Unseen Test Dataset Research Articles

Accelerometer-derived movement features as predictive biomarkers for muscle atrophy in neurocritical care: a prospective cohort study

BackgroundPhysical inactivity and subsequent muscle atrophy are highly prevalent in neurocritical care and are recognized as key mechanisms underlying intensive care unit acquired weakness (ICUAW). The lack of quantifiable biomarkers for inactivity complicates the assessment of its relative importance compared to other conditions under the syndromic diagnosis of ICUAW. We hypothesize that active movement, as opposed to passive movement without active patient participation, can serve as a valid proxy for activity and may help predict muscle atrophy. To test this hypothesis, we utilized non-invasive, body-fixed accelerometers to compute measures of active movement and subsequently developed a machine learning model to predict muscle atrophy.MethodsThis study was conducted as a single-center, prospective, observational cohort study as part of the MINCE registry (metabolism and nutrition in neurointensive care, DRKS-ID: DRKS00031472). Atrophy of rectus femoris muscle (RFM) relative to baseline (day 0) was evaluated at days 3, 7 and 10 after intensive care unit (ICU) admission and served as the dependent variable in a generalized linear mixed model with Least Absolute Shrinkage and Selection Operator regularization and nested-cross validation.ResultsOut of 407 patients screened, 53 patients (age: 59.2 years (SD 15.9), 31 (58.5%) male) with a total of 91 available accelerometer datasets were enrolled. RFM thickness changed − 19.5% (SD 12.0) by day 10. Out of 12 demographic, clinical, nutritional and accelerometer-derived variables, baseline RFM muscle mass (beta − 5.1, 95% CI − 7.9 to − 3.8) and proportion of active movement (% activity) (beta 1.6, 95% CI 0.1 to 4.9) were selected as significant predictors of muscle atrophy. Including movement features into the prediction model substantially improved performance on an unseen test data set (including movement features: R2 = 79%; excluding movement features: R2 = 55%).ConclusionActive movement, as measured with thigh-fixed accelerometers, is a key risk factor for muscle atrophy in neurocritical care patients. Quantifiable biomarkers reflecting the level of activity can support more precise phenotyping of ICUAW and may direct tailored interventions to support activity in the ICU. Studies addressing the external validity of these findings beyond the neurointensive care unit are warranted.Trial registrationDRKS00031472, retrospectively registered on 13.03.2023.

Customized Convolutional Neural Network for Glaucoma Detection in Retinal Fundus Images

Glaucoma is one of the leading causes of permanent blindness and remains a current challenge in the field of ophthalmology. This research aims to present a comprehensive investigation into the development and evaluation of new technology for glaucoma detection in retinal fundus images. The development and evaluation are presented on a customized architecture, using the Convolutional Neural Network (CNN) method. The proposed CNN architecture is designed to address the complex characteristics of glaucoma changes in the identification process. The research dataset consists of 506 retinal images categorized into 117 glaucoma images, 19 suspected glaucoma images, and 370 healthy images. Through our in-depth exploration, we conducted a careful analysis to uncover patterns and fundamental trends related to glaucoma-related features. During the training phase, the proposed CNN achieved outstanding average accuracy, sensitivity, and specificity values of 92.88%, 94.66%, and 89.31%, respectively. In the unseen test dataset, the model demonstrated competitive performance with an accuracy of 80.87%, sensitivity of 85.65%, and specificity of 71.26%. These findings emphasize the potential of the model as a reliable tool for glaucoma detection. The results indicate that the proposed method utilizing a customized CNN architecture is designed for glaucoma detection in retinal fundus images. The presented output results also hold promise for clinical relevance and can be considered an improvement in the care of retinal fundus patients.

IMG-28. AUTOMATIC BRAIN TUMOR VOLUMETRIC ANALYSIS IN MAGNETIC RESONANCE IMAGING GENERALIZABLE TO PEDIATRIC NEURO-ONCOLOGY

Abstract BACKGROUND The prognosis of brain tumors is variable in clinical practice if it only relies on human interpretation of magnetic resonance imaging (MRI). The automatic segmentation of brain tumors in MRI enables quantitative analysis in support of clinical trials and personalized patient care. We developed benchmarked deep learning-based tools that are generalizable to the volumetric quantification of various tumor types across diverse populations. METHODS We participated in the well-established international brain tumor segmentation challenge (BraTS 2023) and benchmarking competition. The challenge made available 4,500 multi-national brain tumor cases with multi-sequence MRIs, including pediatric high-grade gliomas (PED), i.e., high-grade astrocytoma and diffuse midline glioma, and adult gliomas, brain metastases (MET) and intracranial meningiomas (MEN). Each case comprises four MRI volumes: T1, contrast-enhanced T1, T2, and T2-FLAIR. Manual segmentations were provided to establish ground truth for enhancing tumor (ET), tumor core (TC), and whole tumor (WT). Our framework used a model ensemble strategy based on two state-of-the-art deep learning models: a convolutional neural network (nnU-Net) and a vision transformer (Swin UNETR) and was tested for broader applicability across multiple tumor types. The framework was trained on 99, 1,000, and 165 cases and validated on 45, 141, and 31 unseen cases for PED, MEN, and MET, respectively. Automatic segmentations were evaluated by lesion-wise volume overlap (Dice similarity score, DSC) and Hausdorff distance (HD). RESULTS In the evaluation on independent unseen test datasets, our automatic tool was ranked first for PED, third for MEN, and fourth for MET volumetric analysis. Our method resulted in PED lesion-wise DSC of 0.733, 0.782, 0.817 and HD (mm) of 75.93, 25.54, 24.18 for ET, TC, and WT, respectively. CONCLUSIONS These brain tumor volumetric analysis tools are readily available to be efficiently tested on diverse datasets. Automatic MRI analysis provides consistent quantitative data for multi-institutional protocols and clinical trials.

XECGArch: a trustworthy deep learning architecture for interpretable ECG analysis considering short-term and long-term features

Deep learning-based methods have demonstrated high classification performance in the detection of cardiovascular diseases from electrocardiograms (ECGs). However, their blackbox character and the associated lack of interpretability limit their clinical applicability. To overcome existing limitations, we present a novel deep learning architecture for interpretable ECG analysis (xECGArch). For the first time, short- and long-term features are analyzed by two independent convolutional neural networks (CNNs) and combined into an ensemble, which is extended by methods of explainable artificial intelligence (xAI) to whiten the blackbox. To demonstrate the trustworthiness of xECGArch, perturbation analysis was used to compare 13 different xAI methods. We parameterized xECGArch for atrial fibrillation (AF) detection using four public ECG databases (n=9854\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$n = 9854$$\\end{document} ECGs) and achieved an F1 score of 95.43% in AF versus non-AF classification on an unseen ECG test dataset. A systematic comparison of xAI methods showed that deep Taylor decomposition provided the most trustworthy explanations (+24%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$+24\\%$$\\end{document} compared to the second-best approach). xECGArch can account for short- and long-term features corresponding to clinical features of morphology and rhythm, respectively. Further research will focus on the relationship between xECGArch features and clinical features, which may help in medical applications for diagnosis and therapy.

Enhancing Skin Cancer Diagnosis Using Swin Transformer with Hybrid Shifted Window-Based Multi-head Self-attention and SwiGLU-Based MLP.

Skin cancer is one of the most frequently occurring cancers worldwide, and early detection is crucial for effective treatment. Dermatologists often face challenges such as heavy data demands, potential human errors, and strict time limits, which can negatively affect diagnostic outcomes. Deep learning-based diagnostic systems offer quick, accurate testing and enhanced research capabilities, providing significant support to dermatologists. In this study, we enhanced the Swin Transformer architecture by implementing the hybrid shifted window-based multi-head self-attention (HSW-MSA) in place of the conventional shifted window-based multi-head self-attention (SW-MSA). This adjustment enables the model to more efficiently process areas of skin cancer overlap, capture finer details, and manage long-range dependencies, while maintaining memory usage and computational efficiency during training. Additionally, the study replaces the standard multi-layer perceptron (MLP) in the Swin Transformer with a SwiGLU-based MLP, an upgraded version of the gated linear unit (GLU) module, to achieve higher accuracy, faster training speeds, and better parameter efficiency. The modified Swin model-base was evaluated using the publicly accessible ISIC 2019 skin dataset with eight classes and was compared against popular convolutional neural networks (CNNs) and cutting-edge vision transformer (ViT) models. In an exhaustive assessment on the unseen test dataset, the proposed Swin-Base model demonstrated exceptional performance, achieving an accuracy of 89.36%, a recall of 85.13%, a precision of 88.22%, and an F1-score of 86.65%, surpassing all previously reported research and deep learning models documented in the literature.

Learning to Generalize Towards Unseen Domains via a Content-Aware Style Invariant Model for Disease Detection From Chest X-Rays.

Performance degradation due to distribution discrepancy is a longstanding challenge in intelligent imaging, particularly for chest X-rays (CXRs). Recent studies have demonstrated that CNNs are biased toward styles (e.g., uninformative textures) rather than content (e.g., shape), in stark contrast to the human vision system. Radiologists tend to learn visual cues from CXRs and thus perform well across multiple domains. Motivated by this, we employ the novel on-the-fly style randomization modules at both image (SRM-IL) and feature (SRM-FL) levels to create rich style perturbed features while keeping the content intact for robust cross-domain performance. Previous methods simulate unseen domains by constructing new styles via interpolation or swapping styles from existing data, limiting them to available source domains during training. However, SRM-IL samples the style statistics from the possible value range of a CXR image instead of the training data to achieve more diversified augmentations. Moreover, we utilize pixel-wise learnable parameters in the SRM-FL compared to pre-defined channel-wise mean and standard deviations as style embeddings for capturing more representative style features. Additionally, we leverage consistency regularizations on global semantic features and predictive distributions from with and without style-perturbed versions of the same CXR to tweak the model's sensitivity toward content markers for accurate predictions. Our proposed method, trained on CheXpert and MIMIC-CXR datasets, achieves 77.32±0.35, 88.38±0.19, 82.63±0.13 AUCs(%) on the unseen domain test datasets, i.e., BRAX, VinDr-CXR, and NIH chest X-ray14, respectively, compared to 75.56±0.80, 87.57±0.46, 82.07±0.19 from state-of-the-art models on five-fold cross-validation with statistically significant results in thoracic disease classification.

#2067 Prediction of gastrointestinal bleeding hospitalization risk in hemodialysis using machine learning techniques

Abstract Background and Aims INitiativeS on advancing Patients’ outcomes In REnal disease (INSPIRE) is an academia and industry collaboration set forth to identify critical investigations needed to advance the practice of medicine in nephrology. Gastrointestinal bleeding (GIB) is one of the most common types of bleeding events in the kidney dialysis population [1]. We aimed to develop a model to predict GIB hospitalization risk in a hemodialysis (HD) patient within 180 days. We evaluated advanced machine learning algorithm (XGBoost) and compared it to traditional machine learning algorithm (logistic regression) modeling techniques. Method We used data from a kidney care network from 2017 through 2020. We included data from adult dialysis patients (age ≥18 years). GIB related hospitalization was defined based on international classification of diseases (ICD) diagnosis codes recorded as the primary, secondary, or tertiary discharge reason for hospitalization. Two distinct models were created using XGBoost and logistic regression algorithms and using the same dataset. Both models were evaluated using metrics such as area under the receiver operating curve (AUROC), accuracy, sensitivity, and specificity. Missing data was imputed using mean for quantitative data and mode for qualitative data. The dataset then was randomly divided into 60% training, 20% validation and 20% test dataset. The test data, comprising unseen patients and data was used to evaluate the performance of the model. This means that the model had never encountered the test data during its learning phase, which included training the model and then validating the model. Results The incidence of 180-day GIB hospitalization was 1.12% in the HD population (n = 5 116 with GIB hospitalization/ n = 451 653 without GIB hospitalization), and consistent in the unseen test dataset (n = 465 with GIB hospitalization/ n = 38586 without GIB hospitalization). The XGBoost model showed higher predictive performance compared to logistic regression (Table). The AUROC was 0.72 (95% confidence interval (CI) 0.69, 0.74) for the XGBoost model. In comparison, the AUROC was 0.615 (95% CI 0.60, 0.64) for the logistic regression model. Specificity was 67% versus 58% for the XGBoost model versus logistic regression model respectively. However, sensitivity for both the models was 65%. The top predictors for major GIB were consistent in both models for many factors, yet some factors showed different importance on outcome prediction (Fig. 1). Higher risk of GIB hospitalization was associated with older age, lower ferritin levels, and recent all-cause hospitalizations in both models. The XGBoost showed high importance on outcome prediction for lower hemoglobin and higher serum 25 hydroxy (25OH) vitamin D values, and the logistic regression model showed high importance for higher amounts of saline delivered during an HD treatment and lower intact parathyroid hormone (iPTH) levels. Conclusion We found advanced machine learning prediction modeling (XGBoost) appears suitable for identifying a HD patient at risk for a GIB hospitalization in the next 180 days, and outperforms traditional machine learning (logistic regression) modeling techniques. Although both models showed identical sensitivity, the XGBoost model had higher specificity. Prospective testing is needed to confirm the model's performance. The association between bone mineral metabolism markers and GIB hospitalization risk is unexpected and warrants further investigation.

An End-to-End Artificial Intelligence of Things (AIoT) Solution for Protecting Pipeline Easements against External Interference-An Australian Use-Case.

High-pressure pipelines are critical for transporting hazardous materials over long distances, but they face threats from third-party interference activities. Preventive measures are implemented, but interference accidents can still occur, making the need for high-quality detection strategies vital. This paper proposes an end-to-end Artificial Intelligence of Things (AIoT) solution to detect potential interference threats in real time. The solution involves developing a smart visual sensor capable of processing images using state-of-the-art computer vision algorithms and transmitting alerts to pipeline operators in real time. The system's core is based on the object-detection model (e.g., You Only Look Once version 4 (YOLOv4) and DETR with Improved deNoising anchOr boxes (DINO)), trained on a custom Pipeline Visual Threat Assessment (Pipe-VisTA) dataset. Among the trained models, DINO was able to achieve the best Mean Average Precision (mAP) of 71.2% for the unseen test dataset. However, for the deployment on a limited computational-ability edge computer (i.e., the NVIDIA Jetson Nano), the simpler and TensorRT-optimized YOLOv4 model was used, which achieved a mAP of 61.8% for the test dataset. The developed AIoT device captures the image using a camera, processes on the edge using the trained YOLOv4 model to detect the potential threat, transmits the threat alert to a Fleet Portal via LoRaWAN, and hosts the alert on a dashboard via a satellite network. The device has been fully tested in the field to ensure its functionality prior to deployment for the SEA Gas use-case. The AIoT smart solution has been deployed across the 10km stretch of the SEA Gas pipeline across the Murray Bridge section. In total, 48 AIoT devices and three Fleet Portals are installed to ensure the line-of-sight communication between the devices and portals.

Self-supervised learning of shedding droplet dynamics during steam condensation

Knowledge of condensate shedding droplet dynamics provides important information for the characterization of two-phase heat and mass transfer phenomena. Detecting and segmenting the droplets during shedding requires considerable time and effort if performed manually. Here, we developed a self-supervised deep learning model for segmenting shedding droplets from a variety of dropwise and filmwise condensing surfaces. The model eliminates the need for image annotation by humans in the training step and, therefore, reduces labor significantly. The trained model achieved an average accuracy greater than 0.9 on a new unseen test dataset. After extracting the shedding droplet size and speed, we developed a data-driven model for shedding droplet dynamics based on condensation heat flux and surface properties such as wettability and tube diameter. Our results demonstrate that condensate droplet departure size is both heat flux and tube size dependent and follows different trends based on the condensation mode. The results of this work provide an annotation-free methodology for falling droplet segmentation as well as a statistical understanding of droplet dynamics during condensation.

Generalizable Polyp Segmentation via Randomized Global Illumination Augmentation.

Accurately segmenting polyps from colonoscopy images is essential for diagnosing colorectal cancer. Despite the tremendous success of the deep convolutional neural networks in automatic polyp segmentation, it suffers from domain shift issues, where the trained model yields performance deterioration on unseen test datasets. This paper proposes an illumination enhancement-based domain generalization approach to improve the generalization capability of the model on unseen test datasets and alleviate this issue. In particular, an image decomposition module (IDM) was developed to separate colonoscopy images into reflectance, local, and global illumination components. An illumination transform module (ITM) was proposed to augment images with different global illuminations by synthesizing target-like global illumination maps. A novel illumination variance insensitiveness (IViSen) is also introduced to evaluate the robustness of the model against illumination disturbance. IViSen is easy to compute and correlates well with model generalizability. The segmentation performance of the proposed model on four colonoscopy datasets was examined: CVC-ClinicDB, CVC-ColonDB, ETIS-Larib, and Kvasir-SEG. The method outperformed the competitive methods when tested on unseen domains. In particular, the proposed approach yielded 60.82% and 53.19% in terms of mean Dice and IoU, respectively, with 2.06% and 2.31% improvements.

Harnessing the Power of Voice: A Deep Neural Network Model for Alzheimer's Disease Detection.

Voice, reflecting cerebral functions, holds potential for analyzing and understanding brain function, especially in the context of cognitive impairment (CI) and Alzheimer's disease (AD). This study used voice data to distinguish between normal cognition and CI or Alzheimer's disease dementia (ADD). This study enrolled 3 groups of subjects: 1) 52 subjects with subjective cognitive decline; 2) 110 subjects with mild CI; and 3) 59 subjects with ADD. Voice features were extracted using Mel-frequency cepstral coefficients and Chroma. A deep neural network (DNN) model showed promising performance, with an accuracy of roughly 81% in 10 trials in predicting ADD, which increased to an average value of about 82.0%±1.6% when evaluated against unseen test dataset. Although results did not demonstrate the level of accuracy necessary for a definitive clinical tool, they provided a compelling proof-of-concept for the potential use of voice data in cognitive status assessment. DNN algorithms using voice offer a promising approach to early detection of AD. They could improve the accuracy and accessibility of diagnosis, ultimately leading to better outcomes for patients.

Placental vessel segmentation and registration in fetoscopy: Literature review and MICCAI FetReg2021 challenge findings

Fetoscopy laser photocoagulation is a widely adopted procedure for treating Twin-to-Twin Transfusion Syndrome (TTTS). The procedure involves photocoagulation pathological anastomoses to restore a physiological blood exchange among twins. The procedure is particularly challenging, from the surgeon’s side, due to the limited field of view, poor manoeuvrability of the fetoscope, poor visibility due to amniotic fluid turbidity, and variability in illumination. These challenges may lead to increased surgery time and incomplete ablation of pathological anastomoses, resulting in persistent TTTS. Computer-assisted intervention (CAI) can provide TTTS surgeons with decision support and context awareness by identifying key structures in the scene and expanding the fetoscopic field of view through video mosaicking. Research in this domain has been hampered by the lack of high-quality data to design, develop and test CAI algorithms. Through the Fetoscopic Placental Vessel Segmentation and Registration (FetReg2021) challenge, which was organized as part of the MICCAI2021 Endoscopic Vision (EndoVis) challenge, we released the first large-scale multi-center TTTS dataset for the development of generalized and robust semantic segmentation and video mosaicking algorithms with a focus on creating drift-free mosaics from long duration fetoscopy videos. For this challenge, we released a dataset of 2060 images, pixel-annotated for vessels, tool, fetus and background classes, from 18 in-vivo TTTS fetoscopy procedures and 18 short video clips of an average length of 411 frames for developing placental scene segmentation and frame registration for mosaicking techniques. Seven teams participated in this challenge and their model performance was assessed on an unseen test dataset of 658 pixel-annotated images from 6 fetoscopic procedures and 6 short clips. For the segmentation task, overall baseline performed was the top performing (aggregated mIoU of 0.6763) and was the best on the vessel class (mIoU of 0.5817) while team RREB was the best on the tool (mIoU of 0.6335) and fetus (mIoU of 0.5178) classes. For the registration task, overall the baseline performed better than team SANO with an overall mean 5-frame SSIM of 0.9348. Qualitatively, it was observed that team SANO performed better in planar scenarios, while baseline was better in non-planner scenarios. The detailed analysis showed that no single team outperformed on all 6 test fetoscopic videos. The challenge provided an opportunity to create generalized solutions for fetoscopic scene understanding and mosaicking. In this paper, we present the findings of the FetReg2021 challenge, alongside reporting a detailed literature review for CAI in TTTS fetoscopy. Through this challenge, its analysis and the release of multi-center fetoscopic data, we provide a benchmark for future research in this field.

Cross-View Recurrence-Based Self-Supervised Super-Resolution of Light Field

Compared with external-supervised learning-based (ESLB) methods, self-supervised learning-based (SSLB) methods can overcome the domain gap problem caused by different light field (LF) acquisition conditions, which results in the performance degradation of light field super-resolution on unseen test datasets. Current SSLB methods exploit the cross-scale recurrence feature in the single view image for super-resolution, ignoring the correlation information among views. Different from previous works, we propose a cross-view recurrence-based self-supervised mapping framework to correlate complementary information among views in the down-scaled input LF. Specifically, the cross-view recurrence information consists of geometry structure features and similar structure features. The former is to provide sub-pixel information according to disparity correlations among adjacent views, and the latter is to acquire similar color and contour information among arbitrary views, which can compensate for error disparity guidance of geometry structure features in sharp variance areas. Moreover, instead of the widely used “All-to-All” strategy, we propose a “Part-to-Part” mapping strategy, which is better competent for SSLB approaches with limited training examples solely extracted from input LF. Finally, considering that self-supervised methods need to retrain from the beginning toward each test image, based on the proposed “part-to-part” strategy, an efficient end-to-end network is designed to extract these cross-view features for superior SASR performance with less training time. Experiment results demonstrate that our method outperforms other state-of-the-art ESLB methods on both large and small domain gap cases. Compared with the only SSLB method (LFZSSR [1]), our approach achieves better performance with 524 times less training time.

IoT-Based Object-Detection System to Safeguard Endangered Animals and Bolster Agricultural Farm Security

Significant threats to ecological equilibrium and sustainable agriculture are posed by the extinction of animal species and the subsequent effects on farms. Farmers face difficult decisions, such as installing electric fences to protect their farms, although these measures can harm animals essential for maintaining ecological equilibrium. To tackle these essential issues, our research introduces an innovative solution in the form of an object-detection system. In this research, we designed and implemented a system that leverages the ESP32-CAM platform in conjunction with the YOLOv8 object-detection model. Our proposed system aims to identify endangered species and harmful animals within farming environments, providing real-time alerts to farmers and endangered wildlife by integrating a cloud-based alert system. To train the YOLOv8 model effectively, we meticulously compiled diverse image datasets featuring these animals in agricultural settings, subsequently annotating them. After that, we tuned the hyperparameter of the YOLOv8 model to enhance the performance of the model. The results from our optimized YOLOv8 model are auspicious. It achieves a remarkable mean average precision (mAP) of 92.44% and an impressive sensitivity rate of 96.65% on an unseen test dataset, firmly establishing its efficacy. After achieving an optimal result, we employed the model in our IoT system and when the system detects the presence of these animals, it immediately activates an audible buzzer. Additionally, a cloud-based system was utilized to notify neighboring farmers effectively and alert animals to potential danger. This research’s significance lies in its potential to drive the conservation of endangered species while simultaneously mitigating the agricultural damage inflicted by these animals.

CESCProg: a compact prognostic model and nomogram for cervical cancer based on miRNA biomarkers.

Cervical squamous cell carcinoma, more commonly cervical cancer, is the fourth common cancer among women worldwide with substantial burden of disease, and less-invasive, reliable and effective methods for its prognosis are necessary today. Micro-RNAs are increasingly recognized as viable alternative biomarkers for direct diagnosis and prognosis of disease conditions, including various cancers. In this work, we addressed the problem of systematically developing an miRNA-based nomogram for the reliable prognosis of cervical cancer. Towards this, we preprocessed public-domain miRNA -omics data from cervical cancer patients, and applied a cascade of filters in the following sequence: (i) differential expression criteria with respect to controls; (ii) significance with univariate survival analysis; (iii) passage through dimensionality reduction algorithms; and (iv) stepwise backward selection with multivariate Cox modeling. This workflow yielded a compact prognostic DEmiR signature of three miRNAs, namely hsa-miR-625-5p, hs-miR-95-3p, and hsa-miR-330-3p, which were used to construct a risk-score model for the classification of cervical cancer patients into high-risk and low-risk groups. The risk-score model was subjected to evaluation on an unseen test dataset, yielding a one-year AUROC of 0.84 and five-year AUROC of 0.71. The model was validated on an out-of-domain, external dataset yielding significantly worse prognosis for high-risk patients. The risk-score was combined with significant features of the clinical profile to establish a predictive prognostic nomogram. Both the miRNA-based risk score model and the integrated nomogram are freely available for academic and not-for-profit use at CESCProg, a web-app (https://apalania.shinyapps.io/cescprog).

Machine Learning Identifies a Signature of Nine Exosomal RNAs That Predicts Hepatocellular Carcinoma.

Hepatocellular carcinoma (HCC) is the third leading cause of cancer-related death worldwide. Although alpha fetoprotein (AFP) remains a commonly used serological marker of HCC, the sensitivity and specificity of AFP in detecting HCC is often limited. Exosomal RNA has emerged as a promising diagnostic tool for various cancers, but its use in HCC detection has yet to be fully explored. Here, we employed Machine Learning on 114,602 exosomal RNAs to identify a signature that can predict HCC. The exosomal expression data of 118 HCC patients and 112 healthy individuals were stratified split into Training, Validation and Unseen Test datasets. Feature selection was then performed on the initial training dataset using permutation importance, and the predictive performance of the selected features were tested on the validation dataset using Support Vector Machine (SVM) Classifier. A minimum of nine features were identified to be predictive of HCC and these nine features were then evaluated across six different models in an unseen test set. These features, mainly in the immune, platelet/neutrophil and cytoskeletal pathways, exhibited good predictive performance with ROC-AUC from 0.79-0.88 in the unseen test set. Hence, these nine exosomal RNAs have potential to be clinically useful minimally invasive biomarkers for HCC.

Time-series forecasting of a typical PWR system response under Control Element Assembly withdrawal at full power

To expedite the decision-making process under Nuclear Power Plant (NPP) accident conditions, at a reduced computational cost, a Machine Learning (ML) time-series meta-model is proposed to map the relationships between the NPP's real-time parameters to its transient response and hence forecast its future response. To this end, three variations of Recurrent Neural Networks (RNNs), namely: the Long-Short-Term-Memory (LSTM), Gated Recurrent Unit (GRU) and a series combination of a convolutional neural network (CNN) with LSTM were developed. The NPP response was simulated using the thermal hydraulics best estimate code, MARS-KS, and cross-validated against the Design Control Document (DCD) of APR1400. The time-series database required to train, test, and validate the time-series Machine Learning (ML) expert algorithm was acquired by coupling DAKOTA software with the best estimate thermal hydraulics system code, MARS-KS to conduct a Best Estimate Plus Uncertainty (BEPU) analysis. The Monte-Carlo random propagation approach and the Latin Hypercube Sampling technique were used to sample and propagate key uncertain parameters until a statistically significant sample of the NPP transient responses was achieved. All three RNN models were developed and optimized with the help of the Talos tool, and were found to exhibit reasonable performance in forecasting the key most probable NPP transient responses, Furthermore, the uncertainty of the ML predictions was assessed by integrating a Bayesian Neural Network (BNN) with the LSTM model. The prediction accuracy ranged from 86% to 96%, when evaluated on an unseen test dataset. Based on a voting system with five performance metrics (MAE, MSE, RMSE, R2 and Accuracy) the LSTM model and its derivative CNN+LSTM outperformed the GRU model; while the CNN +LSTM is found to be the most computationally efficient.

Predicting Carbonate Chemistry on the Northwest Atlantic Shelf Using Neural Networks

AbstractThe Northwest Atlantic Shelf (NAS) region has experienced accelerated warming, heatwaves, and is susceptible to ocean acidification, yet also suffers from a paucity of carbonate chemistry observations, particularly at depth. We address this critical data gap by developing three different neural network models to predict dissolved inorganic carbon (DIC) and total alkalinity (TA) in the NAS region from more readily available hydrographic and satellite data. The models predicted DIC withr2between 0.913–0.963 and root mean square errors (RMSE) between 15.4–23.7 (μmol kg−1) and TA withr2between 0.986–0.983 and RMSE between 9.0–10.4 (μmol kg−1) on an unseen test data set that was not used in training the models. Cross‐validation analysis revealed that all models were insensitive to the choice of training data and had good generalization performance on unseen data. Uncertainty in DIC and TA were low (coefficients of variation 0.1%–1%). Compared with other predictive models of carbonate system variables in this region, a larger and more diverse data set with full seasonal coverage and a more sophisticated model architecture resulted in a robust predictive model with higher accuracy and precision across all seasons. We used one of the models to generate a reconstructed seasonal distribution of carbonate chemistry fields based on DIC and TA predictions that shows a clear seasonal progression and large spatial gradients consistent with observations. The distinct models will allow for a range of applications based on the predictor variables available and will be useful to understand and address ocean sustainability challenges.

Plasmid permissiveness of wastewater microbiomes can be predicted from 16S rRNA sequences by machine learning.

Wastewater treatment plants (WWTPs) harbor a dense and diverse microbial community. They constantly receive antimicrobial residues and resistant strains, and therefore provide conditions for horizontal gene transfer (HGT) of antimicrobial resistance (AMR) determinants. This facilitates the transmission of clinically important genes between, e.g. enteric and environmental bacteria, and vice versa. Despite the clinical importance, tools for predicting HGT remain underdeveloped. In this study, we examined to which extent water cycle microbial community composition, as inferred by partial 16S rRNA gene sequences, can predict plasmid permissiveness, i.e. the ability of cells to receive a plasmid through conjugation, based on data from standardized filter mating assays using fluorescent bio-reporter plasmids. We leveraged a range of machine learning models for predicting the permissiveness for each taxon in the community, representing the range of hosts a plasmid is able to transfer to, for three broad host-range resistance IncP plasmids (pKJK5, pB10, and RP4). Our results indicate that the predicted permissiveness from the best performing model (random forest) showed a moderate-to-strong average correlation of 0.49 for pB10 [95% confidence interval (CI): 0.44-0.55], 0.43 for pKJK5 (0.95% CI: 0.41-0.49), and 0.53 for RP4 (0.95% CI: 0.48-0.57) with the experimental permissiveness in the unseen test dataset. Predictive phylogenetic signals occurred despite the broad host-range nature of these plasmids. Our results provide a framework that contributes to the assessment of the risk of AMR pollution in wastewater systems. The predictive tool is available as an application at https://github.com/DaneshMoradigaravand/PlasmidPerm.

#2868 PREDICTION OF MAJOR GASTROINTESTINAL BLEEDING EVENTS IN HEMODIALYSIS

Abstract Background and Aims INitiativeS on advancing Patients’ outcomes In REnal disease (INSPIRE) is an academia and industry collaboration set forth to identify critical investigations needed to advance the practice of medicine in nephrology. At the inaugural INSPIRE meeting, the group came to a consensus that major bleeding events represent potentially preventable complications that occur more frequently in people with kidney disease versus the general population [1]. Given the most common class of bleeding event is gastrointestinal bleeding (GIB) [2], the INSPIRE group assessed if a machine learning model could be developed to assist with determining a hemodialysis (HD) patient's 180-day GIB hospitalization risk. Method Model was developed using adult HD patient data from United States (2017-2020). Patient data was randomly split (50% training, 30% validation, and 20% testing). HD treatments ≤180 days before GIB hospitalization were classified as positive observations, and others were negative observations. Datasets were randomly sampled to build an XGBoost model that considered 386 exposures initially and was refined to the top 50 exposures. Unseen testing dataset was used to determine final model performance. Results The incidence of 180-day GIB hospitalization was 1.18% in the HD population (n = 451,579), and 1.16% among patients in the testing dataset (n = 27,991). The model showed an area under the curve = 0.69, sensitivity = 57.9%, specificity = 68.9%, accuracy = 68.8% and balanced accuracy = 63.4%. Exposures with largest effect size observed via Shapley values were older age (group mean GIB event = 68.2 years vs no GIB event = 63.4 years), shorter days since last all-cause hospital admission (group mean GIB event = 203.2 days vs no GIB event = 253.2 days), and higher serum 25-hydroxy (OH) vitamin D levels from most recent lab (group mean GIB event = 33.4 ng/mL vs no GIB event = 30.5 ng/mL). Other important predictors included lower hemoglobin and iron indices, longer dialysis vintage, and proton pump inhibitor use (Fig. 1). Conclusion The machine learning model appears suitable for early detection of GIB event risk in HD, yet prospective testing is needed. The association between higher 25OH vitamin D and GIB events was unexpected and warrants investigation. A consistent signal has been observed in warfarin users without kidney disease among which 25OH vitamin D levels of 30–100 ng/mL were shown to associate with the highest GIB risk [3].