Artificial intelligence (AI)–supported tools are increasingly integrated into online learning environments to promote self-regulated learning (SRL), a critical competency for success in learning. Although recent research has investigated the potential of AI to guide and scaffold SRL processes, a systematic understanding of how AI technologies are adopted and how effectively they support learners remains limited. To address this gap, we conduct a systematic review of empirical studies published between 2020 and 2025 that examined AI-supported SRL in online settings. We systematically analyze studies across several dimensions: (a) types of AI technologies adopted (e.g., predictive modeling, natural language processing, intelligent tutoring systems, conversational agents), (b) functions served (e.g., feedback generation, adaptive support, performance prediction, scaffolding), (c) learning subjects supported, and (d) types of activities facilitated. We further map how these interventions align with specific SRL processes, such as goal setting, planning, monitoring, strategy use, reflection, and motivation regulation. Our analysis highlights which aspects of SRL are currently supported by AI versus those less developed. Evidence of intervention effectiveness is also reported. Findings shed light on how AI applications can be aligned with established SRL theory to provide support that empowers learners in developing their SRL skills. We conclude with recommendations for future research on the use of AI to advance SRL in online learning environments.

This paper presents a hybrid framework that integrates Machine Learning (ML) with the
ALMODES methodology, creating transparent, data-driven dynamic system models. The
framework preserves the interpretable ALMODES ‘white-box’ structure—a discrete-time, directed
graph model governed by the state update equation s(t) = s(t-1)A(t) + u(t) + ε(t), where A is the
real-valued dependency matrix.

ML is systematically integrated to calibrate, refine, and stress-test the model. Its key
applications include learning parameters through constrained regression, refining the model’s
graphical structure when expert knowledge is uncertain, capturing time-varying interactions using a
state-space layer, and optimizing policy decisions via Model-Predictive Control (MPC).
A core principle of this hybrid approach is its commitment to interpretability and
governance. Strict guardrails and domain-mandated constraints on interaction signs and sparsity
ensure all ML-driven outputs are rendered back into a human-readable directed graph. This
preserves the ALMODES causal-map ethos and guarantees the learned model remains conceptually
meaningful.

The framework’s utility is demonstrated by applying it to incorporate Environmental, Social,
and Governance (ESG) objectives into a Balanced Scorecard (BSC) model.

Background: Recent studies have shown that traditional biosensors suffer from low performances in detecting (identifying) and categorizing (classifying) neurotransmitters (NTs) in complex biological fluids (CBFs). To enhance these performances, nanoparticles (NPs) and biocatalysts have been functionalized on the biosensors’ functional surfaces to specifically improve sensitivity and selectivity respectively. However, such enhancements remain limited in efficacity, as the extraction of discernible electrochemical signatures for target NTs are obscured by overlapping signal peaks arising from coexisting NTs, background-current drifts, redox-potential shifts and intrinsic noise in CBFs. Additionally, electrode fouling and signal crosstalk at the biosensors’ functional surfaces further compromise biorecognition accuracy. Objective: To mitigate these problems, this work aims at integrating machine learning classifiers with NP-enhanced biosensors and voltammetry, to improve the detection and categorization of NTs in CBFs. Design: This study proposes an ML framework for the detection and categorization of the NTs; dopamine (DA), serotonin (SE), epinephrine (EP), ascorbic acid (AA) and their mixtures, based on voltammetry data recorded with novel NP-enhanced biosensors and voltammetry. The NPs cerium oxide, gold, gallium and silver were individually immobilized on the biosensors’ functional surfaces through chitosan, a biodegradable and biocompatible polymer matrix. Result: Linear Discriminant analysis (LDA), Ensemble learning (EL) and Support Vector Machines (SVM) were analyzed. SVM was found to accurately detect and categorize the NTs and their mixtures with 80% validation accuracy. Clinical relevance: The outcome of this study has the potential to advance neuromodulation therapies, particularly deep brain stimulation, for the treatment of neurodegenerative disorders such as Parkinson’s disease, where accurate real-time monitoring of DA dynamics is essential for optimizing therapeutic efficacy.

This proof-of-concept study aimed to develop and internally validate a radiomic machine learning model (ARARAT) based on axial T2-weighted MRI for classifying prostate cancer as ISUP < 3 versus ISUP ≥ 3. A total of 112 biopsy-proven malignant lesions from the PROSTATEx dataset were analyzed. Fixed spherical ROIs (5 mm radius) were centered on each lesion, and 107 radiomic features were extracted using PyRadiomics. Data were z-score standardized and split into training and test sets (80/20) with group-aware stratification. Models were trained using 10-fold cross-validation, and performance was evaluated via AUC, average precision (AP), sensitivity, specificity, and calibration. The best-performing model (Random Forest) achieved AUC = 0.84, AP = 0.70, sensitivity = 85.7%, specificity = 78.6%, and accuracy = 81.0%. Calibration improved after isotonic regression (ECE = 0.168 → 0.099). SHAP analysis identified first-order and GLSZM features as most influential. Decision curve analysis demonstrated net benefit over “treat all” and “treat none” strategies between 5–35% thresholds.

The ARARAT model demonstrates promising internal performance for noninvasive risk stratification of prostate cancer using only T2-weighted MRI radiomics. While developed exclusively on malignant lesions, limiting assessment of benign versus malignant discrimination, it highlights the feasibility of radiomic-based ISUP classification. External validation in a representative cohort is warranted to confirm clinical applicability.