data and making association analysis with historical data, the retraining period is adaptively selected to match the new attack interval. Compared with the previous works, our proposed scheme has a lower false positive rate than existing schemes by at least 46.79%, and the false negative rate is reduced by at least 79.85%, which are carried out on three classic IIoT datasets.

Factors of Performance for Application of AI Models in GPU Cloud

Introduction. In machine learning (ML) and artificial intelligence (AI) works, the emphasis is usually on the quality of classification or the accuracy of parameter estimation. If the focus is on performance, then it is also mainly about the performance of the model's training phase. However, with the proliferation of AI applications in real-world problems, the problem of ensuring high data processing performance with ready models becomes more important. By its nature, this problem is fundamentally different from the one of model training: the latter deals with intensive calculations and the former with simple calculations, but large flows of data (files) coming from the network or file system for processing. That is, the typical task of parallel processing with intensive input-output. Besides, in terms of application, the AI module that performs classification, evaluation, or other data processing is a "black box": the cost of developing and training the model, as well as the risks of failure, are too high to handle such tasks in a non-professional manner. Therefore, performance optimization primarily involves the selection and balancing of system parameters. Cloud systems with their flexibility, manageability and easy scaling are the ideal platforms for such tasks. Consider in more detail the task of investigating the factors which affect performance on a single, but notable, pattern recognition sample of a subset of ImageNet image collection [1] classified by the 50-layer deep learning neural network ResNet-50 [2]. The purpose of the paper is to experimentally investigate the factors that influence the performance of a ready-to-use neural network model application in GPU cloud systems of various architectures. Results. Overheads related to microservices and distributed architectures, memory, network, batch size, synchronous and asynchronous interactions are estimated. The complex nonlinear nature of the influence of the system parameters in various combinations is demonstrated.

Research on Dermoscopic Segmentation based on Multi-scale Convolutional Neural Network

Abstract The skin is an important organ of the human body, and the major skin diseases represented by melanoma are difficult to find and difficult to treat, which directly threaten the patient’s life safety. Therefore, rapid and accurate dermatoscopy diagnosis[1] is particularly important. The lack of experienced dermatologists limits the large-scale early diagnosis or screening of melanoma. Based on the above problems, this paper uses the method of multi-scale convolutional neural network to segment the Dermoscopic image by means of image segmentation in artificial intelligence to help medical personnel to achieve reliable diagnostic reference. In the overall segmentation network, this paper first enhances the contrast of the original image through preprocessing, and uses data enhancement and segmentation to increase the dataset. In the model training phase, this paper adopts the basic segmentation framework of U-Net[2] network. Different from U-Net networks, this paper introduces multi-scale features and multi-scale fusion mechanisms to deeply mine and fuse dermoscopic image features. Secondly, the model uses Focal Loss as the fundus segmentation loss function, and finds the optimal parameters of the loss function through a random grid search algorithm. In this paper, for the problem of positive and negative sample imbalance in dermoscopic images, this paper proposes an adaptive Focal Loss loss function and Channel-Wise Attention mechanism to adaptively weight the feature maps.

A Novel Approach for Automatic Detection of Concrete Surface Voids Using Image Texture Analysis and History‐Based Adaptive Differential Evolution Optimized Support Vector Machine

To inspect the quality of concrete structures, surface voids or bugholes existing on a concrete surface after the casting process needs to be detected. To improve the productivity of the inspection work, this study develops a hybrid intelligence approach that combines image texture analysis, machine learning, and metaheuristic optimization. Image texture computations employ the Gabor filter and gray‐level run lengths to characterize the condition of a concrete surface. Based on features of image texture, Support Vector Machines (SVM) establish a decision boundary that separates collected image samples into two categories of no surface void (negative class) and surface void (positive class). Furthermore, to assist the SVM model training phase, the state‐of‐the‐art history‐based adaptive differential evolution with linear population size reduction (L‐SHADE) is utilized. The hybrid intelligence approach, named as L‐SHADE‐SVM‐SVD, has been developed and complied in Visual C#.NET framework. Experiments with 1000 image samples show that the L‐SHADE‐SVM‐SVD can obtain a high prediction accuracy of roughly 93%. Therefore, the newly developed model can be a promising alternative for construction inspectors in concrete quality assessment.

The effects of measurement error and testing frequency on the fitness-fatigue model applied to resistance training: A simulation approach

This study investigated the effects of measurement error and testing frequency on prediction accuracy of the standard fitness-fatigue model. A simulation-based approach was used to systematically assess measurement error and frequency inputs commonly used when monitoring the training of athletes. Two hypothetical athletes (intermediate and advanced) were developed and realistic training loads and daily ‘true’ power values were generated using the fitness-fatigue model across 16 weeks. Simulations were then completed by adding Gaussian measurement errors to true values with mean 0 and set standard deviations to recreate more and less reliable measurement practices used in real-world settings. Errors were added to the model training phase (weeks 1–8) and sampling of data was used to recreate different testing frequencies (every day to once per week) when obtaining parameter estimates. In total, 210 sets of simulations (N = 104 iterations) were completed using an iterative hill-climbing optimisation technique. Parameter estimates were then combined with training loads in the model testing phase (weeks 9–16) to quantify prediction errors. Regression analyses identified positive associations between prediction errors and the linear combination of measurement error and testing frequency ([Formula: see text]=0.87–0.94). Significant model improvements (P < 0.001) were obtained across all scenarios by including an interaction term demonstrating greater deleterious effects of measurement error at low testing frequencies.The findings of this simulation study represent a lower-bound case and indicate that in real-world settings, where a fitness-fatigue model is used to predict training response, measurement practices that generate coefficients of variation greater than [Formula: see text]4% will not provide satisfactory results.

Construction of non-convex fuzzy sets and its application

Although non-convex fuzzy set (FS) has the high potential of great performance in data modeling and controlling, it is seldom used and discussed because the lack of linguistic explanation and normative construction way. To address this problem, we propose a method named “parametric qualitative fuzzy set (PQ FS) plus choice strategy” for the construction and linguistic explanation of non-convex FS, in which PQ FS is a collection of convex FSs with special structure, and choice strategy is an approach to choose convex FSs from PQ FS. Based on this method, a non-convex FS is obtained as the trajectory of a collection of convex FS by choosing specific convex FS under specific situation. Thus, the linguistic explanation of non-convex FS is obtained: using non-convex FSs to represent linguistic variables does not violate the routine of using convex FSs, because it shows that the linguistic variable is just represented by different convex FS at different situation. Theorems are shown to demonstrate that the “PQ FS plus choice strategy” can effectively construct a non-convex FS. Furthermore, “Why a fuzzy logic system (FLS) adopting non-convex FSs may have a higher approximation capability” is discussed by introducing a parametric qualitative FLS (PQ FLS) that is compared with a typical Mamdani FLS as a function approximator. This indicates that non-convex FSs can approximate more extrema in a given universe with smaller partition numbers or fewer rules than convex FSs. Finally, the experimental results verify that a PQ FLS designed with the proposed non-convex FS construction method can outperform traditional convex fuzzy logic controllers (FLCs). Meanwhile, using parallel computing in the model training phase of PQ FLSs can reduce the calculation time compared to single-thread mode.

1521-P: An Efficacy Evaluation on Prediabetes Management Using a Machine Learning-Based Risk Stratification Approach

The progression of prediabetes towards diabetes was heterogeneous, suggesting a demand for individualized management. We aim to investigate whether stratifying participants with prediabetes according to their progression risk could affect the response to interventions. We developed a 1-year diabetes progression risk evaluating model with a panel of diabetes risk factors using a machine learning method (i.e., Bagging algorithm with 300 decision trees) in prediabetes participants in Pinggu Study (a prospective population-based survey in suburban Beijing, n=622). The model training phase was conducted on 2/3 randomly selected participants and the validation phase was conducted on other 1/3. This model successfully predicted the 1-year progression of prediabetes participants in Pinggu study (internal ROC AUC 1.000, external ROC AUC 0.883). In Beijing Prediabetes Reversion Program (BPRP, a multi-central RCT to evaluate the efficacy of lifestyle and/or pioglitazone on prediabetes reversion, n=1897), subjects were divided into tertiles as high, medium and low risk groups according to their progression probability predicted by the model. Pioglitazone plus intensive lifestyle significantly increased reversion and reduced diabetes progression only in high risk group. This study suggests personalized treatment on prediabetes according to their progression risk is necessary. Disclosure X. Zou: None. Z. Zhu: None. Y. Luo: None. Y. Li: None. X. Zhou: None. L. Ji: Advisory Panel; Self; AstraZeneca. Consultant; Self; AstraZeneca, Bayer AG, Boehringer Ingelheim International GmbH, Bristol-Myers Squibb Company, Eli Lilly and Company, Merck KGaA, Merck Sharp & Dohme Corp., Novartis AG, Novo Nordisk A/S, Roche Pharma, Sanofi, Takeda Pharmaceutical Company Limited. Research Support; Self; AstraZeneca, Bristol-Myers Squibb Company, Eli Lilly and Company, Merck Sharp & Dohme Corp., Novartis AG, Roche Pharma, Sanofi. Funding The Major Chronic Non-Communicable Disease Prevention and Control Research of China; Beijing Science and Technology Committee; National Natural Science Foundation of China

New Approach for Sediment Yield Forecasting with a Two-Phase Feedforward Neuron Network-Particle Swarm Optimization Model Integrated with the Gravitational Search Algorithm

Predicting sediment yield is an important task for decision-makers in environmental monitoring and water management since the benefits of applying non-linear, artificial intelligence (AI) models for optimal prediction can be far reaching in real-life decision support systems. AI-based models are considered to be favorable predictive tools since the nonlinear nature of suspended sediment data series warrants the utilization of nonlinear predictive methods for feature extraction, and for accurate simulation of suspended sediment load. In this study, Artificial Neural Network (ANN) approaches are employed to estimate the monthly sediment load where the two-phase Feed-forward Neuron Network Particle Swarm Optimization Gravitational Search Algorithm (FNN-PSOGSA) is developed, and then evaluated in respect to 3 distinct algorithms: the Adaptive Neuro-Fuzzy Inference System (ANFIS), Feed-forward Neuron Network (FNN) and the single-phase Feed-forward Neuron Network Particle Swarm Optimization (FNN-PSO). The study is carried out using the monthly rainfall, runoff and sediment data spanning a 10 year period (2000–2009) where about 75% of data are used in model training phase, 25% in testing phase. Three statistical performance criteria namely: the mean absolute error (MAE), Nash-Sutcliffe coefficient (NSE) and the Willmott’s Index (WI) and diagnostic plots visualizing the tested results are used to evaluate the performance of four AI-based models. The results reveal that the objective model (the two-phase FNN-PSOGSA model) and the single-phase FNN-PSO model yielded more precise results compared to the other forecast models. This result accorded to an NSE value of 0.612 (for the FNN-PSOGSA model) vs. an NS value of 0.500, 0.331 and 0.244 for the FNN-PSO, FNN and ANFIS models, and WI = 0.832 vs. 0.771, 0.692 and 0.726, respectively The study also demonstrated that the FNN model generated slightly better results than the ANFIS model for the estimation of sediment load data but overall, the two-phase FNN-PSOGSA model outperformed all comparison models. In light of the superior performance, this research advocates that the fully-optimized two-phase FNN-PSOGSA model can be explored as a decision-support tool for monthly sediment load forecasting using the rainfall and runoff values as the predictor datasets.