Event-Based Models (EBMs) are widely used to infer latent disease progression from cross-sectional biomarker data, with existing studies employing diverse approaches to characterize uncertainty, stability, and validation. However, comparatively little guidance exists for systematically integrating and reporting these complementary forms of diagnostic evidence. We develop a Reporting Framework for Event-Based Model Diagnostics supported by a theoretical decomposition and empirical characterization of local ordering uncertainty in classical single-order EBMs. Our swap-local decomposition of the Evidence Lower Bound (ELBO) separates each local ordering decision into a posterior visibility term reflecting stage overlap and a subject-level contrast term reflecting directional evidence, enabling low-information ambiguity to be distinguished from genuine competing cohort evidence. We characterize these diagnostics through controlled synthetic experiments and an Alzheimer's disease cerebrospinal fluid proteomic case study, and evaluate complementary uncertainty targets through local optimization perturbations, leave-two-out sensitivity analyses, and repeated full-pipeline ReduXis experiments. The latter assess end-to-end initialization sensitivity and are distinct from validation of the classical fixed-fit swap-local decomposition. Collectively, these theoretical and empirical findings motivate a practical reporting framework that distinguishes broadly applicable core reporting elements from conditional, model-specific diagnostics. By organizing complementary diagnostic evidence according to its uncertainty target and model-specific scope, this work provides a principled foundation for more transparent, interpretable, and reproducible reporting of Event-Based Models.
Background and objectivesIdiopathic normal pressure hydrocephalus (NPH) and Alzheimer’s disease (AD) may present with overlapping clinical and neuroimaging features, complicating differential diagnosis. This study evaluated an explainable, attention-enhanced deep learning framework for distinguishing NPH from AD using two-dimensional brain magnetic resonance imaging (MRI).MethodsThe dataset comprised 96 patients (48 NPH and 48 AD) and 295 MRI images representing axial, sagittal, and coronal anatomical planes. The reported patient-level partition comprised 68 patients (208 images) for training, 14 patients (42 images) for validation, and 14 patients (45 images) for testing. Five ImageNet-pretrained convolutional neural network architectures–EfficientNetB2, ResNet152, VGG19, DenseNet201, and Xception–were evaluated with the Convolutional Block Attention Module (CBAM). EfficientNetB2 without CBAM was included as a baseline. Classification performance was assessed using standard performance metrics, confidence intervals, and paired statistical comparisons. Grad-CAM was used to visualize model activation patterns.ResultsEfficientNetB2 with CBAM achieved the highest observed image-level classification accuracy of 93.33% (42/45; patient-cluster bootstrap 95% CI: 86.05–100.00%) and a receiver operating characteristic area under the curve (ROC-AUC) of 0.9722 (95% CI: 0.9127–1.0000). For NPH classification, sensitivity and specificity were 90.48% and 95.83%, respectively. The corresponding EfficientNetB2 baseline achieved 88.89% accuracy, an exploratory difference of 4.44 percentage points; the paired difference was not statistically significant (exact McNemar test, p = 0.500). Grad-CAM visualizations revealed varying activation patterns across the six configurations without establishing clinical validity.ConclusionAttention-enhanced convolutional neural networks demonstrated promising image-level performance for differentiating NPH from AD on the evaluated MRI dataset. Nevertheless, the limited patient sample, absence of patient-level prediction aggregation, and lack of independent external validation warrant cautious interpretation. Further validation in larger and heterogeneous clinical cohorts is required before clinical application.
ObjectiveInterictal background EEG in childhood absence epilepsy (CAE) may appear normal. We asked whether drug-naive children with CAE differ from controls in alpha frequency and the regional distribution of phase-lagged synchronization, and how these differences depend on participant-specific surrogate baselines and analysis frequency.MethodsWe analyzed eyes-closed resting EEG (19 channels, 26 s) from 49 drug-naive children with CAE and 83 controls. Individual alpha frequency (IAF) was estimated by spectral parameterization of posterior channels. Weighted phase-lag index (wPLI) at 12 Hz was summarized globally and as the frontal–posterior log-ratio λFP = ln(wPLIF/wPLIP), where F and P denote frontal and posterior within-region means. Participant-specific baselines were obtained from 100 circular-shift surrogates, preserving each channel's wavelet-coefficient sequence up to circular translation while altering inter-channel alignment. Group comparisons used studentized permutation tests. The regional contrast was also evaluated at each participant's IAF.ResultsIAF was lower in patients (8.80 versus 9.42 Hz; p = 0.002). At 12 Hz, controls showed posterior predominance of synchronization (λFP = −0.352), which was largely removed by circular shifting (−0.036), whereas patients' mean regional balance was near zero. After baseline subtraction, the patient-minus-control difference in λFP was 0.332 (95 % CI [0.181, 0.483]; p < 0.001; Hedges' g = 0.79). In the 131 participants with identifiable IAF, the contrast was 0.316 at 12 Hz and 0.124 at IAF (p = 0.139). A direct comparison using within-participant changes supported a larger contrast at 12 Hz (difference 0.192, 95 % CI [0.017, 0.367]; exploratory, unadjusted p = 0.032). The global wPLI difference after baseline subtraction remained uncertain (−0.021, 95 % CI [−0.044, 0.001]; p = 0.069).ConclusionsDrug-naive CAE was associated with slower alpha activity and reduced posterior predominance of phase-lagged synchronization at 12 Hz. Circular shifting largely removed the controls' posterior predominance, supporting a difference in regional organization sensitive to inter-channel alignment. The smaller contrast at IAF indicates that analysis frequency materially affects this finding, without establishing that alpha slowing alone explains it.
IntroductionNeuroarchitecture studies how designed space shapes emotion, often with electroencephalography (EEG), and computational models that decode emotion from EEG must be trained on labelled recordings. Yet none of the widely shared EEG emotion datasets was recorded while people moved through designed spaces, and the few spatial recordings we found, all released since 2023, are single studies, each cited at most once. Researchers therefore borrow general emotion datasets, recorded from seated participants watching films, listening to music, playing games or conversing, and must decide whether each can be found, obtained and reused for transfer learning. This review contributes a reusable method for that decision and applies it.MethodsFollowing PRISMA 2020 with the dataset as the unit of analysis, two search passes identified 27 publicly obtainable datasets (22 in the first, 5 in the second), none designed for spatial research. A published codebook of 52 criteria defines three instruments: openness, judged against the FAIR principles with access class recorded separately; AI-readiness, the equally weighted mean of three sub-scores (label information, fit to standard tasks and alignment with other data), which describes usefulness as a transfer source; and ontology maturity, on five levels, which describes how far emotion labels can be mapped onto other vocabularies. Every identifier and access route was checked.ResultsOpenness was weakest in Interoperability (2.044 against a FAIR composite of 2.857), and only 8 datasets were fully open. A dataset-level identifier was missing for 13 of 27 datasets (48.1%), and three widely used datasets could not be reached through their published routes at a re-check on 11 September 2026. AI-readiness averaged 2.684 on a 1.0–4.0 scale (median 2.60, range 2.10–3.33), and ontology levels concentrated at Level 3 (16 of 27), with none at Level 4 or 5. Openness and readiness were nearly unrelated (ρ = 0.08), whereas ontology level tracked readiness (ρ = 0.68), partly because both read the label structure.DiscussionOpen access alone is therefore a poor guide to reuse. The method, codebook and scripts are published so that researchers in fields without dedicated data can test the datasets they plan to reuse.
