Performance Of Model Research Articles

A novel hybrid methodology integrating pixel- and object-based techniques for mapping land use and land cover from high-resolution satellite data

ABSTRACT The classification of very high-resolution satellite imagery remains a focal point in remote sensing, attracting increased attention across diverse scientific disciplines. Various classification methods, including pixel- and object-based techniques, have been proposed, and their performances and limitations have been discussed in the literature. This paper presents a hybrid method that combines the strengths of pixel- and object-based methods in image classification to minimize errors associated with the segmentation process, particularly under-segmentation errors in object-based image analysis. The core concept behind the method lies in categorizing segmented image objects as either homogeneous or heterogeneous based on their class probability. In this process, the estimated possibilities from the object-based classification model are considered, and segments are designated as homogeneous or heterogeneous using a user-defined threshold. The object-based classification model determines the class labels for homogeneous image objects, while the heterogeneous ones, containing pixels representing different land cover classes, are classified using the pixel-based model. The performance of hybrid classification models, created by varying thresholds, is analysed using high-resolution WorldView-3 and WorldView-2 imagery and compared with pixel- and object-based classification results. For the implementation of image classification methods, Canonical Correlation Forest (CCF), Extreme Gradient Boosting (XGBoost), and Categorical Boosting (CatBoost) were employed. The findings indicated that employing the suggested hybrid strategy with a threshold value selected within a specific range (e.g. between 60% and 80%) and employing a robust classification algorithm that provides class probabilities (e.g. CCF) results in a statistically significant improvement in overall accuracy compared to pixel and object-based methods, with gains of 5% and 4%, respectively. Visual analysis of the produced thematic maps revealed that the proposed method minimizes the salt-and-pepper effect associated with pixel-based classification while mitigating segmentation problems arising from object-based classification.

Two Characteristic Patterns of the Summer Indian Ocean Dipole

Abstract This study identifies two distinct patterns of the summer Indian Ocean dipole (IOD)—the coastal IOD and the offshore IOD—named based on the proximity of their eastern pole to Sumatra. Their spatial characteristics, evolutionary mechanisms, relationships with ENSO, impacts on precipitation, and the factors controlling the simulation performances of climate models are discussed. The coastal IOD has the same eastern pole as the conventional IOD, which is located off the coast of Sumatra, but its western pole is situated in the central region of the tropical Indian Ocean (TIO). In contrast, the offshore IOD shares the same western pole as the conventional IOD, which is located off the coast of Somalia, whereas its eastern pole is positioned in the central TIO. Regarding their evolutions, while they initially develop similarly, their later evolutions differ due to their distinct pole locations: The offshore IOD peaks in summer, while the coastal IOD can be sustained into autumn. The coastal IOD correlates to preceding and late ENSO states, but the offshore IOD does not, making it an independent internal mode of TIO. The two IODs affect climate differently, with only the coastal IOD affecting Australian rainfall. Climate models exhibit varied levels of performance in simulating the two IODs. Specifically, a stronger link between spring TIO rainfall and ENSO, as well as stronger southeasterly monsoon winds in the southern TIO, can enhance the coastal IOD modeling, while a stronger summer Somali jet benefits the simulation of the offshore IOD. Distinguishing these two IODs has implications for accurate diagnosis and prediction of the summer climate surrounding the TIO. Significance Statement The Indian Ocean dipole (IOD) significantly influences the monsoon climate. Previous studies use a unified IOD index based on all-season sea surface temperature (SST) data of tropical Indian Ocean (TIO), resulting in a representation biased toward the autumn peak season. However, understanding SST patterns during summer is crucial for Asian summer monsoon precipitation, and what are the characteristic patterns of summer IOD remains unclear. This study fundamentally advances the understanding of the IOD’s seasonal diversity by identifying two distinct dipole patterns in the summer TIO. These patterns differ from the conventional IOD in terms of spatial characteristics, evolutions, ENSO relationships, and precipitation impacts. Distinguishing between these two IODs is anticipated to improve the diagnosis and prediction of summer monsoon rainfall.

MACHINE LEARNING METHODS APPLIED IN AIR QUALITY PREDICTION

Air quality is an important environmental component that has a significant influence on public health and well-being. Poor air quality can cause a variety of health problems, including respiratory and cardiovascular disorders. Therefore, there is a growing demand for air quality prediction tools to enable consumers and authorities to take the best decisions and to implement the necessary actions to reduce air pollution. The present paper describes an innovative application that uses machine learning techniques to supply to the users real-time air quality predictions made on past data from their unique location. The Scikit-learn Python package was used to implement five machine learning algorithms, including K-Nearest Neighbors, Random Forest, Gradient Boosting, Support Vector Regression (SVR) and AdaBoost. To achieve robust model performance, compatibility with cross-validation approaches was evaluated. The obtained results indicate that these machine learning techniques are successful at forecasting air quality. The AdaBoost method emerged as the best accurate predictor after extensive investigation, closely followed by Gradient Boosting, SVR, Random Forest, and K-Nearest Neighbors. Furthermore, the investigation also focused on the adapted handling of inaccurate data and providing graphical visualizations to highlight the algorithm's efficacy.

Artificial intelligence algorithm for neoplastic cell percentage estimation and its application to copy number variation in urinary tract cancer.

Bladder cancer is characterized by frequent mutations, which provide potential therapeutic targets for most patients. The effectiveness of emerging personalized therapies depends on an accurate molecular diagnosis, for which the accurate estimation of the neoplastic cell percentage (NCP) is a crucial initial step. However, the established method for determining the NCP, manual counting by a pathologist, is time-consuming and not easily executable. To address this, artificial intelligence (AI) models were developed to estimate the NCP using nine convolutional neural networks and the scanned images of 39 cases of urinary tract cancer. The performance of the AI models was compared to that of six pathologists for 119 cases in the validation cohort. The ground truth value was obtained through multiplexed immunofluorescence. The AI model was then applied to 41 cases in the application cohort that underwent next-generation sequencing testing, and its impact on the copy number variation (CNV) was analyzed. Each AI model demonstrated high reliability, with intraclass correlation coefficients (ICCs) ranging from 0.82 to 0.88. These values were comparable or better to those of pathologists, whose ICCs ranged from 0.78 to 0.91 in urothelial carcinoma cases, both with and without divergent differentiation/ subtypes. After applying AI-driven NCP, 190 CNV (24.2%) were reclassified with 66 (8.4%) and 78 (9.9%) moved to amplification and loss, respectively, from neutral/minor CNV. The neutral/minor CNV proportion decreased by 6%. These results suggest that AI models could assist human pathologists in repetitive and cumbersome NCP calculations.

Significance of plasma p-tau217 in predicting long-term dementia risk in older community residents: Insights from machine learning approaches.

Whether plasma biomarkers play roles in predicting incident dementia among the general population is worth exploring. A total of 1857 baseline dementia-free older adults with follow-ups up to 13.5 years were included from a community-based cohort. The Recursive Feature Elimination (RFE) algorithm aided in feature selection from 90 candidate predictors to construct logistic regression, naive Bayes, bagged trees, and random forest models. Area under the curve (AUC) was used to assess the model performance for predicting incident dementia. During the follow-up of 12,716 person-years, 207 participants developed dementia. Four predictive models, incorporated plasma p-tau217, age, and scores of MMSE, STICK, and AVLT, exhibited AUCs ranging from 0.79 to 0.96 in testing datasets. These models maintained robustness across various subgroups and sensitivity analyses. Plasma p-tau217 outperforms most traditional variables and may be used to preliminarilyscreen older individuals at high risk of dementia. Plasma p-tau217 showed comparable importance with age and cognitive tests in predicting incident dementia among community older adults. Machine learning models combining plasma p-tau217, age, and cognitive tests exhibited excellent performance in predicting incident dementia. The training models demonstrated robustness in subgroup and sensitivity analysis.

Predicting major clinical events among Canadian adults with laboratory-confirmed influenza infection using the influenza severity scale

We developed and validated the Influenza Severity Scale (ISS), a standardized risk assessment for influenza, to estimate and predict the probability of major clinical events in patients with laboratory-confirmed infection. Data from the Canadian Immunization Research Network’s Serious Outcomes Surveillance Network (2011/2012–2018/2019 influenza seasons) enabled the selecting of all laboratory-confirmed influenza patients. A machine learning-based approach then identified variables, generated weighted scores, and evaluated model performance. This study included 12,954 patients with laboratory-confirmed influenza infections. The optimal scale encompassed ten variables: demographic (age and sex), health history (smoking status, chronic pulmonary disease, diabetes mellitus, and influenza vaccination status), clinical presentation (cough, sputum production, and shortness of breath), and function (need for regular support for activities of daily living). As a continuous variable, the scale had an AU-ROC of 0.73 (95% CI, 0.71–0.74). Aggregated scores classified participants into three risk categories: low (ISS < 30; 79.9% sensitivity, 51% specificity), moderate (ISS ≥ 30 but < 50; 54.5% sensitivity, 55.9% specificity), and high (ISS ≥ 50; 51.4% sensitivity, 80.5% specificity). ISS demonstrated a solid ability to identify patients with hospitalized laboratory-confirmed influenza at increased risk for Major Clinical Events, potentially impacting clinical practice and research.

Explainable artificial intelligence approaches for brain-computer interfaces: a review and design space.

Objective.This review paper provides an integrated perspective of Explainable Artificial Intelligence (XAI) techniques applied to Brain-Computer Interfaces (BCIs). BCIs use predictive models to interpret brain signals for various high-stake applications. However, achieving explainability in these complex models is challenging as it compromises accuracy. Trust in these models can be established by incorporating reasoning or causal relationships from domain experts. The field of XAI has emerged to address the need for explainability across various stakeholders, but there is a lack of an integrated perspective in XAI for BCI (XAI4BCI) literature. It is necessary to differentiate key concepts like explainability, interpretability, and understanding, often used interchangeably in this context, and formulate a comprehensive framework.Approach.To understand the need of XAI for BCI, we pose six key research questions for a systematic review and meta-analysis, encompassing its purposes, applications, usability, and technical feasibility. We employ the PRISMA methodology-preferred reporting items for systematic reviews and meta-analyses to review (n = 1246) and analyse (n = 84) studies published in 2015 and onwards for key insights.Main results.The results highlight that current research primarily focuses on interpretability for developers and researchers, aiming to justify outcomes and enhance model performance. We discuss the unique approaches, advantages, and limitations of XAI4BCI from the literature. We draw insights from philosophy, psychology, and social sciences. We propose a design space for XAI4BCI, considering the evolving need to visualise and investigate predictive model outcomes customised for various stakeholders in the BCI development and deployment lifecycle.Significance.This paper is the first to focus solely on reviewing XAI4BCI research articles. This systematic review and meta-analysis findings with the proposed design space prompt important discussions on establishing standards for BCI explanations, highlighting current limitations, and guiding the future of XAI in BCI.

The circular economy through the prism of machine learning and the YouTube video media platform

The transition to a Circular Economy (CE) is rapidly gaining ground across countries and industries. It is the means of achieving more sustainable development by adopting innovative environmentally friendly strategies and saving primary resources. There are several studies indicating the increasing public and corporate interest in the CE but still remain limited in terms of the multitude and utilization of social media data. This work aims to shed light on the most common topics discussed on the YouTube platform, related to the CE. For this reason, we selected 17 videos including the term “Circular Economy” since these have been the most relevant with a sufficient number of comments and views. The model identified two main topics referring to “Sustainable industry and environmental responsibility” and “Circular Economy and resource management” which is a strong indicator of the people's interest in the utilization of resources alongside industrial and corporate activities. The two-topic configuration presented the highest coherence score; however, five and ten-topic configurations have been deployed since there was no extreme differentiation in the model's performance, which could provide more detailed insights. This work's innovation lies in utilizing Machine Learning techniques and social media data to unravel CE's debates.

An imaging-based model to predict the malignant potential of intraductal papillary mucinous neoplasm of the pancreas.

To develop and validate imaging-based models for predicting the malignancy risk of intraductal papillary mucinous neoplasm (IPMN). We retrospectively analyzed data from 241 IPMN patients who underwent preoperative CT and MRI for model development. Cyst size, presence and size of the enhancing mural nodule (EMN), main pancreatic duct (MPD) diameter, thickened/enhancing cyst wall, abrupt MPD caliber change with distal atrophy, and lymphadenopathy were assessed. Multiple logistic regression models predicting malignancy risk were created using either continuous (Model C) or dichotomized variables (Model D) using these imaging features. Validation included internal (n = 55) and external (n = 43) datasets. Model performance was assessed using the area under the receiver operating characteristic curve (AUC) and compared with that of the international guideline-based model (Model F). Model C identified age, EMN size, MPD diameter, and lymphadenopathy as independent predictors on CT, and age and presence and size of EMN on MRI. Model D identified age ≥ 68, cyst size ≥ 31 mm, EMN ≥ 6 mm, MPD ≥ 7 mm, and lymphadenopathy as independent predictors on CT, and age ≥ 68, EMN ≥ 4.5 mm, and lymphadenopathy on MRI. Model C (AUC, 0.763-0.899) performed slightly better than Model D (AUC, 0.753-0.912) without statistical significance. No significant difference was observed between Models C and F (AUC, 0.729-0.952). Combining Model C with obstructive jaundice improved performance (AUC, 0.802-0.941) without statistical significance. Our imaging-based models effectively predicted the malignancy risk of IPMNs, comparable to international consensus guidelines. Imaging features are important for predicting the malignant potential of IPMNs. Our imaging-based model may help determine surgical candidacy for patients with IPMNs. Non-invasively determining the malignant potential of intraductal papillary mucinous neoplasms (IPMNs) allows for appropriate treatment decision-making We identified multiple imaging features that are associated with malignant transformation and developed models for this prediction. Our model performs comparably with international consensus guidelines in predicting the malignant potential of IPMNs.

Natural Language Processing Pretraining Language Model for Computer Intelligent Recognition Technology

Computer intelligent recognition technology refers to the use of computer vision, Natural Language Processing (NLP), machine learning and other technologies to enable computers to recognize, analyze, understand and answer human language and behavior. The common applications of computer intelligent recognition technology include image recognition, NLP, face recognition, target tracking, and other fields. NLP is a field of computer science, which involves the interaction between computers and natural languages. NLP technology can be used to process, analyze and generate natural language data, such as text, voice and image. Common NLP technology applications include language translation, emotion analysis, text classification, speech recognition and question answering system. Language model is a machine learning model, which uses a large number of text data for training to learn language patterns and relationships in text data. Although the language model has made great progress in the past few years, it still faces some challenges, including: poor semantic understanding, confusion in multilingual processing, slow language processing and other shortcomings. Therefore, in order to optimize these shortcomings, this article would study the pre-training language model based on NLP technology, which aimed at using NLP technology to optimize and improve the performance of the language model, thus optimizing the computer intelligent recognition technology. The model had a higher language understanding ability and more accurate prediction ability. In addition, the model could learn language rules and structures by using a large number of corpus, so as to better understand natural language. Through experiments, it could be known that the data size and total computing time of the traditional Generative Pretrained Transformer-2 (GPT-2) language model were 10 GB and 97 hours respectively. The data size and total computing time of BERT (Bidirectional Encoder Representations from Transformer) were 12 GB and 86 hours respectively. The data size and total computing time of the pre-training language model based on NLP were 18 GB and 71 hours respectively. Obviously, the pre-training language model based on NLP had a larger data size and shorter computing time. The experimental results showed that the NLP technology could better optimize the language model and effectively improve its various capabilities. This article opened up a new development direction for computer intelligent recognition technology and provided excellent technical support for the development of language models.

Architecture Optimization for Hybrid Deep Residual Networks in Liver Tumor Segmentation Using a GA

Nowadays, liver tumor segmentation has been widely applied for vital medical objectives such as illness diagnosis, treatment, and evaluation of liver function. In this work, we aim to improve the performance of hybrid convolution neural network (CNN) models for liver tumor segmentation. A new automatic design method is introduced to build a hybrid CNN model with an optimal architecture using a combination of the original U_Net and the Res-UNet. In the proposed design method, the Res-UNet model is divided into different blocks, and then many different candidate architectures for the hybrid CNN model can be created using gradually these blocks to replace their corresponding blocks of the original U-Net. The replacement is executed if the candidate architecture using a residual block gives better performance than the corresponding one that uses the original U-Net block. We use a genetic algorithm (GA) to get the optimal architecture among the candidate architectures that gives the best performance for the hybrid CNN model. During the optimization process, in addition to the optimal architecture, other crucial configurations such as the learning algorithm, learning rate, and batch size are also determined. We refer to the hybrid CNN model that has the best architecture as GA_Res_UNet. Comparison has been made with other CNN models that are commonly used in image segmentation of liver tumors. We use the 3D-IRCADb01 liver tumor dataset that contains 2800 magnetic resonance (MR) images of 20 patients. Each image has a liver tumor with its corresponding mask. All MR images are resized to 256 × 256 in DICOM format. The findings indicate that GA_Res_UNet outperformed the other compared models over the considered dataset. It achieves 98.5% of dice coefficient and 99.8% of accuracy. Moreover, two other liver tumor segmentation datasets (LiTS17 and CHAOS) are utilized in another comparison to show the superior performance of our proposed model. The reported results show that the proposed model outperformed the other compared state-of-the-art models.

Numerical approximation for algorithmic tangent moduli for nonlinear viscoelastic model with CSDA method

This work revisits the notion of complex step derivative approximation (CSDA) and presents its use in constitutive model of a class of nonlinear viscoelastic materials. The effectiveness of a CSDA is evaluated by putting it through a series of straightforward examples. After that, the idea of the CSDA is put to use in order to carry out a numerical evaluation of the algorithmic tangent moduli of a viscoelastic constitutive model. The performance of the constitutive models is evaluated through the use of three different numerical tests, and the results are compared to those that were achieved by the application of an analytical method. In comparison to other numerical differentiation techniques, It has been found that the CSDA scheme is the most computationally efficient and robust method of numerical differentiation, regardless of the size of the finite difference interval.

A Machine Learning Model for Risk Stratification of Postdiagnosis Diabetic Ketoacidosis Hospitalization in Pediatric Type 1 Diabetes: Retrospective Study.

Diabetic ketoacidosis (DKA) is the leading cause of morbidity and mortality in pediatric type 1 diabetes (T1D), occurring in approximately 20% of patients, with an economic cost of $5.1 billion/year in the United States. Despite multiple risk factors for postdiagnosis DKA, there is still a need for explainable, clinic-ready models that accurately predict DKA hospitalization in established patients with pediatric T1D. We aimed to develop an interpretable machine learning model to predict the risk of postdiagnosis DKA hospitalization in children with T1D using routinely collected time-series of electronic health record (EHR) data. We conducted a retrospective case-control study using EHR data from 1787 patients from among 3794 patients with T1D treated at a large tertiary care US pediatric health system from January 2010 to June 2018. We trained a state-of-the-art; explainable, gradient-boosted ensemble (XGBoost) of decision trees with 44 regularly collected EHR features to predict postdiagnosis DKA. We measured the model's predictive performance using the area under the receiver operating characteristic curve-weighted F1-score, weighted precision, and recall, in a 5-fold cross-validation setting. We analyzed Shapley values to interpret the learned model and gain insight into its predictions. Our model distinguished the cohort that develops DKA postdiagnosis from the one that does not (P<.001). It predicted postdiagnosis DKA risk with an area under the receiver operating characteristic curve of 0.80 (SD 0.04), a weighted F1-score of 0.78 (SD 0.04), and a weighted precision and recall of 0.83 (SD 0.03) and 0.76 (SD 0.05) respectively, using a relatively short history of data from routine clinic follow-ups post diagnosis. On analyzing Shapley values of the model output, we identified key risk factors predicting postdiagnosis DKA both at the cohort and individual levels. We observed sharp changes in postdiagnosis DKA risk with respect to 2 key features (diabetes age and glycated hemoglobin at 12 months), yielding time intervals and glycated hemoglobin cutoffs for potential intervention. By clustering model-generated Shapley values, we automatically stratified the cohort into 3 groups with 5%, 20%, and 48% risk of postdiagnosis DKA. We have built an explainable, predictive, machine learning model with potential for integration into clinical workflow. The model risk-stratifies patients with pediatric T1D and identifies patients with the highest postdiagnosis DKA risk using limited follow-up data starting from the time of diagnosis. The model identifies key time points and risk factors to direct clinical interventions at both the individual and cohort levels. Further research with data from multiple hospital systems can help us assess how well our model generalizes to other populations. The clinical importance of our work is that the model can predict patients most at risk for postdiagnosis DKA and identify preventive interventions based on mitigation of individualized risk factors.

Application of Machine Learning in Nanotoxicology: A Critical Review and Perspective.

The massive production and application of nanomaterials (NMs) have raised concerns about the potential adverse effects of NMs on human health and the environment. Evaluating the adverse effects of NMs by laboratory methods is expensive, time-consuming, and often fails to keep pace with the invention of new materials. Therefore, in silico methods that utilize machine learning techniques to predict the toxicity potentials of NMs are a promising alternative approach if regulatory confidence in them can be enhanced. Previous reviews and regulatory OECD guidance documents have discussed in detail how to build an in silico predictive model for NMs. Nevertheless, there is still room for improvement in addressing the ways to enhance the model representativeness and performance from different angles, such as data set curation, descriptor selection, task type (classification/regression), algorithm choice, and model evaluation (internal and external validation, applicability domain, and mechanistic interpretation, which is key to ensuring stakeholder confidence). This review explores how to build better predictive models; the current state of the art is analyzed via a statistical evaluation of literature, while the challenges faced and future perspectives are summarized. Moreover, a recommended workflow and best practices are provided to help in developing more predictive, reliable, and interpretable models that can assist risk assessment as well as safe-by-design development of NMs.

Denoising Drug Discovery Data for Improved Absorption, Distribution, Metabolism, Excretion, and Toxicity Property Prediction.

Predicting absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties of small molecules is a key task in drug discovery. A major challenge in building better ADMET models is the experimental error inherent in the data. Furthermore, ADMET predictors are typically regression tasks due to the continuous nature of the data, which makes it difficult to apply existing denoising methods from other domains as they largely focus on classification tasks. Here, we develop denoising schemes based on deep learning to address this. We find that the training error (TE) can be used to identify the noise in regression tasks while ensemble-based and forgotten event-based metrics fail to detect the noise. The most significant performance increase occurs when the original model is finetuned with the denoised data using TE as the noise detection metric. Our method has the ability to improve models with medium noise and does not degrade the performance of models with noise outside this range (low noise and high noise regimes). To our knowledge, our denoising scheme is the first to improve model performance for ADMET data and has implications for improving models for experimental assay data in general.

Development and validation of the eMCI-CHD tool: A multivariable prediction model for the risk of mild cognitive impairment in patients with coronary heart disease.

This study aimed to develop and validate an eMCI-CHD tool based on clinical data to predict mild cognitive impairment (MCI) risk in patients with coronary heart disease (CHD). This cross-sectional study prospectively collected data from 400 patients with coronary heart disease (aged 55-90 years, 62% men) from July 2022 to September 2023 and randomized (7:3 ratio) them into training and validation sets. After determining the modeling variables through least absolute shrinkage and selection operator regression analysis, four ML classifiers were developed: logistic regression, extreme gradient boosting (XGBoost), support vector machine, and random forest. The performance of the models was evaluated using area under the ROC curve, accuracy, sensitivity, specificity, and F1 score. Decision curve analysis was used to assess the clinical performance of the established models. The SHapley Additive exPlanations (SHAP) method was applied to determine the significance of the features, the predictive model was visualized with a nomogram, and an online web-based calculator for predicting CHD-MCI risk scores was developed. Of 400 CHD patients (average age 70.86±8.74 years), 220 (55%) had MCI. The XGBoost model demonstrated superior performance (AUC: 0.86, accuracy: 78.57%, sensitivity: 0.74, specificity: 0.84, F1: 0.79) and underwent validation. An online tool (https://mr.cscps.com.cn/mci/index.html) with seven predictive variables (APOE gene typing, age, education, TyG index, NT-proBNP, C-reactive protein, and occupation) assessed MCI risk in CHD patients. This study highlights the potential for predicting MCI risk among CHD patients using an ML model-driven nomogram and risk scoring tool based on clinical data.

Enhancing Clinical Relevance of Pretrained Language Models Through Integration of External Knowledge: Case Study on Cardiovascular Diagnosis From Electronic Health Records.

Despite their growing use in health care, pretrained language models (PLMs) often lack clinical relevance due to insufficient domain expertise and poor interpretability. A key strategy to overcome these challenges is integrating external knowledge into PLMs, enhancing their adaptability and clinical usefulness. Current biomedical knowledge graphs like UMLS (Unified Medical Language System), SNOMED CT (Systematized Medical Nomenclature for Medicine-Clinical Terminology), and HPO (Human Phenotype Ontology), while comprehensive, fail to effectively connect general biomedical knowledge with physician insights. There is an equally important need for a model that integrates diverse knowledge in a way that is both unified and compartmentalized. This approach not only addresses the heterogeneous nature of domain knowledge but also recognizes the unique data and knowledge repositories of individual health care institutions, necessitating careful and respectful management of proprietary information. This study aimed to enhance the clinical relevance and interpretability of PLMs by integrating external knowledge in a manner that respects the diversity and proprietary nature of health care data. We hypothesize that domain knowledge, when captured and distributed as stand-alone modules, can be effectively reintegrated into PLMs to significantly improve their adaptability and utility in clinical settings. We demonstrate that through adapters, small and lightweight neural networks that enable the integration of extra information without full model fine-tuning, we can inject diverse sources of external domain knowledge into language models and improve the overall performance with an increased level of interpretability. As a practical application of this methodology, we introduce a novel task, structured as a case study, that endeavors to capture physician knowledge in assigning cardiovascular diagnoses from clinical narratives, where we extract diagnosis-comment pairs from electronic health records (EHRs) and cast the problem as text classification. The study demonstrates that integrating domain knowledge into PLMs significantly improves their performance. While improvements with ClinicalBERT are more modest, likely due to its pretraining on clinical texts, BERT (bidirectional encoder representations from transformer) equipped with knowledge adapters surprisingly matches or exceeds ClinicalBERT in several metrics. This underscores the effectiveness of knowledge adapters and highlights their potential in settings with strict data privacy constraints. This approach also increases the level of interpretability of these models in a clinical context, which enhances our ability to precisely identify and apply the most relevant domain knowledge for specific tasks, thereby optimizing the model's performance and tailoring it to meet specific clinical needs. This research provides a basis for creating health knowledge graphs infused with physician knowledge, marking a significant step forward for PLMs in health care. Notably, the model balances integrating knowledge both comprehensively and selectively, addressing the heterogeneous nature of medical knowledge and the privacy needs of health care institutions.

Development of machine learning models for forecasting the strength of resilient modulus of subgrade soil: genetic and artificial neural network approaches

Accurately predicting the Modulus of Resilience (MR) of subgrade soils, which exhibit non-linear stress–strain behaviors, is crucial for effective soil assessment. Traditional laboratory techniques for determining MR are often costly and time-consuming. This study explores the efficacy of Genetic Programming (GEP), Multi-Expression Programming (MEP), and Artificial Neural Networks (ANN) in forecasting MR using 2813 data records while considering six key parameters. Several Statistical assessments were utilized to evaluate model accuracy. The results indicate that the GEP model consistently outperforms MEP and ANN models, demonstrating the lowest error metrics and highest correlation indices (R2). During training, the GEP model achieved an R2 value of 0.996, surpassing the MEP (R2 = 0.97) and ANN (R2 = 0.95) models. Sensitivity and SHAP (SHapley Additive exPlanations) analysis were also performed to gain insights into input parameter significance. Sensitivity analysis revealed that confining stress (21.6%) and dry density (26.89%) are the most influential parameters in predicting MR. SHAP analysis corroborated these findings, highlighting the critical impact of these parameters on model predictions. This study underscores the reliability of GEP as a robust tool for precise MR prediction in subgrade soil applications, providing valuable insights into model performance and parameter significance across various machine-learning (ML) approaches.

PSA-FL-CDM: A Novel Federated Learning-Based Consensus Model for Post-Stroke Assessment

The rapid development of Internet of Things (IoT) technologies and the potential benefits of employing the vast datasets generated by IoT devices, including wearable sensors and camera systems, has ushered in a new era of opportunities for enhancing smart rehabilitation in various healthcare systems. Maintaining patient privacy is paramount in healthcare while providing smart insights and recommendations. This study proposed the adoption of federated learning to develop a scalable AI model for post-stroke assessment while protecting patients’ privacy. This research compares the centralized (PSA-MNMF) model performance with the proposed scalable federated PSA-FL-CDM model for sensor- and camera-based datasets. The computational time indicates that the federated PSA-FL-CDM model significantly reduces the execution time and attains comparable performance while preserving the patient’s privacy. Impact Statement—This research introduces groundbreaking contributions to stroke assessment by successfully implementing federated learning for the first time in this domain and applying consensus models in each node. It enables collaborative model training among multiple nodes or clients while ensuring the privacy of raw data. The study explores eight different clustering methods independently on each node, revolutionizing data organization based on similarities in stroke assessment. Additionally, the research applies the centralized PSA-MNMF consensus clustering technique to each client, resulting in more accurate and robust clustering solutions. By utilizing the FedAvg federated learning algorithm strategy, locally trained models are combined to create a global model that captures the collective knowledge of all participants. Comparative performance measurements and computational time analyses are conducted, facilitating a fair evaluation between centralized and federated learning models in stroke assessment. Moreover, the research extends beyond a single type of database by conducting experiments on two distinct datasets, wearable and camera-based, broadening the understanding of the proposed methods across different data modalities. These contributions develop stroke assessment methodologies, enabling efficient collaboration and accurate consensus clustering models and maintaining data privacy.

Integrated machine learning and geospatial analysis enhanced gully erosion susceptibility modeling in the Erer watershed in Eastern Ethiopia

Land degradation from gully erosion poses a significant threat to the Erer watershed in Eastern Ethiopia, particularly due to agricultural activities and resource exploitation. Identifying erosion-prone areas and underlying factors using advanced machine learning algorithms (MLAs) and geospatial analysis is crucial for addressing this problem and prioritizing adaptive and mitigating strategies. However, previous studies have not leveraged machine learning (ML) and GIS-based approaches to generate susceptibility maps identifying these areas and conditioning factors, hindering sustainable watershed management solutions. This study aimed to predict gully erosion susceptibility (GES) and identify underlying areas and factors in the Erer watershed. Four ML models, namely, XGBoost, random forest (RF), support vector machine (SVM), and artificial neural network (ANN), were integrated with geospatial analysis using 22 geoenvironmental predictors and 1,200 inventory points (70% used for training and 30% for testing). Model performance and robustness were validated through the area under the curve (AUC), accuracy, precision, sensitivity, specificity, kappa coefficient, F1 score, and logarithmic loss. The relative slope position is most influential, with 100% importance in SVM and RF and 95% importance in XGBoost, while annual rainfall (AR) dominated ANN (100% importance). Notably, XGBoost demonstrated robustness and superior prediction/mapping, achieving an AUC of 0.97, 91% accuracy, 92% precision, and 81% kappa while maintaining a low logloss (0.0394). However, SVM excelled in classifying gully resistant/susceptible areas (97% sensitivity, 98% specificity, and 91% F1 score). The ANN model predicted the most areas with very high gully susceptibility (13.74%), followed by the SVM (11.69%), XGBoost (10.65%), and RF (7.85%) models, while XGBoost identified the most areas with very low susceptibility (70.19%). The ensemble technique was employed to further enhance GES modeling, and it outperformed the individual models, achieving an AUC of 0.99, 93.5% accuracy, 92.5% precision, 97.5% sensitivity, 95.4% specificity, 85.8% kappa, and 94.9% F1 score. This technique also classified the GES of the watershed as 36.48% very low, 26.51% low, 16.24% moderate, 11.55% high, and 9.22% very high. Furthermore, district-level analyses revealed the most susceptible areas, including the Babile, Fedis, Harar, and Meyumuluke districts, with high GES areas of 32.4%, 21.3%, 14.3%, and 13.6%, respectively. This study offers robust and flexible ML models with comprehensive validation metrics to enhance GES modeling and identify gully prone areas and factors, thereby supporting decision-making for sustainable watershed conservation and land degradation prevention.