Objective Conformance checking in healthcare aims to verify whether patient care traces comply with clinical guidelines, but it typically requires formal, machine-interpretable guideline models, such as Computer-Interpretable Guidelines (CIGs), that are rarely available in practice.Methods This work presents a modular framework that leverages orchestrated Large Language Models (LLMs) to support medical conformance checking directly from unstructured data, without the need for CIGs. The proposed architecture combines multiple LLMs and auxiliary components to extract patient traces from clinical discharge letters, derive normative rules from textual clinical guidelines, formalize rules into executable scripts, and compute a Trace Conformance Indicator that quantifies conformance on the event log.Results The framework has been implemented and evaluated in the stroke care domain at the neurological ward of Alessandria hospital, where hundreds of patient traces were automatically extracted from hospital data. In particular, most of the available traces proved to be conformant to the 50 rules derived from the guideline. Conclusion Our work has demonstrated the feasibility of LLM orchestration in our domain, while, at the same time, verifying the good rule conformance checking results in Alessandria.
Increasing cognitive demand is accompanied by task-related cerebral hemodynamic changes but mean and variance alone may not describe their temporal organization. We assessed whether statistical complexity measures (SCM) derived from bilateral middle cerebral artery blood-velocity (CBv) signals varied across five arithmetic task levels in 26 healthy young adults (mean age 23.4 ± 2.1 years). Behavioral performance and NASA-TLX workload were also collected. For the primary Rényi configuration (embedding dimension m=7, alpha=1.4), SCM varied across baseline and task levels (Jensen-Rényi: Friedman chi-square=56.98, FDR-adjusted p=1.02x10-10, Kendall’s W=0.456; Jensen-Shannon: chi-square=70.23, FDR-adjusted p=3.66x10-13, W=0.562). Paired Wilcoxon tests with Holm correction identified differences for three of four adjacent task-level transitions (levels 2-3, 3-4, and 4-5), whereas levels 1-2 were not different. Conventional bilateral mean CBv also varied by level (chi-square=41.79, FDR-adjusted p=1.11x10-7, W=0.418). The mean of the recorded arterial-pressure channel did not differ after FDR correction (p=0.081), although its variance showed a small effect (W=0.110, FDR-adjusted p=0.032). These findings demonstrate that SCM provides a sensitive complementary characterization of task-related cerebral hemodynamic dynamics and supports its potential for scalable, within-person monitoring of graded cognitive demand.
Purpose Intracranial hemorrhage (ICH) is a time-sensitive emergency requiring rapid diagnosis. This study evaluated the real-world effectiveness of integrating an Artificial Intelligence (AI)-powered triage tool into a teleradiology workflow for detecting and prioritizing acute ICH on non-contrast CT (NCCT) scans, assessing the benefits on PACS-to-assessment time (PTAT) and report turnaround time (RTAT). Methods This retrospective, multi-vendor study compared NCCT interpretation for suspected acute ICH before (group1: pre-AI, June 2020) and after (group2: post-AI, June 2021) AI tool integration into a teleradiology network. In the post-AI phase, teleradiologists had access to the AI outputs. Diagnostic performance was assessed against a reference standard from two board-certified neuroradiologists. PTAT and RTAT were statistically compared between both phases, for all cases, ICH-positive findings and overnight cases. Results Among 572 patients (283 pre-AI, 289 post-AI), ICH prevalence was 12% in both groups. AI achieved 97.3% sensitivity, 96.0% specificity, and 99.6% negative predictive value. Mean PTAT and RTAT were significantly reduced by 13.72 min (p = 0.002) and 26.62 min (p = 0.001), respectively after AI integration. In ICH-positive cases, PTAT dropped by 31.43 min, though this did not reach statistical significance. During overnight hours, AI led to significant reduction of 26 min in both PTAT and RTAT (n = 131 pre-AI; n = 118 post-AI; p = 0.001). Conclusion Integrating an AI-driven tool for the detection of ICH into a teleradiology workflow significantly reduced PTAT and RTAT without compromising diagnostic accuracy. This highlights AI's potential to enhance teleradiology workflow efficiency, accelerate critical decisions, and improve patient outcomes in high-demand, resource-limited settings.
A persistent and underexamined problem in EEG functional connectivity analysis is that identical data can yield opposite interpretations depending on the analytic metric. Metrics such as weighted phase lag index (wPLI) and magnitude-squared coherence (MSC) derive from the same cross-spectral density (CSD) matrix yet capture different coupling properties, and this choice alone can reverse the reported direction of a group difference, a phenomenon we term directional fragility. We introduce Metric-T, a permutation-based diagnostic that quantifies it through a unit-erasure principle: direction consistency (DC) converts each metric into the percentage of channel-pair × frequency-band features for which one group exceeds the other, enabling cross-metric comparison on a common scale (T = DC(wPLI) − DC(MSC)). In a public dementia EEG dataset (OpenNeuro ds004504; N = 88: Alzheimer's n = 36, frontotemporal dementia n = 23, controls n = 29), reversals occurred in 3 of 12 band × comparison conditions (25%), each exceeding 20 percentage points, the largest being frontotemporal dementia vs. control in theta (T = +41.5 pp). None survived multiple-comparison correction (all corrected p > 0.10; Westfall–Young max-T minimum p = 0.234), indicating that connectivity differences in this heterogeneous cohort are directionally fragile and statistically unstable. A 700-condition simulation confirmed that reversal reflects coupling-mode composition rather than noise (p < 0.001). Metric-T requires only the CSD matrix already computed in any pipeline, is implementable in Python, and is fully reproducible. Rather than asserting a biomarker, it flags when metric choice, not neurobiology, may drive a reported group effect.
Assessment of the depth of anesthesia (DOA) primarily relies on electroencephalogram (EEG) signals. However, the high dimensionality of EEG data can increase computational complexity and limit real-time clinical applications. An efficient EEG signal representation is therefore important for practical DOA monitoring. We propose a compact statistical EEG representation that summarizes each one-second EEG segment using a set of descriptors: the mean, standard deviation, maximum, minimum, median, and the first and third quartiles (Q1 and Q3). The resulting representation, together with the electromyogram (EMG) signal, is fed into a hybrid deep learning model that combines a residual neural network (ResNet) variant with a bidirectional long short-term memory (Bi-LSTM) layer and an attention mechanism. Comparative experiments were conducted using both raw EEG signals and the proposed compact representation, and the model’s performance was benchmarked against state-of-the-art approaches. The proposed approach achieved a root mean square error (RMSE) of 5.99 ± 1.10 and an accuracy of 83.73%, outperforming several of these methods and remaining comparable to the strongest baselines, with its clearest advantage in computational efficiency: it attained the lowest inference latency among all evaluated models (1.91 ms per sample on CPU), together with substantially reduced training time and memory usage. K-fold cross-validation confirmed consistent generalization across data partitions. The combination of the compact statistical EEG representation and the hybrid deep learning model predicts DOA with competitive accuracy at low computational cost. These results indicate that the method is a promising and computationally efficient solution for real-time anesthesia monitoring.
Alzheimer's disease (AD) is a neurodegenerative condition that causes cognitive impairment and structural brain changes, and early diagnosis is crucial for treatment. This paper proposes a cloud-based hybrid deep learning architecture for scalable Alzheimer's disease staging based on structural magnetic resonance imaging (MRI). The hybrid architecture leverages convolutional neural networks (CNNs), attention, and transformer-based modeling to capture both local and global spatial features in MRI images. CNN extracts multi-dimensional features, while the attention and transformer layers enrich context representation for more accurate staging of the disease. The framework was trained on a GPU-enabled computer using the PyTorch deep learning framework on a publicly available MRI dataset for Alzheimer's disease from the Kaggle website. The model was cross-validated on the OASIS dataset to assess its generalizability. The experimental findings showed that the proposed approach achieved an accuracy of 99.65% on the Kaggle dataset, which is better than several state-of-the-art deep learning models such as VGG16, ResNet variants, DenseNet121, EfficientNet-B0, and Vision Transformer. This approach also achieved an accuracy of 91.67% on the OASIS dataset, demonstrating the model's ability to generalize across neuroimaging datasets. To enable practical usage and accessibility, the trained model was also deployed in a cloud-based inference environment on Hugging Face Spaces, which allows MRI image upload and prediction from a web browser. The scalable and cloud-deployable design allows scalable medical image processing and integration with cloud-based diagnostic systems as well as telemedicine-based healthcare solutions. These findings demonstrate the potential for robust and accessible staging of Alzheimer's disease using the proposed framework in clinical settings.
Multi-omics analysis of cerebrospinal fluid (CSF) offers a unique window into the pathophysiology of the central nervous system. Yet, the high dimensionality of such data combined with small cohort sizes create a structural imbalance. In this setting, variables far outnumber observations, undermining the performance of supervised classification models. This challenge is especially acute in neurological diseases, where phenotypic heterogeneity demands both large and representative training sets. This review positions generative AI as a foundational response to these compounded limitations. We analyze how generative adversarial networks (GANs), variational autoencoders (VAEs), and diffusion models can augment training datasets, enable and refine supervised classification. By generating synthetic samples and capturing latent disease structure, these approaches improve how accurately classifiers generalize in diagnosis and disease stratification. Generative models support trajectory modeling and the identification of biomarkers through multimodal integration. Initial studies have used GANs to generate synthetic CSF multi-omics samples, directly addressing the scarcity of patient cohorts required to train robust classifiers. VAE-based normative modeling has further demonstrated that generative approaches can capture the biological heterogeneity of CSF profiles without requiring large labeled datasets. Critical challenges related to clinical validation, interpretability, algorithmic bias, and regulatory frameworks are addressed. Generative AI would emerge as a complementary framework for supervised classification to analyze small-sample clinical settings.
Background and Objective: Pre-hospital stroke diagnosis remains a major clinical challenge due to the variability of symptoms, which can often be subtle or atypical. This variability leads to Emergency Medical Services (EMS) personnel frequently misdiagnosing stroke and therefore missing the treatment time window. This study aimed to develop an automated approach to detect strokes using video assessment of neurological tasks.Methods: We created a large database of 300 patients and 86 healthy control subjects, totaling over 50 hours of video recordings. Each subject performed all the tasks of the National Institutes of Health Stroke Scale (NIHSS), which is a standardized scale used to assess the presence and severity of stroke, creating, to our knowledge, the world’s largest video dataset of stroke patients. Using this dataset, we developed preprocessing algorithms and machine learning (ML) models to detect clinically observable stroke-related neurological symptoms in patients. A direct comparison was made between the performance of our approach and the performance of 2,000 EMS personnel trained in stroke recognition. Statistical evaluation included macro F1-score, sensitivity, and specificity metrics.Results: The proposed approach achieved superior performance compared to EMS personnel, with a 7.2% increase in macro F1-score and a 12.0% increase in sensitivity. Specificity values also demonstrated robust classification capability, confirming the reliability of the automated approach.Conclusions: Machine learning models can outperform trained EMS personnel in early stroke recognition by detecting clinically observable stroke-related neurological symptoms in pre-hospital settings. These results suggest promise for pre-hospital stroke screening, but clinical impact requires prospective real-world evaluation.
Machine learning has achieved notable success in medical imaging; however, its reliability remains constrained by the quality and quantity of training data, particularly in specialized domains such as fetal brain ultrasound. Federated learning (FL) offers a collaborative alternative by enabling multiple institutions to train shared models without exposing sensitive data. Yet, its effectiveness often degrades under non-independent and identically distributed (non-IID) data and its communication demands can be prohibitive. To address these challenges, we propose FedFB, a communication-efficient federated learning framework that integrates online ensemble knowledge distillation for privacy-aware collaborative learning. In FedFB, multiple teacher models trained locally on private datasets collectively distill their knowledge into a lightweight dual-branch student network enhanced with an auxiliary attention–convolution module, without increasing model complexity. The distillation process leverages a small auxiliary public dataset for soft-label transfer, reducing data exposure by avoiding the exchange of raw private data or model parameters, while relying on shared supervisory signals for knowledge transfer. Experimental results on fetal brain ultrasound datasets demonstrate that FedFB achieves 91.07% accuracy, 91.28% precision, 91.03% recall, and 91.15% F1-score under the non-IID conditions, while reducing communication overhead by more than 85% compared to traditional FL methods. Furthermore, the framework’s robustness was validated on brain tumor MRI and chest X-ray datasets, confirming its generalization capability across distinct medical imaging modalities.
High-frequency oscillations (HFOs) are vital biomarkers for identifying the seizure onset zone (SOZ) in patients with drug-resistant epilepsy (DRE). However, EEG artifacts especially muscle and power-line noise overlapping with the HFO frequency range (80–250 Hz) pose significant challenges for accurate detection. Traditional artifact removal methods like independent component analysis (ICA) are labor-intensive and subjective, highlighting the need for automated pre-processing techniques. This study introduces CleanEEG, a U-Net based encoder–decoder model designed to automate artifact removal from clinical EEG. CleanEEG was trained on paired noisy and clean sleep EEG segments from 25 DRE patients (177 segment pairs) at a 512 Hz sampling rate, with clean targets generated through ICA pre-processing. Model performance was quantitatively evaluated on an independent validation set comprising 24 segment pairs from six separate patients excluded from training. Evaluation metrics included relative root mean square error (RRMSE), correlation coefficient (CC), and signal-to-noise ratio (SNR). CleanEEG effectively removed muscle and power-line noise artifacts while preserving important clinical features such as interictal epileptiform discharges (IEDs) and brief potentially ictal rhythmic discharges (BIRDs). The model significantly improved signal quality across electrodes, reducing reconstruction errors and increasing SNR. Additionally, CleanEEG preserved neural activity without introducing distortions and qualitatively demonstrated artifact removal capability on unseen awake EEG data. In a representative DRE patient, critical spatial patterns of HFOs were maintained, essential for accurate SOZ localization. Overall, CleanEEG offers an automated, robust, and efficient solution for artifact removal, enhancing diagnostic accuracy in epilepsy monitoring and HFO analysis, particularly in long-term scalp EEG recordings.
Careful evaluation of research methodology is fundamental to scientific progress but represents a significant burden on human experts. The complexity of functional MRI (fMRI) methods makes transparent reporting, as suggested by OHBM COBIDAS guidelines, particularly critical. Large Language Models (LLMs) present a potential solution for rapid, scalable methodological assessment. We evaluated three state-of-the-art LLMs (Gemini 2.5 Pro, Claude 4 Sonnet, ChatGPT-o3-pro) against human expert ratings. Fifty fMRI articles (taken from 2016 to 2025) were independently evaluated by ten human experts and three LLMs using an 82-item COBIDAS based rubric. Human raters demonstrated excellent inter-rater reliability (ICC = 0.801), while LLMs showed poor internal agreement (ICC = 0.254). When comparing total scores across papers, Gemini showed strong positive correlation with human consensus (r = 0.693, p < 0.0001), Claude showed moderate positive correlation (r = 0.394, p = 0.004), while ChatGPT showed negative correlation (r = -0.172, p = 0.233). Gemini maintained high reliability when added to human raters (combined ICC = 0.811), achieving 85.3 % exact agreement and 98.8 % within-1-point agreement. Domain-specific analysis revealed Gemini's consistently high agreement across all six COBIDAS sections (experimental design: 0.915, statistical modeling: 0.880), while ChatGPT and Claude showed weaker, more variable performance. Obvious differences emerged in determining non-applicable items: humans marked 40.5 % as not applicable versus 32.3 % for Gemini, 9.2 % for ChatGPT and 21.1 % for Claude. ChatGPT exhibited extreme score volatility, with papers ranging from 0 to 121 points compared to humans' 44.2-77.7 range. LLM scoring required 1-7 min versus 30-35 min for humans. This proof-of-concept study demonstrates that LLM-assisted methodological evaluation is feasible for complex neuroimaging research and could likely be applied to other research fields.
Background Most neurological care is delivered in outpatient settings without mandated clinical coding. The clinical records remain stored as unstructured text with inconsistent formatting. There is a significant opportunity to increase the value of these data through automated clinical coding utilising natural language processing (NLP). While existing models for full ICD-10 clinical coding lack sufficient accuracy for clinical use, 60% of neurology outpatient cases fall into just five diagnostic categories. This suggests that a simplified coding system could enhance feasibility and serve as a foundation for more complex coding schemes. Objective We propose a simplified coding system of 29 codes for neurology outpatient episodes. We evaluate several machine learning methods in a supervised single-label classification task on real-world outpatient care notes. Methods We collected outpatient care notes created between 15 November 2018 and 2 December 2022. The training dataset included 14,917 care notes, most of which were annotated with ICD-10 codes during routine care and subsequently mapped to 29 simplified diagnostic categories. An external validation set of 1,042 randomly selected encounters was retrospectively coded.Models included logistic regression, support vector machine, bidirectional LSTM, BERT-based models (DistilBERT, RoBERTa), and a generative large language model (LLM), Mistral 7B. All but the LLM were trained via 10-fold stratified cross-validation; final models were trained on the complete dataset. Results DistilBERT and RoBERTa outperformed traditional models, with F1-scores of 81.73 (95% CI: 79.02–84.13) and 81.16 (95% CI: 78.84–83.76), respectively. The LLM–DistilBERT hybrid performed worse than all but BiLSTM and produced “medical hallucinations,” making it unsuitable for clinical use. The training data were highly imbalanced. BERT-based models showed strong performance on high-frequency categories, with F1-scores over 85% for the top five classes. At a 0.85 confidence threshold, DistilBERT achieved 96% accuracy on 64% of the external validation set. Conclusions BERT-based NLP models perform well in classifying neurology outpatient clinic notes when a reduced set of diagnostic categories is used. In a human-in-the-loop workflow, such models can meaningfully reduce the manual coding workload while preserving accuracy. To our knowledge, this is the first applied study of automated clinical coding in neurology outpatient care.
Background Many studies on resting-state functional connectivity (FC) in major depressive disorder (MDD) have investigated FC as a biomarker of disease pathogenesis. However, few studies have examined conditional dependencies among FC, clinical status, and demographic variables. Considering such dependencies allows the identification of direct relationships obscured by spurious correlations. Aim This study aimed to examine the neural mechanisms of MDD and propose a structural relationship between FC and MDD, focusing on sulcal regions. Methods Using a large dataset of 431 healthy controls and 235 MDD patients with MDD, we combined partial least squares (PLS)-based feature extraction with logistic regression and light gradient boosting machine (LightGBM) models for diagnostic classification, followed by Bayesian network (BN) analysis employing a directed acyclic graph. Results The classification models demonstrated moderate accuracy (logistic regression: area under the curve [AUC] = 0.735; LightGBM: AUC = 0.710). Structure learning with the Max–Min Hill-Climbing algorithm revealed direct edges from the MDD diagnosis to variables derived from the BDI and PLS components, but no direct parent nodes of MDD were identified. Intervention simulation showed that the MDD diagnosis significantly reduced FC in the default mode network (DMN), dorsal attention network, and between subcortical structures and cortex. Conclusion MDD diagnosis is associated with disease-specific disruptions not only in the DMN but also across multiple networks, underscoring the need to consider widespread network dysfunction in the pathophysiology of MDD. Future longitudinal and interventional research is required to clarify the causal relationships between the diagnosis and brain function.
Accurate and automated meningioma segmentation remains a biomedical engineering challenge, particularly when relying on single-modality MRI data. We evaluate SegResNet, a U-Net-based deep learning architecture, for meningioma segmentation using 817 T1-contrast-enhanced (T1CE) magnetic resonance imaging (MRI) images from 282 patients from across Auckland, New Zealand. We investigate the effect of incorporating additional images from the 2023 Brain Tumour Segmentation (BraTS) meningioma challenge during training on model performance. The baseline model trained solely on the Auckland dataset achieved 75.67% mean Dice. Incorporating an additional 200 and 400 BraTS images improved segmentation performance to 77.89% and 76.73%, respectively. A separate experiment involving pre-training on BraTS data followed by fine-tuning on Auckland data achieved 75.90% Dice. Our results suggest that while leveraging external datasets can enhance model robustness, the extent of improvement depends on dataset heterogeneity and alignment with the target domain.Analysis of a subset of images unaffected by skull-stripping artifacts indicated notably higher segmentation accuracy (up to 84.02% Dice), highlighting the influence of preprocessing on performance. Evaluations using the 2023 and 2024 BraTS lesion-wise metrics demonstrated that importance of context-appropriate metric selection. Our findings highlight the adaptability of SegResNet to a single-modality T1CE – a widely available sequence in standard clinical protocols – clinical dataset and emphasize how public data integration, careful preprocessing, and task-aligned evaluation can support robust segmentation models for diverse and resource-constrained environments.
In acute ischemic stroke, large vessel occlusions of the anterior circulation are increasingly treated with endovascular therapy (EVT). The efficacy of this therapy depends on adequate treatment selection. Treatment decisions can be based on predictions of functional outcome. Most existing studies predict functional outcomes using clinical parameters. We set out to study functional outcome prediction performance by integrating imaging in a multimodal setting. Using a multi-center dataset containing 2927 patients, we compare the functional outcome prediction performances of clinical baseline models, including the clinically validated MR PREDICTS decision tool, image-based models with deep learning networks, and a multimodal approach combining clinical and imaging information. The predicted outcome measure is dichotomized modified Rankin Scale score 90 days after EVT. We perform sanity checks, hyperparameter optimization, and comparisons of effectiveness of using CTA, NCCT, or both images as input. Our experiments show that information extracted from CTA or NCCT images does not significantly improve the performance, as quantified using AUC, of functional outcome prediction methods compared to a baseline model. The multimodal approach may replace radiologically derived biomarkers, as its performance is non-inferior.
Purpose Diffusion-weighted MRI is widely used to probe brain microstructure, but its signal reflects both diffusion and perfusion effects. Intravoxel Incoherent Motion (IVIM) MRI enables non-contrast separation of these components, offering potential clinical value in neuroimaging. However, clinical translation remains limited due to variability in acquisition and post-processing methods. This systematic review evaluates optimization strategies aimed at improving the accuracy, reproducibility, and clinical utility of IVIM parameters in brain. Methods Registered in PROSPERO and conducted according to PRISMA guidelines, a systematic search across five databases was performed. Original peer-reviewed studies focusing on optimization of IVIM acquisition or post-processing in human brain imaging were included, while reviews and studies lacking methodological detail were excluded. Study quality was assessed using a customized QUADAS-2 tool. Due to methodological heterogeneity, an effect direction plot was applied instead of meta-analysis. Results Out of 1,668 identified records, 14 studies were included. Acquisition strategies such as optimized b-value sampling, cardiac gating, and advanced sequences reduced parameter variability by up to 40%. Post-processing methods, including Bayesian fitting, deep learning–based models, and advanced denoising, improved parameter accuracy by up to 99% and precision by up to 95%. Effect direction analysis demonstrated significant positive effects on accuracy and clinical utility (p < 0.001) and repeatability (p < 0.05), while scan-time reduction showed no significant benefit (p > 0.05). No study reported gold-standard validation. Conclusion Optimization of IVIM acquisition and post-processing enhances parameter robustness and reproducibility in brain MRI, though protocol heterogeneity remains a major obstacle to standardization and clinical adoption.
This study investigated age-related changes of brain fiber pathways from diffusion tensor imaging (DTI) tractograms with directional decomposition. Two hundred subjects were stratified into three age groups. Tractograms were generated at two levels: from individual DTI images (subject-level), and from group-averaged images (group-level). Fiber tracking was performed within the cerebral white matter, brainstem, thalamus, and cerebellum at both the levels. Each tractogram was decomposed into directional tracts. At the subject-level, original and decomposed tracts were used to quantify tract density and correlations with age. Tract density was highest in the thalamus and brainstem, while the cerebellum showed the greatest inter-subject variability. Tract count exhibited some significant correlations with age: in cerebral white matter, it decreased overall, especially along S-I and A-P directions; in thalamus, S-I and A-P tracts decreased, while L-R and mixed-direction tracts increased. The brainstem tracts demonstrated its overall stability during aging. At the group level, ∼60% of brainstem tracts were oriented along the S–I direction, and ∼64% of cerebellar tracts along the A–P direction. Notably, the posterolateral tracts of the cerebellum showed asymmetry, with the left side associated with visuospatial processing, containing fewer tracts than the right side associated with language pathways. These findings highlight region- and direction-specific changes with age, revealing structural patterns that are not captured by conventional scalar measures. They suggested candidate biomarkers for brain aging and provided useful references for longitudinal neuroimaging and brain stimulation studies, with potential applications in the early detection of neurodegeneration and optimization of stimulation strategies.
High-resolution physiological signals, such as intracranial pressure (ICP) and regional cerebral oxygen saturation (rSO2), are critical for managing traumatic brain injury (TBI) by enabling continuous monitoring of cerebral autoregulation and vascular reactivity. These signals provide essential insights into brain perfusion dynamics, supporting timely clinical interventions. However, the high temporal resolution of these data introduces challenges in real-time use, integration into predictive models, and computational efficiency. Consequently, resolution reduction techniques are essential for simplifying the data while retaining critical features necessary for accurate prediction and modeling. Using the Multi-omic Analytics and Integrative Neuroinformatics in the HUman Brain (MAIN-HUB) Lab database, high-frequency cerebral physiologic dataset, we aimed to evaluate the effects of point-sampling resolution reduction on point and interval predictions using the autoregressive integrated moving average (ARIMA) model for both raw signals and derived indices. Temporal resolution was reduced by selecting the first value within non-overlapping intervals, ranging from 1-min (min) to 12-h windows. A total of 125 patient data was analyzed across various temporal resolutions. The results indicated that ARIMA models performed well at higher resolutions (e.g., 1-min), capturing short-term physiological dynamics with lower errors. However, as resolution decreased, errors and variability increased, particularly for signals like mean arterial pressure (MAP) and cerebral perfusion pressure (CPP), which exhibit rapid or complex physiological changes. The findings underscore the need to carefully balance temporal resolution, model performance, and computational efficiency, especially when dealing with high-frequency physiological data in clinical settings.
Ventriculomegaly is a key neuroimaging feature in conditions such as normal pressure hydrocephalus (NPH) and other disorders of cerebrospinal fluid (CSF) dynamics. The Evans Index (EI), defined as the ratio of the maximal frontal horn width to the maximal inner skull diameter, remains a simple and widely used marker for ventricular enlargement. However, manual EI measurement is subject to observer variability and dependent on proper alignment to the anterior commissure-posterior commissure (AC-PC) plane, limiting reproducibility in large-scale and multi-center studies. We present a fully automated deep learning-based pipeline for EI calculation directly from raw T1-weighted MPRAGE MRI scans. The pipeline integrates (i) landmark detection using the BrainSignsNet model, (ii) rigid AC-PC alignment, and (iii) robust segmentation of the lateral ventricles (LV) and intracranial volume (ICV) via nnU-Net models, including a custom ventricular network trained on 1,300 annotated scans enriched for hydrocephalus. The Evans Index is then derived from automated measurement of frontal horn width and inner skull diameter at the aligned axial slice. Internal validation using data from the Baltimore Longitudinal Study of Aging, BIOCARD, and Johns Hopkins cohorts demonstrated high segmentation accuracy (Dice coefficient = 0.98). External validation on the PENS trial, including pre- and post-shunt NPH scans, showed excellent agreement with expert manual EI measurements (mean bias = 0.0068, mean absolute error = 0.0103, r = 0.96, p < 0.001). Bias analyses revealed no significant association between measurement error and age, sex, or ventricular volume This fully automated, orientation-standardized method achieves accurate and reproducible Evans Index measurement across diverse MRI datasets. By eliminating manual intervention, the pipeline enhances scalability for large neuroimaging cohorts and provides a reliable tool for clinical assessment, screening and monitoring of ventriculomegaly, especially in NPH.
