UNASSIGNED: Training data fuel and shape the development of artificial intelligence (AI) models. Intensive data requirements are a major bottleneck limiting the success of AI tools in sectors with inherently scarce data. In health care, training data are difficult to curate, triggering growing concerns that the current lack of access to health care by under-privileged social groups will translate into future bias in health care AIs. In this report, we developed an autoencoder to grow and enhance inherently scarce datasets to alleviate our dependence on big data.
UNASSIGNED: Computational study with open-source data.
UNASSIGNED: The data were obtained from 6 open-source datasets comprising patients aged 40-80 years in Singapore, China, India, and Spain.
UNASSIGNED: The reported framework generates synthetic images based on real-world patient imaging data. As a test case, we used autoencoder to expand publicly available training sets of optic disc photos, and evaluated the ability of the resultant datasets to train AI models in the detection of glaucomatous optic neuropathy.
UNASSIGNED: Area under the receiver operating characteristic curve (AUC) were used to evaluate the performance of the glaucoma detector. A higher AUC indicates better detection performance.
UNASSIGNED: Results show that enhancing datasets with synthetic images generated by autoencoder led to superior training sets that improved the performance of AI models.
UNASSIGNED: Our findings here help address the increasingly untenable data volume and quality requirements for AI model development and have implications beyond health care, toward empowering AI adoption for all similarly data-challenged fields.
UNASSIGNED: The authors have no proprietary or commercial interest in any materials discussed in this article.

Data collection, curation, and cleaning constitute a crucial phase in Machine Learning (ML) projects. In biomedical ML, it is often desirable to leverage multiple datasets to increase sample size and diversity, but this poses unique challenges, which arise from heterogeneity in study design, data descriptors, file system organization, and metadata. In this study, we present an approach to the integration of multiple brain MRI datasets with a focus on homogenization of their organization and preprocessing for ML. We use our own fusion example (approximately 84,000 images from 54,000 subjects, 12 studies, and 88 individual scanners) to illustrate and discuss the issues faced by study fusion efforts, and we examine key decisions necessary during dataset homogenization, presenting in detail a database structure flexible enough to accommodate multiple observational MRI datasets. We believe our approach can provide a basis for future similarly-minded biomedical ML projects.

Worldwide manufacturing industries are significantly affected by COVID-19 pandemic because of their production characteristics with low-cost country sourcing, globalization, and inventory level. To analyze the correlated time series, spatial-temporal model becomes more attractive, and the graph convolution network (GCN) is also commonly used to provide more information to the nodes and its neighbors in the graph. Recently, attention-adjusted graph spatio-temporal network (AGSTN) was proposed to address the problem of pre-defined graph in GCN by combining multi-graph convolution and attention adjustment to learn spatial and temporal correlations over time. However, AGSTN may show potential problem with limited small non-sensor data; particularly, convergence issue. This study proposes several variants of AGSTN and applies them to non-sensor data. We suggest data augmentation and regularization techniques such as edge selection, time series decomposition, prevention policies to improve AGSTN. An empirical study of worldwide manufacturing industries in pandemic era was conducted to validate the proposed variants. The results show that the proposed variants significantly improve the prediction performance at least around 20% on mean squared error (MSE) and convergence problem.

The integration of emotional intelligence in machines is an important step in advancing human-computer interaction. This demands the development of reliable end-to-end emotion recognition systems. However, the scarcity of public affective datasets presents a challenge. In this literature review, we emphasize the use of generative models to address this issue in neurophysiological signals, particularly Electroencephalogram (EEG) and Functional Near-Infrared Spectroscopy (fNIRS). We provide a comprehensive analysis of different generative models used in the field, examining their input formulation, deployment strategies, and methodologies for evaluating the quality of synthesized data. This review serves as a comprehensive overview, offering insights into the advantages, challenges, and promising future directions in the application of generative models in emotion recognition systems. Through this review, we aim to facilitate the progression of neurophysiological data augmentation, thereby supporting the development of more efficient and reliable emotion recognition systems.

OBJECTIVE: To evaluate the potential of synthetic radiomic data generation in addressing data scarcity in radiomics/radiogenomics models.
METHODS: This study was conducted on a retrospectively collected cohort of 386 colorectal cancer patients (n = 2570 lesions) for whom matched contrast-enhanced CT images and gene TP53 mutational status were available. The full cohort data was divided into a training cohort (n = 2055 lesions) and an independent and fixed test set (n = 515 lesions). Differently sized training sets were subsampled from the training cohort to measure the impact of sample size on model performance and assess the added value of synthetic radiomic augmentation at different sizes. Five different tabular synthetic data generation models were used to generate synthetic radiomic data based on \"real-world\" radiomics data extracted from this cohort. The quality and reproducibility of the generated synthetic radiomic data were assessed. Synthetic radiomics were then combined with \"real-world\" radiomic training data to evaluate their impact on the predictive model\'s performance.
RESULTS: A prediction model was generated using only \"real-world\" radiomic data, revealing the impact of data scarcity in this particular data set through a lack of predictive performance at low training sample numbers (n = 200, 400, 1000 lesions with average AUC = 0.52, 0.53, and 0.56 respectively, compared to 0.64 when using 2055 training lesions). Synthetic tabular data generation models created reproducible synthetic radiomic data with properties highly similar to \"real-world\" data (for n = 1000 lesions, average Chi-square = 0.932, average basic statistical correlation = 0.844). The integration of synthetic radiomic data consistently enhanced the performance of predictive models trained with small sample size sets (AUC enhanced by 9.6%, 11.3%, and 16.7% for models trained on n_samples = 200, 400, and 1000 lesions, respectively). In contrast, synthetic data generated from randomised/noisy radiomic data failed to enhance predictive performance underlining the requirement of true signal data to do so.
CONCLUSIONS: Synthetic radiomic data, when combined with real radiomics, could enhance the performance of predictive models. Tabular synthetic data generation might help to overcome limitations in medical AI stemming from data scarcity.

Data-driven machine learning (ML) provides a promising approach to understanding and predicting the rejection of trace organic contaminants (TrOCs) by polyamide (PA). However, various confounding variables, coupled with data scarcity, restrict the direct application of data-driven ML. In this study, we developed a data-knowledge codriven ML model via domain-knowledge embedding and explored its application in comprehending TrOC rejection by PA membranes. Domain-knowledge embedding enhanced both the predictive performance and the interpretability of the ML model. The contribution of key mechanisms, including size exclusion, charge effect, hydrophobic interaction, etc., that dominate the rejections of the three TrOC categories (neutral hydrophilic, neutral hydrophobic, and charged TrOCs) was quantified. Log D and molecular charge emerge as key factors contributing to the discernible variations in the rejection among the three TrOC categories. Furthermore, we quantitatively compared the TrOC rejection mechanisms between nanofiltration (NF) and reverse osmosis (RO) PA membranes. The charge effect and hydrophobic interactions possessed higher weights for NF to reject TrOCs, while the size exclusion in RO played a more important role. This study demonstrated the effectiveness of the data-knowledge codriven ML method in understanding TrOC rejection by PA membranes, providing a methodology to formulate a strategy for targeted TrOC removal.

RNA modification, N4-acetylcytidine (ac4C), is enzymatically catalyzed by N-acetyltransferase 10 (NAT10) and plays an essential role across tRNA, rRNA, and mRNA. It influences various cellular functions, including mRNA stability and rRNA biosynthesis. Wet-lab detection of ac4C modification sites is highly resource-intensive and costly. Therefore, various machine learning and deep learning techniques have been employed for computational detection of ac4C modification sites. The known ac4C modification sites are limited for training an accurate and stable prediction model. This study introduces GANSamples-ac4C, a novel framework that synergizes transfer learning and generative adversarial network (GAN) to generate synthetic RNA sequences to train a better ac4C modification site prediction model. Comparative analysis reveals that GANSamples-ac4C outperforms existing state-of-the-art methods in identifying ac4C sites. Moreover, our result underscores the potential of synthetic data in mitigating the issue of data scarcity for biological sequence prediction tasks. Another major advantage of GANSamples-ac4C is its interpretable decision logic. Multi-faceted interpretability analyses detect key regions in the ac4C sequences influencing the discriminating decision between positive and negative samples, a pronounced enrichment of G in this region, and ac4C-associated motifs. These findings may offer novel insights for ac4C research. The GANSamples-ac4C framework and its source code are publicly accessible at http://www.healthinformaticslab.org/supp/.

Several recent studies have evidenced the relevance of machine-learning for soil salinity mapping using Sentinel-2 reflectance as input data and field soil salinity measurement (i.e., Electrical Conductivity-EC) as the target. As soil EC monitoring is costly and time consuming, most learning databases used for training/validation rely on a limited number of soil samples, which can affect the model consistency. Based on the low soil salinity variation at the Sentinel-2 pixel resolution, this study proposes to increase the learning database\'s number of observations by assigning the EC value obtained on the sampled pixel to the eight neighboring pixels. The method allowed extending the original learning database made up of 97 field EC measurements (OD) to an enhanced learning database made up of 691 observations (ED). Two classification machine-learning models (i.e., Random Forest-RF and Support Vector Machine-SVM) were trained with both OD and ED to assess the efficiency of the proposed method by comparing the models\' outcomes with EC observations not used in the models´ training. The use of ED led to a significant increase in both models\' consistency with the overall accuracy of the RF (SVM) model increasing from 0.25 (0.26) when using the OD to 0.77 (0.55) when using ED. This corresponds to an improvement of approximately 208% and 111%, respectively. Besides the improved accuracy reached with the ED database, the results showed that the RF model provided better soil salinity estimations than the SVM model and that feature selection (i.e., Variance Inflation Factor-VIF and/or Genetic Algorithm-GA) increase both models´ reliability, with GA being the most efficient. This study highlights the potential of machine-learning and Sentinel-2 image combination for soil salinity monitoring in a data-scarce context, and shows the importance of both model and features selection for an optimum machine-learning set-up.

Tumor boundary identification during colorectal cancer surgery can be challenging, and incomplete tumor removal occurs in approximately 10% of the patients operated for advanced rectal cancer. In this paper, a deep learning framework for automatic tumor segmentation in colorectal ultrasound images was developed, to provide real-time guidance on resection margins using intra-operative ultrasound. A colorectal ultrasound dataset was acquired consisting of 179 images from 74 patients, with ground truth tumor annotations based on histopathology results. To address data scarcity, transfer learning techniques were used to optimize models pre-trained on breast ultrasound data for colorectal ultrasound data. A new custom gradient-based loss function (GWDice) was developed, which emphasizes the clinically relevant top margin of the tumor while training the networks. Lastly, ensemble learning methods were applied to combine tumor segmentation predictions of multiple individual models and further improve the overall tumor segmentation performance. Transfer learning outperformed training from scratch, with an average Dice coefficient over all individual networks of 0.78 compared to 0.68. The new GWDice loss function clearly decreased the average tumor margin prediction error from 1.08 mm to 0.92 mm, without compromising the segmentation of the overall tumor contour. Ensemble learning further improved the Dice coefficient to 0.84 and the tumor margin prediction error to 0.67 mm. Using transfer and ensemble learning strategies, good tumor segmentation performance was achieved despite the relatively small dataset. The developed US segmentation model may contribute to more accurate colorectal tumor resections by providing real-time intra-operative feedback on tumor margins.

Large high-quality datasets are essential for building powerful artificial intelligence (AI) algorithms capable of supporting advancement in cardiac clinical research. However, researchers working with electrocardiogram (ECG) signals struggle to get access and/or to build one. The aim of the present work is to shed light on a potential solution to address the lack of large and easily accessible ECG datasets. Firstly, the main causes of such a lack are identified and examined. Afterward, the potentials and limitations of cardiac data generation via deep generative models (DGMs) are deeply analyzed. These very promising algorithms have been found capable not only of generating large quantities of ECG signals but also of supporting data anonymization processes, to simplify data sharing while respecting patients\' privacy. Their application could help research progress and cooperation in the name of open science. However several aspects, such as a standardized synthetic data quality evaluation and algorithm stability, need to be further explored.