Machine Learning Framework Research Articles

A multimodal MRI-based machine learning framework for classifying cognitive impairment in cerebral small vessel disease

The heterogeneity of cerebral small vessel disease (CSVD) with mild cognitive impairment (MCI) presents a challenge for diagnosis and classification. This study aims to propose a multimodal magnetic resonance imaging (MRI)-based machine learning framework to effectively classify MCI and NCI in CSVD patients. We enrolled 165 CSVD patients, categorized into NCI (n = 81) and MCI (n = 84) groups based on neurocognitive assessments. Multimodal MRI data, including T1-weighted, resting-state functional MRI, and diffusion tensor images, were collected. Image preprocessing, feature extraction and selection were applied to obtain MRI features from three modalities. The AutoGluon platform was utilized for model development, and traditional machine learning algorithms were applied for comparison. The models were validated using a validation cohort of 83 CSVD patients, and their performance was assessed via receiver operating characteristic curve analysis. The AutoGluon model to distinguish MCI from NCI based on multimodal MRI features demonstrated high area under the curve (AUC), accuracy, sensitivity, specificity, precision, balanced accuracy, and F1-score in the training cohort (0.926, 88.48%, 88.10%, 88.89%, 89.16%, 88.50%, and 88.63%, respectively) and validation cohort (0.878, 81.93%, 86.36%, 76.92%, 80.85%, 81.64%, and 83.51%, respectively). Other traditional machine learning models had AUCs of 0.755–0.831, and their prediction accuracies were significantly lower than that of AutoGluon model (P < 0.001). Our study provides a multimodal MRI-based machine learning framework, utilizing the AutoGluon platform, that outperforms traditional algorithms in classifying MCI and NCI, offering a promising tool for the early prediction of MCI in CSVD.

Physics-based machine learning for computational fracture mechanics

This study introduces a physics-based machine learning (ϕML) framework for modeling both brittle and ductile fractures in elastic-viscoplastic materials. It integrates physical principles, including governing equations and constraints, directly into the neural network architecture. Specifically, a feedforward neural network is designed to embed physical laws within its architecture, ensuring thermodynamic consistency. Building on this foundation, synthetic datasets generated from finite element-based phase-field fracture simulations are employed to train the proposed framework, focusing on capturing the homogeneous, one-dimensional fracture responses. Detailed analyses are performed on the stored elastic energy and the dissipated work due to plasticity and fracture, demonstrating the capability of the framework to predict essential fracture features. The proposed ϕML framework overcomes the shortcomings of classical machine learning models, which rely heavily on large datasets and lack guarantees of physical principles. By leveraging its physics-integrated design, the ϕML framework demonstrates exceptional performance in predicting key properties of brittle and ductile fractures with limited training data. This ensures reliability, efficiency, and physical consistency, establishing a foundational approach for integrating machine learning with computational fracture mechanics.

Accelerated Design of Fenton-Like Copper Single-Atom Catalysts by Adaptive Learning with Genetic Programming.

Traditional trial-and-error methods for optimizing catalyst synthesis are time-consuming and costly, exploring only a small fraction of the vast combinatorial space. Machine learning (ML) offers a promising alternative but still has the limitation of relying on well-selected initial datasets, which the recent development of active learning (AL) could be addressed. Here, we novelly integrate an AL-derived algorithm, the Active Learning Genetic Algorithm (ALGA), into experimental workflows to optimize the synthesis of Fenton-like single-atom catalysts (SACs). Our results show that the closed-loop ALGA framework effectively learns from limited and sparse datasets, greatly reducing the research cycle compared to traditional ML and AL frameworks. By iteratively retaining better-performing genetic information and proactively expanding the search space through mutation and crossover, ALGA identifies the highest-performing Fenton-like Cu SACs with less than 90 experiments. The maximum phenol degradation rate k-value (0.147 min-1) achieved within the ALGA framework is approximately 3 times higher than that of the initial dataset and surpasses the reported best Fenton-like Cu SACs. Our successful implementation of ALGA signifies an advancement in SACs synthesis assisted by the AL-derived algorithm, offering a guiding methodology for the exploration of other functional materials.

Decoding the Plastic Patch: Exploring the Global Microplastic Distribution in the Surface Layers of Marine Regions with Interpretable Machine Learning.

The marine environment is grappling with microplastic (MP) pollution, necessitating an understanding of its distribution patterns, influencing factors, and potential ecological risks. However, the vast area of the ocean and budgetary constraints make conducting comprehensive surveys to assess MP pollution impractical. Interpretable machine learning (ML) offers an effective solution. Herein, we used four ML algorithms based on MP data calibrated to the size range of 20-5000 μm and considered various factors to construct a robust predictive ML model of marine MP distribution. Interpretation of the ML model indicated that biogeochemical and anthropogenic factors substantially influence global marine MP pollution, while atmospheric and physical factors exert lesser effects. However, the extent of the influence of each factor may vary within specific marine regions and their underlying mechanisms may differ across regions. The predicted results indicated that the global marine MP concentrations ranged from 0.176 to 27.055 particles/m3 and that MPs in the 20-5000-μm size range did not pose a potential ecological risk. The interpretable ML framework developed in this study covered MP data preprocessing, MP distribution prediction, and interpretation of the influencing factors of MPs, providing an essential reference for marine MP pollution management and decision making.

Evaluation of Different Machine Learning Approaches to Predict Antigenic Distance Among Newcastle Disease Virus (NDV) Strains

Newcastle disease virus (NDV) continues to present a significant challenge for vaccination due to its rapid evolution and the emergence of new variants. Although molecular and sequence data are now quickly and inexpensively produced, genetic distance rarely serves as a good proxy for cross-protection, while experimental studies to assess antigenic differences are time consuming and resource intensive. In response to these challenges, this study explores and compares several machine learning (ML) methods to predict the antigenic distance between NDV strains as determined by hemagglutination-inhibition (HI) assays. By analyzing F and HN gene sequences alongside corresponding amino acid features, we developed predictive models aimed at estimating antigenic distances. Among the models evaluated, the random forest (RF) approach outperformed traditional linear models, achieving a predictive accuracy with an R2 value of 0.723 compared to only 0.051 for linear models based on genetic distance alone. This significant improvement demonstrates the usefulness of applying flexible ML approaches as a rapid and reliable tool for vaccine selection, minimizing the need for labor-intensive experimental trials. Moreover, the flexibility of this ML framework holds promise for application to other infectious diseases in both animals and humans, particularly in scenarios where rapid response and ethical constraints limit conventional experimental approaches.

Efficient and Accurate Prediction of Double Perovskite Quasiparticle Band Gaps via Machine Learning and a Descriptor.

Perovskites have attracted considerable attention in materials science due to their promising applications in photovoltaics and photocatalysis. Accurate prediction of their electronic band gap is essential for optimizing the performance. Traditional computational methods for band gap prediction often face a trade-off between accuracy and computational efficiency. General density functional theory (DFT) calculations typically underestimate band gap values, while the more accurate quasi-particle method demands substantial computational resources. In this study, a multistep machine learning framework was developed for efficient screening of semiconductor double perovskites. Furthermore, we proposed an interpretable descriptor that can predict quasi-particle band gaps of perovskites with a precision of over 90% accuracy. Using this approach, we screened 4,507 perovskite candidates and identified 94 structures that have suitable band gaps and are lead-free. Among these, six candidate structures were selected for further verification based on their photocatalytic potential and thermal stability.

Classification of intracranial pressure epochs using a novel machine learning framework

Patients with acute brain injuries are at risk for life threatening elevated intracranial pressure (ICP). External Ventricular Drains (EVDs) are used to measure and treat ICP, which switch between clamped and draining configurations, with accurate ICP data only available during clamped periods. While traditional guidelines focus on mean ICP values, evolving evidence indicates other waveform features may hold prognostic value. However, current machine learning models using ICP waveforms exclude EVD data due to a lack of digital labels indicating the clamped state, markedly limiting their generalizability. We introduce, detail, and validate CICL (Classification of ICP epochs using a machine Learning framework), a semi-supervised approach to classify ICP segments from EVDs as clamped, draining, or noise. This paves the way for multiple applications, including generalizable ICP crisis prediction, potentially benefiting tens of thousands of patients annually and highlights an innovate methodology to label large high frequency physiological time series datasets.

Predicting biomolecule adsorption on nanomaterials: a hybrid framework of molecular simulations and machine learning.

The adsorption of biomolecules on the surface of nanomaterials (NMs) is a critical determinant of their behavior, toxicity, and efficacy in biological systems. Experimental testing of these phenomena is often costly or complicated. Computational approaches, particularly the integrating methods of various theoretical levels, can provide essential insights into nano-bio interactions and bio-corona formation, facilitating the efficient design of nanomaterials for biomedical applications. This study presents a hybrid, meta-modeling approach that integrates physics-based molecular modeling with machine learning algorithms to predict the interaction energy between NMs and biomolecules extracted from the potential of mean force (PMF). Novel descriptors for the surface properties of NMs are developed and combined with structural descriptors of biomolecules to derive quantitative structure-property relationships (QSPRs). The developed QSPR model (training set: R2 = 0.84, RMSE = 1.52, Rcv2 = 0.83, and RMSEcv = 1.34; validation set: R2 = 0.70, RMSE = 1.94, and Rcv2 = 0.72, RMSEcv = 1.88) helps in understanding and predicting the interactions between NMs (including carbon-based materials, metals, metal oxides, metalloids, and cadmium selenide) and biomolecules (including amino acids and amino acid derivatives). The model facilitates safe and sustainable design of nanomaterials for various applications, particularly for nanomedicine, by providing insight into nano-bio interactions, identification of toxicity risk and optimizing functionalization for safety according to the risk mitigation policy of regulatory authorities. Additionally, a dedicated application has been developed and is available on GitHub, enabling researchers to analyze the surface properties of nanomaterials belonging to the above-mentioned groups of NMs.

Learn2Aggregate: Supervised Generation of Chvatal-Gomory Cuts Using Graph Neural Networks

We present Learn2Aggregate, a machine learning (ML) framework for optimizing the generation of Chvatal-Gomory (CG) cuts in mixed integer linear programming (MILP). The framework trains a graph neural network to classify useful constraints for aggregation in CG cut generation. The ML-driven CG separator selectively focuses on a small set of impactful constraints, improving runtimes without compromising the strength of the generated cuts. Key to our approach is the formulation of a constraint classification task which favours sparse aggregation of constraints, consistent with empirical findings. This, in conjunction with a careful constraint labeling scheme and a hybrid of deep learning and feature engineering, results in enhanced CG cut generation across five diverse MILP benchmarks. On the largest test sets, our method closes roughly twice as much of the integrality gap as the standard CG method while running 40% faster. This performance improvement is due to our method eliminating 75% of the constraints prior to aggregation.

Trajectory Tracking on the Optimal Path of Two-dimensional Quadratic Barrier Escaping

Abstract The diffusion trajectory of a Brownian particle passing over the saddle point of a two-dimensional quadratic potential energy surface is tracked in detail according to the deep learning strategies. Generative Adversarial Networks (GANs) emanating in the category of Machine Learning (ML) frameworks are used to generate and assess the rationality of the data. While the optimization of them is based on the Long Short-Term Memory (LSTM) strategies. In addition to drawing a heat map, the optimal path of two-dimensional (2D) diffusion (which was firstly predicted in Ref.[6, 7]) is simultaneously demonstrated in a stereoscopic space. The results of our simulation proved to be completely consistent with the previous theoretical predictions.

Adaptive Quantum Learning Architectures with Superalgebraic Entangled Layers

This paper introduces a novel quantum machine learning framework that integrates adaptive learning capabilities with superalgebraically structured entangled layers. Leveraging the mathematical formalism of Lie superalgebras—specifically osp(1∣2) and sl(2∣1)—the proposed architecture encodes quantum states in graded Hilbert spaces and governs their evolution via symmetry-preserving supercharges. The entangled layers dynamically adapt to task-specific feedback through a hybrid quantum-classical training protocol that ensures algebraic consistency and quantum coherence. Symbolic computation and quantum simulation tools are employed to validate the fidelity and symmetry of entanglement evolution. The system is benchmarked on tasks such as symbolic reasoning, quantum phase classification, and structured regression, outperforming baseline models in interpretability, stability, and generalization. This research lays the foundation for scalable, physically grounded quantum AI systems capable of learning through algebraic alignment with fundamental symmetries. Keywords: quantum machine learning, adaptive learning, entanglement, superalgebra, osp(1∣2), graded Hilbert space, quantum AI, symmetry-preserving learning, Lie superalgebras, supercharges, hybrid quantum-classical architecture, interpretable QML, symbolic reasoning, quantum circuits

AutoSciLab: A Self-Driving Laboratory for Interpretable Scientific Discovery

Advances in robotic control and sensing have propelled the rise of automated scientific laboratories capable of high-throughput experiments. However, automated scientific laboratories are currently limited by human intuition in their ability to efficiently design and interpret experiments in high-dimensional spaces, throttling scientific discovery. We present AutoSciLab, a machine learning framework for driving autonomous scientific experiments, forming a surrogate researcher purposed for scientific discovery in high-dimensional spaces. AutoSciLab autonomously follows the scientific method in four steps: (i) generating high-dimensional experiments (x) using a variational autoencoder (ii) selecting optimal experiments by forming hypotheses using active learning (iii) distilling the experimental results to discover relevant low-dimensional latent variables (z) with a ‘directional autoencoder’ and (iv) learning a human interpretable equation connecting the discovered latent variables with a quantity of interest (y = f (z)), using a neural network equation learner. We validate the generalizability of AutoSciLab by rediscovering a) the principles of projectile motion and b) the phase-transitions within the spin-states of the Ising model (NP-hard problem). Applying our framework to an open-ended nanophotonics problem, AutoSciLab discovers a new way to steer incoherent light emission beyond current state-of-the-art, defining a new structure(material)-property(light-emission) relationship governing the physical process using closed-loop noisy experimental feedback.

Identification of a deubiquitinating gene-related signature in ovarian cancer using integrated transcriptomic analysis and machine learning framework

BackgroundOvarian carcinoma represents an aggressive malignancy with poor prognosis and limited therapeutic efficacy. While deubiquitinating (DUB) genes are known to regulate crucial cellular processes and cancer progression, their specific roles in ovarian carcinoma remain poorly understood.MethodsWe conducted an integrated analysis of single-cell RNA sequencing and bulk transcriptome data from public databases. DUB genes were identified through Genecard database. Using the Seurat package, we performed cell clustering and differential expression analysis. Cell–cell communications were analyzed using CellChat. A DUB-related risk signature (DRS) was developed using machine learning approaches through integration of GEO and TCGA datasets. The prognostic value and immune characteristics of the signature were systematically evaluated.ResultsOur analysis revealed eight distinct cell subtypes in the tumor microenvironment, including epithelial, fibroblast, myeloid, and Treg cells. DUB-high cells were predominantly found in Treg and myeloid populations, exhibiting elevated expression of tumor-related pathways and enhanced cell–cell communication networks, particularly between fibroblasts and myeloid cells. Conversely, DUB-low cells were enriched in epithelial populations with reduced immune activity. The DRS model demonstrated robust prognostic value across multiple independent cohorts. High-risk patients, as classified by the DRS, showed significantly poorer survival outcomes and distinct immune infiltration patterns compared to low-risk patients.ConclusionThis study provides comprehensive insights into DUB gene expression patterns across different cell populations in ovarian carcinoma. The established DRS model offers a promising tool for risk stratification and may guide personalized therapeutic strategies. Our findings highlight the potential role of DUB genes in modulating the tumor immune microenvironment and patient outcomes in ovarian carcinoma.

Advancing Machine Learning-Enhanced Flow Modeling for Collision Phenomena in Total Cavopulmonary Connection

Abstract The Fontan procedure, essential for newborns with single ventricle malformations, establishes a total cavopulmonary connection (TCPC). Understanding TCPC hemodynamics is critical for evaluating surgical outcomes. However, the intricate flow collisions within TCPC present a significant challenge. While computational fluid dynamics (CFD) offers insight into these mechanics, its computational inefficiency limits clinical application. Machine learning (ML) show promise in overcoming this limitation. This study develops an ML framework to analyze TCPC hemodynamics. The ML models are trained to predict the relationship between TCPC flow parameters and two hemodynamic metrics: power loss (PL) and hepatic flow distribution (HFD). Four ML algorithms - Random Forest, Support Vector Regression, Extreme Gradient Boosting (XGBoost), and Artificial Neural Networks (ANNs) - are evaluated for their performance using a training dataset of 7,056 CFD simulations. All ML models demonstrate strong correlations with CFD results (R2 = 0.95). Notably, the ANN model slightly outperforms others in predicting HFD (R2 = 0.98), while both ANN and XGBoost models achieve similar accuracy in predicting PL (R2 = 0.99). Additionally, the study investigates the feasibility of training an ANN model with a reduced dataset of 968 samples, which can still successfully capture the relationship between flow parameters and TCPC hemodynamics. This study underscores the potential of ML-enabled models to enhance the efficiency of hemodynamic assessments in TCPC with flow collision scenarios. Given that flow collision phenomena are common in various physiological systems and engineering contexts, these findings may drive advancements in ML-augmented flow modeling across a broad range of applications.

AI Driven Cardiovascular Risk Prediction Using NLP and Large Language models for Personalized Medicine in Athletes.

The performance and long-term health of athletes are significantly influenced by their cardiovascular resilience and associated risk factors. This study explores the innovative applications of Natural Language Processing (NLP) and Large Language Models (LLMs) in biomedical diagnostics, particularly for AI-driven arrhythmia detection, hypertrophic cardiomyopathy (HCM) in athletes, and personalized medicine. The complexity of analysing diverse biomedical datasets, such as electrocardiograms (ECG), clinical records, genetic screening reports, and imaging results, poses challenges in obtaining precise early diagnoses. To address these issues, we introduce a hybrid machine learning (ML) framework that integrates the Wolf Pack Search Algorithm Dynamic Random Forest (WPSA-DRF) with a RoBERTa-based LLM to enhance the accuracy of cardiovascular disease predictions. Using advanced NLP techniques, including biomedical text mining, entity recognition, and feature extraction, the system processes structured and unstructured clinical data to detect abnormalities associated with sudden cardiac arrest (SCA), arrhythmias, and genetic cardiomyopathies. The proposed system achieves a diagnostic accuracy of 92.5%, precision of 92.7%, recall of 99.23%, and F1-score of 95.6%, outperforming traditional diagnostic methodologies. Furthermore, the research underscores the role of LLMs in personalized medicine, identifying patient-specific risk factors and optimizing treatment pathways for cardiac patients. This work highlights how NLP-driven AI solutions are transforming biomedical research, accelerating early disease detection, and improving clinical decision-making for both athletes and the general population.

Day-Ahead Energy Price Forecasting with Machine Learning: Role of Endogenous Predictors

Accurate Day-Ahead Energy Price (DAEP) forecasting is essential for optimizing energy market operations. This study introduces a machine learning framework to predict the DAEP with a 24 h lead time, leveraging historical data and forecasts available at the prediction time. Hourly DAEP data from the California Independent System Operator (January 2017 to July 2023) were integrated with exogenous and engineered endogenous features. A custom rolling window cross-validation, with 24 h validation blocks sliding daily across 2372 folds, evaluates an Extreme Gradient Boosting (XGBoost) model’s performance under diverse market conditions, achieving a median mean absolute error of 6.26 USD/MWh and root mean squared error of 8.27 USD/MWh, with variability reflecting market volatility. The feature importance analysis using Shapley additive explanations highlighted the dominance of engineered endogenous features in driving the 24 h lead time forecasts under relatively stable market conditions. Forecasting the DAEP at a runtime of 10 AM on the prior day was used to assess model uncertainty. This involved training random forest, support vector regression, XGBoost, and feed forward neural network models, followed by stacking and voting ensembles. The results indicate the need for ensemble forecasting and evaluation beyond a static train–test split to ensure the practical utility of machine learning for DAEP forecasting across varied market dynamics. Finally, operationalizing the forecast model for bidding decisions by forecasting the DAEP and real-time prices at runtime is presented and discussed.

Effective ML-Based Android Malware Detection and Categorization

The rapid proliferation of malware poses a significant challenge regarding digital security, necessitating the development of advanced techniques for malware detection and categorization. In this study, we investigate Android malware detection and categorization using a two-step machine learning (ML) framework combined with feature engineering. The proposed framework first performs binary categorization to detect malware and then applies multi-class categorization to categorize malware into types, such as adware, banking Trojans, SMS malware, and riskware. Feature selection techniques such as chi-squared testing and select-from-model (SFM) were employed to reduce dimensionality and enhance model performance. Various ML classifiers were evaluated, and the proposed model achieved outstanding accuracy, at 97.82% for malware detection and 96.09% for malware categorization. The proposed framework outperforms existing approaches, demonstrating the effectiveness of feature engineering and random forest (RF) models in addressing computational efficiency. This research contributes a robust and interpretable framework for Android malware detection that is resource-efficient and practical for use in real-world applications. It also offers a scalable approach via which practitioners can deploy efficient malware detection systems. Future work will focus on real-time implementation and adaptive methodologies to address evolving malware threats.

A Machine Learning Ensemble Approach for Early Detection of Oral Cancer: Integrating Clinical Data and Imaging Analysis in the Public Health

This study presents an integrated machine learning framework for the early detection of oral cancer, leveraging both clinical data and high-resolution imaging. The research compared several algorithms, including logistic regression, decision trees, random forests, support vector machines, and convolutional neural networks, culminating in an ensemble model that combined clinical indicators with imaging features. Results demonstrate that while traditional models provided moderate diagnostic accuracy, advanced techniques, particularly the ensemble model, achieved superior performance with an accuracy of 91%, sensitivity of 89%, specificity of 92%, and an AUC of 93%. These findings highlight that multimodal data integration significantly enhances early detection capabilities, offering a robust and practical solution for clinical implementation. The proposed framework not only improves diagnostic precision but also supports timely interventions that can potentially reduce the morbidity and mortality associated with late-stage oral cancer.

Unique and shared transcriptomic signatures underlying localized scleroderma pathogenesis identified using interpretable machine learning.

Using transcriptomic profiling at single-cell resolution, we investigated cell-intrinsic and cell-extrinsic signatures associated with pathogenesis and inflammation-driven fibrosis in both adult and pediatric patients with localized scleroderma (LS). We performed single-cell RNA-Seq on adult and pediatric patients with LS and healthy controls. We then analyzed the single-cell RNA-Seq data using an interpretable factor analysis machine learning framework, significant latent factor interaction discovery and exploration (SLIDE), which moves beyond predictive biomarkers to infer latent factors underlying LS pathophysiology. SLIDE is a recently developed latent factor regression-based framework that comes with rigorous statistical guarantees regarding identifiability of the latent factors, corresponding inference, and FDR control. We found distinct differences in the characteristics and complexity in the molecular signatures between adult and pediatric LS. SLIDE identified cell type-specific determinants of LS associated with age and severity and revealed insights into signaling mechanisms shared between LS and systemic sclerosis (SSc), as well as differences in onset of the disease in the pediatric compared with adult population. Our analyses recapitulate known drivers of LS pathology and identify cellular signaling modules that stratify LS subtypes and define a shared signaling axis with SSc.

Integrating Machine Learning and Geospatial Data for Mapping Socioeconomic Vulnerability to Urban Natural Hazard

Rapid urbanization and climate change are increasing the risks associated with natural hazards, especially in cities where socio-economic disparities are significant. Current hazard risk assessment frameworks fail to consider socio-economic factors, which limits their ability to effectively address vulnerabilities at the community level. This study introduces a machine learning framework designed to assess flood susceptibility and socio-economic vulnerability, particularly in urban areas with limited data. Using Kigali, Rwanda, as a case study, we quantified socio-economic vulnerability through a composite index that includes indicators of sensitivity and adaptive capacity. We utilized a variety of data sources, such as demographic, environmental, and remotely sensing datasets, applying machine learning algorithms like Multilayer Perceptron (MLP), Random Forest, Support Vector Machine (SVM), and XGBoost. Among these, MLP achieved the best predictive performance, with an AUC score of 0.902 and an F1-score of 0.86. The findings indicate spatial differences in socio-economic vulnerability, with central and southern Kigali showing greater vulnerability due to a mix of socio-economic challenges and high flood risk. The vulnerability maps created were validated against historical flood records, socio-economic research, and expert insights, confirming their accuracy and relevance for urban risk assessment. Additionally, we tested the framework’s scalability and adaptability in Kampala, Uganda, and Dar es Salaam, Tanzania, showing that making context-specific adjustments to the model improves its transferability. This study offers a solid, data-driven approach for combining assessments of flood susceptibility and socio-economic vulnerability, filling important gaps in urban resilience planning. The results support the advancement of risk-informed decision-making, especially in areas with limited access to detailed socio-economic information.