One or two things we know about concept drift-a survey on monitoring in evolving environments. Part B: locating and explaining concept drift.

In an increasing number of industrial and technical processes, machine learning-based systems are being entrusted with supervision tasks. While they have been successfully utilized in many application areas, they frequently are not able to generalize to changes in the observed data, which environmental changes or degrading sensors might cause. These changes, commonly referred to as concept drift can trigger malfunctions in the used solutions which are safety-critical in many cases. Thus, detecting and analyzing concept drift is a crucial step when building reliable and robust machine learning-driven solutions. In this work, we consider the setting of unsupervised data streams which is highly relevant for different monitoring and anomaly detection scenarios. In particular, we focus on the tasks of localizing and explaining concept drift which are crucial to enable human operators to take appropriate action. Next to providing precise mathematical definitions of the problem of concept drift localization, we survey the body of literature on this topic. By performing standardized experiments on parametric artificial datasets we provide a direct comparison of different strategies. Thereby, we can systematically analyze the properties of different schemes and suggest first guidelines for practical applications. Finally, we explore the emerging topic of explaining concept drift.

Analyzing the impact of heavy metal exposure on osteoarthritis and rheumatoid arthritis: an approach based on interpretable machine learning.

Introduction: This investigation leverages advanced machine learning (ML) techniques to dissect the complex relationship between heavy metal exposure and its impacts on osteoarthritis (OA) and rheumatoid arthritis (RA). Utilizing a comprehensive dataset from the National Health and Nutrition Examination Survey (NHANES) spanning from 2003 to 2020, this study aims to elucidate the roles specific heavy metals play in the incidence and differentiation of OA and RA. Methods: Employing a phased ML strategy that encompasses a range of methodologies, including LASSO regression and SHapley Additive exPlanations (SHAP), our analytical framework integrates demographic, laboratory, and questionnaire data. Thirteen distinct ML models were applied across seven methodologies to enhance the predictability and interpretability of clinical outcomes. Each phase of model development was meticulously designed to progressively refine the algorithm's performance. Results: The results reveal significant associations between certain heavy metals and an increased risk of arthritis. The phased ML approach enabled the precise identification of key predictors and their contributions to disease outcomes. Discussion: These findings offer new insights into potential pathways for early detection, prevention, and management strategies for arthritis associated with environmental exposures. By improving the interpretability of ML models, this research provides a potent tool for clinicians and researchers, facilitating a deeper understanding of the environmental determinants of arthritis.

Soft computing models for prediction of bentonite plastic concrete strength.

Bentonite plastic concrete (BPC) is extensively used in the construction of water-tight structures like cut-off walls in dams, etc., because it offers high plasticity, improved workability, and homogeneity. Also, bentonite is added to concrete mixes for the adsorption of toxic metals. The modified design of BPC, as compared to normal concrete, requires a reliable tool to predict its strength. Thus, this study presents a novel attempt at the application of two innovative evolutionary techniques known as multi-expression programming (MEP) and gene expression programming (GEP) and a boosting-based algorithm known as AdaBoost to predict the 28-day compressive strength ( ) of BPC based on its mixture composition. The MEP and GEP algorithms expressed their outputs in the form of an empirical equation, while AdaBoost failed to do so. The algorithms were trained using a dataset of 246 points gathered from published literature having six important input factors for predicting. The developed models were subject to error evaluation, and the results revealed that all algorithms satisfied the suggested criteria and had a correlation coefficient (R) greater than 0.9 for both the training and testing phases. However, AdaBoost surpassed both MEP and GEP in terms of accuracy and demonstrated a lower testing RMSE of 1.66 compared to 2.02 for MEP and 2.38 for GEP. Similarly, the objective function value for AdaBoost was 0.10 compared to 0.176 for GEP and 0.16 for MEP, which indicated the overall good performance of AdaBoost compared to the two evolutionary techniques. Also, Shapley additive analysis was done on the AdaBoost model to gain further insights into the prediction process, which revealed that cement, coarse aggregate, and fine aggregate are the most important factors in predicting the strength of BPC. Moreover, an interactive graphical user interface (GUI) has been developed to be practically utilized in the civil engineering industry for prediction of BPC strength.

Evidence modeling for reliability learning and interpretable decision-making under multi-modality medical image segmentation.

Reliability learning and interpretable decision-making are crucial for multi-modality medical image segmentation. Although many works have attempted multi-modality medical image segmentation, they rarely explore how much reliability is provided by each modality for segmentation. Moreover, the existing approach of decision-making such as the softmax function lacks the interpretability for multi-modality fusion. In this study, we proposed a novel approach named contextual discounted evidential network (CDE-Net) for reliability learning and interpretable decision-making under multi-modality medical image segmentation. Specifically, the CDE-Net first models the semantic evidence by uncertainty measurement using the proposed evidential decision-making module. Then, it leverages the contextual discounted fusion layer to learn the reliability provided by each modality. Finally, a multi-level loss function is deployed for the optimization of evidence modeling and reliability learning. Moreover, this study elaborates on the framework interpretability by discussing the consistency between pixel attribution maps and the learned reliability coefficients. Extensive experiments are conducted on both multi-modality brain and liver datasets. The CDE-Net gains high performance with an average Dice score of 0.914 for brain tumor segmentation and 0.913 for liver tumor segmentation, which proves CDE-Net has great potential to facilitate the interpretation of artificial intelligence-based multi-modality medical image fusion.

Inference to the best neuroscientific explanation.

Neuroscientists routinely use reverse inference (RI) to draw conclusions about cognitive processes from neural activation data. However, despite its widespread use, the methodological status of RI is a matter of ongoing controversy, with some critics arguing that it should be rejected wholesale on the grounds that it instantiates a deductively invalid argument form. In response to these critiques, some have proposed to conceive of RI as a form of abduction or inference to the best explanation (IBE). We side with this response but at the same time argue that a defense of RI requires more than identifying it as a form of IBE. In this paper, we give an analysis of what determines the quality of an RI conceived as an IBE and on that basis argue that whether an RI is warranted needs to be decided on a case-by-case basis. Support for our argument will come from a detailed methodological discussion of RI in cognitive neuroscience in light of what the recent literature on IBE has identified as the main quality indicators for IBEs.

User-Centered Development of Explanation User Interfaces for AI-Based CDSS: Lessons Learned from Early Phases.

This paper reports lessons learned during the early phases of the user-centered design process for an explanation user interface for an AI-based clinical decision support system for the intensive care unit. This paper focuses on identifying and verifying physicians' explanation needs in a multi-center, multi-country project. The explanation needs identified through context analysis and user requirements prioritization in an initial center differed from those identified through questionnaire responses from N= 9 physicians after a multi-center project workshop. These results highlight the caution that should be taken when eliciting explanation needs during the user-centered design process.

Sensing emotional valence and arousal dynamics through automated facial action unit analysis.

Information about the concordance between dynamic emotional experiences and objective signals is practically useful. Previous studies have shown that valence dynamics can be estimated by recording electrical activity from the muscles in the brows and cheeks. However, whether facial actions based on video data and analyzed without electrodes can be used for sensing emotion dynamics remains unknown. We investigated this issue by recording video of participants' faces and obtaining dynamic valence and arousal ratings while they observed emotional films. Action units (AUs) 04 (i.e., brow lowering) and 12 (i.e., lip-corner pulling), detected through an automated analysis of the video data, were negatively and positively correlated with dynamic ratings of subjective valence, respectively. Several other AUs were also correlated with dynamic valence or arousal ratings. Random forest regression modeling, interpreted using the SHapley Additive exPlanation tool, revealed non-linear associations between the AUs and dynamic ratings of valence or arousal. These results suggest that an automated analysis of facial expression video data can be used to estimate dynamic emotional states, which could be applied in various fields including mental health diagnosis, security monitoring, and education.

Explainable Machine Learning Approach to Prediction of Prolonged Intesive Care Unit Stay in Adult Spinal Deformity Patients: Machine Learning Outperforms Logistic Regression.

STUDY DESIGN: Retrospective cohort study. OBJECTIVES: Prolonged ICU stay is a driver of higher costs and inferior outcomes in Adult Spinal Deformity (ASD) patients. Machine learning (ML) models have recently been seen as a viable method of predicting pre-operative risk but are often 'black boxes' that do not fully explain the decision-making process. This study aims to demonstrate ML can achieve similar or greater predictive power as traditional statistical methods and follows traditional clinical decision-making processes. METHODS: Five ML models (Decision Tree, Random Forest, Support Vector Classifier, GradBoost, and a CNN) were trained on data collected from a large urban academic center to predict whether prolonged ICU stay would be required post-operatively. 535 patients who underwent posterior fusion or combined fusion for treatment of ASD were included in each model with a 70-20-10 train-test-validation split. Further analysis was performed using Shapley Additive Explanation (SHAP) values to provide insight into each model's decision-making process. RESULTS: The model's Area Under the Receiver Operating Curve (AUROC) ranged from 0.67 to 0.83. The Random Forest model achieved the highest score. The model considered length of surgery, complications, and estimated blood loss to be the greatest predictors of prolonged ICU stay based on SHAP values. CONCLUSIONS: We developed a ML model that was able to predict whether prolonged ICU stay was required in ASD patients. Further SHAP analysis demonstrated our model aligned with traditional clinical thinking. Thus, ML models have strong potential to assist with risk stratification and more effective and cost-efficient care.

An early sepsis prediction model utilizing machine learning and unbalanced data processing in a clinical context.

Background: Early and accurate diagnoses of sepsis patients are essential to reduce the mortality. However, the sepsis is still diagnosed in a traditional way in China despite the increasing number of related studies, which may to some extent lead to delays in the treatment. Methods: The study included 2,385 patients, including 364 with sepsis, collected from the First Affiliated Hospital of Anhui Medical University and partner hospitals from April to July 2022. External validation was conducted using the MIMIC-III database (over 60,000 patients from 2001 to 2012) and the eICU Collaborative Research Database (139,000 patients from 2014 to 2015). Multiple algorithm models, along with the SHapley Additive exPlanations (SHAP) analysis, are applied to explore the main risk factors for the accurate prediction of the sepsis. Multiple Imputations for filling missing data and the Synthetic Minority Oversampling (SMOTE) balancing method for balancing data are used for the data processing. Result: Eighteen diagnostic features are used in the predictive model for early sepsis. The Random Forest model has the best performance among all the models, with an Area Under the Curve (AUC) of 87% and an F1-score (F1) of 77%. Moreover, the interpretation from the SHAP analysis is generally consistent with the current clinical situation. Conclusion: The study revealed the relationship between these 18 clinical features and diagnostic outcomes. The results indicate that patients with laboratory values of Systolic Blood Pressure, Albumin, and Heart Rate exceeding certain thresholds are at a high likelihood of developing sepsis.

Manifold-based Shapley explanations for high dimensional correlated features.

Explainable artificial intelligence (XAI) holds significant importance in enhancing the reliability and transparency of network decision-making. SHapley Additive exPlanations (SHAP) is a game-theoretic approach for network interpretation, attributing confidence to inputs features to measure their importance. However, SHAP often relies on a flawed assumption that the model's features are independent, leading to incorrect results when dealing with correlated features. In this paper, we introduce a novel manifold-based Shapley explanation method, termed Latent SHAP. Latent SHAP transforms high-dimensional data into low-dimensional manifolds to capture correlations among features. We compute Shapley values on the data manifold and devise three distinct gradient-based mapping methods to transfer them back to the high-dimensional space. Our primary objectives include: (1) correcting misinterpretations by SHAP in certain samples; (2) addressing the challenge of feature correlations in high-dimensional data interpretation; and (3) reducing algorithmic complexity through Manifold SHAP for application in complex network interpretations. Code is available at https://github.com/Teriri1999/Latent-SHAP.

Exploring high-emission driving behaviors of heavy-duty diesel vehicles based on engine principles under different road grade levels.

To reveal the outstanding high-emission problems that occur when heavy-duty diesel vehicles (HDDV) pass uphill and downhill, this study proposes a method to depict the nitrogen oxides (NOx) and carbon dioxide (CO2) high-emission driving behaviors caused by slopes from the perspective of engine principles. By calculating emission and grade data of HDDV based on on-board diagnostic (OBD) data and digital elevation model (DEM) data, the 262 short trips including uphill, flat-road and downhill are firstly obtained through the rule-based short trip segmentation method, and the significant correlation between the road grade and emissions of the short trips is verified by Kendall's Tau and K-means clustering. Secondly, by comparing the distribution changes of three speed categories (acceleration state, constant speed state and deceleration state), the differences in HDDV operating states under different grade levels are discussed. Finally, the machine learning models (Random Forest, XGBoost and Elastic Net), are used to develop the NOx and CO2 emission estimation model, identifying high-emission driving behaviors, particularly during uphill driving, which showed the highest proportion of high-emission. Explained by the feature importance and SHapley Additive exPlanations (SHAP) model that large accelerator pedal opening, frequent aggressive acceleration, and high engine load have positive effects both on NOx and CO2 emissions. The difference is in the air-fuel ratio that the engine in the rich or slightly lean burning state will increase CO2 emissions and the lean burning state will increase NOx emissions. In addition, due to the uncertainty of the actual uphill, drivers often undergo a rapid "deceleration-uniform-acceleration" process, which significantly contributes to high NOx and CO2 emissions from the engine perspective. The findings provide insights for designing driving strategies in slope scenarios and offer a novel perspective on depicting driving behaviors.

Cognitive Capacities as Functional Natural Kinds.

In this paper, we articulate a functional approach to cognitive capacities. It is a restricted functionalism for various reasons, but especially because it does not claim that all cognitive (and/or mental) entities and processes are functional in the sense of a systemic capacities approach. One of the central aims of a cognitive theory consists in providing explanations of behavioral phenomena of (human and non-human) animals, and of the phenomena that are involved in those explanations. We accept that part of what lies at the heart of these explanations are certain functional entities -we call them "cognitive functional systems" -which in our view stand for most of the cognitive capacities of an organism; that is, systems that are individuated primarily by the main cognitive functions they undertake. Additionally, in the paper, we go into further detail concerning these functional systems, their internal organization, the nature of their causal interactions, etc. We also argue that some of these classes of cognitive functional systems (i.e., cognitive capacities) can be construed as "natural kinds" whenever their kinds of functional organizations are understood as kinds of hierarchically ordered classes of information processing events that are related among each other in regular (often complex) ways.

Spatio-temporal learning and explaining for dynamic functional connectivity analysis: Application to depression.

BACKGROUND: Functional connectivity has been shown to fluctuate over time. The present study aimed to identifying major depressive disorders (MDD) with dynamic functional connectivity (dFC) from resting-state fMRI data, which would be helpful to produce tools of early depression diagnosis and enhance our understanding of depressive etiology. METHODS: The resting-state fMRI data of 178 subjects were collected, including 89 MDD and 89 healthy controls. We propose a spatio-temporal learning and explaining framework for dFC analysis. A yet effective spatio-temporal model is developed to classifying MDD from healthy controls with dFCs. The model is a stacking neural network model, which learns network structure information by a multi-layer perceptron based spatial encoder, and learns time-varying patterns by a Transformer based temporal encoder. We propose to explain the spatio-temporal model with a two-stage explanation method of importance feature extracting and disorder-relevant pattern exploring. The layer-wise relevance propagation (LRP) method is introduced to extract the most relevant input features in the model, and the attention mechanism with LRP is applied to extract the important time steps of dFCs. The disorder-relevant functional connections, brain regions, and brain states in the model are further explored and identified. RESULTS: We achieved the best classification performance in identifying MDD from healthy controls with dFC data. The top important functional connectivity, brain regions, and dynamic states closely related to MDD have been identified. LIMITATIONS: The data preprocessing may affect the classification performance of the model, and this study needs further validation in a larger patient population. CONCLUSIONS: The experimental results demonstrate that the proposed spatio-temporal model could effectively classify MDD, and uncover structural and temporal patterns of dFCs in depression.

Chemistry-informed Machine Learning Explains Calcium-binding Proteins' Fuzzy Shape for Communicating Changes in the Atomic States of Calcium Ions.

Proteins' fuzziness are features for communicating changes in cell signaling instigated by binding with secondary messengers, such as calcium ions, associated with the coordination of muscle contraction, neurotransmitter release, and gene expression. Binding with the disordered parts of a protein, calcium ions must balance their charge states with the shape of calcium-binding proteins and their versatile pool of partners depending on the circumstances they transmit, but it is unclear whether the limited experimental data available can be used to train models to accurately predict the charges of calcium-binding protein variants. Here, we developed a chemistry-informed, machine-learning algorithm that implements a game theoretic approach to explain the output of a machine-learning model without the prerequisite of an excessively large database for high-performance prediction of atomic charges. We used the ab initio electronic structure data representing calcium ions and the structures of the disordered segments of calcium-binding peptides with surrounding water molecules to train several explainable models. Network theory was used to extract the topological features of atomic interactions in the structurally complex data dictated by the coordination chemistry of a calcium ion, a potent indicator of its charge state in protein. With our designs, we provided a framework of explainable machine learning model to annotate atomic charges of calcium ions in calcium-binding proteins with domain knowledge in response to the chemical changes in an environment based on the limited size of scientific data in a genome space.

XAI-TRIS: non-linear image benchmarks to quantify false positive post-hoc attribution of feature importance.

The field of 'explainable' artificial intelligence (XAI) has produced highly acclaimed methods that seek to make the decisions of complex machine learning (ML) methods 'understandable' to humans, for example by attributing 'importance' scores to input features. Yet, a lack of formal underpinning leaves it unclear as to what conclusions can safely be drawn from the results of a given XAI method and has also so far hindered the theoretical verification and empirical validation of XAI methods. This means that challenging non-linear problems, typically solved by deep neural networks, presently lack appropriate remedies. Here, we craft benchmark datasets for one linear and three different non-linear classification scenarios, in which the important class-conditional features are known by design, serving as ground truth explanations. Using novel quantitative metrics, we benchmark the explanation performance of a wide set of XAI methods across three deep learning model architectures. We show that popular XAI methods are often unable to significantly outperform random performance baselines and edge detection methods, attributing false-positive importance to features with no statistical relationship to the prediction target rather than truly important features. Moreover, we demonstrate that explanations derived from different model architectures can be vastly different; thus, prone to misinterpretation even under controlled conditions.

Inside Ockham's razor: A mechanism driving preferences for simpler explanations.

People often prefer simpler explanations, defined as those that posit the presence of fewer causes (e.g., positing the presence of a single cause, Cause A, rather than two causes, Causes B and C, to explain observed effects). Here, we test one hypothesis about the mechanisms underlying this preference: that people tend to reason as if they are using "agnostic" explanations, which remain neutral about the presence/absence of additional causes (e.g., comparing "A" vs. "B and C," while remaining neutral about the status of B and C when considering "A," or of A when considering "B and C"), even in cases where "atheist" explanations, which specify the absence of additional causes (e.g., "A and not B or C" vs. "B and C and not A"), are more appropriate. Three studies with US-based samples (total N = 982) tested this idea by using scenarios for which agnostic and atheist strategies produce diverging simplicity/complexity preferences, and asking participants to compare explanations provided in atheist form. Results suggest that people tend to ignore absent causes, thus overgeneralizing agnostic strategies, which can produce preferences for simpler explanations even when the complex explanation is objectively more probable. However, these unwarranted preferences were reduced by manipulations that encouraged participants to consider absent causes: making absences necessary to produce the effects (Study 2), or describing absences as causes that produce alternative effects (Study 3). These results shed light on the mechanisms driving preferences for simpler explanations, and on when these mechanisms are likely to lead people astray.

Development and validation of an interpretable machine learning model for predicting post-stroke epilepsy.

BACKGROUND: Epilepsy is a serious complication after an ischemic stroke. Although two studies have developed prediction model for post-stroke epilepsy (PSE), their accuracy remains insufficient, and their applicability to different populations is uncertain. With the rapid advancement of computer technology, machine learning (ML) offers new opportunities for creating more accurate prediction models. However, the potential of ML in predicting PSE is still not well understood. The purpose of this study was to develop prediction models for PSE among ischemic stroke patients. METHODS: Patients with ischemic stroke from two stroke centers were included in this retrospective cohort study. At the baseline level, 33 input variables were considered candidate features. The 2-year PSE prediction models in the derivation cohort were built using six ML algorithms. The predictive performance of these machine learning models required further appraisal and comparison with the reference model using the conventional triage classification information. The Shapley additive explanation (SHAP), based on fair profit allocation among many stakeholders according to their contributions, is used to interpret the predicted outcomes of the naive Bayes (NB) model. RESULTS: A total of 1977 patients were included to build the predictive model for PSE. The Boruta method identified NIHSS score, hospital length of stay, D-dimer level, and cortical involvement as the optimal features, with the receiver operating characteristic curves ranging from 0.709 to 0.849. An additional 870 patients were used to validate the ML and reference models. The NB model achieved the best performance among the PSE prediction models with an area under the receiver operating curve of 0.757. At the 20 % absolute risk threshold, the NB model also provided a sensitivity of 0.739 and a specificity of 0.720. The reference model had poor sensitivities of only 0.15 despite achieving a helpful AUC of 0.732. Furthermore, the SHAP method analysis demonstrated that a higher NIHSS score, longer hospital length of stay, higher D-dimer level, and cortical involvement were positive predictors of epilepsy after ischemic stroke. CONCLUSIONS: Our study confirmed the feasibility of applying the ML method to use easy-to-obtain variables for accurate prediction of PSE and provided improved strategies and effective resource allocation for high-risk patients. In addition, the SHAP method could improve model transparency and make it easier for clinicians to grasp the prediction model's reliability.

Subjective Understanding is Reduced by Mechanistic Framing.

People often believe that they have a good understanding of how devices work (e.g., how a ballpoint pen works), despite having poor knowledge of their internal mechanics. We hypothesized that this bias occurs in part because people conflate mechanistic understanding with functional understanding of how devices work (e.g., how to operate a ballpoint pen). In two experiments, we found that increasing the salience of mechanistic information led to lower judgments of understanding for how devices work. In Experiment 1, we did this by showing participants either the internal parts of a device or an external, whole-object view of that same device. Those who saw the internal parts rated their understanding as less than those who saw a whole-object view. In Experiment 2, we removed the pictures and instead tested participants (without feedback) on their mechanistic or functional knowledge using true-or-false questions. Those who were tested on mechanistic knowledge rated their understanding of devices as less than those who were tested on functional knowledge.

Interpretable machine learning models for predicting clinical pregnancies associated with surgical sperm retrieval from testes of different etiologies: a retrospective study.

BACKGROUND: The relationship between surgical sperm retrieval of different etiologies and clinical pregnancy is unclear. We aimed to develop a robust and interpretable machine learning (ML) model for predicting clinical pregnancy using the SHapley Additive exPlanation (SHAP) association of surgical sperm retrieval from testes of different etiologies. METHODS: A total of 345 infertile couples who underwent intracytoplasmic sperm injection (ICSI) treatment with surgical sperm retrieval due to different etiologies from February 2020 to March 2023 at the reproductive center were retrospectively analyzed. The six machine learning (ML) models were used to predict the clinical pregnancy of ICSI. After evaluating the performance characteristics of the six ML models, the Extreme Gradient Boosting model (XGBoost) was selected as the best model, and SHAP was utilized to interpret the XGBoost model for predicting clinical pregnancies and to reveal the decision-making process of the model. RESULTS: Combining the area under the receiver operating characteristic curve (AUROC), accuracy, precision, recall, F1 score, brier score, and the area under the precision-recall (P-R) curve (AP), the XGBoost model has the best performance (AUROC: 0.858, 95% confidence interval (CI): 0.778-0.936, accuracy: 79.71%, brier score: 0.151). The global summary plot of SHAP values shows that the female age is the most important feature influencing the model output. The SHAP plot showed that younger age in females, bigger testicular volume (TV), non-tobacco use, higher anti-müllerian hormone (AMH), lower follicle-stimulating hormone (FSH) in females, lower FSH in males, the temporary ejaculatory disorders (TED) group, and not the non-obstructive azoospermia (NOA) group all resulted in an increased probability of clinical pregnancy. CONCLUSIONS: The XGBoost model predicts clinical pregnancies associated with testicular sperm retrieval of different etiologies with high accuracy, reliability, and robustness. It can provide clinical counseling decisions for patients with surgical sperm retrieval of various etiologies.

A stepwise prediction and interpretation of gestational diabetes mellitus: Foster the practical application of machine learning in clinical decision.

Background: Machine learning has shown to be an effective method for early prediction and intervention of Gestational diabetes mellitus (GDM), which greatly decreases GDM incidence, reduces maternal and infant complications and improves the prognosis. However, there is still much room for improvement in data quality, feature dimension, and accuracy. The contributions and mechanism explanations of clinical data at different pregnancy stages to the prediction accuracy are still lacking. More importantly, current models still face notable obstacles in practical applications due to the complex and diverse input features and difficulties in redeployment. As a result, a simple, practical but accurate enough model is urgently needed. Design and methods: In this study, 2309 samples from two public hospitals in Shenzhen, China were collected for analysis. Different algorithms were systematically compared to build a robust and stepwise prediction system (level A to C) based on advanced machine learning, and models under different levels were interpreted. Results: XGBoost reported the best performance with ACC of 0.922, 0.859 and 0.850, AUC of 0.974, 0.924 and 0.913 for the selected level A to C models in the test set, respectively. Tree-based feature importance and SHAP method successfully identified the commonly recognized risk factors, while indicated new inconsistent impact trends for GDM in different stages of pregnancy. Conclusion: A stepwise prediction system was successfully established. A practical tool that enables a quick prediction of GDM was released at https://github.com/ifyoungnet/MedGDM.This study is expected to provide a more detailed profiling of GDM risk and lay the foundation for the application of the model in practice.