Agastya, Andiar Rinanda (2026) Design And Development Of A Non-Invasive Brain-Computer Interface Based On 14-Channel Electroencephalography (EEG) For Decoding Imagined Speech As A Non-Motor Augmentative and Alternative Communication (AAC). Other thesis, Institut Teknologi Sepuluh Nopember.
![[thumbnail of 5026221219-Undergraduate_Thesis.pdf]](http://repository.its.ac.id/style/images/fileicons/text.png) |
Text
5026221219-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (7MB) | Request a copy |
Abstract
Otak manusia mengatur fungsi bahasa dan wicara melalui jaringan kortikal motorik pada area bahasa, seperti area Broca dan Wernicke, yang berkoordinasi melalui lintasan dorsal dan ventral. Pada beberapa individu, fungsi-fungsi tersebut hilang meskipun kemampuan kognitif mereka tetap utuh, misalnya pada penyintas locked-in syndrome. Kondisi ini mengakibatkan penderita kehilangan kendali motorik, mengalami kesulitan dalam mengekspresikan keinginan mereka, dan menanggung beban medis kronis akibat komplikasi, yang secara keseluruhan memperburuk kualitas hidup mereka. Penelitian ini berfokus pada pengembangan sistem komunikasi asisitif non-motorik yang memanfaatkan aktivitas sinyal otak secara langsung dan non-invasif, tanpa memerlukan fungsi motorik, melalui proses dekode ucapan yang dibayangkan (imagined speech) menjadi kalimat yang dapat dilisankan. Sistem ini menggunakan perangkat elektroensefalografi (EEG) Emotiv EPOC-X 14-kanal (256 Hz), dengan sinyal yang diproses melalui band-pass filter 0,5 hingga 50 Hz dan dievaluasi pada enam paradigma pemodelan. Pemodelan tersebut mencakup pendekatan deep learning (EEGNet) serta pendekatan machine learning klasik (Support Vector Machine yang dilatih menggunakan fitur spektral hand-crafted). Dari evaluasi tersebut, Support Vector Machine berbasis subject-dependent muncul sebagai kandidat klasifikasi datar (flat-classification) yang secara signifikan lebih kuat (Wilcoxon p = 0,0049). Model ini kemudian dikembangkan lebih lanjut menjadi kaskade hierarkis coarse-to-fine yang mendekomposisi dekode suku kata berdasarkan struktur vokal dari kosakata target. Suku kata hasil dekode selanjutnya dirangkai menjadi kata, disempurnakan menjadi kalimat menggunakan penyempurna kalimat berbasis aturan (rule-based sentence-refiner), dan dilisankan secara komunikatif melalui text-to-speech. Cakupan dekode meliputi sepuluh kata dasar bahasa Indonesia yang terdiri dari 19 kelas suku kata. Evaluasi dilakukan terhadap 12 subjek sehat berdasarkan akurasi keluaran suku kata dan kata, cakupan kelas, serta latensi end-to-end yang diukur pada nilai median dan persentil ke-95. Hasil evaluasi menunjukkan bahwa kaskade coarse-to-fine mencapai akurasi end-to-end sebesar 28,57 persen untuk suku kata pertama dan 13,89 persen untuk kata utuh. Pencapaian ini merupakan peningkatan yang signifikan secara statistik dan berdasar secara struktural dibandingkan dengan baseline datar (p = 0,0005) yang dikonfirmasi melalui perbandingan multikriteria yang adil, dengan latensi sistem end-to-end terukur pada median 1058 milidetik dan persentil ke-95 sebesar 1105 milidetik. Temuan dari penelitian ini mewujudkan sebuah sistem komunikasi asisitif non-motorik dengan potensi penerapan yang jelas, termasuk namun tidak terbatas pada individu yang mengalami locked-in syndrome.
==========================================================================================================================================
The human brain orchestrates language and speech functions through motor cortical networks in language areas such as Broca's and Wernicke's areas, which coordinate through dorsal and ventral pathways. In some people, these functions are lost even though their cognitive abilities remain intact, for example in survivors of locked-in syndrome. This condition causes sufferers to lose motor control, experience difficulty expressing their desires, and carry chronic medical burdens due to complications, all of which worsen their quality of life. This research focuses on the development of a non-motor assistive communication system that utilizes direct brain signal activity non-invasively, without requiring motor function, by decoding imagined speech into sentences that can be spoken. The system uses a 14-channel Emotiv EPOC-X electroencephalography (EEG) device (256 Hz), with signals processed through a 0.5 to 50 Hz band-pass filter and evaluated across six modeling paradigms spanning both a deep learning approach (EEGNet) and a classical machine learning approach (Support Vector Machine trained on hand-crafted spectral features), from which the subject-dependent Support Vector Machine emerged as the significantly stronger flat-classification candidate (Wilcoxon p = 0.0049) and was further developed into a coarse-to-fine hierarchical cascade that decomposes syllable decoding along the vowel structure of the target vocabulary, with decoded syllables assembled into words, refined into sentences using a rule-based sentence-refiner, and spoken communicatively through text-to-speech. The decoding scope covers ten basic Indonesian words comprising 19 syllable classes. Evaluation was conducted on 12 healthy subjects based on syllable and word output accuracy, class coverage, and end-to-end latency measured at the median and 95th percentile, showing that the coarse-to-fine cascade achieved 28.57 percent first-syllable and 13.89 percent full-word end-to-end accuracy, a statistically significant and structurally grounded improvement over the flat baseline (p = 0.0005) confirmed through a fair, multi-criteria comparison, with end-to-end system latency measured at a median of 1058 milliseconds and a 95th percentile of 1105 milliseconds. The findings of this study constitute a non-motor assistive communication system with clear potential for application, including but not limited to people with locked-in syndrome.
| Item Type: | Thesis (Other) |
|---|---|
| Uncontrolled Keywords: | imagined speech, elektroensefalografi, sistem komunikasi asisitif, machine learning, deep learning, text-to-speech, locked-in syndrome, imagined speech, electroencephalography, assistive communication system, machine learning, deep learning, text-to-speech, locked-in syndrome |
| Subjects: | T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK5102.9 Signal processing. |
| Divisions: | Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Information System > 57201-(S1) Undergraduate Thesis |
| Depositing User: | Andiar Rinanda Agastya |
| Date Deposited: | 29 Jul 2026 07:07 |
| Last Modified: | 29 Jul 2026 07:07 |
| URI: | http://repository.its.ac.id/id/eprint/139318 |
Actions (login required)
 |
View Item |