Architecture Search Research Articles

RobMedNAS: searching robust neural network architectures for medical image synthesis

Investigating U-Net model robustness in medical image synthesis against adversarial perturbations, this study introduces RobMedNAS, a neural architecture search strategy for identifying resilient U-Net configurations. Through retrospective analysis of synthesized CT from MRI data, employing Dice coefficient and mean absolute error metrics across critical anatomical areas, the study evaluates traditional U-Net models and RobMedNAS-optimized models under adversarial attacks. Findings demonstrate RobMedNAS’s efficacy in enhancing U-Net resilience without compromising on accuracy, proposing a novel pathway for robust medical image processing.

Integrating automated machine learning and metabolic reprogramming for the identification of microplastic in soil: A case study on soybean

The accumulation of polyethylene microplastic (PE-MPs) in soil can significantly impact plant quality and yield, as well as affect human health and food chain cycles. Therefore, developing rapid and effective detection methods is crucial. In this study, traditional machine learning (ML) and H2O automated machine learning (H2O AutoML) were utilized to offer a powerful framework for detecting PE-MPs (0.1 %, 1 %, and 2 % by dry soil weight) and the co-contamination of PE-MPs and fomesafen (a common herbicide) in soil. The development of the framework was based on the results of the metabolic reprogramming of soybean plants. Our study stated that traditional ML exhibits lower accuracy due to the challenges associated with optimizing complex parameters. H2O AutoML can accurately distinguish between clean soil and contaminated soil. Notably, H2O AutoML can detect PE-MPs as low as 0.1 % (with 100 % accuracy) and co-contamination of PE-MPs and fomesafen (with 90 % accuracy) in soil. The VIP and SHAP analyses of the H2O AutoML showed that PE-MPs and the co-contamination of PE-MPs and fomesafen significantly interfered with the antioxidant system and energy regulation of soybean. We hope this study can provide a reliable scientific basis for sustainable development of the environment.

Zero-Shot Proxy with Incorporated-Score for Lightweight Deep Neural Architecture Search

Zero-Shot Proxy with Incorporated-Score for Lightweight Deep Neural Architecture Search

Reassessing the management of uncomplicated urinary tract infection: A retrospective analysis using machine learning causal inference.

Uncomplicated urinary tract infection (UTI) is a common indication for outpatient antimicrobial therapy. National guidelines for the management of uncomplicated UTI were published by the Infectious Diseases Society of America in 2011, however it is not fully known the extent to which they align with current practices, patient diversity, and pathogen biology, all of which have evolved significantly in the time since their publication. We aimed to re-evaluate efficacy and adverse events for first-line antibiotics (nitrofurantoin, and trimethoprim-sulfamethoxazole), versus second-line antibiotics (fluoroquinolones) and versus alternative agents (oral β-lactams) for uncomplicated UTI in contemporary clinical practice by applying machine learning algorithms to a large claims database formatted into the Observational Medical Outcomes Partnership (OMOP) common data model. Our primary outcome was a composite endpoint for treatment failure, defined as outpatient or inpatient re-visit within 30 days for UTI, pyelonephritis or sepsis. Secondary outcomes were the risk of 4 common antibiotic-associated adverse events: gastrointestinal symptoms, rash, kidney injury and C. difficile infection. We adjusted for covariate-dependent censoring and treatment indication using a broad set of domain-expert derived features. Sensitivity analyses were conducted using OMOP-learn, an automated feature engineering package for OMOP datasets. Our study included 57,585 episodes of UTI from 49,037 patients. First-line antibiotics were prescribed in 35,018 (61%) episodes, second-line antibiotics were prescribed in 21,140 (37%) episodes and alternative antibiotics were prescribed in 1,427 (2%) episodes. After adjustment, patients receiving first-line therapies had an absolute risk difference of -2.1% [95% CI -2.9% to -1.6%] for having a revisit for UTI within 30 days of diagnosis relative to second-line antibiotics. First-line therapies had an absolute risk difference of -6.6% [95% CI -9.4% to -3.8%] for 30-day revisit compared to alternative β-lactam antibiotics. Differences in adverse events were clinically similar between first and second line agents, but lower for first-line agents relative to alternative antibiotics (-3.5% [95% CI -5.9% to -1.2%]). Results were similar for models built with OMOPlearn. Our study provides support for the continued use of first-line antibiotics for the management of uncomplicated UTI. Our results also provide proof-of-principle that automated feature extraction methods for OMOP formatted data can emulate manually curated models, thereby promoting reproducibility and generalizability.

Machine learning‐driven implementation of workflow optimization in cloud computing for IoT applications

AbstractThe optimization of workflow scheduling in Internet of Things (IoT) environments presents significant challenges due to the dynamic and heterogeneous nature of these systems. Traditional techniques must often adapt to fluctuating network conditions and varying data loads. To address these limitations, we propose a novel approach that leverages Automated Machine Learning (AutoML) integrated with cloud computing to optimize workflow scheduling for IoT applications. Our solution automates machine learning model selection, training, and tuning, significantly enhancing computational efficiency and adaptability. Through extensive experimentation, we demonstrate that our AutoML‐driven approach surpasses conventional algorithms across several key metrics, including accuracy, computational efficiency, adaptability to dynamic environments, and communication efficiency. Specifically, our method achieves a scheduling accuracy improvement of up to 25%, a reduced computational overhead by 30%, and a 40% enhancement in adaptability under dynamic conditions. Furthermore, the scalability of our solution is critical in cloud computing contexts, enabling efficient handling of large‐scale IoT deployments by leveraging cloud resources for distributed processing. This scalability ensures that our approach can effectively manage increasing data volumes and device heterogeneity inherent in modern IoT systems.

A physics-guided automated machine learning approach for obtaining surface radiometric temperatures on sunny days based on UAV-derived images

Urban surface radiometric temperatures, approximate to the surface kinetic temperatures, are predominantly retrieved using satellites or unmanned aerial vehicles (UAVs) and exhibit pronounced spatiotemporal variations. Despite numerous methods ranging from empirical to physical models for obtaining urban microscale surface radiometric temperatures via UAVs, challenges remain given the limited physical significance and substantial professional barriers to method application. Against this background, this study introduces a novel and straightforward approach for acquiring spatially distributed radiometric temperatures on sunny days without understanding the complex radiative transfer process as well as acquiring low-altitude atmospheric parameters. An automated machine learning was employed to train a model capable of efficiently estimating radiometric temperatures. Training and testing datasets were created based on the urban radiative transfer equation, incorporating three independent variables: UAV-measured surface brightness temperature, broadband emissivity, and sky view factor, which collectively represent the diverse thermal environments across different surface characteristics and urban layouts during sunny transitional and summer seasons. The model's accuracy was subsequently confirmed through direct comparisons with radiometric temperatures retrieved from UAV-collected multimodal images and kinetic temperatures synchronously collected on the ground across four periods. The results indicate that AutoGluon achieved high accuracy (MAE: 0.04 K; RMSE: 0.06 K; R2: 0.99). Additional ground measurement validations further demonstrated the model's reliability, with absolute biases on sunlit surfaces maintained within 1.25 K. Given its capability for real-time, high-spatial-resolution mapping of radiometric temperatures (April test: 8.70 cm, July test: 6.89 cm) in urban microscales with considerable heterogeneity, such a method is envisioned to be an effective tool for the dynamic monitoring and management of thermal environments at the microscale level in urban settings.

Turbofan engine health status prediction with neural network pattern recognition and automated feature engineering

PurposeThis study aims to present the concept of aircraft turbofan engine health status prediction with artificial neural network (ANN) pattern recognition but augmented with automated features engineering (AFE).Design/methodology/approachThe main concept of engine health status prediction was based on three case studies and a validation process. The first two were performed on the engine health status parameters, namely, performance margin and specific fuel consumption margin. The third one was generated and created for the engine performance and safety data, specifically created for the final test. The final validation of the neural network pattern recognition was the validation of the proposed neural network architecture in comparison to the machine learning classification algorithms. All studies were conducted for ANN, which was a two-layer feedforward network architecture with pattern recognition. All case studies and tests were performed for both simple pattern recognition network and network augmented with automated feature engineering (AFE).FindingsThe greatest achievement of this elaboration is the presentation of how on the basis of the real-life engine operational data, the entire process of engine status prediction might be conducted with the application of the neural network pattern recognition process augmented with AFE.Practical implicationsThis research could be implemented into the engine maintenance strategy and planning. Engine health status prediction based on ANN augmented with AFE is an extremely strong tool in aircraft accident and incident prevention.Originality/valueAlthough turbofan engine health status prediction with ANN is not a novel approach, what is absolutely worth emphasizing is the fact that contrary to other publications this research was based on genuine, real engine performance operational data as well as AFE methodology, which makes the entire research very reliable. This is also the reason the prediction results reflect the effect of the real engine wear and deterioration process.

Quantum Architecture Search with Neural Predictor Based on Graph Measures

AbstractQuantum architecture search (QAS) has attracted increasing attention owing to its remarkable ability to automate the design of quantum circuits for variational quantum algorithms (VQAs). However, evaluating the performance of numerous quantum circuits is essential to provide feedback for the search strategy, which inevitably renders QAS computationally expensive. Performance predictors have emerged as highly efficient evaluation methods to mitigate this challenge. However, the performance predictor faces a critical challenge in reducing the required number of circuit‐performance pairs for training. This study encodes circuit architecture by representing a quantum circuit as a relational graph that emphasizes message exchange. Subsequently, valuable information about circuit architecture is extracted through three types of graph measures, including distance‐based, degree‐based, and cluster‐based measures. The graph measures define a smooth space related to circuit performance, facilitating the training of the performance predictor. The effectiveness of the proposed method is assessed across three tasks within variational quantum eigensolvers (VQE): identifying the ground states of the Transverse Field Ising Model (TFIM), the Heisenberg model, and the molecule. The simulation results demonstrate notable enhancements in predictive accuracy achieved by our method, coupled with a substantial reduction in the required number of training samples for the predictor.

Automatic machine learning model for enhanced partition and identification of breast disorders in breast MRI scan

ABSTRACT The rapid identification and categorisation of breast cancers using low-contrast MRI images presents a significant challenge due to the disease’s prevalence among women of all ages. Nowadays, it is very difficult to classify and generalise models from low-contrast MRI datasets due to the prevalence of class imbalance caused by a wide range of symptoms and untrustworthy data sources. Using ensemble learning with optimised k-means helps with these problems; it reduces overfitting and improves reliability by using methods like k-means clustering, frog feap algorithm (FLA), boosting, and bagging to balance datasets and show the complexity of symptoms. In this research, we present a new strategy that makes use of a hybrid optimised K-Means algorithm combined with the FLA, along with thresholding and morphological approaches, and ensemble learning. There are two main stages to our work. In the initial phase, we preprocess low-contrast MRI images and delineate the breast tumour area from the segmented MRI images. Furthermore, we employ Discrete Wavelet Transformation (DWT) to extract the most salient features. In the next step, the retrieved features will be used as input parameters for an ensemble-learned breast tumour classifier. To determine whether a low-contrast breast tumour MRI picture has a benign or malignant tumour after model training, we use a variety of classifiers, such as k-nearest neighbours (KNN), decision trees (DT), gradient boosting (GB), random forests (RF), and artificial neural networks (ANN). Through a voting mechanism that averages the accuracy of all models, our suggested framework achieves a remarkable 98.8% accuracy rate. In addition, proposed model is quite efficient & reliable to detect the kind of breast tumour or cancer with a sensitivity of 98.02% and a specificity of 97.5%.

AI-Based Visual Early Warning System

Facial expressions are a universally recognised means of conveying internal emotional states across diverse human cultural and ethnic groups. Recent advances in understanding people’s emotions expressed through verbal and non-verbal communication are particularly noteworthy in the clinical context for the assessment of patients’ health and well-being. Facial expression recognition (FER) plays an important and vital role in health care, providing communication with a patient’s feelings and allowing the assessment and monitoring of mental and physical health conditions. This paper shows that automatic machine learning methods can predict health deterioration accurately and robustly, independent of human subjective assessment. The prior work of this paper is to discover the early signs of deteriorating health that align with the principles of preventive reactions, improving health outcomes and human survival, and promoting overall health and well-being. Therefore, methods are developed to create a facial database mimicking the underlying muscular structure of the face, whose Action Unit motions can then be transferred to human face images, thus displaying animated expressions of interest. Then, building and developing an automatic system based on convolution neural networks (CNN) and long short-term memory (LSTM) to recognise patterns of facial expressions with a focus on patients at risk of deterioration in hospital wards. This research presents state-of-the-art results on generating and modelling synthetic database and automated deterioration prediction through FEs with 99.89% accuracy. The main contributions to knowledge from this paper can be summarized as (1) the generation of visual datasets mimicking real-life samples of facial expressions indicating health deterioration, (2) improvement of the understanding and communication with patients at risk of deterioration through facial expression analysis, and (3) development of a state-of-the-art model to recognize such facial expressions using a ConvLSTM model.

Gendered Extremism in the Pacific on 4chan: A Mixed-methods Exploration of Australian and New Zealanders’ Concepts of Women, Gender, and Sexual Violence on /Pol/

ABSTRACT The association between 4chan and online extremist subcultures has seen increasing academic scrutiny—particularly following the 2019 Christchurch attack by a right-wing terrorist who frequented the anonymous forums. Gender-based extremism features as one (of many) critical subcultures that commands our academic attention, though few studies to date have sought to capture and assess the entire landscape of this phenomenon on 4chan’s most notorious board: /pol/. Drawing on a pre-Christchurch attack dataset extracted from Papasavva et al. (2020), this study investigates how Australians and New Zealanders (ANZ) broadly conceptualize—and debate—women, gender, and sexual violence on 4chan’s /pol/ board. We apply a mixed-methods approach, combining automated machine learning tools alongside expert qualitative analysis. Across nearly 300,000 posts and comments, we show how gender is constructed within this community, and the conjugal order they demand as a result. This order racially and sexually defines gender identities and norms, which are perceived as mechanisms to restore power and dominance to an ethnically and ideologically conforming in-group. Those that violate or disrupt the conjugal order are legitimized as targets of sexual, and gender-based violence. This normalizes far/extreme right gendered constructs across ANZ contexts in support of exclusivist far/extreme right ideological positions.

Evaluating the impact of nonverbal behavior on language ability ratings

Nonverbal behavior can impact language proficiency scores in speaking tests, but there is little empirical information of the size or consistency of its effects or whether language proficiency may be a moderating variable. In this study, 100 novice raters watched and scored 30 recordings of test takers taking an international, high stakes proficiency test. The speech samples were each 2 minutes long and ranged in proficiency levels. The raters scored each sample on fluency, vocabulary, grammar, and comprehensibility using 7-point semantic differential scales. Nonverbal behavior was extracted using an automated machine learning software called iMotions, and data was analyzed with ordinal mixed effects regression. Results showed that attentional variance predicted fluency, vocabulary, and grammar scores, but only when accounting for proficiency. Higher standard deviations of attention corresponded with lower scores for the lower-proficiency group, but not the mid/higher-proficiency group. Comprehensibility scores were only predicted by mean valence when proficiency was an interaction term. Higher mean valence, or positive emotional behavior, corresponded with higher scores in the lower-proficiency group, but not the mid/higher-proficiency group. Effect sizes for these predictors were quite small, with small amounts of variance explained. These results have implications for construct representation and test fairness.

Automated external defibrillator location selection considering myocardial infarction risk and medical resources

AbstractThe effective deployment of medical emergency equipment, such as automated external defibrillator (AED), is essential to myocardial infarction (MI) patients. However, there are shortcomings in current studies that simultaneously consider the risk of MI and the availability of medical resources when siting the AEDs. In this study, an AED site recommendation framework was proposed to address the lack of consideration for both the MI risk and medical resources when siting the AEDs. It conducts the AED sitting under different scenarios considering the spatial distribution of MI risk and healthcare accessibility in Chinese cities. First, an automated machine learning framework data is proposed to estimate the MI risk at the community scale based on multi‐source spatio‐temporal. Second, the accessibility of medical resources was calculated by an improved Gaussian two‐step moving search algorithm. Finally, the AED siting in multiple scenarios is conducted based on the coverage model. The performance of the AED siting model was evaluated at Wuhan city. The results show that MI risk is impacted by both socioeconomic and cultural characteristics (municipal utilities, streetscape environment, educational and commercial facilities). There is a strong spatial heterogeneity in the distribution of both MI risk and medical resources in Wuhan, and an unreasonable match between the two was detected in some regions. Medical resources need to be strengthened in some high‐risk areas, such as rural areas and tourist attractions. In addition, 1015 AED candidate sites were identified by the location set covering problem model, with a 15‐min accessibility rate of 96.5%. Given the limited resources, mobile AEDs which have about 15‐min service range can be deployed based on the maximum covering location problem model to meet the demand in central urban areas efficiently. This study can contribute to more rational selection of AED sites and the prevention of myocardial infarction among residents, particularly when supported by policies that promote balanced regional development of pre‐hospital medical emergency networks.

A quantum information theoretic analysis of reinforcement learning-assisted quantum architecture search

In the field of quantum computing, variational quantum algorithms (VQAs) represent a pivotal category of quantum solutions across a broad spectrum of applications. These algorithms demonstrate significant potential for realising quantum computational advantage. A fundamental aspect of VQAs involves formulating expressive and efficient quantum circuits (namely ansatz), and automating the search of such ansatz is known as quantum architecture search (QAS). Recently reinforcement learning (RL) techniques is utilized to automate the search for ansatzes, know as RL-QAS. This study investigates RL-QAS for crafting ansatz tailored to the variational quantum state diagonalisation problem. Our investigation includes a comprehensive analysis of various dimensions, such as the entanglement thresholds of the resultant states, the impact of initial conditions on the performance of RL-agent, the phase transition behaviour of correlation in concurrence bounds, and the discrete contributions of qubits in deducing eigenvalues through conditional entropy metrics. We leverage these insights to devise an entanglement-guided admissible ansatz in QAS to diagonalise random quantum states using optimal resources. Furthermore, the methodologies presented herein offer a generalised framework for constructing reward functions within RL-QAS applicable to variational quantum algorithms.

Automated sex and age partitioning for the estimation of reference intervals using a regression tree model

Abstract Objectives Reference intervals (RI) play a decisive role in the interpretation of medical laboratory results. An important step in the determination of RI is age- and sex specific partitioning, which is usually based on an empirical approach by graphical representation. In this study, we evaluate an automated machine learning approach. Methods This study uses pediatric data from the CALIPER RI (Canadian Laboratory Initiative on Pediatric Reference Intervals) study. The calculation of potential partitions is carried out using a regression tree model included in the rpart package of the statistical programming language R. The Harris & Boyd method is used to compare the corresponding partitions suggested by rpart and CALIPER. For better comparability, the reference ranges of the partitions of both approaches are then calculated using reflimR. Results Most of the partitions suggested by rpart or CALIPER show sufficient heterogeneity among themselves to justify age- and/or sex-specific RI partitioning. With only few individual exceptions, both methods yield comparable results. The partitions of both approaches for albumin and γ-glutamyltransferase are very similar to each other. For creatinine rpart suggests a slightly earlier distinction between the sexes. Alkaline phosphatase shows the most pronounced differences. In addition to a considerable earlier sex split, rpart suggests different age intervals for both sexes, resulting in three partitions for females and four partitions for males. Conclusions Our findings indicate that the automated analysis provided by rpart yields results that comparable to traditional methods. Nevertheless, the medical plausibility of the automatic suggestions needs to be validated by human experts.

Depth-adaptive graph neural architecture search for graph classification

In recent years, graph neural networks (GNNs) based on neighborhood aggregation schemes have become a promising method in various graph-based applications. To solve the expert-dependent and time-consuming problem in human-designed GNN architectures, graph neural architecture search (GNAS) has been popular. However, as the mainstream GNAS methods automatically design GNN architectures with fixed GNN depth, they cannot mine the true potential of GNN architectures for graph classification. Although a few GNAS methods have explored the importance of adaptive GNN depth based on fixed GNN architectures, they have not designed a general search space for graph classification, which limits the discovery of excellent GNN architectures. In this paper, we propose Depth-Adaptive Graph Neural Architecture Search for Graph Classification (DAGC), which systemically constructs and explores the search space for graph classification, rather than studying individual designs. Through decoupling the graph classification process, DAGC proposes a complete and flexible search space, including GNN depth, aggregation function, and pooling operation components. To this end, DAGC adopts a learnable agent based on reinforcement learning to effectively guide the search for depth-adaptive GNN architectures. Extensive experiments on five real-world datasets demonstrate that DAGC outperforms the state-of-the-art human-designed GNN architectures and GNAS methods. The code is available at: https://github.com/Zhen-Peng-Wu/DAGC.

PSO-based lightweight neural architecture search for object detection

In recent years, Neural Architecture Search (NAS) has received widespread attention for its remarkable ability to design neural networks. However, existing NAS works have mainly focused on network search, with limited emphasis on downstream applications after discovering efficient neural networks. In this paper, we propose a lightweight search strategy based on the particle swarm optimization algorithm and apply the searched network as backbone for object detection tasks. Specifically, we design a lightweight search space based on Ghostconv modules and improved Mobileblocks, achieving comprehensive exploration within the search space using variable-length encoding strategy. During the search process, to balance network performance and resource consumption, we propose a multi-objective fitness function and incorporated the classification accuracy, parameter size, and FLOPs of candidate individuals into optimization. For particle performance evaluation, we propose a new strategy based on weight sharing and dynamic early stopping, significantly accelerating the search process. Finally, we fine-tune the globally optimal particle decoded as the backbone, adding Ghost PAN feature fusion modules and detection heads to build an object detection model, and we achieve a 17.01% mAP on the VisDrone2019 dataset. Experimental results demonstrate the competitiveness of our algorithm in terms of search time and the balance between accuracy and efficiency, and also confirm the effectiveness of object detection models designed through NAS methods.

Efficient predictive modeling of resilient modulus in stabilized clayey soil using automated machine learning

Pavement subgrade design relies on the resilient modulus (Mr) to analyze structural response to vehicle-like loading. Adding stabilizers to the soil subgrade makes estimating Mr difficult and resource-intensive. This study uses an automated machine learning (ML) strategy to predict the Mr of stabilized clayey soil using recycled plastic waste. The proposed method automates model selection and hyperparameter tuning, making it a feasible alternative to tedious ML modeling and costly laboratory testing. From extensive laboratory investigation involving 3285 experimental data points, the automated ML model using Bayesian optimization evaluates ensembles, support vector machine (SVM), neural network (NET), decision trees, (TREE), and Gaussian process (GP) regression models, identifying the best model based on cross-validation mean squared error (MSE). Bayesian optimization explores hyperparameter spaces to find optimal configurations, enhancing the accuracy, scalability, and reliability of the prediction model. The optimization process yielded the best results for the ensemble least square boost (LSBoost) model with a cross-validation mean squared error (MSE) value of 6.723×10−29. The optimized ML model’s performance is measured using R2 and adjusted R2. The LSboost model’s R2 and adjusted R2 values of 0.9999 suggested overfitting, prompting further investigations using performance metrics like root mean squared error (RMSE), mean absolute error (MSE), and probability density function (PDF) for normalized absolute error (NAE) for training and testing datasets for the predictive ML model. The small RMSE and MAE (0.0049 and 0.0005) values and symmetrical NAE distribution of the proposed ML model demonstrate its high accuracy and generalization capabilities. The proposed model was subsequently tested on the new, unseen data and achieved predictions with an error rate of 0.24 %. This confirms the proposed ML model’s superiority over conventional deformation models, making it ideal for reliable geotechnical engineering applications.

Advances in neural architecture search.

Automated machine learning (AutoML) has achieved remarkable success in automating the non-trivial process of designing machine learning models. Among the focal areas of AutoML, neural architecture search (NAS) stands out, aiming to systematically explore the complex architecture space to discover the optimal neural architecture configurations without intensive manual interventions. NAS has demonstrated its capability of dramatic performance improvement across a large number of real-world tasks. The core components in NAS methodologies normally include (i) defining the appropriate search space, (ii) designing the right search strategy and (iii) developing the effective evaluation mechanism. Although early NAS endeavors are characterized via groundbreaking architecture designs, the imposed exorbitant computational demands prompt a shift towards more efficient paradigms such as weight sharing and evaluation estimation, etc. Concurrently, the introduction of specialized benchmarks has paved the way for standardized comparisons of NAS techniques. Notably, the adaptability of NAS is evidenced by its capability of extending to diverse datasets, including graphs, tabular data and videos, etc., each of which requires a tailored configuration. This paper delves into the multifaceted aspects of NAS, elaborating on its recent advances, applications, tools, benchmarks and prospective research directions.

Application Value of the Automated Machine Learning Model Based on Modified Computed Tomography Severity Index Combined With Serological Indicators in the Early Prediction of Severe Acute Pancreatitis.

Machine learning (ML) algorithms are widely applied in building models of medicine due to their powerful studying and generalizing ability. To assess the value of the Modified Computed Tomography Severity Index (MCTSI) combined with serological indicators for early prediction of severe acute pancreatitis (SAP) by automated ML (AutoML). The clinical data, of the patients with acute pancreatitis (AP) hospitalized in Hospital 1 and hospital 2 from January 2017 to December 2021, were retrospectively analyzed. Serological indicators within 24 hours of admission were collected. MCTSI score was completed by noncontrast computed tomography within 24 hours of admission. Data from the hospital 1 were adopted for training, and data from the hospital 2 were adopted for external validation. The diagnosis of AP and SAP was based on the 2012 revised Atlanta classification of AP. Models were built using traditional logistic regression and AutoML analysis with 4 types of algorithms. The performance of models was evaluated by the receiver operating characteristic curve, the calibration curve, and the decision curve analysis based on logistic regression and decision curve analysis, feature importance, SHapley Additive exPlanation Plot, and Local Interpretable Model Agnostic Explanation based on AutoML. A total of 499 patients were used to develop the models in the training data set. An independent data set of 201 patients was used to test the models. The model developed by the Deep Neural Net (DL) outperformed other models with an area under the receiver operating characteristic curve (areas under the curve) of 0.907 in the test set. Furthermore, among these AutoML models, the DL and gradient boosting machine models achieved the highest sensitivity values, both exceeding 0.800. The AutoML model based on the MCTSI score combined with serological indicators has good predictive value for SAP in the early stage.