Skip to Content
  • 6.5
    Impact Factor
  • 7.3
    CiteScore
  • 21 days
    Submission to First Decision
  • 6 days
    Acceptance to Publication

AI

AI is an international, peer-reviewed, open access journal on artificial intelligence (AI), including broad aspects of cognition and reasoning, perception and planning, machine learning, intelligent robotics, and applications of AI, published monthly online by MDPI.

Get Alerted

Add your email address to receive forthcoming issues of this journal.

All Articles (1,161)

  • Article
  • Open Access

This study examines teachers’ perceptions of support, responsibility, and influence in generative artificial intelligence (GenAI) use. A cross-sectional survey included 436 teachers recruited through a Slovenian digital education project network. The analysis plan was registered after data collection and descriptive access; the approximate encouragement–governance correlation was already known. General encouragement received higher normalised ratings than AI-specific rules, procedures, and support (mean difference on the percentage-of-maximum-possible scale = 38.9, 95% confidence interval [36.9, 40.9]; paired n = 344). Compliance and problem-solving responsibility were rated above rule clarity and procedure knowledge, whereas expected personal responsibility for harm was rated below classroom influence. These non-parallel item contrasts do not establish a general responsibility–influence imbalance. Procedural support was associated with lower endorsement of a combined unclear-rules-and-avoidance statement (odds ratio = 0.66, 95% confidence interval [0.53, 0.82]), which does not isolate avoidance behaviour. In a teacher-prepared-materials vignette, 79% selected the teacher as primarily responsible. Convenience sampling, unavailable school identifiers, non-equivalent measures, and an unconfirmed governance model limit interpretation. The contribution is an item-specific account of jointly measured perceptions, without causal or population-level inference.

AI

20 November 2026

Leadership encouragement and governance infrastructure items on the percentage-of-maximum-possible (POMP) scale, shown as means with 95% confidence intervals.
  • Article
  • Open Access

Schizophrenia is a severe neuropsychiatric disorder that affects cognition, perception, and behaviour and requires objective diagnostic methods. Electroencephalography (EEG) provides a non-invasive approach to detecting neural abnormalities associated with schizophrenia. This study presents an artificial intelligence framework for systematically evaluating multi-domain EEG feature combinations and lobe-wise analysis for schizophrenia classification under conventional and leakage-free subject-wise validation protocols. EEG recordings with schizophrenia are obtained from a publicly available dataset. Time-domain, frequency-domain, entropy-based, and non-linear features are extracted to capture complementary neurophysiological information. Experiments evaluated the individual and combined feature domains across individual and multiple brain-lobe combinations. Classification performance is assessed using multiple machine-learning classifiers under conventional stratified 10-fold cross-validation, followed by subject-wise stratified 10-fold and leave-one-subject-out validation to eliminate subject-level data leakage. Under conventional stratified 10-fold cross-validation, the best-performing configuration using the frontal-temporal-central-occipital (FTCO) lobe combination with frequency-domain, entropy-based, and non-linear features achieved an accuracy of 99.76%. The proposed framework is further evaluated under subject-wise 10-fold and leave-one-subject-out validation protocols, confirming its robustness under leakage-free evaluation with mean accuracies of 81.88% and 80.95%, respectively. The combination of complementary EEG feature domains and lobe-wise information improved schizophrenia classification while providing insights into the contribution of different brain regions.

AI

19 November 2026

Proposed EEG-based schizophrenia detection framework.
  • Review
  • Open Access

Magnetic resonance-guided focused ultrasound (MRgFUS) is an incisionless platform that combines transcranial acoustic energy delivery with MR imaging and MR thermometry. In this narrative review, we summarize the physical principles that constrain transcranial MRgFUS, the clinical evidence for its neurological applications, and the emerging role of artificial intelligence (AI) across the treatment workflow. Unilateral thalamotomy for medication-refractory essential tremor and tremor-dominant Parkinson’s disease, staged bilateral thalamotomy for essential tremor, and pallidal or pallidothalamic ablation for motor complications of Parkinson’s disease are supported by randomized or pivotal trials and regulatory approval, whereas applications in neuropathic pain, epilepsy, and blood–brain barrier opening remain investigational. AI methods have been explored for synthetic computed tomography, acoustic simulation, automated planning, prediction of sonication parameters and temperature, accelerated or referenceless MR thermometry, cavitation-based monitoring of blood–brain barrier opening, outcome prediction, and digital biomarkers; we tabulate the principal studies by task, dataset, validation strategy, performance, clinical endpoint, and technology readiness. Most clinical AI studies are retrospective, single-center, and small (typically 20–120 patients, with some test sets of fewer than ten patients), and many technical studies rely on simulations, phantoms, or animal models; external validation is rare, and no study has yet shown that AI improves clinical outcomes. AI should therefore currently be regarded as an early-stage decision-support tool that operates under clinician oversight. We also discuss the applicable regulatory frameworks, liability, and the validation steps required before clinical adoption.

AI

19 November 2026

Overview of focused ultrasound modalities used in neurology. (a) High-intensity focused ultrasound: a hemispherical phased array (≈650 kHz, 1024 elements) focuses energy through the skull, with CT-based phase correction, to produce a thermal lesion (≈55–60 °C) monitored by MR thermometry. (b) Low-intensity pulsed ultrasound with intravenous microbubbles: stable cavitation transiently increases blood–brain barrier permeability (clinical systems mainly 220 kHz; closure typically documented within 24 h), whereas broadband emissions from inertial cavitation signal a risk of microhemorrhage. (c) Low-intensity focused ultrasound neuromodulation: reversible excitation or inhibition without a lesion, depending on pulse parameters. In (a), gray squares represent the transducer elements, the light-blue band the chilled, degassed water, and the thin colored lines the beam paths converging through the skull (gray ring) onto the focal lesion (orange) in the brain (pink). In (b), the arcs represent pulsed ultrasound wavefronts, the pink blocks endothelial cells lining the vessel, the circles with dotted outlines oscillating microbubbles, and the purple arrow and dots the passage of therapeutic agents through an opened tight junction. In (c), the arcs represent ultrasound wavefronts, and the inset, linked by a dotted line, magnifies the neuronal membrane (dotted rows) and its ion channels (dark bars). Thin gray lines link labels to structures, and “→” means “leads to”. The colored bars at the bottom give the clinical status of each modality. Schematic, not to scale; parameters are from [5,47,94,97,99]. The figure was drawn by the authors as a vector schematic.
  • Article
  • Open Access

Per-station energy metering in Industrial IoT (IIoT) environments makes it possible to instrument flexible production lines, yet turning such measurements into design decisions remains an open problem. This paper presents a learning-augmented energy digital twin of a real flexible manufacturing line (FML) with seven workstations, industrial robots and radio-frequency identification (RFID) traceability producing three product types. The twin combines per-station time and power surrogate models, learned from metered data at four to five operating points through regression with model selection, with a discrete-event simulator derived from the timed Petri-net model of the zero-buffer, re-entrant line. A multi-layered Grey Wolf Optimizer (AGWO), whose second layer calibrates the compute budget, reaches the optimum in all 30 independent runs within a few milliseconds, and its search core is benchmarked against particle swarm optimization (PSO) and a genetic algorithm (GA) on 23 test functions. The computed set-points were implemented on the physical line. The twin quantifies the classical improvement levers: conveyor-speed tuning yields at most 2.7% energy-per-part savings, job-mix sequencing of 0.6%, and buffer installation of 0%, because the bottleneck station is utilized 99.2% of the time. In contrast, a model-based what-if analysis projects that reducing its cycle time by 25–50% would lower the energy per part by 24–48%. The results show that an inexpensive, measurement-grounded twin can exclude unprofitable design directions and target investment where it matters.

AI

18 November 2026

Conceptual model of the FML.

Featured Articles of Last Quarter

Highly Accessed Articles

News & Conferences

Latest Issues

Open for Submission

Journal Sections

XFacebookLinkedIn
AI - ISSN 2673-2688