Learning from Imbalanced Data

The potential associations between acupuncture sensation and brain functional network: a EEG study

Acupuncture has been widely used as an effective treatment for post-stroke rehabilitation. However, the potential association between acupuncture sensation, an important factor influencing treatment efficacy, and brain functional network is unclear. This research sought to reveal and quantify the changes in brain functional network associated with acupuncture sensation. So multi-channel EEG signals were collected from 30 healthy participants and the Massachusetts General Hospital Acupuncture Sensation Scale (MASS) was utilized to assess their needling sensations. Phase Lag Index (PLI) was used to construct the brain functional network, which was analyzed with graph theoretic methods. It showed that in the needle insertion (NI) state the MASS Index was significantly higher than in the needle retention (NR) state (P < 0.001), and the mean values of PLI were also higher than in the Pre-Rest state and NR state significantly (P < 0.01). In the NI state global efficiency, local efficiency, nodal efficiency, and degree centrality were significantly higher than in the Pre-Rest state and the NR state (P < 0.05), while the opposite is true for the shortest path length (P < 0.01). Then Pearson correlation analysis showed a correlation between MASS Index and graph theory metrics (P < 0.05). Finally, Support Vector Regression (SVR) was used to predict the MASS Index with a minimum mean absolute error of 0.65. These findings suggest that the NI state of acupuncture treatment changes the structure of the brain functional network and affects the graph theory metrics of the brain functional network, which may be an objective biomarker for quantitative evaluation of acupuncture sensation.

Supplementary Information

The online version contains supplementary material available at 10.1007/s11571-025-10233-1.

Dissolved organic carbon estimation in lakes: Improving machine learning with data augmentation on fusion of multi-sensor remote sensing observations

This paper presents a novel approach for estimating Dissolved Organic Carbon (DOC) concentrations in lakes considering both carbon sources and sink operators. Despite the critical role of DOC, the combined application of machine learning, as a robust predictor, and remote sensing technology, which reduces costly and time-intensive in-situ sampling, has been underexplored in DOC research. Focusing on lakes over the states of New York, Vermont and Maine (United States, U.S.), this study integrates in-situ DOC measurements with surface reflectance bands obtained from Landsat satellites between 2000 and 2020. Using these bands as inputs of the Random Forest (RF) predictive model, the introduced methodology aims to explore the ability of remote sensing data for large-scale DOC simulation. Initial results indicate low accuracy metrics and significant under-estimation due to the imbalance distribution of DOC samples. Statistical analysis showed that the mean DOC concentration was 5.37±3.37 mg/L (mean±one standard deviation), with peak up to 25 mg/L. A highly skewed distribution of chemical components towards the lower ranges can lead to model misrepresentation of extreme and hazardous events, as they are clouded by unimportant events due to significantly lower occurrence rates. To address this issue, the Synthetic Minority Over-sampling Technique (SMOTE) was applied as a key innovation, generating synthetic samples that enhance RF accuracy and reduce the associated errors. Fusion of in-situ and remote sensing data, combined with machine learning and data augmentation, significantly enhances DOC estimation accuracy, especially in high concentration ranges which are critical for environmental health. With prediction metrics of RMSE = 1.75, MAE = 1.09, and R2 = 0.74, RF-SMOTE significantly improve the metrics obtained from stand-alone RF, particularly in estimating high DOC concentrations. Considering the product spatial resolution of 30 m, the model's output provides potential revenue for global application in lake monitoring, even in remote regions where direct sampling is limited. This novel fusion of remote sensing, machine learning and data augmentation offers valuable insights for water quality management and understanding carbon cycling in aquatic ecosystems.

Machine Learning Classification of Chirality and Optical Rotation Using a Simple One-Hot Encoded Cartesian Coordinate Molecular Representation

Absolute stereochemical configurations and optical rotations were computed for 121,416 molecular structures from the QM9 quantum chemistry data set using density functional theory. A representation for the molecules was developed using Cartesian coordinate geometries and encoded atom types to serve as input for various machine learning algorithms. Classifiers were developed and trained to predict the chirality and signs of optical rotations using a variety of machine learning methods. These methods are compared, and the results demonstrate that machine learning is a viable tool for making predictions of the stereochemical properties of molecules.

From pixels to prognosis: leveraging radiomics and machine learning to predict IDH1 genotype in gliomas

Gliomas are the most common primary tumors of the central nervous system, and advances in genetics and molecular medicine have significantly transformed their classification and treatment. This study aims to predict the IDH1 genotype in gliomas using radiomics and machine learning (ML) methods. Retrospective data from 108 glioma patients were analyzed, including MRI data supported by demographic details such as age, sex, and comorbidities. Tumor segmentation was manually performed using 3D Slicer software, and 112 radiomic features were extracted with the PyRadiomics library. Feature selection using the mRMR algorithm identified 17 significant radiomic features. Various ML algorithms, including KNN, Ensemble, DT, LR, Discriminant and SVM, were applied to predict the IDH1 genotype. The KNN and Ensemble models achieved the highest sensitivity (92-100%) and specificity (100%), emerging as the most successful models. Comparative analyses demonstrated that KNN achieved an accuracy of 92.59%, sensitivity of 92.38%, specificity of 100%, precision of 100%, and an F1-score of 95.02%. Similarly, the Ensemble model achieved an accuracy of 90.74%, sensitivity of 90.65%, specificity of 100%, precision of 100%, and an F1-score of 95.13%. To evaluate their effectiveness, KNN and Ensemble models were compared with commonly used machine learning approaches in glioma classification. LR, a conventional statistical approach, exhibited lower predictive performance with an accuracy of 79.63%, while SVM, a frequently utilized ML model for radiomics-based tumor classification, achieved an accuracy of 85.19%. Our findings are consistent with previous research indicating that radiomics-based ML models achieve high accuracy in IDH1 mutation prediction, with reported performances typically exceeding 80%. These findings suggest that KNN and Ensemble models are more effective in capturing the non-linear radiomic patterns associated with IDH1 status, compared to traditional ML approaches. Our findings indicate that radiomic analyses provide comprehensive genotypic classification by assessing the entire tumor and present a safer, faster, and more patient-friendly alternative to traditional biopsies. This study highlights the potential of radiomics and ML techniques, particularly KNN, Ensemble, and SVM, as powerful tools for predicting the molecular characteristics of gliomas and developing personalized treatment strategies.

Development and validation of a machine learning risk prediction model for asthma attacks in adults in primary care

Primary care consultations provide an opportunity for patients and clinicians to assess asthma attack risk. Using a data-driven risk prediction tool with routinely collected health records may be an efficient way to aid promotion of effective self-management, and support clinical decision making. Longitudinal Scottish primary care data for 21,250 asthma patients were used to predict the risk of asthma attacks in the following year. A selection of machine learning algorithms (i.e., Naïve Bayes Classifier, Logistic Regression, Random Forests, and Extreme Gradient Boosting), hyperparameters, training data enrichment methods were explored, and validated in a random unseen data partition. Our final Logistic Regression model achieved the best performance when no training data enrichment was applied. Around 1 in 3 (36.2%) predicted high-risk patients had an attack within one year of consultation, compared to approximately 1 in 16 in the predicted low-risk group (6.7%). The model was well calibrated, with a calibration slope of 1.02 and an intercept of 0.004, and the Area under the Curve was 0.75. This model has the potential to increase the efficiency of routine asthma care by creating new personalized care pathways mapped to predicted risk of asthma attacks, such as priority ranking patients for scheduled consultations and interventions. Furthermore, it could be used to educate patients about their individual risk and risk factors, and promote healthier lifestyle changes, use of self-management plans, and early emergency care seeking following rapid symptom deterioration.

Exploring species taxonomic kingdom using information entropy and nucleotide compositional features of coding sequences based on machine learning methods

The flow of genetic information from DNA to protein is governed by the central dogma of molecular biology. Genetic drift and mutations usually lead to changes in DNA composition, thereby affecting the coding sequences (CDS) that encode functional proteins. Analyzing the nucleotide distribution in the coding regions of species is crucial for understanding their evolution. In this study, we applied Markov processes to analyze codon formation in 37,031,061 CDSs across 3,735 species genomes, spanning viruses, archaea, bacteria, and eukaryotes, to explore compositional changes. Our results revealed species preferences for different nucleotides. Information entropies and Markov information densities show that eukaryotes exhibit higher redundancy, followed by viruses, suggesting more gene duplication in eukaryotes and high mutation rates in viruses. Evolutionary trends showed an increase in information entropy and a decrease in Markov entropy, with negative correlations between first- and second-order Markov information densities. Furthermore, uniform manifold approximation and projection (UMAP) was used to reduce information redundancy for revealing unique evolutionary patterns in species classification. The machine learning methods demonstrated excellent performance in species classification accuracy, providing profound insights into CDS evolution and protein synthesis.

Fast and effective assessment for individuals with special needs form optimization and prediction models

The aim of this study was to determine which items in the psychological assessment forms used by counselling and research centres for individuals with special needs are effective in classifying individuals into special needs diagnostic categories. Data were obtained from the psychological assessment request forms of 1814 individuals aged 6 to 12 years who were referred to the centre between 2019 and 2023 with suspected special needs and who were classified as having special needs based on psychological and developmental assessments. In addition, we wanted to develop new predictive models using machine learning methods based on these items. Optimizing the psychological assessment application form so that it contains fewer questions may help to make the assessment process faster and more effective. It is expected that the results of this study will make an important contribution to saving time and energy for experts and individuals.

Ecological niche modeling for surveillance of foot-and-mouth disease in South Asia

Control of transboundary diseases at a regional level is commended over the country level due to its inherent complexities. World Organization for Animal Health (WOAH) has established different zones worldwide to control such contagious diseases as foot-and-mouth disease (FMD). Controlling FMD is difficult because of the complicated connection between FMD risk factors, and the deficits of surveillance activities in countries. We used an ecological niche model (ENM) that accounts for the under-reporting of outbreaks to determine FMD risk and risk factors in South Asian countries India, Bangladesh, and Sri Lanka. Centered on known outbreak information, we predicted high-risk areas using similar regional ecological features. Using a multi-algorithm machine-learning ensemble that includes random forest, support vector, and gradient boosting, 15 predictive variables (i.e., livestock densities, land cover, and climate), 660 FMD outbreaks from 13 years (2009-2022) in the region including the outbreaks from India, Bangladesh, and Sri Lanka we identified that Sri Lanka and Bangladesh appeared to have low to medium outbreak risk in the range of 0.04 to 0.55. India was used to fit the model. The machine learning models demonstrated high predictive performance (accuracy >0.87) through cross-validation. Production systems, isothermality, cattle density (per Km2), and mean diurnal range was identified as the most important predictors of FMD outbreaks. These models help to determine FMD low-risk areas to minimize FMD surveillance activities and high-risk areas to focus on performing additional confirmatory testing, and improve surveillance in a regional context.

Validation of machine learning-based models to predict and explain the risk of ovarian cancer: a multicentric study on BRCA-mutated patients undergoing risk-reducing salpingo-oophorectomy

Objective

BRCA-mutated women are recommended to undergo bilateral risk-reducing salpingo-oophorectomy (RRSO) after childbearing, due to the lack of effective methods that could be able to early detect the occurrence of ovarian cancer. Thus, predictive machine learning (ML) techniques could be crucial to aid clinicians in identifying high-risk BRCA-mutated patients and determining the appropriate timing for performing RRSO.

Methods

In this work, we addressed this task by developing explainable ML models using clinical data referred to a multicentric cohort of 694 BRCA-mutated patients from six Italian centers (Policlinico Gemelli, IRCCS San Gerardo, Policlinico Bari, Istituto Tumori Regina Elena, Istituto Tumori Giovanni Paolo II, Ospedale F. Miulli), who performed salpingo-oophorectomy, out of which 39 patients showed tumor (5.6%). Data from Istituto Tumori Regina Elena and Policlinico Bari were used as External Validation Cohort (EVC). The other data were employed as Investigational Cohort (IC). Resampling and ensemble techniques were implemented to handle dataset imbalance. Explainable techniques enabled us to identify some protective and risk factors predicted by the models with respect to the task under study.

Results

The best ML model achieved an AUC value of 79.3% (95% CI: 75.3% - 83.0%), an accuracy value of 73.8% (95% CI: 69.6% - 78.2%), a sensitivity value of 66.7% (95% CI: 58.1% - 75.3%), a specificity value of 74.3% (95% CI: 68.7% - 80.0%), and a G-mean value of 70.4% (95% CI: 63.0% - 76.0%) on EVC. Although the model demonstrated good overall performance, its limited sensitivity reduces its effectiveness in this high-risk population. The variables CA125, age and MatoRRSO were found to be the most significant risk factors, in agreement with the clinical perspective. Conversely, variables such as Estroprogestinuse and PregnancyNfdt played a protective factor role.

Conclusion

Our ML proposal explores the intricate relationships between multiple clinical variables, with a particular emphasis on understanding their non-linear associations. However, while our approach provides valuable insights into risk assessment for BRCA-mutated patients, its current predictive capacity does not significantly improve upon existing clinical models.

A predictive model for damp risk in english housing with explainable AI

Damp in residential buildings poses risks to indoor air quality, occupant health, and structural integrity, and affects up to 27% of homes in the England. This study develops a predictive model for damp risk, using 2,073 inspection records from a housing association across 125 local authorities. Homes were labelled as damp (1,630) or non-damp (443), with data supplemented by national Energy Performance Certificate (EPC) records, incorporating building characteristics and energy efficiency indicators. To evaluate model performance, both a balanced dataset (869 homes, 426 damp, 443 non-damp) and a larger imbalanced dataset (2,073 homes) were used. Seven machine learning algorithms were deployed, with the best-performing model achieving 0.636 accuracy on balanced data and 0.793 on imbalanced data. SHAP (SHapley Additive exPlanations) analysis identified heating cost, energy consumption, and wall energy efficiency as the strongest predictors of damp. Statistical tests and causal analysis were applied to interpret SHAP results, offering insights into potential damp risk and mitigations. The findings suggest that machine learning can support early identification of homes likely to develop damp, helping housing managers prioritise interventions before damp issues escalate.

Integrative approaches for predicting protein network perturbations through machine learning and structural characterization

Chromatin remodeling complexes, such as the Saccharomyces cerevisiae INO80 complex, exemplify how dynamic protein interaction networks govern cellular function through a balance of conserved structural modules and context-dependent functional partnerships, as revealed by integrative machine learning and structural mapping approaches. In this study, we explored the INO80 complex using machine learning to predict network changes caused by genetic deletions. Tree-based models outperformed linear approaches, highlighting non-linear relationships within the interaction network. Feature selection identified key INO80 components (e.g., Arp5, Arp8) and cross-compartment features from other remodeling complexes like SWR1 and NuA4, emphasizing shared functional pathways. Perturbation patterns aligned with biological modules, particularly those linked to telomere maintenance and aging, underscoring the functional coherence of these networks. Structural mapping revealed that not all interactions are predictable through proximity alone, particularly with Arp5 and Yta7. By combining structural insights with machine learning, we enhanced predictions of genetic perturbation effects, providing a template for analyzing cross-species homologs (e.g., human INO80) and their disease-associated variants. This integrative approach bridges the gap between static structural data and dynamic functional networks, offering a pathway to disentangle conserved mechanisms from context-dependent adaptations in chromatin biology. SIGNIFICANCE: By leveraging an innovative, integrative machine learning approach, we have successfully predicted and analyzed perturbations in the INO80 network with good accuracy and depth. Our novel combination of machine learning, perturbation analysis, and structural investigation approach has provided crucial insights into the complex's structure-function relationships, shedding new light on its pivotal roles in affected pathways such as telomere maintenance. Our findings not only enhance our understanding of the INO80 complex but also establish a powerful framework for future studies in chromatin biology and beyond. This work represents a step forward in our understanding of chromatin remodeling complexes and their diverse cellular functions, laying the groundwork for future studies that can further refine our computational approaches and experimental techniques in this field.

Evaluating the three-level approach of the U-smile method for imbalanced binary classification

Real-life binary classification problems often involve imbalanced datasets, where the majority class outnumbers the minority class. We previously developed the U-smile method, which comprises the U-smile plot and the BA, RB and I coefficients, to assess the usefulness of a new variable added to a reference prediction model and validated it under class balance. In this study, we evaluated the U-smile method under class imbalance, proposed a three-level approach of the U-smile method, and used the I coefficients as a weighting factor for point size in the U-smile plots of the BA and RB coefficients. Using real data from the Heart Disease dataset and generated random variables, we built logistic regression models to assess four new variables added to the reference model (nested setting). These models were evaluated at seven pre-defined imbalance levels of 1%, 10%, 30%, 50%, 70%, 90% and 99% of the event class. The results of the U-smile method were compared to those of certain traditional measures: Brier skill score, net reclassification index, difference in F1-score, difference in Matthews correlation coefficient, difference in the area under the receiver operating characteristic curve of the new and reference models, and the likelihood-ratio test. The reference model overfitted to the majority class at higher imbalance levels. The BA-RB-I coefficients of the U-smile method identified informative variables across the entire imbalance range. At higher imbalance levels, the U-smile method indicated both prediction improvement in the minority class (positive BA and I coefficients) and reduction in overfitting to the majority class (negative RB coefficients). The U-smile method outperformed traditional evaluation measures across most of the imbalance range. It proved highly effective in variable selection for imbalanced binary classification, making it a useful tool for real-life problems, where imbalanced datasets are prevalent.

A methodological systematic review of validation and performance of sepsis real-time prediction models

Sepsis real-time prediction models (SRPMs) provide timely alerts and may improve patient outcomes but face limited clinical adoption due to inconsistent validation methods and potential biases. Comprehensive evaluation, including external full-window validation with model- and outcome-level metrics, is crucial for real-world effectiveness, yet performance evidence remains scarce. This study systematically reviewed SRPM performance across validation methods, analyzing 91 studies from multiple databases. Only 54.9% applied full-window validation with both metric types. Performance decreased under external and full-window validation, with median AUROCs of 0.886 and 0.861 at 6- and 12-hours pre-onset, dropping to 0.783 in full-window external validation. Median Utility Scores declined from 0.381 in internal to -0.164 in external validation. Combining AUROC and Utility Score identified top-performing SRPMs in 18.7% of studies. Hand-crafted features significantly improved performance. Future research should focus on multi-center datasets, hand-crafted features, multi-metric full-window validation, and prospective trials to support clinical implementation.

Which pathway of the possible sarcopenia algorithm of the AWGS 2019 guideline is the best in Korean community-dwelling older men and women?

OBJECTIVE

To compare the diagnostic accuracy of possible sarcopenia identification pathways, as suggested by Asian Working Group for Sarcopenia (AWGS) in 2019, by gender among Korean community-dwelling older adults.

DESIGN

Cross-sectional analysis of data from 2,129 community-dwelling adults (70-84 years, 50.4% men) enrolled in Korean Frailty and Aging Cohort Study.

METHODS

Based on AWGS 2019 guideline, possible sarcopenia was defined by low handgrip strength (HGS) or slow five-times chair stand test (5CST) time, referred to as "assessments." "Case-findings" (low calf circumference [CC], SARC-F ≥4, or SARC-CalF ≥11) were recommended for screening 'possible sarcopenia' before assessment. For the six 'possible sarcopenia' pathways (combining three case-finding and two assessment tools), area under the curve (AUC) and F1 score are compared.

RESULTS

For case-finding in men, CC demonstrated the highest AUC (0.657) and F1 score (0.504) for predicting sarcopenia compared with SARC-F and SARC-CalF (p <0.001, =0.001). Among men with low CC, ΔAUC between HGS and 5CST was not significant as assessment (p=0.079) (AUCs: 0.763 vs. 0.707; F1 scores: 0.713 vs. 0.650). For case-finding in women, SARC-CalF demonstrated the highest AUC (0.631) and F1 score (0.389) compared with CC and SARC-F (p=0.012, <0.001). Subsequently, ΔAUC between HGS and 5CST was not significant in women (p=0.069) (AUCs: 0.566 vs. 0.636; F1 scores: 0.387 vs. 0.514).

CONCLUSIONS

Based on AWGS 2019 guideline, CC in men and SARC-CalF in women was the best case-finding tool for community-dwelling older adults. After the best case-finding in each gender, two assessment pathways demonstrated insignificant difference in both genders.

BRIEF SUMMARY

For case-finding of possible sarcopenia, using calf circumference in older men and using SARC-CalF in older women demonstrated the highest diagnostic accuracy for predicting sarcopenia. After the best case-finding in each gender, two assessment pathways (handgrip strength and five-times chair stand test) of possible sarcopenia demonstrated insignificant difference in both genders.

GQEO: Nearest neighbor graph-based generalized quadrilateral element oversampling for class-imbalance problem

The class imbalance problem is one of the difficult factors affecting the performance of traditional classifiers. The oversampling technique is the most common way to solve the class imbalance problem. They alleviate the performance impact of the class imbalance problem on traditional machine learning by augmenting minority instance feature representation. However, many SMOTE-based oversampling techniques perform linear interpolation on the line segment between the anchor instance and its nearest neighbor. This type of method only uses local information and ignores the impact of the global neighborhood relationship on the anchor instance. Therefore, inspired by finite element interpolation, a novel generalized quadrilateral element oversampling technique (GQEO) based on k-nearest neighbor graphs is proposed. First, GQEO uses the k-nearest neighbor to search the global neighbor relationship and build the global neighbor relationship graph. Then, the global neighbor graph is searched for nodes forming generalized quadrilateral elements, using planar quadrilaterals as constraints. Finally, in generalized quadrilateral elements, we use one-dimensional shape functions to synthesize minority instances in quadrilateral elements. Experimental results on 30 imbalanced datasets show that GQEO can alleviate the impact of the class imbalance problem and prevent noise from participating in the synthesis process. GQEO obtains competitive results compared to state-of-the-art oversampling techniques that consider minority noises.

Using explainable machine learning to predict the irritation and corrosivity of chemicals on eyes and skin

Contact with specific chemicals often results in corrosive and irritative responses in the eyes and skin, playing a pivotal role in assessing the potential hazards of personal care products, cosmetics, and industrial chemicals to human health. While traditional animal testing can provide valuable information, its high costs, ethical controversies, and significant demand for animals limit its extensive use, particularly during preliminary screening stages. To address these issues, we adopted a computational modeling approach, integrating 3316 experimental data points on eye irritation and 3080 data points on skin irritation, to develop various machine learning and deep learning models. Under the evaluation of the external validation set, the best-performing models for the two tasks achieved balanced accuracies (BAC) of 73.0 % and 75.1 %, respectively. Furthermore, interpretability analyses were conducted at the dataset level, molecular level, and atomic level to provide insights into the prediction outcomes. Analysis of substructure frequencies identified structural alert fragments within the datasets. This information serves as a reference for identifying potentially irritating chemicals. Additionally, a user-friendly visualization interface was developed, enabling non-specialists to easily predict eye and skin irritation potential. In summary, our study provides a new avenue for the assessment of irritancy potential in chemicals used in pesticides, cosmetics, and ophthalmic drugs.

International external validation of the SORG machine learning algorithm for predicting sustained postoperative opioid prescription after anterior cervical discectomy and fusion using a Taiwanese cohort of 1,037 patients

BACKGROUND CONTEXT

Anterior cervical discectomy and fusion (ACDF) is widely performed for cervical spine disorders, with opioids commonly prescribed postoperatively for pain management. However, prolonged opioid use carries significant risks such as dependency and adverse health effects. Predictive models like the SORG machine learning algorithm (SORG-MLA) have been developed to forecast prolonged opioid use post-ACDF. External validation is essential to ensure their effectiveness across different healthcare settings and populations.

PURPOSE

The study aimed to assess the generalizability of the SORG-MLA to a Taiwanese patient cohort for predicting prolonged opioid use after ACDF.

STUDY DESIGN

Retrospective cohort study utilizing data from a tertiary care center in Taiwan.

PATIENT SAMPLE

1,037 patients who underwent ACDF between 2010 and 2018 were included.

OUTCOME MEASURES

The primary outcome was sustained postoperative opioid prescription defined as continuous opioid use for at least 90 days following ACDF.

METHODS

The performance of the SORG-MLA in the validation cohort was assessed using discrimination measures (area under the receiver operating characteristic curve [AUROC] and the area under the precision-recall curve [AUPRC]), calibration, overall performance (Brier Score), and decision curve analysis. Comparing the validation cohort to the developmental revealed significant differences in demographic profiles, medicolegal frameworks, ethnic cultural contexts and key predictors of postoperative opioid use identified by the SORG-MLA. The Taiwanese cohort was characterized by an older age demographic, a lower proportion of female participants, higher smoking prevalence, higher incidence of preoperative myelopathy and radiculopathy, and more frequent use of antidepressants prior to surgery. Conversely, these patients were less likely to have extended preoperative opioid prescriptions beyond 180 days, undergo multilevel ACDF procedures, or be treated with concurrent medications such as Beta-2 agonists, Gabapentin, and ACE inhibitors. This study had no funding source or conflict of interests.

RESULTS

The model demonstrated good discriminative ability, with an AUROC of 0.78 and an AUPRC of 0.35. Calibration curves indicated that the model overestimated the risk of prolonged opioid use. This discrepancy may be attributed to the significantly higher incidence of sustained opioid consumption in the American development cohort, spanning from 2000 to 2018, which was threefold higher than that in the Taiwanese validation cohort between 2010 and 2018 (9.9% [270/2737] vs. 3.3% [34/1037]; p < .01). The Brier score was 0.033, which improved upon the null model's score of 0.040, indicating robust overall performance. Decision curve analysis confirmed the model's clinical utility, demonstrating net benefits across various decision thresholds.

CONCLUSIONS

The SORG-MLA has demonstrated robust discriminative abilities and overall performance when applied to a unique Taiwanese cohort. However, the model exhibited an overestimation of the risk of prolonged opioid use, suggesting the need for recalibration with more contemporary data to reflect current opioid prescription practices, ethnic and cultural differences, and opioid regulations. Following recalibration, integration and prospective validation within the electronic healthcare system should be pursued. This will enable clinicians to proactively identify patients at heightened risk of prolonged opioid use following ACDF.

Knowledge Graph-Based Few-Shot Learning for Label of Medical Imaging Reports

BACKGROUND

The application of artificial intelligence (AI) in the field of automatic imaging report labeling faces the challenge of manually labeling large datasets.

PURPOSE

To propose a data augmentation method by using knowledge graph (KG) and few-shot learning.

METHODS

A KG of lumbar spine X-ray images was constructed, and 2000 data were annotated based on the KG, which were divided into training, validation, and test sets in a ratio of 7:2:1. The training dataset was augmented based on the synonym/replacement attributes of the KG and was the augmented data was input into the BERT (Bidirectional Encoder Representations from Transformers) model for automatic annotation training. The performance of the model under different augmentation ratios (1:10, 1:100, 1:1000) and augmentation methods (synonyms only, replacements only, combination of synonyms and replacements) was evaluated using the precision and F1 scores. In addition, with the augmentation ratio was fixed, iterative experiments were performed by supplementing the data of nodes that perform poorly in the validation set to further improve model's performance.

RESULTS

Prior to data augmentation, the precision was 0.728 and the F1 score was 0.666. By adjusting the augmentation ratio, the precision increased from 0.912 at a 1:10 augmentation ratio to 0.932 at a 1:100 augmentation ratio (P<.05), while F1 score improved from 0.853 at a 1:10 augmentation ratio to 0.881 at a 1:100 augmentation ratio (P<.05). Additionally, the effectiveness of various augmentation methods was compared at a 1:100 augmentation ratio. The augmentation method that combined synonyms and replacements (F1=0.881) was superior to the methods that only used synonyms (F1=0.815) and only used replacements (F1=0.753) (P<.05). For nodes that exhibited suboptimal performance on the validation set, supplementing the training set with target data improved model performance, increasing the average F1 score to 0.979 (P<.05).

CONCLUSION

Based on the KG, this study trained an automatic labeling model of radiology reports using a few-shot data set. This method effectively reduces the workload of manual labeling, improves the efficiency and accuracy of image data labeling, and provides an important research strategy for the application of AI in the domain of automatic labeling of image reports.

Advancing preeclampsia prediction: a tailored machine learning pipeline integrating resampling and ensemble models for handling imbalanced medical data

BACKGROUND

Constructing a predictive model is challenging in imbalanced medical dataset (such as preeclampsia), particularly when employing ensemble machine learning algorithms.

OBJECTIVE

This study aims to develop a robust pipeline that enhances the predictive performance of ensemble machine learning models for the early prediction of preeclampsia in an imbalanced dataset.

METHODS

Our research establishes a comprehensive pipeline optimized for early preeclampsia prediction in imbalanced medical datasets. We gathered electronic health records from pregnant women at the People's Hospital of Guangxi from 2015 to 2020, with additional external validation using three public datasets. This extensive data collection facilitated the systematic assessment of various resampling techniques, varied minority-to-majority ratios, and ensemble machine learning algorithms through a structured evaluation process. We analyzed 4,608 combinations of model settings against performance metrics such as G-mean, MCC, AP, and AUC to determine the most effective configurations. Advanced statistical analyses including OLS regression, ANOVA, and Kruskal-Wallis tests were utilized to fine-tune these settings, enhancing model performance and robustness for clinical application.

RESULTS

Our analysis confirmed the significant impact of systematic sequential optimization of variables on the predictive performance of our models. The most effective configuration utilized the Inverse Weighted Gaussian Mixture Model for resampling, combined with Gradient Boosting Decision Trees algorithm, and an optimized minority-to-majority ratio of 0.09, achieving a Geometric Mean of 0.6694 (95% confidence interval: 0.5855-0.7557). This configuration significantly outperformed the baseline across all evaluated metrics, demonstrating substantial improvements in model performance.

CONCLUSIONS

This study establishes a robust pipeline that significantly enhances the predictive performance of models for preeclampsia within imbalanced datasets. Our findings underscore the importance of a strategic approach to variable optimization in medical diagnostics, offering potential for broad application in various medical contexts where class imbalance is a concern.

DASNet: A Convolutional Neural Network with SE Attention Mechanism for ccRCC Tumor Grading

Clear cell renal cell carcinoma (ccRCC) is the most common form of renal cell carcinoma in adults, comprising approximately 80% of cases. The lethality of ccRCC rises significantly at stage III or beyond, emphasizing the need for early detection to enable timely therapeutic interventions. This study introduces a non-invasive and efficient classification method, Domain Adaptive Squeeze-and-Excitation Network (DASNet), for grading ccRCC through Computed Tomography (CT) images using advanced deep learning and machine learning techniques. The dataset is enhanced using MedAugment technology and balanced to improve generalization and classification performance. To mitigate overfitting, renal angiomyolipoma (AML) samples are incorporated, increasing data diversity and model robustness. EfficientNet and RegNet serve as foundational models, leveraging local feature extraction and Squeeze-and-Excitation (SE) attention mechanisms to enhance recognition accuracy across grades. Furthermore, Domain-Adversarial Neural Networks (DANNs) are employed to maintain consistency between source and target domains, bolstering the model's generalization ability. The proposed model achieves a classification accuracy of 97.50%, demonstrating efficacy in early ccRCC grade identification. These findings not only offer valuable clinical insights but also establish a foundation for broader application of deep learning in tumor detection.

Automated Foveal Avascular Zone Segmentation in Optical Coherence Tomography Angiography Across Multiple Eye Diseases Using Knowledge Distillation

Optical coherence tomography angiography (OCTA) is a noninvasive imaging technique used to visualize retinal blood flow and identify changes in vascular density and enlargement or distortion of the foveal avascular zone (FAZ), which are indicators of various eye diseases. Although several automated FAZ detection and segmentation algorithms have been developed for use with OCTA, their performance can vary significantly due to differences in data accessibility of OCTA in different retinal pathologies, and differences in image quality in different subjects and/or different OCTA devices. For example, data from subjects with direct macular damage, such as in age-related macular degeneration (AMD), are more readily available in eye clinics, while data on macular damage due to systemic diseases like Alzheimer's disease are often less accessible; data from healthy subjects may have better OCTA quality than subjects with ophthalmic pathologies. Typically, segmentation algorithms make use of convolutional neural networks and, more recently, vision transformers, which make use of both long-range context and fine-grained detail. However, transformers are known to be data-hungry, and may overfit small datasets, such as those common for FAZ segmentation in OCTA, to which there is limited access in clinical practice. To improve model generalization in low-data or imbalanced settings, we propose a multi-condition transformer-based architecture that uses four teacher encoders to distill knowledge into a shared base model, enabling the transfer of learned features across multiple datasets. These include intra-modality distillation using OCTA datasets from four ocular conditions: healthy aging eyes, Alzheimer's disease, AMD, and diabetic retinopathy; and inter-modality distillation incorporating color fundus photographs of subjects undergoing laser photocoagulation therapy. Our multi-condition model achieved a mean Dice Index of 83.8% with pretraining, outperforming single-condition models (mean of 83.1%) across all conditions. Pretraining on color fundus photocoagulation images improved the average Dice Index by a small margin on all conditions except AMD (1.1% on single-condition models, and 0.1% on multi-condition models). Our architecture demonstrates potential for broader applications in detecting and analyzing ophthalmic and systemic diseases across diverse imaging datasets and settings.

An Imbalance Regression Approach to Toxicity Prediction of Chemicals for Potential Use in Environmentally Acceptable Lubricants

Lubricants are complex mixtures of chemicals that help machines function at the right level of friction and wear. Lubricant formulation methods are based on empirical experience of chemical substances that have been used as lubricants for decades. In the last years, the discussion about their environmental problem has triggered new legislations resulting in the search for Environmentally Acceptable Lubricants, which should be biodegradable, minimally toxic, and nonbioaccumulative. Finding new chemicals that comply with these three criteria is a long and expensive process that can be boosted by machine learning (ML). In this paper, we are addressing toxicity prediction with machine learning models by exploring the application of ensemble learners to chemicals having imbalanced data distribution. We investigated the effectiveness of sampling techniques to balance the data and improve the performance of the ensemble learning model. The model can predict toxicity for nonundersampled groups, which in our case corresponds to the moderately to highly toxic groups. The results of this work are useful for lubricant formulators since regulations accept moderate-to-highly toxic chemicals in lubricants if their concentration is below 20 wt %.

Impact of imbalanced features on large datasets

The exponential growth of image and video data motivates the need for practical real-time content-based searching algorithms. Features play a vital role in identifying objects within images. However, feature-based classification faces a challenge due to uneven class instance distribution. Ideally, each class should have an equal number of instances and features to ensure optimal classifier performance. However, real-world scenarios often exhibit class imbalances. Thus, this article explores the classification framework based on image features, analyzing balanced and imbalanced distributions. Through extensive experimentation, we examine the impact of class imbalance on image classification performance, primarily on large datasets. The comprehensive evaluation shows that all models perform better with balancing compared to using an imbalanced dataset, underscoring the importance of dataset balancing for model accuracy. Distributed Gaussian (D-GA) and Distributed Poisson (D-PO) are found to be the most effective techniques, especially in improving Random Forest (RF) and SVM models. The deep learning experiments also show an improvement as such.

A Novel Explainable Attention-Based Meta-Learning Framework for Imbalanced Brain Stroke Prediction

The accurate prediction of brain stroke is critical for effective diagnosis and management, yet the imbalanced nature of medical datasets often hampers the performance of conventional machine learning models. To address this challenge, we propose a novel meta-learning framework that integrates advanced hybrid resampling techniques, ensemble-based classifiers, and explainable artificial intelligence (XAI) to enhance predictive performance and interpretability. The framework employs SMOTE and SMOTEENN for handling class imbalance, dynamic feature selection to reduce noise, and a meta-learning approach combining predictions from Random Forest and LightGBM, and further refined by a deep learning-based meta-classifier. The model uses SHAP (Shapley Additive Explanations) to provide transparent insights into feature contributions, increasing trust in its predictions. Evaluated on three datasets, DF-1, DF-2, and DF-3, the proposed framework consistently outperformed state-of-the-art methods, achieving accuracy and F1-Score of 0.992189 and 0.992579 on DF-1, 0.980297 and 0.981916 on DF-2, and 0.981901 and 0.983365 on DF-3. These results validate the robustness and effectiveness of the approach, significantly improving the detection of minority-class instances while maintaining overall performance. This work establishes a reliable solution for stroke prediction and provides a foundation for applying meta-learning and explainable AI to other imbalanced medical prediction tasks.

scCrab: A Reference-Guided Cancer Cell Identification Method based on Bayesian Neural Networks

Cancer is a significant global public health concern, where early detection can greatly enhance curative outcomes. Therefore, the identification of cancer cells holds significant importance as the primary method for cancer diagnosis. The advancement of single-cell RNA sequencing (scRNA-seq) technology has made it possible to address the problem of cancer cell identification at the single-cell level more efficiently with computational methods, as opposed to the time-consuming and less reproducible manual identification methods. However, existing computational methods have shown suboptimal identification performance and a lack of capability to incorporate external reference data as prior information. Here, we propose scCrab, a reference-guided automatic cancer cell identification method, which performs ensemble learning based on a Bayesian neural network (BNN) with multi-head self-attention mechanisms and a linear regression model. Through a series of experiments on various datasets, we systematically validated the superior performance of scCrab in both intra- and inter-dataset predictions. Besides, we demonstrated the robustness of scCrab to dropout rate and sample size, and conducted ablation experiments to investigate the contributions of each component in scCrab. Furthermore, as a dedicated model for cancer cell identification, scCrab effectively captures cancer-related biological significance during the identification process.

Explainable deep learning algorithm for identifying cerebral venous sinus thrombosis-related hemorrhage (CVST-ICH) from spontaneous intracerebral hemorrhage using computed tomography

Background

Misdiagnosis of hemorrhage secondary to cerebral venous sinus thrombosis (CVST-ICH) as arterial-origin spontaneous intracerebral hemorrhage (sICH) can lead to inappropriate treatment and the potential for severe adverse outcomes. The current practice for identifying CVST-ICH involves venography, which, despite being increasingly utilized in many centers, is not typically used as the initial imaging modality for ICH patients. The study aimed to develop an explainable deep learning model to quickly identify ICH caused by CVST based on non-contrast computed tomography (NCCT).

Methods

The study population included patients diagnosed with CVST-ICH and other spontaneous ICH from January 2016 to March 2023 at the Second Affiliated Hospital of Zhejiang University, Taizhou First People's Hospital, Taizhou Hospital, Quzhou Second People's Hospital, and Longyan First People's Hospital. A transfer learning-based 3D U-Net with segmentation and classification was proposed and developed only on admission plain CT. Model performance was assessed using the area under the curve (AUC), sensitivity, and specificity metrics. For further evaluation, the average diagnostic performance of nine doctors on plain CT was compared with model assistance. Interpretability methods, including Grad-CAM++, SHAP, IG, and occlusion, were employed to understand the model's attention.

Findings

An internal dataset was constructed using propensity score matching based on age, initially including 102 CVST-ICH patients (median age: 44 [29, 61] years) and 683 sICH patients (median age: 65 [52, 73] years). After matching, 102 CVST-ICH patients and 306 sICH patients (median age: 50 [40, 62] years) were selected. An external dataset consisted of 38 CVST-ICH and 119 sICH patients from four other hospitals. Validation showed AUC 0·94, sensitivity 0·96, and specificity 0·8 for the internal testing subset; AUC 0·85, sensitivity 0·87, and specificity 0·82 for the external dataset, respectively. The discrimination performance of nine doctors interpreting CT images significantly improved with the assistance of the proposed model (accuracy 0·79 vs 0·71, sensitivity 0·88 vs 0·81, specificity 0·75 vs 0·68, p < 0·05). Interpretability methods highlighted the attention of model to the features of hemorrhage edge appearance.

Interpretation

The present model demonstrated high-performing and robust results on discrimination between CVST-ICH and spontaneous ICH, and aided doctors' diagnosis in clinical practice as well. Prospective validation with larger-sample size is required.

Funding

The work was funded by the National Key R&D Program of China (2023YFE0118900), National Natural Science Foundation of China (No.81971155 and No.81471168), the Science and Technology Department of Zhejiang Province (LGJ22H180004), Medical and Health Science and Technology Project of Zhejiang Province (No.2022KY174), the 'Pioneer' R&D Program of Zhejiang (No. 2024C03006 and No. 2023C03026) and the MOE Frontier Science Center for Brain Science & Brain-Machine Integration, Zhejiang University.

Thy-DAMP: deep artificial neural network model for prediction of thyroid cancer mortality

PURPOSE

Despite the rising incidence of differentiated thyroid cancer (DTC), mortality rates have remained relatively low yet crucial for effective patient management. This study aims to develop a deep neural network capable of predicting mortality in patients with differentiated thyroid cancer.

METHODS

Leveraging data from the Surveillance, Epidemiology, and End Results (SEER) database, we developed Thy-DAMP (Deep Artificial Neural Network Model for Prediction of Thyroid Cancer) to forecast mortality in DTC patients. The dataset comprised demographic, histologic, and staging information. Following data normalization and feature encoding, the dataset was partitioned into training, testing, and validation subsets, with model hyperparameters fine-tuned via cross-validation.

RESULTS

Among the 63,513 patients, the mean age was 48.22 years (SD = 14.8), with 77.32% being female. Papillary carcinoma emerged as the predominant subtype, representing 62.94% of cases. The majority presented with stage I disease (73.96%). Thy-DAMP demonstrated acceptable performance metrics on both the test and validation sets. Sensitivity was 83.24% (95% CI 76.95-88.40%), specificity was 93.53% (95% CI 93.01-94.02%), and accuracy stood at 93.33% (95% CI 92.82-93.83%). The model exhibited a positive predictive value of 19.76% (95% CI 18.20-21.42%) and a negative predictive value of 99.66% (95% CI 99.53-99.75%). Additionally, Thy-DAMP demonstrated a positive likelihood ratio of 12.86 (95% CI 11.62-14.23), a negative likelihood ratio of 0.18 (95% CI 0.13-0.25), and an area under the receiver operating characteristic curve (AUROC) of 0.95. The model was externally validated on a separate dataset with nearly identical performance.

CONCLUSION

Thy-DAMP showcases considerable promise in accurately predicting mortality in DTC patients, leveraging limited set of patient data.

Visualizing risk modification of hypertensive disorders of pregnancy: development and validation of prediction model for personalized interpregnancy weight management

The growing recognition of the importance of interpregnancy weight management in reducing hypertensive disorders of pregnancy (HDP) underscores the importance of effective preventive strategies. However, developing effective systems remains a challenge. We aimed to bridge this gap by constructing a prediction model. This study retrospectively analyzed the data of 1746 women who underwent two childbirths across 14 medical facilities, including both tertiary and primary facilities. Data from 2009 to 2019 were used to create a derivation cohort (n = 1746). A separate temporal-validation cohort was constructed by adding data between 2020 and 2024 (n = 365). Furthermore, the external-validation cohort was constructed using the data from another tertiary center between 2017 and 2023 (n = 340). We constructed a prediction model for HDP development in the second pregnancy by applying logistic regression analysis using 5 primary clinical information: maternal age, pre-pregnancy body mass index, and HDP history; and pregnancy interval and weight change velocity between pregnancies. Model performance was assessed across all three cohorts. HDP in the second pregnancy occurred 7.3% in the derivation, 10.1% in the temporal-validation, and 7.9% in the external-validation cohorts. This model demonstrated strong discrimination, with c-statistics of 0.86, 0.88, and 0.86 for the respective cohorts. Precision-recall area under the curve values were 0.90, 0.85, and 0.91, respectively. Calibration showed favorable intercepts (-0.02 to -0.00) and slopes (0.96-1.02) for all cohorts. In conclusion, this externally validated model offers a robust basis for personalized interpregnancy weight management goals for women planning future pregnancies.

Multimodal deep learning for predicting in-hospital mortality in heart failure patients using longitudinal chest X-rays and electronic health records

Amid an aging global population, heart failure has become a leading cause of hospitalization among older people. Its high prevalence and mortality rates underscore the importance of accurate mortality prediction for swift disease progression assessment and better patient outcomes. The evolution of artificial intelligence (AI) presents new avenues for predicting heart failure mortality. Yet current research has predominantly leveraged structured data and unstructured clinical notes from electronic health records (EHR), underutilizing the prognostic value of chest X-rays (CXRs). This study aims to harness deep learning methodologies to explore the feasibility of enhancing the precision of predicting in-hospital all-cause mortality in heart failure patients using CXRs data. We propose a novel multimodal deep learning network based on the spatially and temporally decoupled Transformer (MN-STDT) for in-hospital mortality prediction in heart failure by integrating longitudinal CXRs and structured EHR data. The MN-STDT captures spatial and temporal information from CXRs through a Hybrid Spatial Encoder and a Distance-Aware Temporal Encoder, ultimately fusing features from both modalities for predictive modeling. Initial pre-training of the spatial encoder was conducted on CheXpert, followed by full model training and evaluation on the MIMIC-IV and MIMIC-CXR datasets for mortality prediction tasks. In a comprehensive view, the MN-STDT demonstrated the best performance, with an AUC-ROC of 0.8620, surpassing all baseline models. Comparative analysis revealed that the AUC-ROC of the multimodal model (0.8620) was significantly higher than that of models using only structured data (0.8166) or chest X-ray data alone (0.7479). This study demonstrates the value of CXRs in the prognosis of heart failure, showing that the combination of longitudinal CXRs with structured EHR data can significantly improve the accuracy of mortality prediction in heart failure. Feature importance analysis based on SHAP provides interpretable decision support, paving the way for potential clinical applications.

Evaluation of machine learning models for predicting xerostomia in adults with head and neck cancer during proton and heavy ion radiotherapy

BACKGROUND AND PURPOSE

Few studies have examined the factors associated with xerostomia during proton and carbon ion radiotherapy for head and neck cancer (HNC), which are reported to have fewer toxic effects compared to traditional photon-based radiotherapy. This study aims to evaluate the performance of machine learning approaches in predicting grade 2 + xerostomia in adults with HNC receiving proton and carbon ion radiotherapy.

MATERIALS AND METHODS

A retrospective study involving 1,769 adults with HNC who completed proton or carbon ion radiotherapy was conducted. Xerostomia was graded using the Radiation Therapy Oncology Group criteria. Eight machine learning models with different combinations sampling methods and class weights were compared to identify the model with the highest balanced accuracy.

RESULTS

The mean age of patients was 47.8 years (range 18-80), with 33.5 % female. The average total radiation dose was 71.0 GyE (SD = 5.7). Grade 1 xerostomia was recorded in 572 patients (32.3 %) and grade 2 in 103 patients (5.8 %). No cases of grade 3 or higher xerostomia were reported. A support vector machine with a linear kernel, a 1:2 positive-to-negative class weight, and SMOTE oversampling achieved the highest balanced accuracy (0.66) and AUC-ROC (0.69) for predicting grade 2 xerostomia, outperforming the logistic regression model (balanced accuracy:0.50, AUC-ROC. 0.67).

CONCLUSION

The prevalence of grade 2 radiation-induced xerostomia during proton and carbon ion radiotherapy was low in adults with HNC, posing challenges for accurate prediction. Further research is needed to develop improved methods for predicting xerostomia during proton and carbon ion radiotherapy.

Advanced Predictive Analytics for Fetal Heart Rate Variability Using Digital Twin Integration

Fetal heart rate variability (FHRV) is a critical indicator of fetal well-being and autonomic nervous system development during labor. Traditional monitoring methods often provide limited insights, potentially leading to delayed interventions and suboptimal outcomes. This study proposes an advanced predictive analytics approach by integrating approximate entropy analysis with a hidden Markov model (HMM) within a digital twin framework to enhance real-time fetal monitoring. We utilized a dataset of 469 fetal electrocardiogram (ECG) recordings, each exceeding one hour in duration, to ensure sufficient temporal information for reliable modeling. The FHRV data were preprocessed and partitioned into parasympathetic and sympathetic components based on downward and non-downward beat detection. Approximate entropy was calculated to quantify the complexity of FHRV patterns, revealing significant correlations with umbilical cord blood gas parameters, particularly pH levels. The HMM was developed with four hidden states representing discrete pH levels and eight observed states derived from FHRV data. By employing the Baum-Welch and Viterbi algorithms for training and decoding, respectively, the model effectively captured temporal dependencies and provided early predictions of the fetal acid-base status. Experimental results demonstrated that the model achieved 85% training and 79% testing accuracy on the balanced dataset distribution, improving from 78% and 71% on the imbalanced dataset. The integration of this predictive model into a digital twin framework offers significant benefits for timely clinical interventions, potentially improving prenatal outcomes.

Structural Similarity, Activity, and Toxicity of Mycotoxins: Combining Insights from Unsupervised and Supervised Machine Learning Algorithms

A large number of mycotoxins and related fungal metabolites have not been assessed in terms of their toxicological impacts. Current methodologies often prioritize specific target families, neglecting the complexity and presence of co-occurring compounds. This work addresses a fundamental question: Can we assess molecular similarity and predict the toxicity of mycotoxins in silico using a defined set of molecular descriptors? We propose a rapid nontarget screening approach for multiple classes of mycotoxins, integrating both unsupervised and supervised machine learning models, alongside molecular and physicochemical descriptors to enhance the understanding of structural similarity, activity, and toxicity. Clustering analyses identify natural clusters corresponding to the known mycotoxin families, indicating that mycotoxins belonging to the same cluster share similar molecular properties. However, topological descriptors play a significant role in distinguishing between acutely toxic and nonacutely toxic compounds. Random forest (RF) and neural networks (NN), combined with molecular descriptors, contribute to improved knowledge and predictive capability regarding mycotoxin toxicity profiles. RF allows the prediction of toxicity using data reflecting mainly structural features and performs well in the presence of descriptors reflecting biological activity. NN models prove to be more sensitive to biological activity descriptors than RF. The use of descriptors encompassing structural complexity and diversity, chirality and symmetry, connectivity, atomic charge, and polarizability, together with descriptors representing lipophilicity, absorption, and permeation of molecules, is crucial for predicting toxicity, facilitating broader toxicological evaluations.

Curvature estimation techniques for advancing neurodegenerative disease analysis: a systematic review of machine learning and deep learning approaches

Neurodegenerative diseases present complex challenges that demand advanced analytical techniques to decode intricate brain structures and their changes over time. Curvature estimation within datasets has emerged as a critical tool in areas like neuroimaging and pattern recognition, with significant applications in diagnosing and understanding neurodegenerative diseases. This systematic review assesses state-of-the-art curvature estimation methodologies, covering classical mathematical techniques, machine learning, deep learning, and hybrid methods. Analysing 105 research papers from 2010 to 2023, we explore how each approach enhances our understanding of structural variations in neurodegenerative pathology. Our findings highlight a shift from classical methods to machine learning and deep learning, with neural network regression and convolutional neural networks gaining traction due to their precision in handling complex geometries and data-driven modelling. Hybrid methods further demonstrate the potential to merge classical and modern techniques for robust curvature estimation. This comprehensive review aims to equip researchers and clinicians with insights into effective curvature estimation methods, supporting the development of enhanced diagnostic tools and interventions for neurodegenerative diseases.

Prediction of tuberculosis treatment outcomes using biochemical makers with machine learning

BACKGROUND

Tuberculosis (TB) continues to pose a significant threat to global public health. Enhancing patient prognosis is essential for alleviating the disease burden.

OBJECTIVE

This study aims to evaluate TB prognosis by incorporating treatment discontinuation into the assessment framework, expanding beyond mortality and drug resistance.

METHODS

Seven feature selection methods and twelve machine learning algorithms were utilized to analyze admission test data from TB patients, identifying predictive features and building prognostic models. SHapley Additive exPlanations (SHAP) were applied to evaluate feature importance in top-performing models.

RESULTS

Analysis of 1,086 TB cases showed that a K-Nearest Neighbor classifier with Mutual Information feature selection achieved an area under the receiver operation curve (AUC) of 0.87 (95% CI: 0.83-0.92). Key predictors of treatment failure included elevated levels of 5'-nucleotidase, uric acid, globulin, creatinine, cystatin C, and aspartate transaminase. SHAP analysis highlighted 5'-nucleotidase, uric acid, and globulin as having the most significant influence on predicting treatment discontinuation.

CONCLUSION

Our model provides valuable insights into TB outcomes based on initial patient tests, potentially guiding prevention and control strategies. Elevated biomarker levels before therapy are associated with increased risk of treatment discontinuation, indicating their potential as early warning indicators.

Recurrent Neural Network/Machine Learning Predictions of Reactive Channels in H+ + C2H4 at ELab = 30 eV: A Prototype of Ion Cancer Therapy Reactions

We present a simplest-level electron nuclear dynamics/machine learning (SLEND/ML) approach to predict chemical properties in ion cancer therapy (ICT) reactions. SLEND is a time-dependent, variational, on-the-fly, and nonadiabatic method. In SLEND, nuclear and electronic parameters determine reactants-to-products trajectories in a quantum phase space; this establishes a mapping between reactants' initial conditions and products' properties. To accelerate simulations, SLEND/ML utilizes a modicum of SLEND trajectories to train ML methods on the aforesaid mapping and employs them to predict chemical properties. We employ SLEND/ML to predict reaction types and products' charges in H+ + C2H4 at ELab = 30 eV, a prototype of ICT reactions involving double-bonded compounds. For reaction predictions, a recurrent neural network (RNN) and k-nearest neighbor method are the best models with 98.23% and 95.13% accuracy. RNN correctly predicts frequent and infrequent reaction types and generalizes over data sets. For charge predictions, the RNN exhibits low mean absolute errors of 0.02-0.07.

Tradeoffs Between Richness and Bias of Augmented Data in Long-Tail Recognition

In long-tail scenarios, models have a very high demand for high-quality data. Information augmentation, as an important class of data-centric methods, has been proposed to improve model performance by expanding the richness and quantity of samples in tail classes. However, the underlying mechanisms behind the effectiveness of information augmentation methods remain underexplored. This has led to reliance on empirical and intricate fine-tuning in the use of information augmentation for long-tail recognition tasks. In this work, we simultaneously consider the richness gain and distribution shift introduced by information augmentation methods and propose effective information gain (EIG) to explore the mechanisms behind the effectiveness of these methods. We find that when the value of the effective information gain appropriately balances the richness gain and distribution shift, the performance of information augmentation methods is fully realized. Comprehensive experiments on long-tail benchmark datasets CIFAR-10-LT, CIFAR-100-LT, and ImageNet-LT demonstrate that using effective information gain to filter augmented data can further enhance model performance without any modifications to the model's architecture. Therefore, in addition to proposing new model architectures, data-centric approaches also hold significant potential in the field of long-tail recognition.

Development of an AI Model for Predicting Methacholine Bronchial Provocation Test Results Using Spirometry

Background/Objectives: The methacholine bronchial provocation test (MBPT) is a diagnostic test frequently used to evaluate airway hyper-reactivity. MBPT is essential for diagnosing asthma; however, it can be time-consuming and resource-intensive. This study aimed to develop an artificial intelligence (AI) model to predict the MBPT results using forced expiratory volume in one second (FEV1) and bronchodilator test measurements from spirometry. Methods: a dataset of spirometry measurements, including Pre- and Post-bronchodilator FEV1, was used to train and validate the model. Results: Among the evaluated models, the multilayer perceptron (MLP) achieved the highest area under the curve (AUC) of 0.701 (95% CI: 0.676-0.725), accuracy of 0.758, and an F1-score of 0.853. Logistic regression (LR) and a support vector machine (SVM) demonstrated comparable performance with AUC values of 0.688, while random forest (RF) and extreme gradient boost (XGBoost) achieved slightly lower AUC values of 0.669 and 0.672, respectively. Feature importance analysis of the MLP model identified key contributing features, including Pre-FEF25-75 (%), Pre-FVC (L), Post FEV1/FVC, Change-FEV1 (L), and Change-FEF25-75 (%), providing insight into the interpretability and clinical applicability of the model. Conclusions: These results highlight the potential of the model to utilize readily available spirometry data, particularly FEV1 and bronchodilator responses, to accurately predict MBPT results. Our findings suggest that AI-based prediction can improve asthma diagnostic workflows by minimizing the reliance on MBPT and enabling faster and more accessible assessments.

Addressing imbalanced data classification with Cluster-Based Reduced Noise SMOTE

In recent years, the challenge of imbalanced data has become increasingly prominent in machine learning, affecting the performance of classification algorithms. This study proposes a novel data-level oversampling method called Cluster-Based Reduced Noise SMOTE (CRN-SMOTE) to address this issue. CRN-SMOTE combines SMOTE for oversampling minority classes with a novel cluster-based noise reduction technique. In this cluster-based noise reduction approach, it is crucial that samples from each category form one or two clusters, a feature that conventional noise reduction methods do not achieve. The proposed method is evaluated on four imbalanced datasets (ILPD, QSAR, Blood, and Maternal Health Risk) using five metrics: Cohen's kappa, Matthew's correlation coefficient (MCC), F1-score, precision, and recall. Results demonstrate that CRN-SMOTE consistently outperformed the state-of-the-art Reduced Noise SMOTE (RN-SMOTE), SMOTE-Tomek Link, and SMOTE-ENN methods across all datasets, with particularly notable improvements observed in the QSAR and Maternal Health Risk datasets, indicating its effectiveness in enhancing imbalanced classification performance. Overall, the experimental findings indicate that CRN-SMOTE outperformed RN-SMOTE in 100% of the cases, achieving average improvements of 6.6% in Kappa, 4.01% in MCC, 1.87% in F1-score, 1.7% in precision, and 2.05% in recall, with setting SMOTE's neighbors' number to 5.

Accurate identification of oxygen desaturation status in COPD by using classifier ensemble

The accurate identification of oxygen desaturation (OD) status plays critical role in the clinic diagnosis of chronic obstructive pulmonary disease (COPD), which is a common disease related to the lungs and respiratory tract of the human body. This paper focuses on a specific type of OD status, i.e., exercise-induced oxygen desaturation (EIOD) status in COPD, and try to further improve the performance of EIOD status identification. We propose a new and effective EIOD status identification method by using classifier ensemble strategy. In the proposed method, five different features of each data point from the time series of SpO2 and pulse are extracted and then combined to form the discriminative feature of the corresponding data point; then, multiple base classifiers with different balanced training subsets are trained and then integrated by using AdaBoost Algorithm. The comparative computational results on the 6-min walk test (6MWT) of the recruited participants show that the proposed method achieved the best global performance with AUC (Area Under Curve) value of 0.8532, indicating that the proposed method can be effectively used for the identification of EIOD and could assist the clinic diagnosis of COPD.

Machine learning (ML) techniques to predict breast cancer in imbalanced datasets: a systematic review

Knowledge discovery in databases (KDD) is crucial in analyzing data to extract valuable insights. In medical outcome prediction, KDD is increasingly applied, particularly in diseases with high incidence, mortality, and costs, like cancer. ML techniques can develop more accurate predictive models for cancer patients' clinical outcomes, aiding informed healthcare decision-making. However, cancer prediction modeling faces challenges because of the unbalanced nature of the datasets, where there is a small minority category of patients with a cancer diagnosis compared to a majority category of cancer-free patients. Imbalanced datasets pose statistical hurdles like bias and overfitting when developing accurate prediction models. This systematic review focuses on breast cancer prediction articles published from 2008 to 2023. The objective is to examine ML methods used in three critical steps of KDD: preprocessing, data mining, and interpretation which address the imbalanced data problem in breast cancer prediction. This work synthesizes prior research in ML methods for breast cancer prediction. The findings help identify effective preprocessing strategies, including balancing and feature selection methods, robust predictive models, and evaluation metrics of those models. The study aims to inform healthcare providers and researchers about effective techniques for accurate breast cancer prediction.

Clinical Prediction Models in Neurocritical Care: An Overview of the Literature and Example Application to Prediction of Hospital Mortality in Traumatic Brain Injury

Clinical prediction models serve as valuable instruments for assessing the risk of crucial outcomes and facilitating decision-making in clinical settings. Constructing these models requires nuanced analytical decisions and expertise informed by the current statistical literature. Access and thorough understanding of such literature may be limited for neurocritical care physicians, which may hinder the interpretation of existing predictive models. The present emphasis is on narrowing this knowledge gap by providing neurocritical care specialists with methodological guidance for interpreting predictive models in neurocritical care. Presented are the statistical learning principles integral to constructing a model predicting hospital mortality (nonsurvival during hospitalization) in patients with moderate and severe blunt traumatic brain injury using components of the IMPACT-Core model. Discussion encompasses critical elements such as model flexibility, hyperparameter selection, data imbalance, cross-validation, model assessment (discrimination and calibration), prediction instability, and probability thresholds. The intricate interplay among these components, the data set, and the clincal context of neurocritical care is elaborated. Leveraging this comprehensive exploration of statistical learning can enhance comprehension of articles encompassing model generation, tailored clinical care, and, ultimately, better interpretation and clinical applicability of predictive models.

Identifying stigmatizing and positive/preferred language in obstetric clinical notes using natural language processing

OBJECTIVE

To identify stigmatizing language in obstetric clinical notes using natural language processing (NLP).

MATERIALS AND METHODS

We analyzed electronic health records from birth admissions in the Northeast United States in 2017. We annotated 1771 clinical notes to generate the initial gold standard dataset. Annotators labeled for exemplars of 5 stigmatizing and 1 positive/preferred language categories. We used a semantic similarity-based search approach to expand the initial dataset by adding additional exemplars, composing an enhanced dataset. We employed traditional classifiers (Support Vector Machine, Decision Trees, and Random Forest) and a transformer-based model, ClinicalBERT (Bidirectional Encoder Representations from Transformers) and BERT base. Models were trained and validated on initial and enhanced datasets and were tested on enhanced testing dataset.

RESULTS

In the initial dataset, we annotated 963 exemplars as stigmatizing or positive/preferred. The most frequently identified category was marginalized language/identities (n = 397, 41%), and the least frequent was questioning patient credibility (n = 51, 5%). After employing a semantic similarity-based search approach, 502 additional exemplars were added, increasing the number of low-frequency categories. All NLP models also showed improved performance, with Decision Trees demonstrating the greatest improvement (21%). ClinicalBERT outperformed other models, with the highest average F1-score of 0.78.

DISCUSSION

Clinical BERT seems to most effectively capture the nuanced and context-dependent stigmatizing language found in obstetric clinical notes, demonstrating its potential clinical applications for real-time monitoring and alerts to prevent usages of stigmatizing language use and reduce healthcare bias. Future research should explore stigmatizing language in diverse geographic locations and clinical settings to further contribute to high-quality and equitable perinatal care.

CONCLUSION

ClinicalBERT effectively captures the nuanced stigmatizing language in obstetric clinical notes. Our semantic similarity-based search approach to rapidly extract additional exemplars enhanced the performances while reducing the need for labor-intensive annotation.

GraphCPP: The new state-of-the-art method for cell-penetrating peptide prediction via graph neural networks

BACKGROUND AND PURPOSE

Cell-penetrating peptides (CPPs) are short amino acid sequences that can penetrate cell membranes and deliver molecules into cells. Several models have been developed for their discovery, yet these models often face challenges in accurately predicting membrane penetration due to the complex nature of peptide-cell interactions. Hence, there is a need for innovative approaches that can enhance predictive performance.

EXPERIMENTAL APPROACH

In this study, we present the application GraphCPP, a novel graph neural network (GNN) for the prediction of membrane penetration capability of peptides.

KEY RESULTS

A new comprehensive dataset-dubbed CPP1708-was constructed resulting in the largest reliable database of CPPs to date. Comparative analyses with previous methods, such as MLCPP2, C2Pred, CellPPD and CellPPD-Mod, demonstrated the superior predictive performance of our model. Upon testing against other published methods, GraphCPP performs exceptionally, achieving 0.5787 Matthews correlation coefficient and 0.8459 area under the curve (AUC) values on one dataset. This means a 92.8% and 23.3% improvement in Matthews correlation coefficient and AUC measures respectively compared with the next best model. The capability of the model to effectively learn peptide representations was demonstrated through t-distributed stochastic neighbour embedding plots. Additionally, the uncertainty analysis revealed that GraphCPP maintains high confidence in predictions for peptides shorter than 40 amino acids. The source code is available at https://github.com/attilaimre99/GraphCPP.

CONCLUSION AND IMPLICATIONS

These findings indicate the potential of GNN-based models to improve CPP penetration prediction and it may contribute towards the development of more efficient drug delivery systems.

Detecting Opioid Use Disorder in Health Claims Data With Positive Unlabeled Learning

Accurate detection and prevalence estimation of behavioral health conditions, such as opioid use disorder (OUD), are crucial for identifying at-risk individuals, determining treatment needs, monitoring prevention and intervention efforts, and recruiting treatment-naive participants for clinical trials. The availability of extensive health data, combined with advancements in machine learning (ML) frameworks, has enabled researchers to employ various ML techniques to predict or identify OUD within patient health data. Ideally, we could directly estimate the prevalence, or the proportion of a population with a condition over time. However, underdiagnosis and undercoding of conditions in patient health records make it challenging to determine the true prevalence of these conditions and to identify at-risk patients with less severe conditions who are more likely to be missed. Consequently, patients without diagnoses may comprise positive and negative examples for a given condition. Treating all undiagnosed (uncoded) patients as negative when applying ML methods can introduce bias into models, affecting their predictive power. To address this issue, we employed Positive Unlabeled Learning Selected Not At Random (PULSNAR), a Positive and Unlabeled (PU) learning technique, to estimate the probability of a given patient having OUD during a time window and the overall population prevalence of OUD. In a sample of 3,342,044 commercially insured US patients with at least one opioid prescription filled, PULSNAR estimated that 5.08% of patients have a cumulative prevalence of OUD over a 2-5 a observation period, compared to the 1.35% with a recorded OUD diagnosis, with 73.5% of cases not diagnosed/coded. The prevalence estimates provided by PULSNAR are consistent with those reported in other studies.

Lung CT-based multi-lesion radiomic model to differentiate between nontuberculous mycobacteria and Mycobacterium tuberculosis

BACKGROUND

Nontuberculous mycobacterial lung disease (NTM-LD) and Mycobacterium tuberculosis lung disease (MTB-LD) are difficult to distinguish based on conventional imaging examinations. In recent years, radiomics has been used to discriminate them. However, existing radiomic methods mainly focus on specific lesion types, and have limitations in handling the presence of multiple lesion types that vary among different patients.

PURPOSE

We aimed to establish a radiomic model based on multiple lesion types in the patient's CT scans, and analyzed the importance of different lesion types in distinguishing the two diseases.

METHODS

120 NTM-LD and 120 MTB-LD patients were retrospectively enrolled in this study and randomly split into the training (168) and testing (72) sets. A total of 1037 radiomic features were extracted separately for each lesion type. The univariate analysis, least absolute shrinkage, and selection operator were used to select the significant radiomic features. The radiomic signature score (Radscore) from each lesion type was estimated and aggregated to construct the multi-lesion feature vector for each patient. A multi-lesion radiomic (MLR) model was then established using the random forest classifier, which can estimate importance coefficients for different lesion types. The performances of the MLR model and single radomic models were investigated by the receiver operating characteristic curve (ROC). The impact of the predicted lesion importance was also evaluated in subjective imaging diagnosis.

RESULTS

The MLR model achieved an area under the curve (AUC) of 90.2% (95% CI: 86.2% 94.1%) in differentiating NTM-LD and MTB-LD, outperforming the models using specific lesion types following existing radiomic models by 1% to 13%. Among different lesion types, tree-in-bud pattern demonstrated the highest distinguishing value, followed by consolidation, nodules, and lymph node enlargement. Given the estimated lesion importance, two senior radiologists exhibited improved accuracy in diagnosis, with an increased accuracy of 8.33% and 8.34%, respectively.

CONCLUSIONS

This is the first radiomic study to use multiple lesion types to distinguish NTM-LD and MTB-LD. The developed MLR model performed well in differentiating the two diseases, and the lesion types with high importance exhibited the potential to assist experienced radiologists in clinical decision-making.

Feature selection enhances peptide binding predictions for TCR-specific interactions

Introduction

T-cell receptors (TCRs) play a critical role in the immune response by recognizing specific ligand peptides presented by major histocompatibility complex (MHC) molecules. Accurate prediction of peptide binding to TCRs is essential for advancing immunotherapy, vaccine design, and understanding mechanisms of autoimmune disorders.

Methods

This study presents a theoretical approach that explores the impact of feature selection techniques on enhancing the predictive accuracy of peptide binding models tailored for specific TCRs. To evaluate our approach across different TCR systems, we utilized a dataset that includes peptide libraries tested against three distinct murine TCRs. A broad range of physicochemical properties, including amino acid composition, dipeptide composition, and tripeptide features, were integrated into the machine learning-based feature selection framework to identify key properties contributing to binding affinity.

Results

Our analysis reveals that leveraging optimized feature subsets not only simplifies the model complexity but also enhances predictive performance, enabling more precise identification of TCR peptide interactions. The results of our feature selection method are consistent with findings from hybrid approaches that utilize both sequence and structural data as input as well as experimental data.

Discussion

Our theoretical approach highlights the role of feature selection in peptide-TCR interactions, providing a quantitative tool for uncovering the molecular mechanisms of the T-cell response and assisting in the design of more advanced targeted therapeutics.

An Efficient Dual-Sampling Approach for Chest CT Diagnosis

Background

This paper aimed to enhance the diagnostic process of lung abnormalities in computed tomography (CT) images, particularly in distinguishing cancer cells from normal chest tissue. The rapid and uneven growth of cancer cells, presenting with variable symptoms, necessitates an advanced approach for accurate identification.

Objective

To develop a dual-sampling network targeting lung infection regions to address the diagnostic challenge. The network was designed to adapt to the uneven distribution of infection areas, which could be predominantly minor or major in different regions.

Methods

A total of 150 CT images were analyzed using the dual-sampling network. Two sampling approaches were compared: the proposed dual-sampling technique and a uniform sampling method.

Results

The dual-sampling network demonstrated superior performance in detecting lung abnormalities compared to uniform sampling. The uniform sampling method, the network results: an F1-Score of 94.2%, accuracy of 94.5%, sensitivity of 93.5%, specificity of 95.4%, and an area under the curve (AUC) of 98.4%. However, with the proposed dual-sampling method, the network reached an F1-score of 94.9%, accuracy of 95.2%, specificity of 96.1%, sensitivity of 94.2%, and an AUC of 95.5%.

Conclusion

This study suggests that the proposed dual-sampling network significantly improves the precision of lung abnormality diagnosis in CT images. This advancement has the potential to aid radiologists in making more accurate diagnoses, ultimately benefiting patient treatment and contributing to better overall population health. The efficiency and effectiveness of the dual-sampling approach in managing the uneven distribution of lung infection areas are key to its success.

Advanced Mass-Spectra-Based Machine Learning for Predicting the Toxicity of Traditional Chinese Medicines

Traditional Chinese medicine (TCM) has been a cornerstone of health care for centuries, valued for its preventive and therapeutic properties. However, recent decades have revealed significant toxicological concerns associated with TCMs due to their complex chemical compositions. Traditional QSAR (quantitative structure-activity relationships) models, which predict toxicity based on chemical structures, face challenges with the intricate nature of TCM compounds. In this study, we effectively resolved this issue by correlating the toxicity of TCMs with advanced analytical descriptors from electron ionization mass spectra (EI-MS) data. The optimal classification model achieved a balanced accuracy of over 0.74. Through interpretable machine learning models, we identified specific toxic components, such as 13-hexyloxacyclotridec-10-en-2-one and loliolide. We applied molecular dynamics (MD) simulations to explore the interactions of identified toxic components with crucial protein targets, using hepatic cytochrome P450 3A4 as an example. This novel approach not only enhances our understanding of the toxicological profiles of TCMs but also maximizes their therapeutic benefits while minimizing adverse effects. More importantly, our findings support the application of analytical descriptor-based machine learning in predicting the toxicity of unknown mixtures in the real environment.

Transfer Learning with a Graph Attention Network and Weighted Loss Function for Screening of Persistent, Bioaccumulative, Mobile, and Toxic Chemicals

In silico methods for screening hazardous chemicals are necessary for sound management. Persistent, bioaccumulative, mobile, and toxic (PBMT) chemicals persist in the environment and have high mobility in aquatic environments, posing risks to human and ecological health. However, lack of experimental data for the vast number of chemicals hinders identification of PBMT chemicals. Through an extensive search of measured chemical mobility data, as well as persistent, bioaccumulative, and toxic (PBT) chemical inventories, this study constructed comprehensive data sets on PBMT chemicals. To address the limited volume of the PBMT chemical data set, a transfer learning (TL) framework based on graph attention network (GAT) architecture was developed to construct models for screening PBMT chemicals, designating the PBT chemical inventories as source domains and the PBMT chemical data set as target domains. A weighted loss (LW) function was proposed and proved to mitigate the negative impact of imbalanced data on the model performance. Results indicate the TL-GAT models outperformed GAT models, along with large coverage of applicability domains and interpretability. The constructed models were employed to identify PBMT chemicals from inventories consisting of about 1 × 106 chemicals. The developed TL-GAT framework with the LW function holds broad applicability across diverse tasks, especially those involving small and imbalanced data sets.

Task relevant autoencoding enhances machine learning for human neuroscience

In human neuroscience, machine learning can help reveal lower-dimensional neural representations relevant to subjects' behavior. However, state-of-the-art models typically require large datasets to train, and so are prone to overfitting on human neuroimaging data that often possess few samples but many input dimensions. Here, we capitalized on the fact that the features we seek in human neuroscience are precisely those relevant to subjects' behavior rather than noise or other irrelevant factors. We thus developed a Task-Relevant Autoencoder via Classifier Enhancement (TRACE) designed to identify behaviorally-relevant target neural patterns. We benchmarked TRACE against a standard autoencoder and other models for two severely truncated machine learning datasets (to match the data typically available in functional magnetic resonance imaging [fMRI] data for an individual subject), then evaluated all models on fMRI data from 59 subjects who observed animals and objects. TRACE outperformed alternative models nearly unilaterally, showing up to 12% increased classification accuracy and up to 56% improvement in discovering "cleaner", task-relevant representations. These results showcase TRACE's potential for a wide variety of data related to human behavior.