Published January 1, 2026 | Version v1

Early Migraine Prediction Using TabNet And TabTransformer: A Comparative Deep Learning Study On Clinical Data

Description

Migraine is a prevalent neurological disorder affecting approximately one billion people worldwide, causing considerable disability and significant socioeconomic consequences. Developing effective treatment options requires accurate bracketing of migraine subtypes; yet, this process often requires specialized expertise that may be lacking in many healthcare settings. Although there are yet few comparative assessments of these infrastructures, attention-grounded deep literacy approaches for irregular clinical data suggest a potential path for automated bracket support. Using structured clinical data, this work provides a systematic relative analysis of TabNet and TabTransformer for early migraine prediction. The Kaggle Migraine Dataset, which included 100,000 case records with 24 clinical characteristics and seven migraine subtypes, was used to create the dataset. To address class imbalance, the approach used balanced class weighting, categorical encoding, and stratified train-test separation. F1-score, ROC-AUC, recall, precision, and delicacy were used to measure performance. With a test accuracy of 99.50 as opposed to TabTransformer\\\'s 92.38, experimental data show that TabNet works noticeably better than TabTransformer. Perfection (99.54), recall (99.50), F1-score (99.51), and near-perfect ROC-AUC (0.9999) were all enhanced by TabNet, with consistent performance across all seven migraine subtypes. Direct point significance visualization was made possible by TabNet\\\'s effective attention medium, providing clinically interpretable perceptivity that met ICHD-3 individual criteria. Because of its quick confluence and memory efficiency, it is useful in environments with limited resources. With significant ramifications for automated clinical decision assistance, these results validate TabNet as a model for migraine bracket tasks that is generally efficient and comprehensible.

Files

IJSET_V12_issue2_650.pdf

Files (463.9 kB)

Name Size
md5:2ffe5b5f8a3bb2385963d603c44cabd7
463.9 kB Preview Download

Additional details