Training Phase Research Articles

Multi-view Test-time Adaptation for Semantic Segmentation in Clinical Cataract Surgery.

Cataract surgery, a widely performed operation worldwide, is incorporating semantic segmentation to advance computer-assisted intervention. However, the tissue appearance and illumination in cataract surgery often differ among clinical centers, intensifying the issue of domain shifts. While domain adaptation offers remedies to the shifts, the necessity for data centralization raises additional privacy concerns. To overcome these challenges, we propose a Multi-view Test-time Adaptation algorithm (MUTA) to segment cataract surgical scenes, which leverages multi-view learning to enhance model training within the source domain and model adaptation within the target domain. In the training phase, the segmentation model is equipped with multi-view decoders to boost its robustness against variations in cataract surgery. During the inference phase, test-time adaptation is implemented using multi-view knowledge distillation, enabling model updates in clinics without data centralization or privacy concerns. We conducted experiments in a simulated cross-center scenario using several cataract surgery datasets to evaluate the effectiveness of MUTA. Through comparisons and investigations, we have validated that MUTA effectively learns a robust source model and adapts the model to target data during the practical inference phase. Code and datasets are available at https://github.com/liamheng/CAI-algorithms.

An Efficient 3D Convolutional Neural Network for Dose Prediction in Cancer Radiotherapy from CT Images.

Introduction: Cancer is a highly lethal disease with a significantly high mortality rate. One of the most commonly used methods for treatment is radiation therapy. However, cancer treatment using radiotherapy is a time-consuming process that requires significant manual work from planners and doctors. In radiation therapy treatment planning, determining the dose distribution for each of the regions of the patient's body is one of the most difficult and important tasks. Nowadays, artificial intelligence has shown promising results in improving the quality of disease treatment, particularly in cancer radiation therapy. Objectives: The main objective of this study is to build a high-performance deep learning model for predicting radiation therapy doses for cancer and to develop software to easily manipulate and use this model. Materials and Methods: In this paper, we propose a custom 3D convolutional neural network model with a U-Net-based architecture to automatically predict radiation doses during cancer radiation therapy from CT images. To ensure that the predicted doses do not have negative values, which are not valid for radiation doses, a rectified linear unit (ReLU) function is applied to the output to convert negative values to zero. Additionally, a proposed loss function based on a dose-volume histogram is used to train the model, ensuring that the predicted dose concentrations are highly meaningful in terms of radiation therapy. The model is developed using the OpenKBP challenge dataset, which consists of 200, 100, and 40 head and neck cancer patients for training, testing, and validation, respectively. Before the training phase, preprocessing and augmentation techniques, such as standardization, translation, and flipping, are applied to the training set. During the training phase, a cosine annealing scheduler is applied to update the learning rate. Results and Conclusions: Our model achieved strong performance, with a good DVH score (1.444 Gy) on the test dataset, compared to previous studies and state-of-the-art models. In addition, we developed software to display the dose maps predicted by the proposed model for each 2D slice in order to facilitate usage and observation. These results may help doctors in treating cancer with radiation therapy in terms of both time and effectiveness.

Phonological Feature Abstraction Before 6 Months: Amodal Recognition of Place of Articulation Across Multiple Consonants.

The classical view is that perceptual attunement to the native language, which emerges by 6-10 months, developmentally precedes phonological feature abstraction abilities. That assumption is challenged by findings from adults adopted into a new language environment at 3-5 months that imply they had already formed phonological feature abstractions about their birth language prior to 6 months. As phonological feature abstraction had not been directly tested in infants, we examined 4-6-month-olds' amodal abstraction of the labial versus coronal place of articulation distinction between consonants. In the training phase, infants heard a series of labial non-words paired with an animal image and a series of coronal non-words (multisyllabic) paired with another image. At test, they viewed a silent video of a talker producing coronal and labial words, paired with either the familiarised image or the contrary image. The infants looked significantly longer on matching trials than mismatching trials, suggesting amodal abstraction of this consonantal place of articulation distinction by 4-6 months. These findings provide direct evidence for the inference from the adoptee findings that phonological feature abstraction emerges prior to perceptual attunement.

Does curcumin supplementation affect inflammation, blood count and serum brain-derived neurotropic factor concentration in amateur long-distance runners?

Curcumin is known for its potential health benefits; however, the evidence remains inconclusive regarding its necessity as a supplement for athletes during the preparatory phase of training. This study aimed to assess the effect of 6-week curcumin supplementation at a dose of 2g/day on selected inflammatory markers, blood count, and brain-derived neurotropic factor (BDNF) levels in middle-aged amateur long-distance runners during the preparatory period of a macrocycle. Thirty runners were randomly assigned to either a curcumin-supplemented group (CUR, n = 15) or a placebo group (PLA, n = 15). Venous blood samples were collected at rest, immediately post-exercise, and 1h post-exercise. The participants underwent a graded exercise stress test, with an increasing inclination angle after reaching a speed of 14 km/h, both before and after the 6-week supplementation period. Blood samples were collected at rest, 3 minutes post-stress test, and after 1 hour of recovery. The results showed no significant changes in C-reactive protein (CRP), interleukin-6 (IL-6), tumor necrosis factor α (TNF-α), interleukin-1 β (IL-1β), or blood morphology due to curcumin supplementation. However, BDNF levels increased by 21% in the CUR group post-supplementation, while a 5% decrease was observed in the PLA group. These findings do not support a significant effect of curcumin supplementation on inflammatory markers, blood count, or BDNF concentration. Further research is warranted to determine the potential benefits of curcumin supplementation for endurance athletes during the preparatory period for a training cycle.

Influencing Factors Analysis and Prediction of Gas Emission in Mining Face

Mine gas emission is one of the main causes of gas disasters. In order to achieve the accurate prediction of gas emission, a gas emission prediction model based on the random forest (RF) method was proposed in combination with the analysis of its influencing factors. The prediction results were compared with the support vector regression (SVR) and BP neural network (BPNN) methods, and then they were verified and analyzed through the Dongqu coal mine. The results show that the gas emission prediction model based on random forest has strong generalization and robustness, and RF has a wide range of parameter adaptation during the modeling process. When the number of trees (ntree) exceeds 100, its training error tends to stabilize, and changes in ntree have no substantial impact on the prediction performance. The SVR prediction model has significant bias in both the training and testing stages. Meanwhile, the BPNN model has excellent prediction results in the training phase, but there is a large error in the testing stage, which indicates that there is an “overfitting” phenomenon in the training stage, resulting in weak generalization. The evaluation of variable importance shows that the extraction rate, coal seam depth, daily production, gas content in adjacent layers, and coal seam thickness have a significant impact on gas emission. Meanwhile, through application analysis, it is further demonstrated that the random forest method has high accuracy, strong stability, and universality, and it can achieve good predictive performance without the need for complex parameter settings and optimization, making it is very suitable for predicting gas emission.

A Novel Method to Forecast Nitrate Concentration Levels in Irrigation Areas for Sustainable Agriculture

This study developed an ANN-based model to predict nitrate concentrations in drainage waters using parameters that are simpler and more cost-effective to measure within the Lower Seyhan Basin, a key agricultural region in Turkey. For this purpose, daily water samples were collected from a drainage measurement station during the 2022 and 2023 water years, and nitrate concentrations were determined in the laboratory. In addition to nitrate concentrations, other parameters, such as flow rate, EC, pH, and precipitation, were also measured simultaneously. The complex relationship between measured nitrate values and other parameters, which are easier and less costly to measure, was used in two different scenarios during the training phase of the ANN-Nitrate model. After the model was trained, nitrate values were estimated for the two scenarios using only the other parameters. In Scenario I, random values from the dataset were predicted, while in Scenario II, predictions were made as a time series, and model results were compared with measured values for both scenarios. The proposed model reliably fills dataset gaps (Scenario I) and predicts nitrate values in time series (Scenario II). The proposed model, although based on an artificial neural network (ANN), also has the potential to be adapted for methods used in machine learning and artificial intelligence, such as Support Vector Machines, Decision Trees, Random Forests, and Ensemble Learning Methods.

PREmbed: Balancing Conditional Generative Models with Embedding Pretraining and Regularization

Recent advancements in conditional generative models, e.g., Conditional Variational AutoEncoder (CVAE) and Conditional Denoising Diffusion Probabilistic Model (CDDPM), which utilize class-specific embeddings for generating images in specific classes, have demonstrated exceptional capabilities in producing high-quality images on balanced datasets. However, their performance decreases with imbalanced real-world data, affecting the fidelity and diversity of the generated images. This paper discusses the reasons and solutions for the imbalance issue in CVAE and CDDPM. By selectively reweighting the gradients of the embedding layer and main network, we identify the embedding layer’s bias as a critical bottleneck in these models. Motivated by this, we propose Embedding Pretraining and Regularization (PREmbed), a training methodology for addressing biased learning in the embedding layer of these models. PREmbed consists of two key components: (i) Masked Autoencoder pretraining on the original imbalanced dataset to obtain a robust initial embedding, and (ii) an embedding regularization loss to preserve class-level distances during the training phase to improve imbalanced embedding learning. Implementing PREmbed leads to the enhanced convergence of a robust embedding layer. Our experiments on imbalanced CIFAR-10 and CUB-200 datasets demonstrate that PREmbed enhances the performance of CVAE and CDDPM; it outperforms the baseline CVAE by 22.3% (FID 18.63) and the CDDPM by 30.3% (FID 12.70) in imbalanced CIFAR-10; 33.7% (FID 24.08) and 26.5% (FID 13.90) in CUB-200; and 16.4% (FID 49.41) and 15.2% (FID 33.05) in ImageNet-LT (which is where it achieves the state-of-the-art performance in imbalanced conditional image generation task).

Efficient Multi-Task Training with Adaptive Feature Alignment for Universal Image Segmentation.

Universal image segmentation aims to handle all segmentation tasks within a single model architecture and ideally requires only one training phase. To achieve task-conditioned joint training, a task token needs to be used in the multi-task training to condition the model for specific tasks. Existing approaches generate the task token from a text input (e.g., "the task is panoptic"). However, such text-based inputs merely serve as labels and fail to capture the inherent differences between tasks, potentially misleading the model. In addition, the discrepancy between visual and textual modalities limits the performance gains in existing text-involved segmentation models. Nevertheless, prevailing modality-alignment methods rely on large-scale uni-modal encoders for both modalities and an extremely large amount of paired data for training, and therefore it is hard to apply these existing models to lightweight segmentation models and resource-constrained devices. In this paper, we propose Adaptive Feature Alignment (AFA) integrated with a learnable task token to address these issues. The learnable task token automatically captures inter-task differences from both image features and text queries during training, providing a more effective and efficient solution than a predefined text-based token. To efficiently align the two modalities without introducing extra complexity, we reconsider the differences between a text token and an image token and replace image features with class-specific means in our proposed AFA. We evaluate our model performance on the ADE20K and Cityscapes datasets. Experimental results demonstrate that our model surpasses baseline models in both efficiency and effectiveness, achieving state-of-the-art performance among segmentation models with a comparable amount of parameters.

A machine-learning reconstruction of sea surface pCO2 in the North American Atlantic Coastal Ocean Margin from 1993 to 2021

Abstract. Insufficient spatiotemporal coverage of observations of the surface partial pressure of CO2 (pCO2) has hindered precise carbon cycle studies in coastal oceans and justifies the development of spatially and temporally continuous pCO2 data products. Earlier pCO2 products have difficulties in capturing the heterogeneity of regional variations and decadal trends of pCO2 in the North American Atlantic Coastal Ocean Margin (NAACOM). This study developed a regional reconstructed pCO2 product for the NAACOM (Reconstructed Coastal Acidification Database-pCO2, or ReCAD-NAACOM-pCO2) using a two-step approach combining random forest regression and linear regression. The product provides monthly pCO2 data at 0.25° spatial resolution from 1993 to 2021, enabling investigation of regional spatial differences, seasonal cycles, and decadal changes in pCO2. The observation-based reconstruction was trained using Surface Ocean CO2 Atlas (SOCAT) observations as observational values, with various satellite-derived and reanalysis environmental variables known to control sea surface pCO2 as model inputs. The product shows high accuracy during the model training, validation, and independent test phases, demonstrating robustness and a capability to accurately reconstruct pCO2 in regions or periods lacking direct observational data. Compared with all the observation samples from SOCAT, the pCO2 product yields a determination coefficient of 0.92, a root-mean-square error of 12.70 µatm, and an accumulative uncertainty of 23.25 µatm. The ReCAD-NAACOM-pCO2 product demonstrates its capability to resolve seasonal cycles, regional-scale variations, and decadal trends of pCO2 along the NAACOM. This new product provides reliable pCO2 data for more precise studies of coastal carbon dynamics in the NAACOM region. The dataset is publicly accessible at https://doi.org/10.5281/zenodo.14038561 (Wu et al., 2024a) and will be updated regularly.

The effect of reward and voluntary choice on the motor learning of serial reaction time task.

Reward and voluntary choice facilitate motor skill learning through motivation. However, it remains unclear how their combination influences motor skill learning. The purpose of the present study is to investigate the effects of reward and voluntary choice on motor skill learning in a serial reaction time task (SRTT). Participants completed six parts of SRTT, including pre-test, training phase, immediate post-test, a random session, delayed post-test, and retention test on the following day. During the training phase, participants were divided into four groups (reward_choice, reward_no-choice, no-reward_choice, no-reward_no-choice). In the reward condition, participants received reward for correct and faster (than a baseline) responses while those in the no-reward groups did not. For the choice manipulation, participants in the voluntary choice groups chose the color of the target, whereas in the forced choice groups, the same color was assigned by the computer. The results showed that the four groups did not exhibit any significant differences in reaction time and error rate in the pre-test phase. Importantly, both reward and voluntary choice significantly enhanced sequence-specific learning effects, while no interaction was found. No significant effects of reward and voluntary choice were observed in the retention test. These findings suggest that reward and voluntary choice enhance motor skill performance and training independently, potentially at the action-selection level, which implies different mechanisms underlying the influences of reward and voluntary choice.

AWGR-based all-optical switching network for distributed machine learning.

With the increase sizes of training datasets and models, the bottleneck in distributed machine learning (DML) training has shifted from computation to communication. To address this bottleneck, we propose an all-optical switching network architecture for accelerating the communication phase of DML training. Experimental results validate packets with error-free and 385 ns server-to-server low-latency communication at traffic load of 0.9. Small-scale DML training experiment deployed in the proposed architecture shows that Resnet50, Resnet101, and Vgg19 can be accelerated by 1.16x to 1.48x compared to electrical switching network. The proposed architecture demonstrates a 58.9% enhancement in cost efficiency and a 60.9% improvement in power efficiency compared to the 3-tier fat-tree architecture.

A Preliminary Study of Nutrients Related to the Risk of Relative Energy Deficiency in Sport (RED-S) in Top-Performing Female Amateur Triathletes: Results from a Nutritional Assessment.

Background/Objectives: As an endurance multi-sport race, triathlon places significant energy demands on athletes during performance and training. Insufficient energy intake from food can lead to low energy availability (LEA) and Relative Energy Deficiency in Sport (RED-S). We aimed to measure symptoms related to LEA, examine the risk of RED-S, and find how diet relates to the risk of RED-S in highly trained female amateur triathletes. Methods: Our sample was 20 top-performing female triathletes competing in Quarter Ironman (IM), Half IM, IM, or Double IM triathlons for 5.5 ± 2.5 y who were during the preparatory phase of training (training load 11 ± 3.76 h/week, a single workout 84 ± 25 min). Triathletes completed 3-day food diaries, training diaries, and the Low Energy Availability in Females Questionnaire (LEAF-Q). Exercise energy expenditure was estimated using wrist-worn activity trackers. To examine dietary patterns related to the first signs of LEA, predating RED-S, we created two groups: the L-LEA group (LEAF-Q score 0-5, no symptoms related to LEA, low risk of RED-S, n = 10) and the H-LEA group (LEAF-Q ≥ 6, at least one LEA-related symptom, high risk of RED-S, n = 10). Results: The risk of RED-S was prevalent in 30% of female triathletes, and 50% showed at least one symptom related to LEA. Macronutrient intake was similar in all participants, but triathletes from the H-LEA group tended to eat more plant-sourced protein and fiber. They consumed less saturated fatty acids but ingested more significant amounts of n-6 polyunsaturated fatty acids (PUFAn6). Conclusions: We conclude that foods higher in plant proteins, fiber, and PUFAn6 might predispose female triathletes to LEA by reducing the diet's energy density.

Training biomechanics into emotion recognition of athletes based on time series change characteristics of expression

Emotion management is a critical psychological skill for athletes, deserving significant attention during the foundational training phase. Early intervention in athletes’ emotional states allows coaches to devise effective training plans, ultimately enhancing competition performance. Emotions can be evaluated through facial expressions and physiological signals, providing a basis for detecting athletes’ training emotions. This study explores the biomechanical characteristics of facial expressions to recognize emotions. First, the relationship between facial expressions and facial Action Units (AUs) is analyzed, identifying AU combinations that effectively represent facial expressions. Considering the temporal and spatial dynamics of facial expressions, AUs are used as feature inputs for an SVM model to classify athletes’ training emotions. Subsequently, emotional expressions are mapped to corresponding emotional states, introducing an emotion index to quantify athletes’ training emotions. Finally, quantitative emotion recognition is achieved based on transient facial expressions observed during training. This research provides a theoretical framework for coaches and athletes to enhance emotional management through biomechanically informed training approaches.

Analysis of Operant Self-administration Behaviors with Supervised Machine Learning: Protocol for Video Acquisition and Pose Estimation Analysis Using DeepLabCut and Simple Behavioral Analysis.

The use of supervised machine learning to approximate poses in video recordings allows for rapid and efficient analysis of complex behavioral profiles. Currently, there are limited protocols for automated analysis of operant self-administration behavior. We provide a methodology to (1) obtain videos of training sessions via Raspberry Pi microcomputers or GoPro cameras, (2) obtain pose estimation data using the supervised machine learning software packages DeepLabCut (DLC) and Simple Behavioral Analysis (SimBA) with a local high-performance computer cluster, (3) compare standard Med-PC lever response versus quadrant time data generated from pose estimation regions of interest, and (4) generate predictive behavioral classifiers. Overall, we demonstrate proof of concept to use pose estimation outputs from DLC to both generate quadrant time results and obtain behavioral classifiers from SimBA during operant training phases.

A priori physical information to aid generalization capabilities of neural networks for hydraulic modeling

The application of Neural Networks to river hydraulics and flood mapping is fledgling, despite the field suffering from data scarcity, a challenge for machine learning techniques. Consequently, many purely data-driven Neural Networks have shown limited capabilities when tasked with predicting new scenarios. In this work, we propose introducing physical information into the training phase in the form of a regularization term. Whereas this idea is formally borrowed from Physics-Informed Neural Networks, the proposed methodology does not necessarily resort to PDEs, making it suitable for scenarios with significant epistemic uncertainties, such as river hydraulics. The method enriches the information content of the dataset and appears highly versatile. It shows improved predictive capabilities for a highly controllable, synthetic hydraulic problem, even when extrapolating beyond the boundaries of the training dataset and in data-scarce scenarios. Therefore, our study lays the groundwork for future employment on real datasets from complex applications.

Low and high frequency isometric handgrip exercise training similarly reduce resting blood pressure in young normotensive adults: A randomised controlled trial

ABSTRACT We investigated the effects of low and high frequency isometric handgrip exercise training (IHGT) on resting blood pressure, and the affective/perceptual responses during training. Sixty young normotensive adults were randomised to either a no-intervention control group (CON: n = 20; 12 female) or a group performing either two (LOW: n = 20; 18 female) or four (HIGH: n = 20; 13 female) sessions/week of IHGT for 4 weeks. IHGT involved 4 × 2-min holds at 30% maximal voluntary contraction using the dominant hand. Resting blood pressure was measured before and after training. Affective valence was measured during the first session of each training week. Systolic blood pressure was reduced following both LOW (adjusted mean change [95% CI]: −4.5 [−6.8, −2.2] mmHg) and HIGH (−5.3 [−7.6, −3.0] mmHg) frequency IHGT groups compared to CON (+0.5 [−1.8, 2.8] mmHg; p < 0.01), with no difference between LOW and HIGH. There were no changes in diastolic blood pressure. During the first session, affective valence decreased by 2.5 ± 2.6 units and became negative (lowest affect: −0.75 ± 1.84 units). However, affective responses improved as training progressed. Low and high frequency IHGT similarly reduce resting blood pressure in young normotensive adults. Negative affective responses in the early phase of training improve as the intervention progresses.

Using Multiple Exemplar Training to Teach Basic Theory-of-Mind Skills to Adolescents With Autism Spectrum Disorder

Adolescents with autism spectrum disorder (ASD) seem to have difficulty dealing with complex social situations, which is likely related to their challenges in understanding others’ perspectives, according to the Theory of Mind (ToM). The purpose of this preliminary study was to examine the effect of multiple exemplar training (MET) on the acquisition, maintenance, and generalization of basic theory-of-mind skills in three Chinese adolescents with ASD by employing a single-case, multiple-probe design. Results suggest that MET was effective in the acquisition and maintenance phases of training for improving all three participants’ target ToM skills and their ToM assessment scores. The participants also demonstrated increased appropriate social interactions and decreased inappropriate social interactions with peers during indoor recess, indicating that MET is effective in helping adolescents with ASD generalize ToM skills to natural settings.

Multi-Agent Reinforcement Learning for Efficient Resource Allocation in Internet of Vehicles

The Internet of Vehicles (IoV), a burgeoning technology, merges advancements in the internet, vehicle electronics, and wireless communications to foster intelligent vehicle interactions, thereby enhancing the efficiency and safety of transportation systems. Nonetheless, the continual and high-frequency communications among vehicles, coupled with regional limitations in system capacity, precipitate significant challenges in allocating wireless resources for vehicular networks. In addressing these challenges, this study formulates the resource allocation issue as a multi-agent deep reinforcement learning scenario and introduces a novel multi-agent actor-critic framework. This framework incorporates a prioritized experience replay mechanism focused on distributed execution, which facilitates decentralized computing by structuring the training processes and defining specific reward functions, thus optimizing resource allocation. Furthermore, the framework prioritizes empirical data during the training phase based on the temporal difference error (TD error), selectively updating the network with high-priority data at each sampling point. This strategy not only accelerates model convergence but also enhances the learning efficacy. The empirical validations confirm that our algorithm augments the total capacity of vehicle-to-infrastructure (V2I) links by 9.36% and the success rate of vehicle-to-vehicle (V2V) transmissions by 6.74% compared with a benchmark algorithm.

Meta-analyses of the effects of high-intensity interval training in elite athletes-part I: mean effects on various performance measures.

Meta-analysts have found that high-intensity interval training (HIIT) improves physical performance, but limited evidence exists regarding its effects on highly trained athletes, measures beyond maximum oxygen uptake ( O2max), and the moderating effects of different types of HIIT. In this study, we present meta-analyses of the effects of HIIT focusing on these deficits. The effects of 6 types of HIIT and other moderators were derived from 34 studies involving highly trained endurance and elite athletes in percent units via log-transformation from separate meta-regression mixed models for sprint, time-trial, aerobic/anaerobic threshold, peak speed/power, repeated-sprint ability, O2max, and exercise economy. The level of evidence for effect magnitudes was evaluated based on the effect uncertainty and the smallest important change of 1%. Compared with control training, HIIT showed good to excellent evidence for the substantial enhancement of most measures for some athlete subgroups in practically important study settings defined by effect moderators (maximum of 12.6%, for endurance female athletes after 6 weeks of aerobic traditional long intervals). The assessment of the moderators indicated good evidence of greater effects as follows: with more aerobic types of HIIT for O2max (+2.6%); with HIIT added to conventional training for most measures (+1.1-2.3%); during the competition phase for O2max (+4.3%); and with tests of longer duration for sprint (+5.5%) and time trial (+4.9%). The effects of sex and type of athlete were unclear moderators. The heterogeneity of HIIT effects within a given type of setting varied from small to moderate (standard deviations of 1.1%-2.3%) and reduced the evidence of benefit in some settings. Although athletes in some settings can be confident of the beneficial effects of HIIT on some measures related to competition performance, further research is needed. There is uncertainty regarding the mean effects on exercise economy and the modifying effects of sex, duration of intervention, phase of training, and type of HIIT for most measures. https://www.crd.york.ac.uk/PROSPERO/display_record.php?RecordID=236384.

Status of a New Model of ERCP Training in China

Aims: ERCP training should adapt to the rapid development of ERCP technology and indications. China has a large population and vast land area, but ERCP training resources are insufficient. To make effective use of scarce ERCP training resources, the Standardized Training for ERCP Procedure (STEP) program was launched as a pilot trial. We aimed to report the graduation and competence situations and analyze the predictive factors. Methods: This was a prospective multicenter study from January 2015 to May 2021 containing 2 stages: the training program stage and the graduate questionnaire survey stage. The graduation and competence rates were investigated and analyzed. Results: In the training stage, 146 trainees completed the training, 113 (77.4%) graduated. The gender ratio (men vs. women) was 6.7:1. The mean age was 40.14±4.13 years. In total, 11 trainees failed the institutional mentored training phase (IMTP), 18 failed the practice strengthen training phase (PSTP), and 4 failed the comprehensive theoretical test. In the questionnaire survey stage, the response rate was 92%, of which 76.9% (80/104) achieved competence. ERCP cases performed independently in IMTP (OR=1.129, P=0.025) and ERCP volume in training years (OR=1.018, P=0.026) were predictive factors for graduation. Job title (senior physician) (OR=78.811, P=0.018) and main practice form of PSTP (independently) (OR=17.062, P=0.031) were predictive factors for competence in available graduates. Conclusions: The STEP achieved an acceptable graduation and competence rate. We gained valuable experience in framework organization and multiple assessment checkpoints. Therefore, this standardized training model is suitable for developing countries with rapid progress in ERCP.