Equation-Level Parameterized Fusion Reformulation for Multimodal Epileptic Seizure Detection Using Interaction Control and Data-Quality Screening

ABDUL-MUMIN KHALID^1*Musah Sulemana¹Iddrisu Wahab Abdul²

• ¹University for Development Studies, Tamale, Ghana
• ²Ghana Communication Technology University, Accra, Ghana

Epileptic seizure detection remains especially difficult to perform reliably due to noise, inter-subject variability, and the poor generalization ability of unimodal learning models. To tackle these problems, this paper proposes an equation-level transformation mechanism for multimodal epileptic seizure detection that aggregates EEG, ECG, EMG, and ACC signals via parameterized fusion and interaction management. To do this, the framework presents four adaptive parameters (a fusion exponent, 𝜌, an interaction weight, 𝛿), a stabilization factor (𝜆) and a synergy amplifier (𝜂) that tightly control different aspects of modality contribution or nonlinearity, cross-modal dynamics, and synergistic amplification under one coherent mathematical formalization for both traditional and deep learning models. Their study is based on a multimodal dataset recording 120 patients with clinically diagnosed epilepsy, comprising 60 from the Tamale Teaching Hospital and 60 from public datasets. The signals were sampled at 512Hz and divided into 2s windows with 50% overlap, yielding approximately 1,024,000 labelled samples. A formal DQA model and an NCS index were used to verify signal reliability and cross-source alignment before fusion. Twelve machine learning classifiers were trained and tested using a strict patient-wise data split to avoid data leakage. Experiments show that after equation-level reformulation, clear and consistent performance gains are achieved across all models. Classifiers using traditional machine learning improved from 55–67% the range of 82–92%, while those using deep learning models improved from 70–82% to the range of 89%-97.9%%, with the Transformer-based model achieving the highest accuracy. Software efficiency analysis shows that the gains from our framework do not come at an unreasonable computational cost. Altogether, the findings corroborate equation-level multimodal fusion as a practical approach and establish a generalizable, interpretable, and reproducible basis for robust seizure detection, one and beyond, into the realm of biomedical time-series analysis.