Large-scale Datasets Research Articles

AI-based image classification algorithms for Leukaemia diagnostics and hematologic cytomorphology: From single cells to molecular features

Due to the progress of image analysis and classification systems in recent years, algorithms have been developed that support morphologic examination of both single cells and tissue samples. These algorithms are typically developed using data-driven strategies, which require comprehensive, large-scale datasets. In the diagnostic workup of hematopoietic malignancies, cytomorphologic examination and differentiation represents a key first step. In recent years, the availability of large-scale, high-quality datasets of single leukocytes from peripheral blood and bone marrow has led to the development of diagnostic support algorithms for this modality. These methods not only allow a faster and more consistent classification of diagnostically relevant cell types, but also pave the way for integrated analysis of cytomorphologic and molecular findings.

Improving Anticancer Drug Selection and Prioritization via Neural Learning to Rank.

Personalized cancer treatment requires a thorough understanding of complex interactions between drugs and cancer cell lines in varying genetic and molecular contexts. To address this, high-throughput screening has been used to generate large-scale drug response data, facilitating data-driven computational models. Such models can capture complex drug-cell line interactions across various contexts in a fully data-driven manner. However, accurately prioritizing the most effective drugs for each cell line still remains a significant challenge. To address this, we developed multiple neural ranking approaches that leverage large-scale drug response data across multiple cell lines from diverse cancer types. Unlike existing approaches that primarily utilize regression and classification techniques for drug response prediction, we formulated the objective of drug selection and prioritization as a drug ranking problem. In this work, we proposed multiple pairwise and listwise neural ranking methods that learn latent representations of drugs and cell lines and then use those representations to score drugs in each cell line via a learnable scoring function. Specifically, we developed neural pairwise and listwise ranking methods, Pair-PushC and List-One on top of the existing methods, pLETORg and ListNet, respectively. Additionally, we proposed a novel listwise ranking method, List-All, that focuses on all the effective drugs instead of the top effective drug, unlike List-One. We also provide an exhaustive empirical evaluation with state-of-the-art regression and ranking baselines on large-scale data sets across multiple experimental settings. Our results demonstrate that our proposed ranking methods mostly outperform the best baselines with significant improvements of as much as 25.6% in terms of selecting truly effective drugs within the top 20 predicted drugs (i.e., hit@20) across 50% test cell lines. Furthermore, our analyses suggest that the learned latent spaces from our proposed methods demonstrate informative clustering structures and capture relevant underlying biological features. Moreover, our comprehensive evaluation provides a thorough and objective comparison of the performance of different methods (including our proposed ones).

Efficient Classification on Remote Sensing Image Using Transfer Learning

With a rapid development in aerial technology, applications of Remote Sensing Images (RSI) have become more diverse. Remote sensing image classification plays a crucial role in analyzing and interpreting Earth observation data for various applications, such as land cover mapping, environmental monitoring, and urban planning. However, accurately classifying remote sensing images poses significant challenges due to their complex spatial and spectral characteristics. In recent years, transfer learning has emerged as a promising technique to improve the classification accuracy by leveraging the knowledge learned from pre-trained models on large-scale datasets. The proposed model explores different transfer learning strategies employed in remote sensing image classification, including fine-tuning, feature extraction, and domain adaptation. It discusses popular pre-trained models, such as VGG16, VGG19, and Inceptionv3, and their applicability to remote sensing datasets. The advantages and limitations of each strategy are analyzed, providing insights into their suitability for various remote sensing applications. A comparative study is done on all these techniques to evaluate the performance measures like Accuracy and Loss.

A retrospective prognostic evaluation using unsupervised learning in the treatment of COVID-19 patients with hypertension treated with ACEI/ARB drugs

Introduction This study aimed to evaluate the prognosis of patients with COVID-19 and hypertension who were treated with angiotensin-converting enzyme inhibitor (ACEI)/angiotensin receptor B (ARB) drugs and to identify key features affecting patient prognosis using an unsupervised learning method. Methods A large-scale clinical dataset, including patient information, medical history, and laboratory test results, was collected. Two hundred patients with COVID-19 and hypertension were included. After cluster analysis, patients were divided into good and poor prognosis groups. The unsupervised learning method was used to evaluate clinical characteristics and prognosis, and patients were divided into different prognosis groups. The improved wild dog optimization algorithm (IDOA) was used for feature selection and cluster analysis, followed by the IDOA-k-means algorithm. The impact of ACEI/ARB drugs on patient prognosis and key characteristics affecting patient prognosis were also analysed. Results Key features related to prognosis included baseline information and laboratory test results, while clinical symptoms and imaging results had low predictive power. The top six important features were age, hypertension grade, MuLBSTA, ACEI/ARB, NT-proBNP, and high-sensitivity troponin I. These features were consistent with the results of the unsupervised prediction model. A visualization system was developed based on these key features. Conclusion Using unsupervised learning and the improved k-means algorithm, this study accurately analysed the prognosis of patients with COVID-19 and hypertension. The use of ACEI/ARB drugs was found to be a protective factor for poor clinical prognosis. Unsupervised learning methods can be used to differentiate patient populations and assess treatment effects. This study identified important features affecting patient prognosis and developed a visualization system with clinical significance for prognosis assessment and treatment decision-making.

Recurrent event query decoder for document-level event extraction

Document-level event extraction is a challenging task in natural language processing, as it involves multiple events within a document and scattered event arguments across sentences. To tackle these challenges, we propose a recurrent event query decoder, i.e., a recurrent module that dynamically updates event queries to capture cross-event dependencies. Our approach then generates arguments by extracting role-argument relations using bilinear mapping, which helps address the issue of scattered arguments. Experimental results demonstrate that our proposed approach outperforms state-of-the-art models on a large-scale public dataset and actual application data, achieving significant improvements in F1-score. In domain-specific event extraction applications, our method achieves higher accuracy with fewer resources compared to general-purpose large language models.

Efficient Geometric Correction Workflow for Airborne Hyperspectral Images through DEM-Driven Correction Techniques

Abstract. Geometric correction, a pivotal step in the preprocessing of airborne remote sensing imagery, is critical for ensuring the accuracy of subsequent quantitative analyses. Achieving precise and efficient geometric correction for airborne hyperspectral data remains a significant challenge in the field. This study presents a new method for system-level and fine-scale geometric correction of uncontrolled airborne images utilizing DEM data, which integrates forward and inverse transformation algorithms. Furthermore, an optimized workflow is proposed to facilitate the processing of large-scale hyperspectral datasets. The effectiveness of the proposed method is demonstrated through an application analysis using airborne HyMap imagery, with experimental outcomes indicating high application accuracy and enhanced processing efficiency.

Thermodiffusively unstable laminar hydrogen flame in a sufficiently large 3D computational domain – Part II: NOx formation mechanism and flamelet modeling

In this work, the NOx formation mechanism in three-dimensional (3D) thermodiffusively unstable premixed hydrogen flames is investigated through direct numerical simulations (DNS) and a reaction path analysis based on the DNS dataset. In part I of this study (Wen et al., 2024), the characteristic patterns of the instabilities were studied using the same dataset. Here, first, the effects of the computational setup (2D vs. 3D) and curvature on the flux ratio of nitrogen atom are quantified. In addition, the composition space model (CSM) proposed in previous works is extended to account for the NOx chemical reaction mechanism to predict NOx formation in thermodiffusively unstable premixed hydrogen flames. The flamelet solutions are obtained from the premixed flamelet equations in composition space so that the wide range of curvatures associated with the strongly corrugated flame front can be considered. The performance of the CSM in predicting the NOx species and the important radicals involved in the NOx formation pathways is evaluated through an a priori analysis. To investigate the effects of curvature on the NOx formation pathways and the performance of the composition space model, the positively- and negatively-curved regions in the DNS are studied separately. The results show that different from the 2D simulation, the NNH reaction pathway becomes dominant in the 3D simulation due to the increased range of curvature, which promotes the accumulation of the highly diffusive H radical in the positively-curved regions. The NNH reaction pathway is dominant in the positively-curved regions, while the N2O reaction pathway is more important in the negatively-curved regions. The flamelet model based on the composition space solutions that consider the effects of curvature yields accurate predictions for the radicals that are sensitive to the effects of curvature.Novelty and significance statementThe novelty of the present work includes the following aspects,(i) Analysis of NOx formation mechanism based on the large-scale 3D DNS dataset of a laminar fuel-lean premixed hydrogen flame on a sufficiently large domain;(ii) Quantification of the effects of computational setup (2D and 3D) and curvature on the NOx formation pathways;(iii) Extension of the previous composition space model (CSM) by incorporating the NOx chemical reaction mechanism to consider the effects of curvature on the NOx formation.

Efficient Feature Matching and Pose-graph Initialization for SfM

Abstract. Over the last decades, Structure from Motion (or image orientation) has been widely studied in the fields of photogrammetry and computer vision, mainly thanks to its feasibility for dealing with various image datasets (such as crowdsourced or UAV images). However, due to the fact that images are becoming easy to obtain, nowadays, it is challenging to deal with large scale of datasets, wherein the feature matching and pose-graph generation are the key limitations in terms of time efficiency. In this work, we proposed an efficient method to accelerate the generation of correspondences and two-view geometries. Specifically, based on some already estimated two-view geometries, unknown two-view geometries can be computed via A* algorithm. Then, the corresponding feature matching can be perform in a guided way using an epiploar-hash bins that is derived from the estimated two-view epipolar geometries. The experimental results demonstrated that, our method can improve the speed of generating pose-graph by 3–4 times comparing to two popular packages (colmap and OpenMVG) and is also faster than one SOTA method of Barath et al., (2021), yet the results of SfM are typically on par with them and reprojection error of our works are even better.

General practitioners’ management of mastitis in breastfeeding women: a mixed method study in Australia

BackgroundMastitis is a common reason new mothers visit their general practitioner (GP). In Australia, the Therapeutic Guidelines: Antibiotic provides practical advice to GPs managing a range of infections, including mastitis. It is not known if Australian GPs prescribe antibiotics and order investigations as recommended for the management of mastitis.MethodsA convergent mixed methods design integrated quantitative analysis of a general practice dataset with analysis of interviews with GPs. Using the large-scale primary care dataset, MedicineInsight, (2021–2022), antibiotics prescribed and investigations ordered for mastitis encounters were extracted. Mastitis encounters were identified by searching ‘Encounter reason’, ‘Test reason’ and ‘Prescription reason’ free text field for the term ‘mastitis’; ‘granulomatous mastitis’ was excluded. Clinical encounters for mastitis occurring within 14 days of a previous mastitis encounter were defined as belonging to the same treatment episode. Semi-structured interviews were conducted with 14 Australian GPs using Zoom or telephone in 2021–2022, and analysed thematically. The Pillar Integration Process was used to develop a joint display table; qualitative codes and themes were matched with the quantitative items to illustrate similarities/contrasts in findings.ResultsDuring an encounter for mastitis, 3122 (91.7%) women received a prescription for an oral antibiotic; most commonly di/flucloxacillin ([59.4%]) or cefalexin (937 [27.5%]). Investigations recorded ultrasound in 303 (8.9%), blood tests (full blood examination [FBE]: 170 [5.0%]; C-reactive protein [CRP]: 71 [2.1%]; erythrocyte sedimentation rate [ESR]: 34 [1.0%]) and breast milk or nipple swab cultures in approximately 1% of encounters. Analysis using pillar integration showed consistency between quantitative and qualitative data regarding mastitis management. The following themes were identified:- GPs support continued breastfeeding.- Antibiotics are central to GPs' management.- Antibiotics are mostly prescribed according to Therapeutic Guidelines.- Analgesia is a gap in the Therapeutic Guidelines.- Low use of breast milk culture.ConclusionsPrescribing antibiotics for mastitis remains central to Australian GPs’ management of mastitis. Interview data clarified that GPs were aware that antibiotics might not be needed in all cases of mastitis and that delayed prescribing was not uncommon. Overall, GPs followed principles of antibiotic stewardship, however there is a need to train GPs about when to consider ordering investigations.

Structured Generation Method of 3D Synthetic Tree Models for Precision Assessment

Abstract. The technology for 3D reconstruction of tree models based on point clouds has been extensively researched, necessitating effective datasets for the study of branch and leaf separation, skeleton point extraction, and tree parameter extraction methods. However, existing datasets for 3D tree models face several challenges, including insufficient data volume for deep learning network training, low accuracy of model ground truth impeding effective method precision evaluation, and a lack of dataset richness to satisfy the needs of multi-type method assessments. In response to these challenges, This paper introduces, for the first time, a fully automated method for generating structured three-dimensional synthetic tree models, and constructs a large-scale 3D synthetic tree dataset enriched with comprehensive structural information. This method facilitates automated computation across several processes, including the mass generation of simulated trees, separation of branches and leaves, noise generation, extraction of skeleton points, and volume calculation. To validate the usability of this dataset across various applications, this paper employs state-of-the-art (SoTA) algorithms to verify the accuracy of methods in 3D tree model reconstruction and carbon stock calculation, thereby thoroughly demonstrating the dataset’s effectiveness.

Machine learning based DNA melt curve profiling enables automated novel genotype detection

Surveillance for genetic variation of microbial pathogens, both within and among species, plays an important role in informing research, diagnostic, prevention, and treatment activities for disease control. However, large-scale systematic screening for novel genotypes remains challenging in part due to technological limitations. Towards addressing this challenge, we present an advancement in universal microbial high resolution melting (HRM) analysis that is capable of accomplishing both known genotype identification and novel genotype detection. Specifically, this novel surveillance functionality is achieved through time-series modeling of sequence-defined HRM curves, which is uniquely enabled by the large-scale melt curve datasets generated using our high-throughput digital HRM platform. Taking the detection of bacterial genotypes as a model application, we demonstrate that our algorithms accomplish an overall classification accuracy over 99.7% and perform novelty detection with a sensitivity of 0.96, specificity of 0.96 and Youden index of 0.92. Since HRM-based DNA profiling is an inexpensive and rapid technique, our results add support for the feasibility of its use in surveillance applications.

A novel method for multiple phenotype association studies based on genotype and phenotype network.

Joint analysis of multiple correlated phenotypes for genome-wide association studies (GWAS) can identify and interpret pleiotropic loci which are essential to understand pleiotropy in diseases and complex traits. Meanwhile, constructing a network based on associations between phenotypes and genotypes provides a new insight to analyze multiple phenotypes, which can explore whether phenotypes and genotypes might be related to each other at a higher level of cellular and organismal organization. In this paper, we first develop a bipartite signed network by linking phenotypes and genotypes into a Genotype and Phenotype Network (GPN). The GPN can be constructed by a mixture of quantitative and qualitative phenotypes and is applicable to binary phenotypes with extremely unbalanced case-control ratios in large-scale biobank datasets. We then apply a powerful community detection method to partition phenotypes into disjoint network modules based on GPN. Finally, we jointly test the association between multiple phenotypes in a network module and a single nucleotide polymorphism (SNP). Simulations and analyses of 72 complex traits in the UK Biobank show that multiple phenotype association tests based on network modules detected by GPN are much more powerful than those without considering network modules. The newly proposed GPN provides a new insight to investigate the genetic architecture among different types of phenotypes. Multiple phenotypes association studies based on GPN are improved by incorporating the genetic information into the phenotype clustering. Notably, it might broaden the understanding of genetic architecture that exists between diagnoses, genes, and pleiotropy.

Efficient Feature Matching for Large-scale Images based on Cascade Hash and Local Geometric Constraint

Abstract. Feature matching plays a crucial role in 3D reconstruction to provide correspondences between overlapped images. The accuracy and efficiency of feature matching significantly impact the performance of 3D reconstruction. The widely used framework with the exhaustive nearest neighbor searching (NNS) between descriptors and RANSAC-based geometric estimation is, however, low-efficient and unreliable for large-scale UAV images. Inspired by indexing-based NNS, this paper implements an efficient feature matching method for large-scale images based on Cascade Hashing and local geometric constraints. Our proposed method improves upon traditional feature matching approaches by introducing a combination of image retrieval, data scheduling, and GPU-accelerated Cascade Hashing. Besides, it utilizes a local geometric constraint to filter matching results within a matching framework. On the one hand, the GPU-accelerated Cascade Hashing technique generates compact and discriminative hash codes based on image features, facilitating the rapid completion of the initial matching process, and significantly reducing the search space and time complexity. On the other hand, after the initial matching is completed, the method employs a local geometric constraint to filter the initial matching results, enhancing the accuracy of the matching results. This forms a three-tier framework based on data scheduling, GPU-accelerated Cascade Hashing, and local geometric constraints. We conducted experiments using two sets of large-scale UAV image data, comparing our method with SIFTGPU to evaluate its performance in initial matching, outlier rejection, and 3D reconstruction. The results demonstrate that our method achieves a feature matching speed 2.0 times that of SIFTGPU while maintaining matching accuracy and producing comparable reconstruction results. This suggests that our method holds promise for efficiently addressing large-scale image matching.

Transcriptome dynamics of Gossypium purpurascens in response to abiotic stresses by Iso-seq and RNA-seq data

Gossypium purpurascens is a member of the Malvaceae family, holds immense economic significance as a fiber crop worldwide. Abiotic stresses harm cotton crops, reduce yields, and cause economic losses. Generating high-quality reference genomes and large-scale transcriptomic datasets across diverse conditions can offer valuable insights into identifying preferred agronomic traits for crop breeding. The present research used leaf tissues to conduct PacBio Iso-seq and RNA-seq analysis. We carried out an in-depth analysis of DEGs using both correlations with cluster analysis and principal component analysis. Additionally, the study also involved the identification of both lncRNAs and CDS. We have prepared RNA-seq libraries from 75 RNA samples to study the effects of drought, salinity, alkali, and saline-alkali stress, as well as control conditions. A total of 454.06 Gigabytes of transcriptome data were effectively validated through the identification of differentially expressed genes and KEGG and GO analysis. Overwhelmingly, gene expression profiles and full-length transcripts from cotton tissues will aid in understanding the genetic mechanism of abiotic stress tolerance in G. purpurascens.

A comparison of RNA-Seq data preprocessing pipelines for transcriptomic predictions across independent studies

BackgroundRNA sequencing combined with machine learning techniques has provided a modern approach to the molecular classification of cancer. Class predictors, reflecting the disease class, can be constructed for known tissue types using the gene expression measurements extracted from cancer patients. One challenge of current cancer predictors is that they often have suboptimal performance estimates when integrating molecular datasets generated from different labs. Often, the quality of the data is variable, procured differently, and contains unwanted noise hampering the ability of a predictive model to extract useful information. Data preprocessing methods can be applied in attempts to reduce these systematic variations and harmonize the datasets before they are used to build a machine learning model for resolving tissue of origins.ResultsWe aimed to investigate the impact of data preprocessing steps—focusing on normalization, batch effect correction, and data scaling—through trial and comparison. Our goal was to improve the cross-study predictions of tissue of origin for common cancers on large-scale RNA-Seq datasets derived from thousands of patients and over a dozen tumor types. The results showed that the choice of data preprocessing operations affected the performance of the associated classifier models constructed for tissue of origin predictions in cancer.ConclusionBy using TCGA as a training set and applying data preprocessing methods, we demonstrated that batch effect correction improved performance measured by weighted F1-score in resolving tissue of origin against an independent GTEx test dataset. On the other hand, the use of data preprocessing operations worsened classification performance when the independent test dataset was aggregated from separate studies in ICGC and GEO. Therefore, based on our findings with these publicly available large-scale RNA-Seq datasets, the application of data preprocessing techniques to a machine learning pipeline is not always appropriate.

Enhancing surgical instrument segmentation: integrating vision transformer insights with adapter.

In surgical image segmentation, a major challenge is the extensive time and resources required to gather large-scale annotated datasets. Given the scarcity of annotated data in this field, our work aims to develop a model that achieves competitive performance with training on limited datasets, while also enhancing model robustness in various surgical scenarios. We propose a method that harnesses the strengths of pre-trained Vision Transformers (ViTs) and data efficiency of convolutional neural networks (CNNs). Specifically, we demonstrate how a CNN segmentation model can be used as a lightweight adapter for a frozen ViT feature encoder. Our novel feature adapter uses cross-attention modules that merge the multiscale features derived from the CNN encoder with feature embeddings from ViT, ensuring integration of the global insights from ViT along with local information from CNN. Extensive experiments demonstrate our method outperforms current models in surgical instrument segmentation. Specifically, it achieves superior performance in binary segmentation on the Robust-MIS 2019 dataset, as well as in multiclass segmentation tasks on the EndoVis 2017 and EndoVis 2018 datasets. It also showcases remarkable robustness through cross-dataset validation across these 3 datasets, along with the CholecSeg8k and AutoLaparo datasets. Ablation studies based on the datasets prove the efficacy of our novel adapter module. In this study, we presented a novel approach integrating ViT and CNN. Our unique feature adapter successfully combines the global insights of ViT with the local, multi-scale spatial capabilities of CNN. This integration effectively overcomes data limitations in surgical instrument segmentation. The source code is available at: https://github.com/weimengmeng1999/AdapterSIS.git .

Federated 3D multi-organ segmentation with partially labeled and unlabeled data.

This paper considers a new problem setting for multi-organ segmentation based on the following observations. In reality, (1) collecting a large-scale dataset from various institutes is usually impeded due to privacy issues; (2) many images are not labeled since the slice-by-slice annotation is costly; and (3) datasets may exhibit inconsistent, partial annotations across different institutes. Learning a federated model from these distributed, partially labeled, and unlabeled samples is an unexplored problem. To simulate this multi-organ segmentation problem, several distributed clients and a central server are maintained. The central server coordinates with clients to learn a global model using distributed private datasets, which comprise a small part of partially labeled images and a large part of unlabeled images. To address this problem, a practical framework that unifies partially supervised learning (PSL), semi-supervised learning (SSL), and federated learning (FL) paradigms with PSL, SSL, and FL modules is proposed. The PSL module manages to learn from partially labeled samples. The SSL module extracts valuable information from unlabeled data. Besides, the FL module aggregates local information from distributed clients to generate a global statistical model. With the collaboration of three modules, the presented scheme could take advantage of these distributed imperfect datasets to train a generalizable model. The proposed method was extensively evaluated with multiple abdominal CT datasets, achieving an average result of 84.83% in Dice and 41.62mm in 95HD for multi-organ (liver, spleen, and stomach) segmentation. Moreover, its efficacy in transfer learning further demonstrated its good generalization ability for downstream segmentation tasks. This study considers a novel problem of multi-organ segmentation, which aims to develop a generalizable model using distributed, partially labeled, and unlabeled CT images. A practical framework is presented, which, through extensive validation, has proved to be an effective solution, demonstrating strong potential in addressing this challenging problem.

Pretraining of attention-based deep learning potential model for molecular simulation

Machine learning-assisted modeling of the inter-atomic potential energy surface (PES) is revolutionizing the field of molecular simulation. With the accumulation of high-quality electronic structure data, a model that can be pretrained on all available data and finetuned on downstream tasks with a small additional effort would bring the field to a new stage. Here we propose DPA-1, a Deep Potential model with a gated attention mechanism, which is highly effective for representing the conformation and chemical spaces of atomic systems and learning the PES. We tested DPA-1 on a number of systems and observed superior performance compared with existing benchmarks. When pretrained on large-scale datasets containing 56 elements, DPA-1 can be successfully applied to various downstream tasks with a great improvement of sample efficiency. Surprisingly, for different elements, the learned type embedding parameters form a spiral in the latent space and have a natural correspondence with their positions on the periodic table, showing interesting interpretability of the pretrained DPA-1 model.

Scene-aware data augmentation for ship detection in SAR images

ABSTRACT Ship detection in synthetic aperture radar (SAR) images is widely applied in marine monitoring. In recent years, convolutional neural networks (CNNs) have made significant advances in SAR ship detection. However, the scene imbalance problem, i.e. much more offshore scenes than inshore scenes in SAR images, severely hampers the improvement of detection accuracy. In this article, we propose scene-aware data augmentation to address the problem. First, a novel concept, inshore rate, is introduced to estimate the land area percentage of various scenes in SAR images. We can divide image samples into inshore and offshore ones based on the inshore rate without manual category annotations. Second, to augment inshore samples, we synthesize dummy inshore images by cropping the ground truth bounding boxes (GT bboxes) from offshore images and pasting them onto inshore images. This offline data augmentation strategy can address the scene imbalance problem and improve detection accuracy. The experiments conducted on SAR Ship Detection Dataset (SSDD) and Large-Scale SAR Ship Detection Dataset (LS-SSDD) demonstrate the effectiveness of the proposed method in improving detection performance.

Reconstructing long-term natural flows by ensemble machine learning

Large-scale human activities have caused significant disruption to the natural flow regime of rivers worldwide, leading to substantial deviations from river flows’ natural state. The natural hydrological regime of rivers holds great significance for the conservation and management of riverine ecosystems, which relies on the accurate reconstruction of natural river flows. However, the application of existing high spatiotemporal resolution large-scale natural flow datasets at the regional scale has not been adequately assessed, thereby limiting their regional-scale utility. In this study, utilizing global reanalysis datasets and observation data, in combination with ensemble machine learning models, we constructed long-term natural flow time series for inflows into the Pearl River Delta and evaluated the performance of these reconstructed inflows based on observed flow records during undisturbed periods. The reconstruction shows that the ensemble model, the AutoGluon model, indicates the best performance when compared with other machine learning models and outperforms the existing publicly available large-scale datasets in terms of evaluation metrics for the corresponding natural flows. The results of Indicator of Hydrologic Alteration and Range of Variability Analysis (IHA-RVA) indicate that climate variability, reservoir regulation and human activities have significantly altered the natural flow regime of the Pearl River Delta. Specifically, human activities generally tend to smooth natural flow variability in the Pearl River Delta. This research is expected to provide an efficient and reliable method for reconstructing natural flows and offers guidance for water resource management and ecosystem conservation.