An Artificial Intelligence-Based Decision- Support Tool to Predict Outcomes in Low Back Pain

Presenter: S. Singh

This study developed an artificial intelligence (AI)-based decision-support tool that combines clinical data and radiomics to generate personalized treatment recommendations for patients with low back pain (LBP). The tool uses deep learning to automatically segment mid-sagittal T2-weighted magnetic resonance imaging (MRI) scans and extract anatomical measurements, including disc height and width, vertebral height and width, spinal canal diameter, disc height index, signal intensity, and disc volume. The segmentation model was trained using 402 patient MRI scans, while 4,260 disc slices with radiological geometric parameters were used for spondylosis grading. The dataset was labeled using expert-verified Pfirrmann grades and 11-grade spondylosis severity classifications to reduce variability associated with manual grading. The imaging data were combined with Oswestry Disability Index (ODI) and Numerical Rating Scale (NRS) scores to generate individualized treatment recommendations. The DeepLabV3+ segmentation model with a ResNet50 encoder achieved an accuracy of 95.5%, which improved to 98.7% after 8-fold cross-validation. After cross-validation, the model also achieved 96.95% precision, 97.1% recall, a Dice coefficient of 96.9%, and an intersection over union (IoU) of 94.8%. For spondylosis severity prediction, the convolutional neural network (CNN) with MobileNetV2 achieved 97.84% accuracy and 96.76% IoU after cross-validation. Using geometric imaging data, the Gradient Boost classifier achieved 91.65% accuracy and 84.59% IoU.

Overall, the study demonstrates the potential of an AI-based tool that integrates imaging features and clinical assessments to support personalized treatment planning for patients with low back pain.

Beyond Traditional Screening: AI-Driven Early Detection of Cognitive Disorders and Dementia

Presenter: Z. Gelencsér

This study developed and evaluated PreDEM, an artificial intelligence (AI)-based platform designed to identify early signs of dementia by analyzing digital behavioral patterns. The aim was to support earlier diagnosis, as dementia is often detected late because initial symptoms are subtle. Data were collected from 4,334 participants between September 2021 and March 2025. During this period, participants completed 171,577 cognitive tests, generating 6,049,916 data points. A standardized and normalized evaluation system was used to enable objective comparisons across different cognitive tests. The AI algorithm analyzed participants' digital behavioral patterns to detect early indicators of dementia. Based on the AI analysis, participants were classified into "potentially treated" and "potentially healthy" groups. Density function and cluster analyses demonstrated clear differences between these groups. Within the affected group, the AI identified a lower-performing subgroup, and subsequent neurological and neuropsychological assessments confirmed the presence of cognitive impairment. In a higher-performing affected subgroup, cognitive dysfunction was frequently associated with organic causes. Treatment of these underlying conditions, together with continued use of the PreDEM platform, was associated with significant cognitive improvement.

Overall, the findings suggest that the PreDEM AI platform can identify early signs of dementia from digital behavioral data and may support earlier diagnosis and intervention. In pilot studies, the AI system reportedly outperformed traditional diagnostic approaches in identifying dementia.

Artificial Intelligence Analysis of Patient Videos for Neurological Diagnosis: A Systematic Review

Presenter: C. Tatit

This systematic review evaluated the use of artificial intelligence (AI)-based video analysis for the detection, classification, and quantification of neurological disorders. Following a registered PROSPERO protocol, databases were searched from inception to June 2025. Studies using AI to analyze video recordings of neurological signs were included. Data on clinical applications, video type, AI model, validation methods, diagnostic performance, and risk of bias were extracted. Study quality was assessed using the Newcastle–Ottawa Scale. Due to heterogeneity among studies, a meta-analysis was not performed. A total of 30 studies involving 22,532 participants met the inclusion criteria. AI performance varied across neurological conditions. For acute stroke and transient ischemic attack (TIA) detection, AI achieved an accuracy of 85.8%. For post-stroke motor deficit quantification, reported accuracies ranged from 78% to 83%, while facial weakness detection achieved an accuracy of 94.3%. Neonatal seizure detection demonstrated specificity greater than 90%. For movement disorders, including bradykinesia, tremor, gait abnormalities, and tics, reported accuracies ranged from 73% to 97%. In pediatric and fetal neurological applications, AI models achieved area under the curve (AUC) values of 0.80–0.90 for identifying early cerebral palsy and congenital malformations. Most included studies were of moderate quality, with limited external validation.

Overall, the review suggests that AI-based video analysis has the potential to support neurological assessment across a range of acute, chronic, and developmental disorders. However, further multicenter studies with standardized methods and robust external validation are needed before widespread clinical implementation.

Comparing AI with Specialist Ground Truth on Each Headache Attack Using Headache Diary App

Presenter: M. Katsuki

This study evaluated whether artificial intelligence (AI) can accurately classify individual migraine attacks compared with headache specialists. The International Classification of Headache Disorders, 3rd edition (ICHD-3) is designed to diagnose migraine at the patient level but is not intended to classify individual headache attacks. The study aimed to create an expert-reviewed reference dataset and compare the performance of three AI-based classification methods. Patients attending headache specialty clinics in Japan completed standardized questionnaires and recorded their headache attacks using the Migraine Buddy smartphone application for 3 months. Three board-certified headache specialists independently assessed 10 attacks from each of 10 patients. Each attack was assigned a migraine probability score, and majority agreement was used to classify attacks as migraine or non-migraine, providing the reference standard. Of the 100 recorded attacks, 98 were eligible for analysis. Specialists classified 82 attacks as migraine and 16 as non-migraine. The ICHD-3 rule-based decision tree showed high specificity (0.81) but low sensitivity (0.43). Large language models (LLMs) demonstrated high sensitivity (0.95) and moderate specificity (0.69), although their performance varied across repeated analyses. The XGBoost machine learning model provided the most balanced results, with a sensitivity of 0.84 and specificity of 0.88.

Overall, AI-based approaches showed good agreement with specialist assessments for classifying individual migraine attacks. Among the methods evaluated, the XGBoost model achieved the most balanced performance, while the authors emphasized that expert clinical judgment remains important.

EAN 2026; June 27-30, Geneva, Switzerland.