Open Access

ARTICLE

A Multi-Specialist Stacking Decoder of Cognitive Workload and a Decomposition of the Limits of Cross-Dataset Transfer in Electroencephalography

Sugeng Rifqi Mubaroq^1,*, Rolly Maulana Awangga², Tegar Ditya Pragama¹, Sidiq Fathummubin³, Ali Yusuf Abdulhaq¹

1 Digital Business Study Program, Akademi Digital Bandung, Jl. Rancamekar Blok Lio, Bandung, West Java, Indonesia
2 Department of Informatics Engineering, Universitas Logistik dan Bisnis Internasional, Jl. Sariasih No. 54, Sarijadi, Sukasari, Bandung, West Java, Indonesia
3 Software Engineering Technology Study Program, Akademi Digital Bandung, Jl. Rancamekar Blok Lio, Bandung, West Java, Indonesia

* Corresponding Author: Sugeng Rifqi Mubaroq. Email:

Intelligent Automation & Soft Computing 2026, 41, 27-46. https://doi.org/10.32604/iasc.2026.088039

Received 27 June 2026; Accepted 21 July 2026; Issue published 11 August 2026

Abstract

Decoding cognitive workload from electroencephalography (EEG) underpins passive brain–computer interfaces and adaptive learning technology, yet practical decoders share two weaknesses: they rely on a single family of features, and their accuracy collapses on recordings from an unfamiliar device, montage, or task. We address both. We first build a multi-specialist stacking decoder that fuses complementary spectral, Riemannian, and spatial views through a meta-learner. Evaluated across two public corpora under the leave-one-subject-out MOABB benchmarking protocol, it outperforms the best single specialist on the binary workload contrasts, and the strongest of these effects survives family-wide false-discovery-rate correction. The leading view switches between datasets, which is why single-view decoders are brittle. Effect-size profiling further shows that the decoders track the canonical neural signature of load: rising theta and falling alpha power. We then ask how workload models transfer between corpora. Naive pooling degrades accuracy, and we trace this negative transfer to two causes: the corpora label workload differently, and their signals differ at the device level, a component we corroborate with a third research-grade dataset that varies task and site while holding the device class fixed. Harmonizing the labels and applying class-conditional alignment then restores a small positive transfer, though only in one direction: COG-BCI benefits from Simultaneous Task EEG Workload (STEW), not the reverse. Rather than a headline accuracy figure, our contribution is an evidence-based account of when, and why, pooling EEG datasets is worthwhile.

---

Keywords

---