Small Datasets Research Articles

Estimation of minimal data sets sizes for machine learning predictions in digital mental health interventions

Artificial intelligence promises to revolutionize mental health care, but small dataset sizes and lack of robust methods raise concerns about result generalizability. To provide insights on minimal necessary data set sizes, we explore domain-specific learning curves for digital intervention dropout predictions based on 3654 users from a single study (ISRCTN13716228, 26/02/2016). Prediction performance is analyzed based on dataset size (N = 100–3654), feature groups (F = 2–129), and algorithm choice (from Naive Bayes to Neural Networks). The results substantiate the concern that small datasets (N ≤ 300) overestimate predictive power. For uninformative feature groups, in-sample prediction performance was negatively correlated with dataset size. Sophisticated models overfitted in small datasets but maximized holdout test results in larger datasets. While N = 500 mitigated overfitting, performance did not converge until N = 750–1500. Consequently, we propose minimum dataset sizes of N = 500–1000. As such, this study offers an empirical reference for researchers designing or interpreting AI studies on Digital Mental Health Intervention data.

An explainable analysis of diabetes mellitus using statistical and artificial intelligence techniques

BackgroundDiabetes mellitus (DM) is a chronic disease prevalent worldwide, requiring a multifaceted analytical approach to improve early detection and subsequent mitigation of morbidity and mortality rates. This research aimed to develop an explainable analysis of DM by combining sociodemographic and clinical data with statistical and artificial intelligence (AI) techniques.MethodsLeveraging a small dataset that includes sociodemographic and clinical profiles of diabetic and non-diabetic individuals, we employed a diverse set of statistical and AI models for predictive purposes and assessment of DM risk factors. The statistical tests used were Student’s t-test and Chi-square, while the AI techniques were fuzzy cognitive maps (FCM), artificial neural networks (ANN), support vector machines (SVM), and XGBoost.ResultsOur statistical models facilitated an in-depth exploration of variable associations, while the resulting AI models demonstrated exceptional efficacy in DM classification. In particular, the XGBoost model showed superior performance in accuracy, sensitivity and specificity with values of 1 for each of these metrics. On the other hand, the FCM stood out for its explainability capabilities by allowing an analysis of the variables involved in the prediction using scenario-based simulations.ConclusionsAn integrated analysis of DM using a variety of methodologies is critical for timely detection of the disease and informed clinical decision-making.

Quantum neural network-assisted learning for small medical datasets: a case study in emphysema detection

Quantum neural network-assisted learning for small medical datasets: a case study in emphysema detection

Small RNA sequencing data of plasma extracellular vesicles in a breast cancer screening population

The pathogenesis of breast cancer is still unclear. Small RNAs associated with extracellular vesicles (EVs) have been found to be involved in tumor development. It is important to explore the role of EVs small RNAs in breast cancer. In this study, we established a plasma EVS-associated small RNA dataset that included 120 women who were positive for breast cancer screening and 60 women who were negative. These small RNA included 2656 miRNAs, 728 piRNAs, and 154 tsRNAs. These data provide a reference for researchers to explore molecular diagnostic biomarkers for early breast lesions.

3D radiative transfer modeling of almond canopy for nitrogen estimation by hyperspectral imaging

Nitrogen (N) is vital for plant growth, but its imbalance can negatively affect crop yields, the environment, and water quality. This is especially crucial for California’s almond orchards, which are the most N-hungry nut crop and require substantial N for high productivity. The current practices of uniform and extensive N application lead to N leaching into the groundwater, creating environmental hazards. Traditional remote sensing methods often rely on data-driven approaches that work well statistically (achieving a high R2 value) with one dataset but aren’t adaptable across different datasets. To create a more robust, data-driven model, one would typically need a vast and varied collection of datasets. Our goal, however, is to develop a more universally applicable model using smaller datasets, typical of commercial orchards, that can accurately estimate N content in tree canopies, regardless of differences in spatial, spectral, and temporal data. In this study, we investigate and evaluate multiple remote sensing approaches for estimating N concentration in Californian almonds, utilizing hyperspectral imaging at the canopy level. We assess various classical vegetation indices, machine learning models, and a physics-informed 3D radiative transfer model. While cross-validated results show comparable results for radiative transfer models and best-performing machine learning models, most single vegetation indices are not capable of exceeding the baseline model and thus had R2 value less than 0. Despite being less commonly used, 3D radiative transfer modeling shows promise as a strong and adaptable method, producing results that are comparable to the best machine learning models.

DIAFM: An Improved and Novel Approach for Incremental Frequent Itemset Mining

Traditional approaches to data mining are generally designed for small, centralized, and static datasets. However, when a dataset grows at an enormous rate, the algorithms become infeasible in terms of huge consumption of computational and I/O resources. Frequent itemset mining (FIM) is one of the key algorithms in data mining and finds applications in a variety of domains; however, traditional algorithms do face problems in efficiently processing large and dynamic datasets. This research introduces a distributed incremental approximation frequent itemset mining (DIAFM) algorithm that tackles the mentioned challenges using shard-based approximation within the MapReduce framework. DIAFM minimizes the computational overhead of a program by reducing dataset scans, bypassing exact support checks, and incorporating shard-level error thresholds for an appropriate trade-off between efficiency and accuracy. Extensive experiments have demonstrated that DIAFM reduces runtime by 40–60% compared to traditional methods with losses in accuracy within 1–5%, even for datasets over 500,000 transactions. Its incremental nature ensures that new data increments are handled efficiently without needing to reprocess the entire dataset, making it particularly suitable for real-time, large-scale applications such as transaction analysis and IoT data streams. These results demonstrate the scalability, robustness, and practical applicability of DIAFM and establish it as a competitive and efficient solution for mining frequent itemsets in distributed, dynamic environments.

Implementing a Bayesian approach using Stan with Torsten: Population pharmacokinetics analysis of somatrogon.

Fully Bayesian approaches are not commonly implemented for population pharmacokinetic (PK) modeling. In this paper, we evaluate the use of Stan with R and Torsten for population PK modeling of somatrogon, a recombinant long-acting growth hormone approved for the treatment of growth hormone deficiency. As a software for Bayesian inference, Stan provides an easy way to conduct MCMC sampling for a wide range of models with efficient sampling algorithms, and there are several diagnostic tools to evaluate the MCMC convergence and other potential issues. Three different sets of priors were evaluated for estimation and prediction: a weakly informative uniform set, a moderately informative set, and a very informative set of priors. All three prior sets showed good performance and all chains mixed well. There were some minor differences in the final parameter posterior distributions while implementing different prior sets, but the posterior predictions covered the observations nicely, not only for the individuals included in posterior sampling but also for new individuals. The impact of a centered versus non-centered parameterization were evaluated, with the non-centered approach improving the estimation time, but it was still computationally intensive. Computational resources had the biggest impact on sampling time. Stan took approximately 2.5 h total for the MCMC sampling on a high-performance computing platform (6 cores) and may be reduced further with additional computational resources. The model and comparisons presented show that with adequate computational resources, the Bayesian approaches using Stan and Torsten are useful for population PK analysis, especially for the analysis of special populations, small sample datasets, and when complex model structures are needed.

Enhancing Software Effort Estimation with Pre-Trained Word Embeddings: A Small-Dataset Solution for Accurate Story Point Prediction

Traditional software effort estimation methods, such as term frequency–inverse document frequency (TF-IDF), are widely used due to their simplicity and interpretability. However, they struggle with limited datasets, fail to capture intricate semantics, and suffer from dimensionality, sparsity, and computational inefficiency. This study used pre-trained word embeddings, including FastText and GPT-2, to improve estimation accuracy in such cases. Seven pre-trained models were evaluated for their ability to effectively represent textual data, addressing the fundamental limitations of TF-IDF through contextualized embeddings. The results show that combining FastText embeddings with support vector machines (SVMs) consistently outperforms traditional approaches, reducing the mean absolute error (MAE) by 5–18% while achieving accuracy comparable to deep learning models like GPT-2. This approach demonstrated the adaptability of pre-trained embeddings for small datasets, balancing semantic richness with computational efficiency. The proposed method optimized project planning and resource allocation while enhancing software development through accurate story point prediction while safeguarding privacy and security through data anonymization. Future research will explore task-specific embeddings tailored to software engineering domains and investigate how dataset characteristics, such as cultural variations, influence model performance, ensuring the development of adaptable, robust, and secure machine learning models for diverse contexts.

Axial-UNet++ Power Line Detection Network Based on Gated Axial Attention Mechanism

The segmentation and recognition of power lines are crucial for the UAV-based inspection of overhead power lines. To address the issues of class imbalance, low sample quantity, and long-range dependency in images, a specialized semantic segmentation network for power line segmentation called Axial-UNet++ is proposed. Firstly, to tackle the issue of long-range dependencies in images and low sample quantity, a gated axial attention mechanism is introduced to expand the receptive field and improve the capture of relative positional biases in small datasets, thereby proposing a novel feature extraction module termed axial-channel local normalization module. Secondly, to address the imbalance in training samples, a new loss function is developed by combining traditional binary cross-entropy loss with focal loss, enhancing the precision of image semantic segmentation. Lastly, ablation and comparative experiments on the PLDU and Mendeley datasets demonstrate that the proposed model achieves 54.7% IoU and 80.1% recall on the PLDU dataset, and 79.3% IoU and 93.1% recall on the Mendeley dataset, outperforming other listed models. Additionally, robustness experiments show the adaptability of the Axial-UNet++ model under extreme conditions and the augmented image dataset used in this study has been open sourced.

Word embeddings on ideology and issues from Swedish parliamentarians’ motions: a comparative approach

ABSTRACT Quantitative analysis of large-scale political text data in the form of word embeddings has great potential for systematising differences between political parties. We examine the differences between embeddings obtained from speakers from the two competitors for the PM position in Sweden (Social Democrats and Moderates) over a 30-year period. The goal is to compare how off-the-shelf general pre-trained models perform relative to pre-training on a smaller dataset from the same domain. In the analysis, we focus on two types of concepts: issues and ideological terms. We find that generally, the off-the-shelf pre-trained models lead to more reliable results and greater emphasis on ideological differences between the studied parties.

A novel knowledge distillation framework for enhancing small object detection in blurry environments with unmanned aerial vehicle-assisted images

Deep learning-based object detectors excel on mobile devices but often struggle with blurry images that are common in real-world scenarios, like unmanned aerial vehicle (UAV)-assisted images. Current models are designed for sharp images, leading to potential detection failures in blurry images. Using image deblurring before object detection is an option, but it demands significant computing power and relies heavily on the accuracy of the deblurring algorithms. Another common issue is the suitable dataset for the specific problem. To address the aforementioned issues, we develop a UAV-assisted small object detection dataset and propose a novel knowledge distillation method for object detection in blurry images in complex environments. Following this, we employ a technique known as self-supervised knowledge distillation, where we introduce a deblurring subnet module with the help of two attention modules, where both networks are trained in a fully-supervised manner. Based on the experiment results, our proposed model achieves an improvement of 4.3% accuracy in the VisDrone synthetic motion blur dataset and 4.6% in detecting objects within synthetic blurry images in our developed small object detection dataset (SOD-Dataset), as well as competitive results compared with other state-of-the-art methods. Meanwhile, ablation experiments and a visualization analysis validate the contributions of each component of the model.

Detection and Identification of Stuttering Types Using Siamese Network

Stuttering is a complex speech disorder characterized by disruptions in the fluency of verbal expression, often leading to challenges in communication for those affected. Accurate identification and classification of stuttering types can greatly benefit persons who stutter (PWS), especially in an era where voice technologies are becoming increasingly ubiquitous and integrated into daily life. In this work, we adapt a simple yet effective Siamese network architecture, known for its capability to learn from paired speech segments, to extract novel features from audio speech data. Our approach leverages these features to enhance the detection and identification of stuttering events. For our experiments, we rely on a subset of the SEP-28k stuttering dataset, initially implementing a single-task model and gradually evolving it into a more sophisticated multi-task model. Our results demonstrate that transitioning the network from a single-task learner to a multi-task learner, coupled with the integration of auxiliary classification heads, significantly improves the identification of stuttering types, even with a relatively small dataset.

Application of artificial intelligence-based detection of furcation involvement in mandibular first molar using cone beam tomography images- a preliminary study

BackgroundRadiographs play a key role in diagnosis of periodontal diseases. Deep learning models have been explored for image analysis in periodontal diseases. However, there is lacuna of research in the deep learning model-based detection of furcation involvements [FI]. The objective of this study was to determine the accuracy of deep learning model in the detection of FI in axial CBCT images.MethodologyWe obtained initial dataset 285 axial CBCT images among which 143 were normal (without FI) and 142 were abnormal (with FI). Data augmentation technique was used to create 600(300 normal and 300 abnormal) images by using 200 images from the training dataset. Remaining 85(43 normal and 42 abnormal) images were kept for testing of model. ResNet101V2 with transfer learning was used employed for the analysis of images.ResultsTraining accuracy of model is 98%, valid accuracy is 97% and test accuracy is 91%. The precision and F1 score were 0.98 and 0.98 respectively. The Area under curve (AUC) was reported at 0.98. The test loss was reported at 0.2170.ConclusionThe deep learning model (ResNet101V2) can accurately detect the FI in axial CBCT images. However, since our study was preliminary in nature and carried out with relatively smaller dataset, a study with larger dataset will further confirm the accuracy of deep learning models.

Chain-structured neural architecture search for financial time series forecasting

AbstractNeural architecture search (NAS) emerged as a way to automatically optimize neural networks for a specific task and dataset. Despite an abundance of research on NAS for images and natural language applications, similar studies for time series data are lacking. Among NAS search spaces, chain-structured are the simplest and most applicable to small datasets like time series. We compare three popular NAS strategies on chain-structured search spaces: Bayesian optimization (specifically Tree-structured Parzen Estimator), the hyperband method, and reinforcement learning in the context of financial time series forecasting. These strategies were employed to optimize simple well-understood neural architectures like the MLP, 1D CNN, and RNN, with more complex temporal fusion transformers (TFT) and their own optimizers included for comparison. We find Bayesian optimization and the hyperband method performing best among the strategies, and RNN and 1D CNN best among the architectures, but all methods were very close to each other with a high variance due to the difficulty of working with financial datasets. We discuss our approach to overcome the variance and provide implementation recommendations for future users and researchers.

Leveraging deep transfer learning and explainable AI for accurate COVID-19 diagnosis: Insights from a multi-national chest CT scan study

The COVID-19 pandemic has emerged as a global health crisis, impacting millions worldwide. Although chest computed tomography (CT) scan images are pivotal in diagnosing COVID-19, their manual interpretation by radiologists is time-consuming and potentially subjective. Automated computer-aided diagnostic (CAD) frameworks offer efficient and objective solutions. However, machine or deep learning methods often face challenges in their reproducibility due to underlying biases and methodological flaws. To address these issues, we propose XCT-COVID, an explainable, transferable, and reproducible CAD framework based on deep transfer learning to predict COVID-19 infection from CT scan images accurately. This is the first study to develop three distinct models within a unified framework by leveraging a previously unexplored large dataset and two widely used smaller datasets. We employed five known convolutional neural network architectures, both with and without pretrained weights, on the larger dataset. We optimized hyperparameters through extensive grid search and 5-fold cross-validation (CV), significantly enhancing the model performance. Experimental results from the larger dataset showed that the VGG16 architecture (XCT-COVID-L) with pretrained weights consistently outperformed other architectures, achieving the best performance, on both 5-fold CV and independent test. When evaluated with the external datasets, XCT-COVID-L performed well with data with similar distributions, demonstrating its transferability. However, its performance significantly decreased on smaller datasets with lower-quality images. To address this, we developed other models, XCT-COVID-S1 and XCT-COVID-S2, specifically for the smaller datasets, outperforming existing methods. Moreover, eXplainable Artificial Intelligence (XAI) analyses were employed to interpret the models’ functionalities. For prediction and reproducibility purposes, the implementation of XCT-COVID is publicly accessible at https://github.com/cbbl-skku-org/XCT-COVID/.

Hierarchical Bayesian modeling for Inverse Uncertainty Quantification of system thermal-hydraulics code using critical flow experimental data

The best estimate plus uncertainty methodology in nuclear system thermal-hydraulic studies necessitates a comprehensive understanding of uncertainties in system code predictions. The forward uncertainty quantification (UQ) process involves the propagation of input uncertainties through the computational models to obtain uncertainties in the outputs. To this end, achieving an accurate estimation of input uncertainties is important, which is the focus of inverse UQ (IUQ). Traditionally, research in Bayesian IUQ within the nuclear engineering domain has largely relied on single-level Bayesian inference. While being effective for relatively small datasets, this approach encounters limitations for cases with large datasets. The use of a single-level model may prove inefficient, as the resultant posterior distributions can significantly differ when distinct subsets of data are employed. To address this issue, we employ an hierarchical Bayesian model for IUQ. This approach involves organizing observations into different groups based on the test conditions, thereby accommodating varying calibration parameters across these distinct groups. In this study, we developed and implemented a hierarchical Bayesian IUQ method to consider the grouping effect of critical flow measurement data from various geometries. Comparing the outcomes of IUQ under different selections of test data using hierarchical Bayesian IUQ against those obtained from single-level Bayesian IUQ, the forward propagation of hierarchical Bayesian IUQ results demonstrates a notably improved agreement with the experimental data.

Unveiling Lung Diseases in CT Scan Images With a Hybrid Bio‐Inspired Mutated Spider‐Monkey and Crow Search Algorithm

ABSTRACTBio‐inspired computer‐aided diagnosis (CAD) has garnered significant attention in recent years due to the inherent advantages of bio‐inspired evolutionary algorithms (EAs) in handling small datasets with elevated precision and reduced computational complexity. Traditional CAD models face limitations as they can only be developed post‐outbreak, relying on datasets that become available during such events such as the COVID‐19 pandemic. The scarcity of data for emerging diseases poses a substantial challenge to achieving elevated precision with conventional deep‐learning algorithms. Furthermore, even when datasets are available, employing deep learning for class‐based classification is arduous, necessitating model retraining, in this paper, we propose a novel hybrid algorithm that leverages the strengths of the crow search algorithm (CSA) and the spider monkey optimization (SMO) algorithm to create an optimised spider monkey crow search (OSM‐CS) algorithm. We developed a CAD tool that maps each input CT image to a high‐dimensional vector by extracting four categories of features: high contrast, polynomial decomposition, textural, and pixel statistics. The proposed OSM‐CS algorithm is employed as a feature selection method. Our experimental results demonstrate the effectiveness of the OSM‐CS algorithm, achieving an impressive accuracy of 98.2% when coupled with an AdaBoost classifier for multi‐class classification and 99.93% for binary classification. This performance surpasses that of state‐of‐the‐art (SOTA) deep learning models and recently published hybrid algorithms, underscoring the potential of the OSM‐CS algorithm as a powerful tool in the realm of CAD.

Beware of diffusion models for synthesizing medical images - A comparison with GANs in terms of memorizing brain MRI and chest x-ray images

Abstract Diffusion models were initially developed for text-to-image generation and are now being utilized to generate high quality synthetic images. Preceded by GANs, diffusion models have shown impressive results using various evaluation metrics. However, commonly used metrics such as FID and IS are not suitable for determining whether diffusion models are simply reproducing the training images. Here we train StyleGAN and a diffusion model, using BRATS20, BRATS21 and a chest x-ray pneumonia dataset, to synthesize brain MRI and chest x-ray images, and measure the correlation between the synthetic images and all training images. Our results show that diffusion models are more likely to memorize the training images, compared to StyleGAN, especially for small datasets and when using 2D slices from 3D volumes. Researchers should be careful when using diffusion models (and to some extent GANs) for medical imaging, if the final goal is to share the synthetic images.

Advancements and Challenges: A Comprehensive Review of GAN-based Models for the Mitigation of Small Dataset and Texture Sticking Issues in Fake License Plate Recognition

This review paper critically examines the recent advancements in refining Generative Adversarial Networks (GANs) to address the challenges posed by small datasets and the persisting issue of texture sticking in the domain of fake license plate recognition. Recognizing the limitations posed by insufficient data, the survey begins with an exploration of various GAN architectures, including pix2pix_GAN, CycleGAN, and SRGAN, that have been employed to synthesize diverse and realistic license plate images. Notable achievements include high accuracy in License Plate Character Recognition (LPCR), advancements in generating new format license plates, and improvements in license plate detection using YOLO. The second focal point of this review centers on mitigating the texture sticking problem, a crucial concern in GAN-generated content. Recent enhancements, such as the integration of StyleGAN2-ADA and StyleGAN3, aim to address challenges related to texture dynamics during video generation. Additionally, adaptive data augmentation mechanisms have been introduced to stabilize GAN training, particularly when confronted with limited datasets. The synthesis of these findings provides a comprehensive overview of the evolving landscape in mitigating challenges associated with small datasets and texture sticking in fake license plate recognition. The review not only underscores the progress made but also identifies emerging trends and areas for future exploration. These insights are vital for researchers, practitioners, and policymakers aiming to bolster the effectiveness and reliability of GAN-based models in the critical domain of license plate recognition.

DESRGAN: Detail-enhanced generative adversarial networks for small sample single image super-resolution

Image super-resolution involves reconstructing a blurry, low-resolution image with limited information into a clear, high-resolution image containing more detailed information. The images generated by super-resolution reconstruction can enhance the performance of downstream computer vision tasks, and hold wide application prospects in fields such as industrial fault detection, plant phenotype parameter extraction, medical imaging, and more. High-frequency components in images, such as edges and texture details, typically require more attention. However, when the training samples are limited, effectively recovering clear high-frequency details of images becomes highly challenging. Therefore, this paper proposes a single-image super-resolution method based on generative adversarial networks, named DESRGAN. Compared to existing methods, DESRGAN achieves better reconstruction of image details even with a limited number of training samples. DESRGAN introduces several key innovations: a shallow generator structure to address overfitting issues in small sample scenarios, a dual-stream feature extraction network with dilated convolutions to capture multi-scale contextual information and expand the receptive field, and an artifact loss designed to eliminate artifacts and preserve the true high-frequency details of the super-resolved images. Extensive ablation experiments and comparative studies with multiple state-of-the-art models are conducted on two small sample datasets, "Root" and "Leaves," as well as five publicly available datasets. The results demonstrate that the proposed DESRGAN achieves superior performance in small sample single image super-resolution tasks, with improvements of 1.39 dB in PSNR and 0.013 in SSIM. The generated high-resolution images exhibit clear texture and edge structures, presenting favorable subjective visual effects. Moreover, the model displays strong generalization capabilities.