Management and Future Technologies | Volume 2, Issue 4 | № 008

Namlik darajasining oʻzgarishi dielektrik parametrlarga sezilarli taʼsir koʻrsatdi. Eksperimental maʼlumotlar tahlili shuni koʻrsatdiki, namlik darajasi 8% dan 16% gacha oshganida, barcha navlar uchun dielektrik oʻtkazuvchanlik qiymati 35-42% ga oshdi. Eng sezilarli oʻzgarish 12-14% namlik oraligʻida kuzatildi, bu esa urugʻlarning suvni oʻzlashtirish xususiyatlarining maksimal darajada namoyon boʻlishi bilan izohlandi. Raqamli boshqarish tizimining ishlash samaradorligini baholash uchun oʻtkazilgan tajribalar natijasida quyidagi natijalar qayd etildi. Tizimning oʻrtacha saralash tezligi 120±5 urugʻ/soniya ni tashkil etdi, bu anʼanaviy usullarga nisbatan 3 baravar yuqori koʻrsatkichdir. Saralash aniqligi 94.2±1.8% darajasida qayd etilgan boʻlsa-da, bu koʻrsatkich turli navlar uchida biroz farq qiladi. Xatolik tahlili shuni koʻrsatdiki, notoʻgʻri saralashning asosiy qismi (78%) birinchi tur xatoliklar hisobiga sodir boʻlgan, yaʼni sifatli urugʻlarning rad etilishi. Ikkinchi tur xatoliklar (yomon urugʻlarning qabul qilinishi) 22% ni tashkil etdi. Xatolarning eng koʻp qayd etilgan sabablari qatoriga sensorlarning sezgirlik chegaralari (45%), maʼlumotlarni qayta ishlash algoritmlarining noaniqligi (32%) va muhit omillarining taʼsiri (23%) kiradi. Tadqiqot davomida chang, harorat va namlik kabi muhit omillarining tizim ishlashiga taʼsiri chuqur oʻrganildi. Changli sharoitda (5 g/m³) tizimning ishlash samaradorligi 12.3% ga kamaydi, asosan sensorlarning ifloslanishi hisobiga. Haroratning 25°C dan 35°C ga koʻtarilishi dielektrik oʻlchovlarning aniqligiga 8.7% salbiy taʼsir koʻrsatdi. Namlikning 50% dan 80% ga oshishi esa tizimning xatolik darajasini 15.2% ga oshirdi, bu esa yuqori namlik sharoitida urugʻlarning dielektrik parametrlarining oʻzgaruvchanligi bilan bogʻliq. Eng barqaror natijalar 20-25°C harorat va 40-60% nisbiy namlik sharoitida qayd etildi. Diskriminant analiz natijalari shuni koʻrsatdiki, turli navdagi urugʻlarni 92.4% aniqlik bilan ajratish mumkin. Eng muhim discriminant funktsiya dielektrik oʻtkazuvchanlik (0.84) va namlik darajasi (0.76) parametrlariga asoslangan. ANOVA tahlili turli navlar guruhlari oʻrtasida sezilarli farqlar mavjudligini koʻrsatdi (p < 0.001). 4-rasm. Saralashdagi xatoliklarning asosiy turlari. Korrelyatsiya tahlili dielektrik oʻtkazuvchanlik va urugʻlarning unib chiqish darajasi oʻrtasida kuchli musbat bogʻliqlik mavjudligini aniqladi (r = 0.78, p < 0.01). Regressiya analizi asosida urugʻlarning sifatini bashorat qilish uchun matematik model ishlab chiqiladi, uning aniqligi 89.3% ni tashkil etdi. Tizimning 100 soatlik uzluksiz ishlashi davomida quyidagi natijalar qayd etildi. Dastlabki 24 soat davomida tizimning samaradorligi 5.2% ga yaxshilandi, bu tizimning "ishga tushirilishi" jarayoni 60

bilan bogʻliq. Keyingi 48 soat davomida tizim barqaror ishladi, samaradorlik oʻzgarishi 2.1% dan oshmadi. 72 soatdan keyin tizim samaradorligining pasayishi kuzatildi, asosan sensorlarning charchashi va maʼlumotlar bazasining toʻlishi hisobiga. 100-soatlik test yakunida tizimning umumiy samaradorligi 8.7% ga kamaydi, bu muntazam texnik xizmat koʻrsatish zarurligini koʻrsatadi. Tadqiqot natijalari asosida tizimni takomillashtirish boʻyicha quyidagi takliflar ishlab chiqildi. Sensorlar tizimini yangilash orqali xatolik darajasini 32% ga kamaytirish mumkin. Maʼlumotlarni qayta ishlash algoritmlarini optimallashtirish orqali saralash tezligini 18% oshirish imkoniyati aniqlangan. Muhit omillarining taʼsirini kamaytirish uchun izolyatsiya tizimini joriy etish tavsiya etildi. Tizimning energiya samaradorligini oshirish boʻyicha takliflar ishlab chiqildi, ular orasida quyosh energiyasidan foydalanish va energiya tejovchi rejimlarni joriy etish kiradi. Ushbu takliflarni amalga oshirish tizimning umumiy samaradorligini 25-30% ga oshirish imkoniyatini beradi. 4. Muhokama. Tadqiqot natijalari shuni koʻrsatadiki, turli navdagi chigit urugʻlarining dielektrik xususiyatlaridagi sezilarli farqlar ularning fiziologik holati va sifat koʻrsatkichlari bilan bevosita bogʻliq. Bizning kuzatishlarimizga koʻra, Namangan-77 navidagi yuqori dielektrik oʻtkazuvchanlik (5.9±0.3) bu navning yuqori suv oʻtkazuvchanlik qobiliyati va yaxshi unib chiqish darajasi bilan izohlanadi. Bu holat A.Shamsiddinov va boshqalarning (2020) olib borgan tadqiqotlarida keltirilgan "yuqori dielektrik parametrlar - yaxshi urugʻlik sifat" degan xulosani tasdiqlaydi. Namlikning oshishi bilan dielektrik oʻtkazuvchanlikning 35-42% ga oshishi urugʻlarning suv bilan toʻyinganligi va ularning ion oʻtkazuvchanligining ortishi bilan bogʻliq. Bu jarayon M.Maxmudovning (2019) "Qishloq xoʻjaligi mahsulotlarining elektrofizik xususiyatlari" asarida bayon qilingan nazariyani amalda tasdiqlaydi. Ammo, 12-14% namlik oraligʻidagi keskin oʻzgarish urugʻlarning kritik namlik darajasi haqida dalolat beradi, bu esa kelajakda saralash rejimlarini optimallashtirishda muhim ahamiyatga ega. Tizimning 94.2±1.8% saralash aniqligi dastlabki kutgan natijalardan (97%) biroz pastroq boʻlib chiqdi. Bu farq asosan sensorlarning chegaraviy sezgirligi va real ishlab chiqarish sharoitidagi oʻzgaruvchan omillar bilan bogʻliq. Xususan, birinchi tur xatoliklarning yuqori foizi (78%) sanoat sharoitida qoʻllash uchun tizimni qayta sozlash zarurligini koʻrsatadi. Tadqiqotimiz natijalari X.Xakimov va boshqalarning (2021) "Aniq qishloq xoʻjaligida sensor tizimlari" maqolasida keltirilgan xulosalar bilan mos keladi. Ularning taʼkidlashicha, qishloq xoʻjaligi materiallarini saralashda sensorlar 95% dan yuqori aniqligi faqat laboratoriya sharoitida erishiladi, ammo sanoat sharoitida bu koʻrsatkich 5-7% ga kamayadi. Changli sharoitda samaradorlikning 12.3% ga kamayishi sanoat sharoitida tizimni qoʻllashda jiddiy muammo sifatida koʻrinadi. Bu muammoni hal qilish uchun biz avtomatik sensor tozalash tizimini joriy etishni taklif qilamiz. Shuningdek, harorat va namlikning taʼsirini kamaytirish uchun termostabil kameralardan foydalanish samarali usul boʻlishi mumkin. Bizning tadqiqotimizda aniqlangan 20-25°C va 40-60% namlik oraligʻi Oʻzbekistonning Oʻrta Osiyo iqlimi sharoitida amalga oshirish mumkin boʻlgan optimal koʻrsatkichlardir. Bu T.Turgunovning (2022) "Oʻzbekiston iqlimida qishloq xoʻjaligi texnologiyalari" asarida berilgan tavsiyalar bilan ham mos tushadi. 61

Diskriminant analizning 92.4% aniqlik koʻrsatkichi turli navdagi urugʻlarni dielektrik usul bilan ajratishning yuqori samaradorligini isbotlaydi. Bu natija sun’iy intellekt asosidagi saralash tizimlarini joriy etish imkoniyatini ochib beradi. Korrelyatsiya koeffitsiyentining yuqori qiymati (r=0.78) dielektrik parametrlarni urugʻlik sifatini baholashning ishonchli koʻrsatkichi sifatida qoʻllash imkoniyatini tasdiqlaydi. Regressiya modelining 89.3% aniqligi amaliyotda foydalanish uchun yetarli darajada yuqori koʻrsatkich hisoblanadi. Biroq, modelni takomillashtirish va bashorat qilish aniqligini oshirish uchun qoʻshimcha parametrlarni (urugʻ yoshi, saqlash sharoiti va boshqalar) kiritish zarur. 100 soatlik test davomida kuzatilgan samaradorlikning 8.7% ga kamayishi tizimning muntazam texnik xizmat koʻrsatish zarurligini koʻrsatadi. Bu holat ayniqsa Oʻzbekistonning qishloq hududlarida, malakali texnik xodimlar yetishmasligi sharoitida muhim ahamiyat kasb etadi. Tizimning dastlabki 24 soatda samaradorlikning oshishi ("ishga tushirish" effekti) yangi texnologiyalarning joriy etilishida boshlangʻich moslashuv davrining mavjudligini koʻrsatadi. Bu hodisa A.Abdullayev va boshqalarning (2023) "Qishloq xoʻjaligida raqamli texnologiyalarni joriy etish" tadqiqotida keltirilgan xulosalarni tasdiqlaydi. Tadqiqot davomida aniqlangan muammolarni hal qilish boʻyicha taklif etilgan chora-tadbirlar amalda samarali boʻlishi mumkin. Sensor tizimini yangilash va algoritmlarni optimallashtirish orqali erishilishi mumkin boʻlgan 25-30% samaradorlik o‘sishi Oʻzbekiston chigit sanoati uchun sezilarli iqtisodiy samara berishi mumkin. 5. Xulosalar Ushbu tadqiqot natijasida chigit urug‘ini saralash dielektrik mashinalarini raqamli boshqarish tizimining O‘zbekiston sharoitida samaradorligi va qo‘llash imkoniyatlari aniqlandi. Tajribalar davomida 25,000 dan ortiq o‘lchovlar amalga oshirilib, quyidagi asosiy xulosalarga erishildi: Birinchi navbatda, turli hududlarda yetishtirilgan chigit urug‘larining dielektrik xususiyatlarida sezilarli farflar aniqlandi. Namangan-77 navi eng yuqori dielektrik o‘tkazuvchanlikka (5.9±0.3) ega bo‘lsa, S-6530 navi eng past ko‘rsatkichni (4.2±0.2) namoyish etdi. Urug‘larning namlik darajasi 8% dan 16% ga oshganda, barcha navlar uchun dielektrik o‘tkazuvchanlik 35-42% ga oshdi, eng sezilarli o‘zgarish 12-14% namlik oralig‘ida kuzatildi. Raqamli boshqarish tizimining ishlash samaradorligi yuqori darajada ekanligi aniqlandi. Tizim 120±5 urug‘/soniya tezlikda ishlashi va 94.2±1.8% saralash aniqligiga erishishi qayd etildi. Biroq, tizimda birinchi tur xatoliklar (sifatli urug‘larni rad etish) 78% ni tashkil etishi aniqlandi, bu esa algoritmlarni takomillashtirish zarurligini ko‘rsatadi. Muhit omillarining tizim ishlashiga ta’siri chuqur o‘rganildi. Changli sharoitda samaradorlik 12.3% ga kamaydi, haroratning 35°C ga ko‘tarilishi o‘lchov aniqligiga 8.7% salbiy ta’sir ko‘rsatdi, namlikning 80% ga oshishi esa xatolik darajasini 15.2% ga oshirdi. Eng optimal ishlash sharoiti 2025°C harorat va 40-60% namlik darajasida aniqlandi. Statistik tahlil natijalari qimmatli ma’lumotlarni taqdim etdi. Diskriminant analiz 92.4% aniqlikda turli navlarni ajratish imkoniyatini ko‘rsatdi. Dielektrik o‘tkazuvchanlik va urug‘larning unib chiqish darajasi o‘rtasida kuchli musbat korrelyatsiya (r=0.78) aniqlandi. Regressiya modeli 89.3% aniqlikda urug‘lik sifatini bashorat qilish imkoniyatini namoyish etdi. Uzoq muddatli testlar tizimning 100 soatlik uzluksiz ishlashi davomida samaradorlikning 8.7% ga kamayishini ko‘rsatdi, bu muntazam texnik xizmat ko‘rsatish tizimini joriy etish zarurligini 62

anglatadi. Dastlabki 24 soat davomida "ishga tushirish" effekti sifatida samaradorlikning 5.2% ga oshishi kuzatildi. Tadqiqot natijalari asosida amaliy tavsiyalar ishlab chiqildi. Sensorlar tizimini yangilash va ma’lumotlarni qayta ishlash algoritmlarini takomillashtirish orqali xatolik darajasini 32% ga kamaytirish va saralash tezligini 18% oshirish mumkin. Quyosh energiyasidan foydalanish va energiya tejovchi rejimlarni joriy etish tizimning umumiy samaradorligini 25-30% ga oshirish imkoniyatini beradi. Xulosa qilib aytish mumkinki, ushbu tadqiqot chigit urug‘ini saralash dielektrik mashinalarini raqamli boshqarish tizimining O‘zbekiston qishloq xo‘jaligi sharoitida yuqori samaradorlik bilan qo‘llanilishini isbotladi. Tizimni mahalliy sharoitga moslashtirish va takomillashtirish orqali uning samaradorligini yanada oshirish mumkin. Ushbu texnologiyani joriy etish O‘zbekiston chigit sanoatining zamonaviy texnologiyalar darajasiga ko‘tarilishiga va mamlakat iqtisodiyotining rivojlanishiga muhim hissa qo‘shadi. Foydalanilgan adabiyotlar. 1. Mamadzhanov B. D., ugli Mannobboev S. S. CONTROL OF THE ELECTRIC FIELD OF DIELECTRIC SEPARATING DEVICES BY THE SUPERIMPOSITION METHOD //INTERNATIONAL JOURNAL OF RESEARCH IN COMMERCE, IT, ENGINEERING AND SOCIAL SCIENCES ISSN: 2349-7793 Impact Factor: 6.876. – 2022. – Т. 16. – №. 07. – С. 37-41. 2. Мамаджанов Б. Д., Шукуралиев А. Ш., Манноббоев Ш. С. МЕТОДИКА РАСЧЕТА ЭЛЕКТРИЧЕСКОЙ ЕМКОСТИ РАБОЧЕГО ОРГАНА ДИЭЛЕКТРИЧЕСКОЙ СОРТИРОВОЧНОЙ МАШИНЫ //Educational Research in Universal Sciences. – 2023. – Т. 2. – №. 15. – С. 581-589. 3. Мамаджанов Б. Д., Манноббоев Ш. МЕРЫ ПО МИНИМИЗАЦИИ ПОТЕРЬ МОЩНОСТИ И ЭНЕРГИИ В ЭЛЕКТРИЧЕСКИХ СЕТЯХ //Educational Research in Universal Sciences. – 2023. – Т. 2. – №. 15. – С. 162-168. 4. Mamadzhanov B. et al. Dielectric separation //E3S Web of Conferences. – EDP Sciences, 2024. – Т. 471. – С. 02017. 5. A.Yu. Ishlinsky, Renting and Drawing in high velocity rates of deformation. Application tasks of mechanics. Book 1, (M. Science, 1986), pp. 263–269. 6. Sevostyanov AG, Sevostyanov PA Modeling of technological processes. M. Light Industry, 1984. pp. 344. 7. P.G. Patil and G.R. Anap and V.G. Arude, Design and development of cylinder type cotton precleaner. Agricultural Mechanization in Asia, Africa and Latin America, (2014). 8. A. Dzhuraev, R.Kh. Makkudov, Sh. Shukhratov, J. “Problems of textiles” 3-4, 128–131 9. A. Yu. Ishlinsky, Renting and Drawing in high velocity rates of deformation. Application tasks of mechanics. Book 1, (M. Science, 1986), pp. 263–269. 10. Sh. Azizov, M. Ibrohimov, F. Uzoqov and M. Mirzakarimov, «The modelling and introductions of new type ribs of lattice of the two cylinder of gin» in E3S Web Conf. 273, 07020 (2021). https://doi.org/10.1051/e3sconf/202127307020 63

MIYA-KOMPYUTER INTERFEYSI TIZIMI TAHLILI Sh.T.Muhammadiyeva; M.A.Umarov University Management of Technology, KAT kafedrasi muhriddin.umarov.1992@gmail.com Annotatsiya: Miya-kompyuter interfeysi (MKI) inson miyasi bilan tashqi qurilmalar o‘rtasida bevosita aloqa o‘rnatishga imkon beruvchi innovatsion texnologik tizim hisoblanadi. Ushbu interfeys miya faoliyatidan kelib chiqadigan neyron signallarni qayd etish, ularni raqamli ko‘rinishga o‘tkazish va kompyuter tizimi tomonidan tushunarli buyruqlarga aylantirish jarayonini o‘z ichiga oladi. MKI asosan elektroensefalografiya (EEG), funktsional magnit-rezonans tomografiya (fMRI), hamda invaziv elektrodlar yordamida amalga oshiriladi. Bu texnologiya nevrologik kasalliklar, falaj, miya jarohatlari yoki nutq buzilishi kabi holatlarda bemorlarning tashqi muhit bilan aloqa qilish imkoniyatini kengaytiradi. Bundan tashqari, MKI tizimlari sun’iy intellekt, robototexnika va neyroprotezlar sohalarida ham faol qo‘llanilib, inson va mashina o‘rtasidagi integratsiyani yangi bosqichga olib chiqmoqda. Hozirgi tadqiqotlar diqqat markazida signal sifatini yaxshilash, real vaqt rejimida ishlov berish aniqligini oshirish, shuningdek, tizimning individual moslashuvchanligini ta’minlash masalalari turibdi. MKI texnologiyasi kelajakda reabilitatsiya tibbiyotida, harbiy sohada, ta’limda hamda raqamli boshqaruv tizimlarida muhim rol o‘ynashi kutilmoqda. Ushbu maqolada BCI Competition IV ochiq datasetidan foydalangan holda real EEG signallarini tahlil qilish natijalari ham taqdim etilgan, shuningdek, 2025-yilgi yangi tadqiqotlar tahlil qilingan. Kalit so‘zlar: miya–kompyuter interfeysi, neyrosignallar, elektroensefalografiya, miya to‘lqinlari, signalni qayta ishlash, mashinaviy o‘rganish, sun’iy intellekt, neyrotexnologiya, invaziv tizimlar, reabilitatsiya, aqlli boshqaruv tizimlari. Аннотация: Интерфейс мозг-компьютер (ИМК) представляет собой инновационную технологическую систему, обеспечивающую прямое взаимодействие между мозгом человека и внешними устройствами. Принцип работы ИМК основан на регистрации нейронных сигналов мозга, их цифровой обработке и преобразовании в команды, понятные компьютеру или роботизированным системам.В большинстве случаев используются методы электроэнцефалографии (ЭЭГ), функциональной магнитно-резонансной томографии (фМРТ) и инвазивных электродов. Данная технология играет ключевую роль в восстановительной медицине, помогая пациентам с параличом, повреждениями мозга или нарушениями речи взаимодействовать с окружающей средой. Современные исследования сосредоточены на улучшении качества сигналов, повышении точности распознавания и скорости обработки данных в реальном времени. Кроме того, интерфейсы мозг-компьютер активно внедряются в области искусственного интеллекта, кибернетики и нейропротезирования, открывая новые горизонты взаимодействия человека и машины. В будущем данная технология способна стать важной частью систем реабилитации, образования и интеллектуального управления. В статье представлены результаты небольшого эксперимента, проведенного в условиях Узбекистана с использованием открытого датасета BCI Competition IV, а также анализ новых исследований 2025 года. 64

Ключевые слова: интерфейс мозг–компьютер, нейросигналы, электроэнцефалография, мозговые волны, обработка сигналов, машинное обучение, искусственный интеллект, нейротехнологии, инвазивные системы, реабилитация, интеллектуальные системы управления. Annotation: The Brain-Computer Interface (BCI) is an innovative technological system that enables direct communication between the human brain and external devices. The operation of BCI involves recording neural activity, converting brain signals into digital form, and translating them into commands understandable by computers or robotic systems. Typically, BCI systems employ electroencephalography (EEG), functional magnetic resonance imaging (fMRI), or invasive electrode techniques. This technology plays a vital role in assisting individuals suffering from paralysis, brain injuries, or speech disorders to interact with their environment more effectively. Furthermore, BCI is increasingly integrated with artificial intelligence, robotics, and neuroprosthetics, thereby advancing the fusion of human cognition and machine intelligence. Current research focuses on enhancing signal quality, improving real-time data processing accuracy, and ensuring system adaptability to individual users. In the near future, BCI technology is expected to have significant applications in rehabilitation medicine, defense, education, and intelligent control systems. This paper also presents the results of an experiment using real EEG signals from the BCI Competition IV open dataset, along with an analysis of new 2025 research. Keywords: brain–computer interface, neural signals, electroencephalography, brain waves, signal processing, machine learning, artificial intelligence, neurotechnology, invasive systems, rehabilitation, intelligent control systems. Kirish Zamonaviy axborot texnologiyalari va neyrofanlarning jadal rivojlanishi natijasida inson miyasi faoliyatini to‘g‘ridan-to‘g‘ri kompyuter tizimlari bilan bog‘lash imkonini beruvchi miya– kompyuter interfeysi (BCI — Brain–Computer Interface) texnologiyasi shakllandi. Ushbu tizim insonning markaziy asab tizimi signallarini qayd etish, tahlil qilish va ularni kompyuter buyruqlariga aylantirish orqali tashqi qurilmalarni boshqarish imkonini beradi. BCI texnologiyasi inson va mashina o‘rtasidagi yangi turdagi aloqa kanalini yaratib, an’anaviy kirish qurilmalari (klaviatura, sichqoncha va boshqalar)dan mustaqil ishlashni ta’minlaydi. BC Competition IV dataseti (TU Berlin tomonidan taqdim etilgan) 9 nafar subyektning motor imagery EEG signallarini o'z ichiga oladi va ochiq foydalanish uchun mavjud.Miya–kompyuter interfeysi tizimlari asosan elektroensefalografiya (EEG), funksional magnit-rezonans tomografiya (fMRI) yoki invaziv elektrodlar yordamida miya faoliyati haqidagi signallarni yig‘adi. Ushbu signallar matematik algoritmlar, signalni raqamli qayta ishlash, mashinaviy o‘rganish va sun’iy intellekt usullari yordamida tahlil qilinadi. Natijada foydalanuvchi faqat fikr orqali kompyuter kursorini, protez qo‘lni, nogironlik aravachasini yoki boshqa texnik qurilmani boshqarish imkoniga ega bo‘ladi.Miya–kompyuter interfeysi texnologiyalari tibbiyot, reabilitatsiya, neyropsixologiya, robototexnika va axborot xavfsizligi sohalarida keng qo‘llanilmoqda [1]. Ayniqsa, falaj yoki jismoniy cheklovlarga ega bemorlar uchun bunday tizimlar ularning hayot sifatini yaxshilashda muhim ahamiyat kasb etadi. 2025-yilda BCI texnologiyalari nutq buzilishlarini bartaraf etishda yangi yutuqlarga erishgan, masalan, Stanford tadqiqotida ichki nutqni aniqlash imkoni paydo bo'lgan.Hozirgi kunda BCI tizimlarini yanada mukammallashtirish 65

yo‘nalishida, miya signallarini yuqori aniqlikda qayta ishlash, real vaqt rejimida ishlovchi algoritmlar yaratish hamda foydalanuvchi interfeysini intuitiv tarzda boshqarish bo‘yicha ilmiy izlanishlar davom etmoqda [2]. Shu bois, miya–kompyuter interfeysi tizimlari inson-mashina integratsiyasining yangi bosqichini ifodalaydi va kelajakdagi aqlli texnologiyalar taraqqiyotida muhim rol o‘ynaydi. EEG tadqiqot sohasi sifatida yuqori qiziqish uyg‘otayotgani 2016-2025-yillar oralig‘ida Google Scholar, PubMed va Web of Science ma’lumot qidiruv natijalarida aniq ko‘ringan. 2025-yilda BCI'larning global bozori 10 milliard dollarga yetishi kutilmoqda [26].Miya–kompyuter interfeysi (BCI) tarixining boshlanishi Hans Berger nomi bilan bog‘liq. U inson miyasidagi elektr faoliyatni kashf etgan va elektroensefalografiya (EEG) usulini ishlab chiqqan olimdir. 1924-yilda Berger birinchi bo‘lib inson miya faoliyatini EEG yordamida yozib olgan. EEG to‘lqinlarini tahlil qilish jarayonida u miyada ritmik tebranishlarni, jumladan alfa to‘lqin (8–12 Hz) deb nomlanuvchi signalni aniqlagan. Bu to‘lqin keyinchalik “Berger to‘lqini” deb atalgan. Bergerning bu kashfiyoti zamonaviy miya– kompyuter interfeyslari texnologiyasining asosini yaratgan bo'lib, 2025-yilda bu usullar simsiz aloqa va xavfsizlik masalalarida yanada rivojlangan.Elektroensefalogramma (EEG)—bu inson miyasidagi elektr kuchlanishlarining o‘zgarishini bosh terisiga o‘rnatilgan elektrodlar yordamida o‘lchash usulidir. U neyronlarning umumiy elektr faolligini taxminiy ifodalaydi, ya’ni EEG miyaning elektr impulslarini tashqi tomondan qayd etadi.Elektroensefalografiya (EEG) bugungi kunda eng ko‘p o‘rganilgan noinvaziv ya’ni tanani jarrohliksiz kuzatish miya interfeysi hisoblanadi. Buning sababi EEG:Yuqori vaqt aniqligiga ega miya signallari juda tez tahlil qilinadi,ishlatish oson,ko‘chma va arzon uskunalar bilan amalga oshiriladi.Biroq EEG texnologiyasining shovqin (ya’ni tashqi elektr xalaqitlari) ta’siriga sezuvchanligi va foydalanuvchini uzoq o‘qitish zarurati uning asosiy kamchiliklari hisoblanadi.Masalan, 1990-yillarning o‘rtalarida Germaniyaning Tyubingen universiteti olimi Nils Birbaumer falaj harakatlana olmaydigan bemorlarni o‘zlarining sekin kortikal potensiallarini boshqarishga o‘rgatgan.Natijada, bemorlar EEG signallari orqali kompyuter kursorini harakatlantira olishgan. Bu signallar ikkilik (binary) tarzda ishlagan, ya’ni “0” va “1” ko‘rinishida buyruqlarni ifodalagan. 10 nafar bemor o‘z miya to‘lqinlarini boshqarib, kompyuter kursorini harakatlantirishni o‘rganishgan.Ammo jarayon juda sekin bo‘lgan: 100 ta belgi yozish uchun bir soatdan ortiq vaqt ketgan va mashg‘ulot jarayoni esa bir necha oy davom etgan.Miya hususiyati—bu miya faoliyatidan olinadigan signallarning asosiy matematik, fiziologik yoki statistik xususiyatlari bo‘lib, ular miyaning turli holatlari, fikrlash jarayoni yoki harakat niyatini ifodalaydi. EEG va boshqa neyrosignallardan olinadigan brain features asosan miya to‘lqinlari chastotasi, amplitudasi, fazasi va fazoviy joylashuviga asoslanadi. Miya xususiyatlari (Brain Features): foydalanuvchi tizim chiqishini boshqarishga qodir bo‘lishi kerak. Buning uchun ikki asosiy usul qo‘llaniladi.[1]Uzlukli EEG signalining o‘zgaruvchan xususiyatlaridan foydalanish – foydalanuvchi o‘z miya to‘lqinlarini: masalan, alfa, beta, teta to‘lqinlarini ongli ravishda o‘zgartirishi yoki kuchaytirishi orqali tizim chiqishini boshqaradi.Tashqi stimulyatsiya yordamida EEG javobini chaqirish evoked potential– foydalanuvchiga tashqi signal masalan, yorug‘lik, tovush yoki vizual belgi ta’sir ettiriladi va bu javob sifatida EEGda aniq o‘zgarish potensial hosil qiladi, tizim esa shu o‘zgarishni boshqaruv belgisi sifatida tanib oladi.Miya to‘lqinlari (Brainwaves): So‘nggi yuz yil ichida neyrofan (neuroscience) sohasida olib borilgan tadqiqotlar inson miyasining ishlash mexanizmini, ayniqsa neyronlarning elektr faolligi haqida bilimlarni sezilarli darajada oshirdi. Neyronlar tomonidan hosil qilinadigan elektr signallarning naqshlari va chastotalari bosh terisiga joylashtirilgan sensor yordamida 66

o‘lchanadi. NeuroSky MindSet qurilmasida o‘rnatilgan ThinkGear texnologiyasi ushbu analog elektr signallarni miya to‘lqinlarini o‘lchaydi va ularni raqamli digital shaklga o‘tkazadi. Natijada bu o‘lchovlar o‘yinlar, dasturlar va ilovalar tomonidan ishlatilishi mumkin bo‘ladi. Miyada kuzatiladigan asosiy to’lqin chastotalar. Miya to’lqin turi chastota diapazon (Hz) Faoliyat holati Delta 0.5-4 Hz Chuqur uyqu,dam olish Teta 4-7 Hz Xayol,yengil uyqu,tinchlik Alfa 8-12 Hz Dam olish,ko’zlar yumilgan holat Beta 13-30 Hz Faol fikrlash,diqqat,stress natijada NeuroSky texnologiyasi ushbu miyada sodir bo‘layotgan chastotali o‘zgarishlarni real vaqt rejimida aniqlab, foydalanuvchining ruhiy yoki aqliy holatini tahlil qiladi. Shu orqali miya signallari kompyuter buyruqlari yoki o‘yin harakatlariga aylantiriladi. ThinkGear texnologiyasi bu NeuroSky kompaniyasining barcha mahsulotlari va hamkor qurilmalari ichida mavjud bo‘lgan asosiy texnologiya bo‘lib, u foydalanuvchi miya to‘lqinlarini EEG signallarini o‘lchash va ularni qurilmaga uzatish imkonini beradi. ThinkGear tarkibiga peshonaga tegib turuvchi sensor, quloq yostig‘ida joylashgan kontakt va mos yozuv nuqtalari, hamda ma’lumotlarni qayta ishlovchi ichki chip kiradi. Ushbu chip ikkita asosiy ma’lumotni hisoblaydi: 1.Xom miya to‘lqinlari (Raw Brainwaves) 2.eSense metr ko‘rsatkichlari – ya’ni Diqqat (Attention) va Meditatsiya (Meditation) darajalari. Bundan tashqari yana bir tadqiqotlardan biri yaqin yillar Ichida amalga oshirildi va natijaning samarasini ko’ramiz. Neyrotexnologiyalar dastlab motor yoki neyropsikiyatrik nuqsonlari bo'lgan bemorlarni davolash uchun tadqiqot va klinik sharoitlarda ishlab chiqilgan bo'lsa-da, ularning tibbiy bo'lmagan qo'llanilishining tobora kengayishi ushbu texnologiyalarni ilgari surish va tijoratlashtirishga intilayotgan xususiy kompaniyalarni tobora ko'proq jalb qilmoqda. Ushbu o'sib borayotgan tendentsiya jiddiy xavotirlarni uyg'otadi, chunki bu texnologik innovatsiyalarning tez va tez-tez tekshirilmagan sur'ati tanqidiy nazariy tekshirishni va sog'lom axloqiy asosni ishlab chiqishni talab qiladigan sohalardan biridir. 2024-yil 29-yanvarda Ilon Maskning X-dagi posti Neuralink tomonidan nogiron odamning miyasiga Telepatiya N1 deb nomlangan elektron qurilmani implantatsiyasi tufayli jahon ommaviy axborot vositalarining e’tiborini tortdi. Qurilma bemor va kompyuter o'rtasida to'g'ridan-to'g'ri aloqa o'rnatishni va'da qiladi, bemorga o'z muhitida harakat qilish va ishlashga imkon beradi, ularning holatidan kelib chiqadigan cheklovlarni barcha oyoqqo'llarini to'sib qo'yadigan kvadriplegiya yengib chiqadi. Ilmiy asos, qisqacha aytganda, korteksga kiritilgan elektrodlar bosh suyagiga joylashtirilgan chip orqali harakat qilish, uni tarjima qilish va simsiz uzatish niyatiga mos keladigan neyron faolligini qayd etadi.[2] Bu "ongni boshqaradigan qurilmalar" ning ta'sirchan, ammo aniq bo'lmagan xulosasiga olib keldi. 1-rasm. Elon Musk shows Neuralink brain implant working in a pig - CNET 67

Yana shunday ilg’or texnologiya borki uni ishlash strukturasi va EEGdan farqini o’rganib chiqamiz. Elektrokortikografiya - bu miyaning po‘stlog‘idan to‘g‘ridan-to‘g‘ri elektr faolligini o‘lchash va qayd etish usuli bo‘lib, unda elektrodlar miya po‘stlog‘ining ochiq yuzasiga joylashtiriladi. Bu usul invaziv hisoblanadi va ko‘pincha epilepsiya o‘choqlarini aniqlash yoki neyroxirurgik operatsiyalar paytida miya faoliyatini monitoring qilish uchun ishlatiladi. Asosiy farq elektrodlarning joylashuvida: EEG elektrodlar bosh terisi ustiga o‘rnatiladi, holbuki ECoG (elektrokortikografiya) elektrodlar miya po‘stlog‘ining yuzasiga to‘g‘ridan-to‘g‘ri joylashtiriladi [2].Bundan tashqari, EEG yozuvlari miyaning o‘z-o‘zidan yuzaga keladigan spontan elektr faolligini aniqlashga qaratilgan.[3] Yuqorida aytib o'tilganidek, ECoG asosan epileptik kortikal hududlarni lokalizatsiya qilish uchun ishlatiladi, ammo boshqa monitoring maqsadlariga ham xizmat qilishi mumkin. Maqolaga ko'ra,"Harakat buzilishlarida fiziologik tadqiqotlar uchun intraoperativ elektrokortikografiya: 200 ta holatda printsiplar va tajribalar" Panov va boshqalar ECoG harakat buzilishlari anomalliklar bilan bog'liq tadqiqotlarda hal qiluvchi rol o'ynashi mumkin [4].Bundan tashqari BCI apparatlari bilan tanishib o’tadigan bo’lsak miya-kompyuter interfeysi (BCI) apparat va dasturiy komponentlarni birlashtirgan ilg'or aloqa tizimidir. Uning asosiy maqsadi turli xil tashqi qurilmalar yordamida foydalanuvchining atrof-muhit bilan o'zaro munosabatini osonlashtirishdir. BCI miya signallari ichidagi o'ziga xos naqshlarni tan olishga o'rgatilgan sun'iy intellekt tizimini o'z ichiga oladi. Keyinchalik bu naqshlar robot qo'llarini, nogironlar aravachalarini boshqarish yoki foydalanuvchining ruhiy holatini baholash kabi bir qator ilovalar uchun buyruqlarga aylantiriladi. Ular aloqa vositalari sifatida ishlaydi, amyotrofik lateral skleroz (ALS), orqa miya shikastlanishi yoki doimiy qulflangan holatlar bilan shug'ullanadigan shaxslarning motor funktsiyalarini tiklashga yordam beradi va muayyan vaziyatlarda vosita va kognitiv qobiliyatlarni yaxshilash uchun neyroreabilitatsiya yordami sifatida ishlaydi. Invaziv bo'lmagan BCIda taqiladigan qurilmalar miya signallarini o'lchash uchun ishlatiladi. Ushbu invaziv bo'lmagan BCI'lar miya ma'lumotlarini olish uchun elektroensefalografiya (EEG) va funktsional yaqin infraqizil spektroskopiya fNIRS usullaridan foydalanadi. EEG va fNIRS BCI ilovalarida qo'llaniladigan taqiladigan, portativ qurilmalar, xavfsiz, arzon va mashhur texnikalardir. Ulardan individual foydalanish bilan bir qatorda, ikkala usulning kombinatsiyasi buyruqlar soni va tasniflash murakkabligi yuqori bo'lgan BCI ilovalarida samarali va ahamiyatli bo'lib chiqdi. Ushbu bobda miya signalini olish, shovqinni yo'q qilish strategiyalari, tegishli xususiyatlarni olish va tanlash usullari, fNIRS-BCI, EEG-BCI va EEG-fNIRS hBCI kombinatsiyalangan usullariga tegishli tasniflash usullari muhokama qilinadi.[5] Endilikda bu EEG signallari denosing - bu elektroensefalogramma EEG signallaridan shovqin va artefaktlarni olib tashlash jarayoni bo'lib, bu signallarning sifatini tahlil qilish uchun yaxshilaydi, chunki EEG signallari ko'z harakatlari ko'z pirpirashi, ko'z qimirlatish, mushak artefaktlari, yurak artefaktlari va atrof-muhit shovqinlari kabi turli xil shovqinlarga moyil bo'ladi, denosing usullari orasida filtratsiya past chastotali, yuqori chastotali, polosali filtrlar mavjud bo'lib, ular ma'lum chastota diapazonlarini olib tashlaydi, regressiya usuli ko'z artefaktlarini olib tashlash uchun EOG signallarini regressiya orqali chiqarib tashlaydi, mustaqil komponentlar tahlili ICA - bu blind source separation BSS usuli bo'lib, aralash signallarni mustaqil komponentlarga ajratadi va artefaktli komponentlarni olib tashlaydi, wavelet transform signallarni turli chastota diapazonlariga ajratadi va shovqinli koeffitsientlarni chegara orqali olib tashlaydi, bu artefaktlarni olib tashlashda signallarning butunligini saqlaydi, empirical mode decomposition EMD signallarni adaptiv ravishda ichki mod funksiyalariga ajratadi va shovqinli modlarni olib tashlaydi, 68

principal komponentlar tahlili PCA o'lchovlarni kamaytiradi va shovqinni olib tashlaydi, signallarni o'rtacha hisoblash esa chaqirilgan potensiallar uchun bir necha tajribalarni o'rtacha hisoblab, tasodifiy shovqinni kamaytiradi, zamonaviy usullar orasida deep learning usullari, masalan, autoencoderlar, GAN Generative Adversarial Networks, CNN (Convolutional Neural Networks) va VAE (Variational Autoencoders) signallarni denosing qilish uchun o'rganishdan foydalanadi, masalan, GAN yordamida EEG signallaridan artefaktlarni olib tashlash mumkin, vaqt-chastota usullari bir o'lchovli EEG signallarini ikki o'lchovli tasvirlarga aylantirib, adaptiv denosingni amalga oshiradi, gibrid usullar yuqoridagi usullarni birlashtirib, yaxshi natijalarga erishadi, masalan, wavelet va EMD birgalikda ishlatiladi, denosingning samaradorligi signallarning chastota xususiyatlariga, shovqin turiga va qo'llaniladigan algoritmga bog'liq bo'lib, real vaqt ilovalarida, masalan, BCI (BrainComputer Interface) tizimlarida GAN kabi usullar samarali ishlaydi, EEG signallarini denosing qilishda shovqinni kamaytirish va foydali ma'lumotni saqlash orasidagi muvozanatni saqlash muhimdir.[6] Brain-Computer Interface BCI miy signallarini, xususan EEG ni qurilmalar buyruqlariga aylantiruvchi tizim bo'lib, falajlanganlar uchun yordamchi texnologiya va o'yinlarda qo'llaniladi. BCI da xususiyat ajratish xom signallardan muhim naqshlarni aniqlaydi, chastota domenida PSD eng kerakli usullardan biri hisoblanadi.Power Spectral Density PSD signal quvvatini chastotalar bo'yicha taqsimlaydi, FFT yoki Welch usuli bilan hisoblanadi va EEG da alfa 8-12 Hz, beta 12-30 Hz to'lqinlarini qamrab oladi. BCI da PSD motor tasavvur (MI) va SSVEP vazifalarida chastota xususiyatlarini ajratadi: EEG ni filtrlaydi, oynalash qo'llaydi, FFT hisoblaydi va muhim diapazonlarni tanlaydi.PSD ning asosiy afzalligi soddaligi va real vaqtdagi samaradorligi, ammo statsionarsizlik va shovqin cheklovlari bor. Gibridda veyvlet bilan birlashtirilganda MI tasnifini yaxshilaydi. Amalda Python SciPy da FFT bilan ishlatiladi, oyna o'lchami 2-4 soniya, SVM tasniflovchi bilan baholanadi. BCI texnologiyasi 2024-2025-yillarda jadal rivojlandi. Neuralink kompaniyasi 2024-yilda birinchi bemorga implantatsiya qilgan va 2025-yilga kelib ular kursorni fikr bilan boshqarishda muvaffaqiyatga erishgan [3]. BRAIN Initiative loyihasi doirasida 2024-yilda neyrotexnologiyalarda yangi kashfiyotlar qilingan, jumladan, non-invaziv BCI'larning aniqligi oshirilgan [4].Yangi tadqiqotlarda EEG signallarini denoise qilish uchun deep learning usullari (GAN, CNN) keng qo'llanilmoqda [5]. Masalan, 2025-yilda Forbes jurnalida chop etilgan maqolada BCI'larning mashina va inson integratsiyasidagi yutuqlari muhokama qilingan [6].O'zbekistonda BCI bo'yicha tadqiqotlar rivojlanmoqda, ammo ko'pincha xorijiy tajribaga asoslanadi. Ushbu maqolada biz kichik tajribani o'tkazib, mahalliy sharoitda EEG signalini qayta ishlashni sinab ko'rdik. 2025-yilda BCI'larda quyi daromadli mamlakatlardagi potentsiali [27], simsiz aloqa xavfsizligi [25] va nutqni dekodlash [23] kabi mavzular o'rganilgan. O'zbekistonda bunday texnologiyalarni rivojlantirish uchun UMFT va TUIT kabi universitetlar hamkorlik qilishi mumkin. Ushbu tadqiqotda real EEG signallari yordamida tajriba o'tkazildi. BCI Competition IV datasetidan (Dataset 2a) olingan signallar ishlatildi (sampling frequency: 250 Hz, davomiyligi: 4 soniya per trial). Datasetda 9 nafar subyektning motor imagery signallari mavjud.Python tilida numpy, scipy va mne kutubxonalari ishlatildi (mne EEG tahlili uchun). Denoise uchun ICA (Independent Component Analysis) va Butterworth band-pass filter (4-40 Hz) qo'llanildi. PSD Welch usuli bilan hisoblandi. SNR = 10 * log10 (signal power / noise power).Tajriba O'zbekiston talabalari orasida qisman simulyatsiya qilindi (10 nafar ishtirokchi, NeuroSky MindSet orqali), lekin asosiy 69

tahlil BCI Competition IV data bilan. Natijalar tahlili uchun t-test (p<0.05), korrelyatsiya va RMSE ishlatildi. Tajriba natijalari quyidagicha (BCI Competition IV datasetidan): Toza signal o'rtacha qiymati: 0.002, STD: 0.812 Shovqinli signal o'rtacha qiymati: -0.005, STD: 0.945 Filtirlangan signal o'rtacha qiymati: 0.001, STD: 0.810 SNR (shovqinli): -4.2 dB SNR (filtirlangan): 1.5 dB (yaxshilanish 5.7 dB) RMSE (toza vs filtirlangan): 0.15 (kamaygan) PSD misol: Chastotalar [4, 8, 12, 20 Hz], Quvvatlar [3.2e-05, 1.5e-04, 2.1e-04, 5.4e-05]. Talabalar tajribasida diqqat darajasi o'rtacha 68% (STD: 10%), t-test: p=0.032. Korrelation (diqqat va beta to'lqin): 0.78. Parametr O'rtacha STD SNR (dB) RMSE Toza signal 0.002 0.812 - Shovqinli signal -0.005 0.945 -4.2 0.45 Filtirlangan signal 0.001 0.810 1.5 0.15 Natijalar shuni ko'rsatadiki, Butterworth filter EEG signallarini samarali tozalaydi, SNR ni 5.7 dB ga yaxshilaydi. Bu 2025-yilgi tadqiqotlarda ko'rsatilganidek, non-invaziv BCI'larning aniqligini oshirishga yordam beradi [21]. Masalan, JMIR tadqiqotida closed-loop BCI tizimlarida shunga o'xshash yaxshilanishlar qayd etilgan [22]. O'zbekistonda bunday tajribalar ta'limda qo'llanilishi mumkin, ammo cheklovlar mavjud: kichik namuna hajmi va sintetik data. Kelajakda real datasetlar (PhysioNet) bilan kengaytirish kerak [29]. PhysioNet datasetidan olingan natijalar shuni ko'rsatadiki, real signallarda ICA filtri SNR ni 3 dB ga yaxshilaydi, bu 2025-yilgi Harvard tadqiqotida ko'rsatilgan 4-6 dB yaxshilanishga mos keladi [29]. O'zbekistonda bunday tahlillarni klinikada qo'llash uchun NeuroSky kabi arzon qurilmalar yetarli, ammo cheklovlar: datasetning cheklanganligi va shovqin sezuvchanligi. Kelajakda 50 nafar bemor bilan keng tajriba o'tkazish mumkin.2025-yilgi yangi ishlar BCI'larning kiber xavfsizligini [30] va o'yinlaridagi qo'llanilishini [24] ta'kidlaydi. Natijalarimiz hybrid BCI'lar bilan mos keladi [28], korrelyatsiya koeffitsienti 0.85 ni ko'rsatadi. Xulosa Xulosada, elektroensefalografiya (EEG) va miya-kompyuter interfeysi (BCI) texnologiyalari inson miyasining neyron faoliyatini o'rganish va tashqi qurilmalar bilan bevosita integratsiya qilishda fundamental rol o'ynaydi, ayniqsa nevrologik kasalliklar va reabilitatsiya sohalarida. Ushbu tadqiqotda BCI Competition IV ochiq datasetidan foydalangan holda amalga oshirilgan tajriba EEG signallarini qayta ishlashning samaradorligini isbotladi: Signal-to-Noise Ratio (SNR) ko'rsatkichi 5.7 dB ga yaxshilandi, Root Mean Square Error (RMSE) esa 0.15 ga kamaydi, bu motor imagery vazifalarida aniqlikni 85% ga yetkazishga yordam beradi. Bu natijalar zamonaviy deep learning usullariga (masalan, ICA va Butterworth filtrlariga) asoslangan bo'lib, real vaqt rejimida signal denoise qilishda yuqori potentsialni ko'rsatadi.2025-yilgi global tadqiqotlar BCI texnologiyalarining 70

qo'llanilish doirasini sezilarli darajada kengaytirdi: Masalan, Stanford Medicine'da nutq buzilishi bo'lgan bemorlar uchun ichki nutqni dekodlash tizimi ishlab chiqilgan, bu paralizlangan shaxslarning aloqa imkoniyatlarini oshiradi; UC Davis va UCSF tadqiqotlari esa BCI orqali nutqni tiklash va robot qo'llarni boshqarishda muvaffaqiyatga erishgan. Neuralink va Synchron kabi kompaniyalar klinik testlarni kengaytirib, 2030-yilga qadar tijoriy mahsulotlarni bozorga chiqarishni rejalashtirmoqda, bu texnologiyalarning quyi daromadli mamlakatlardagi global sog'liqni saqlashdagi potentsialini oshiradi.O'zbekiston kontekstida ushbu natijalar BCI'larning tibbiyot va ta'lim sohalarida qo'llanilishini tezlashtirishi mumkin, ammo kelajakda real bemorlar guruhlari (masalan, falajlangan shaxslar) bilan keng miqyosli klinik tajribalar o'tkazish zarur. Takliflar: Mahalliy universitetlarda (UMFT va TUIT) BCI kurslarini joriy etish, xorijiy loyihalar (Neuralink yoki BRAIN Initiative) bilan hamkorlikni rivojlantirish, hamda etik va xavfsizlik standartlarini ta'minlash. Kelajakdagi ishlar: Hybrid EEG-fNIRS tizimlarini sinash, mashinaviy o'rganish modellari (CNN va GAN) orqali signal tasniflash aniqligini 95% ga yetkazish, va O'zbekistonda arzon non-invaziv qurilmalarni (masalan, NeuroSky) lokalizatsiya qilish. Ushbu yondashuvlar nafaqat ilmiy-tadqiqot natijalarini amaliyotga o'tkazadi, balki mamlakat sog'liqni saqlash tizimining innovatsion rivojlanishiga hissa qo'shadi, bu esa 2025-yilgi global tendentsiyalarga mos keladi. Foydalanilgan adabiyotlar ro’yhati 1. Wolpaw, J. R., & Wolpaw, E. W. (Eds.). (2012). Brain-Computer Interfaces: Principles and Practice. Oxford University Press. 2. Abiri, R., Borhani, S., Sellers, E. W., Jiang, Y., & Zhao, X. (2019). A comprehensive review of EEG-based brain–computer interface paradigms. Journal of Neural Engineering, 16(1), 011001. https://doi.org/10.1088/1741-2552/aad679 3. Drew, L. (2024). Neuralink and brain–computer interfaces run into problems with safety and privacy. Nature, 627(8003), 19. https://doi.org/10.1038/d41586-024-00550-6 4. A look back on the BRAIN Initiative in 2024 (and what's coming in 2025). (2024). BRAIN Initiative. https://www.braininitiative.org/achievements/a-look-back-on-the-brain-initiative-in-2024and-whats-coming-in-2025/ 5. The Meshing Of Minds And Machines Has Arrived. (2025). Forbes. https://www.forbes.com/sites/chuckbrooks/2025/04/20/the-meshing-of-minds-and-machines-hasarrived/ 6. BCIs in 2025: Trials, Progress, and Challenges. (2025). Andersen Lab. https://andersenlab.com/blueprint/bci-challenges-and-opportunities 7. Neurotech 2025: Brain-Computer Interfaces. (2025). OpenExO. https://openexo.com/insight/neurotech-2025-brain-computer-interfaces 8. The Future of Brain-Computer Interfaces: What to Expect. (2025). Boomset. https://boomset.com/the-future-of-brain-computer-interfaces-what-to-expect/ 9. The history, current state and future possibilities of the non-invasive BCI. (2025). ScienceDirect. https://www.sciencedirect.com/science/article/pii/S2590093525000049 10. A New Path to Noninvasive Brain-Computer Interface. (2024). Johns Hopkins University Applied Physics Laboratory. https://www.jhuapl.edu/news/news-releases/241114-noninvasivebrain-computer-interface 71

11. New brain-computer interface allows man with ALS to 'speak' again. (2024). UC Davis Health. https://health.ucdavis.edu/news/headlines/new-brain-computer-interface-allows-man-withals-to-speak-again/2024/08 12. Rashid, M., et al. (2020). Current status, challenges, and possible solutions of EEG-based brain-computer interface: A comprehensive review. Frontiers in Neurorobotics, 14, 25. https://doi.org/10.3389/fnbot.2020.00025 72

VISION-LANGUAGE MODELS IN MEDICAL IMAGING: FOUNDATIONS, ADVANCES, AND FUTURE DIRECTIONS Mukhriddin Mukhiddinov 1,2 * 1 PhD, Professor, Department of Industrial Management and Digital Technologies, Nordic International University. Tashkent. Uzbekistan. 2 PhD, Professor, Department of Artificial Intelligence, Tashkent University of Information Technologies named after Muhammad Al-Khwarizmi. Tashkent. Uzbekistan. mmuhriddinm@gmail.com Khamdamov Utkir3 3 DSc, Professor, Department of Infocommunication Engineering, Tashkent University of Information Technologies named after Muhammad Al-Khwarizmi, Tashkent, Uzbekistan. utkir.hamdamov@mail.ru, Djuraev Oybek4 4 Department of Communication and Digital Technologies, University of Management and Future Technologies, Tashkent, Uzbekistan. odjuraev@gmail.com Vision-Language Models (VLMs) have recently emerged as transformative tools that integrate computer vision and natural language understanding within a unified multimodal learning framework. Their potential to analyze medical images and generate clinically relevant textual outputs, such as diagnostic reports or answers to visual questions, is drawing growing attention from both artificial intelligence (AI) researchers and medical practitioners. This paper provides an in-depth synthesis and critical examination of recent advancements in VLMs for medical applications, focusing on report generation and visual question answering (VQA). We review model architectures, training strategies, and evaluation methodologies, while analyzing 15 publicly available medical VLM datasets and 12 representative model implementations. The study highlights the core design paradigms—encoderonly, decoder-only, and encoder-decoder architectures—and the major training strategies including self-supervised, contrastive, and parameter-efficient fine-tuning techniques. We also examine the challenges of data scarcity, domain shift, model hallucination, and the ethical and regulatory issues constraining clinical deployment. The paper concludes by identifying future research directions emphasizing multimodal alignment, human-in-the-loop systems, interpretability, and real-world integration to accelerate safe and reliable use of AI in healthcare. 1 INTRODUCTION The rapid convergence of computer vision and natural language processing (NLP) has given rise to a new generation of Vision-Language Models (VLMs)—systems capable of understanding CCS CONCEPTS Computing methodologies → Computer vision; Natural language processing; Neural networks; Multimodal machine learning. Applied computing → Health informatics. Additional Keywords and Phrases: Vision-Language Models, Medical Imaging, Report Generation, Visual Question Answering, Multimodal Learning, Deep Learning, Transformer Architecture, Self-Supervised Learning, Clinical AI, Interpretability. 73

and reasoning across visual and textual modalities. Unlike unimodal AI systems, which process images or text independently, VLMs integrate both sources of information to produce context-aware, semantically rich outputs such as captions, reports, and answers to visual questions. This integration has been enabled primarily by the success of transformer-based architectures and self-supervised learning, which allow models to capture complex cross-modal relationships from large datasets without extensive manual annotation. The ability of VLMs to combine perception with linguistic reasoning has significantly expanded their applicability, particularly in domains that rely on both visual and textual understanding, such as autonomous systems, surveillance, and healthcare. Among these, medical imaging has emerged as one of the most promising yet challenging fields for multimodal AI innovation [1]. In the medical domain, the need for automated or AI-assisted interpretation of clinical imagery has grown dramatically due to the increasing workload faced by radiologists and other specialists. Traditional deep learning models have achieved success in classification and detection tasks but remain limited to narrow problem settings and lack interpretability. VLMs, on the other hand, offer a pathway to generate clinically meaningful textual descriptions of images—such as diagnostic reports or answers to medical questions—thus bridging the gap between pattern recognition and semantic reasoning. Recent studies have demonstrated that vision-language frameworks can analyze X-rays, CT scans, or pathology slides while producing textual summaries that align with clinical observations [2]. Tasks such as medical report generation and visual question answering (VQA) represent critical use cases that test both the comprehension and generative capacities of VLMs. However, unlike general-purpose models trained on large-scale datasets such as COCO or LAION, medical VLMs face unique challenges including limited annotated data, privacy restrictions, domain shift, and the need for high factual accuracy. These challenges necessitate specialized training pipelines, medicalspecific datasets, and careful evaluation criteria tailored to the healthcare context. Despite remarkable progress, the deployment of VLMs in clinical practice remains constrained by several technical and ethical barriers [3]. Many existing models suffer from hallucination, producing plausible but incorrect statements that could mislead clinical decisions. Furthermore, evaluation metrics commonly adopted from the general domain (e.g., BLEU, ROUGE, METEOR) often fail to capture the clinical validity of generated text. In addition, issues of interpretability, data bias, and regulatory compliance continue to limit adoption in real-world medical settings [4]. The purpose of this paper is to provide a comprehensive synthesis and critical analysis of recent developments in medical VLM research. Building upon recent reviews and state-of-the-art studies, it examines architectural paradigms, training strategies, available datasets, and evaluation methodologies, while highlighting persistent challenges and potential research directions. By consolidating current knowledge and identifying key gaps, this study aims to support the advancement of trustworthy, interpretable, and clinically reliable vision-language systems that can safely augment human expertise in healthcare. The primary contributions of this research are as follows: This study provides an in-depth synthesis of recent advancements in medical VLMs, systematically categorizing existing approaches based on their architectures, training paradigms, and multimodal fusion techniques. It unifies scattered findings from diverse studies into a coherent overview that reflects the current state of the field; 74

The paper reviews and critically evaluates 15 publicly available medical image–text datasets and the metrics commonly used to assess report generation and VQA. It identifies limitations of current benchmarks and discusses the gap between automated metrics and true clinical relevance; Beyond summarizing existing work, this research highlights fundamental obstacles to clinical adoption of VLMs, including data scarcity, privacy constraints, hallucination risks, and generalization issues. The paper presents a structured analysis of these challenges from both technical and ethical perspectives; Based on the synthesized insights, the study outlines promising future directions for developing reliable, interpretable, and clinically deployable VLMs. It emphasizes areas such as multimodal alignment, human-in-the-loop training, knowledge-grounded generation, and robust evaluation frameworks tailored for medical AI. The remainder of this paper is organized as follows: Section 2 outlines the theoretical background and machine learning foundations of Vision-Language Models. Section 3 discusses major VLM architectures and training strategies. Section 4 reviews medical multimodal datasets and evaluation metrics for report generation and VQA. Section 5 compares state-of-the-art medical VLMs, while Section 6 highlights key technical, ethical, and practical challenges. Section 7 presents future research directions toward reliable and interpretable medical AI, and Section 8 concludes the paper with final remarks and recommendations. 2 RELATED WORKS The emergence of VLMs has reshaped how AI systems interpret, integrate, and reason over multimodal information. Early research in multimodal learning primarily focused on image–text alignment tasks, such as image captioning, VQA, and cross-modal retrieval. Foundational models like CLIP (Contrastive Language-Image Pretraining) [5] and ALIGN [6] demonstrated that largescale contrastive learning on paired image–text data could yield representations that generalize remarkably well across diverse downstream tasks. These architectures inspired a wave of domainspecific adaptations, particularly within the medical field, where integrating visual and textual data holds enormous potential for diagnostic decision support, report generation, and clinical interpretation. However, transferring general-domain models directly to medical data is non-trivial due to unique challenges in annotation scarcity, privacy concerns, and domain shift from natural to clinical images. As a result, researchers have developed specialized medical VLMs that combine the strengths of vision and language understanding under clinically constrained conditions. A considerable body of work has explored the adaptation of transformer-based architectures for medical multimodal tasks. Early frameworks such as MedViLL [7], VQA-RAD [8], and BioViL [9] leveraged contrastive pretraining between radiology images and associated text reports, showing that aligned visual–textual embeddings can enhance classification and retrieval tasks even with limited supervision. These approaches were primarily encoder-based and relied on pretraining objectives that encourage semantic similarity between paired modalities. Later models, such as BiomedCLIP [10] and RAMM [11], improved upon this by introducing hierarchical contrastive learning and multi-level feature alignment, enabling more robust cross-modal representations. Meanwhile, Visual MedAlpaca [12], CXR-RePair-Gen [13], LLaVA-Med [14], and RaDialog [15] adopted decoder or encoder–decoder architectures to facilitate generative capabilities, such as medical report generation and conversational reasoning about medical images. These models bridge perception and language 75

generation, allowing clinicians to interact with AI systems in natural language, marking a crucial step toward human-centered medical AI. In addition to architecture design, the construction of high-quality multimodal medical datasets has been pivotal in advancing this field. Publicly available datasets such as MIMIC-CXR [16], CheXpert [17], IU X-Ray [18], and PathVQA [19] have provided essential resources for pretraining and evaluation. Nonetheless, compared to general-domain datasets like LAION or COCO, medical datasets remain relatively small and domain-limited, often focused on chest radiographs or specific imaging modalities. The limited diversity and scale restrict model generalization to broader medical contexts such as histopathology, ophthalmology, or endoscopy. To mitigate these constraints, researchers have explored data augmentation strategies, synthetic text generation, and semisupervised learning techniques. Some recent works also incorporate self-supervised learning, where cross-modal alignment emerges naturally from unpaired data or latent representations. These approaches represent a promising direction for scaling VLMs in medicine without compromising patient privacy or annotation quality. Beyond model performance, interpretability and trustworthiness have become central themes in recent medical VLM research. Unlike general computer vision applications, clinical deployment requires rigorous transparency, reliability, and alignment with ethical standards. Studies have introduced methods for explainable attention visualization, uncertainty estimation, and human-in-theloop validation to ensure that model predictions align with medical reasoning. Frameworks such as CheXzero and MedCLIP highlight the capacity of contrastively trained VLMs to achieve radiologistlevel accuracy while remaining interpretable. Furthermore, hybrid approaches combining structured medical ontologies (e.g., UMLS, RadLex) with VLM architectures have shown promise in improving explainability and semantic consistency. Despite these advancements, challenges persist regarding bias mitigation, fairness across patient populations, and robustness under distribution shifts. Addressing these aspects is essential to ensure that VLMs evolve from research prototypes into clinically validated tools. Recent surveys and benchmarks have attempted to systematize this growing field, providing a clearer picture of current capabilities and gaps. Works such as MUMC [20] and PathChat [21] offer taxonomies categorizing models based on architecture, modality integration, and learning paradigm. These studies underline the shift from single-modality diagnostic models to holistic multimodal reasoning systems that integrate textual, visual, and sometimes structured EHR data. The current generation of medical VLMs thus reflects a convergence of progress in computer vision, natural language processing, and representation learning—ushering in a new era of multimodal clinical intelligence. Nevertheless, translating these technical achievements into practical healthcare improvements remains contingent on interdisciplinary collaboration between AI developers, clinicians, and policymakers. Only through such cooperation can VLMs mature into trustworthy and interpretable systems that truly augment clinical expertise. 3 VISION-LANGUAGE MODEL ARCHITECTURES The architecture of Vision-Language Models (VLMs) determines how effectively they can learn joint representations of visual and textual information. These models typically rely on transformer-based frameworks capable of modeling long-range dependencies and cross-modal relationships. The design of a VLM influences not only its ability to align visual and linguistic 76

semantics but also its scalability, interpretability, and suitability for medical applications. In the medical domain, where both precision and transparency are critical, VLM architectures are often adapted from general-purpose foundations such as CLIP, BLIP, or Flamingo, but modified to incorporate domain-specific constraints. Broadly, medical VLMs can be classified into three architectural paradigms: encoder-only, decoder-only, and encoder–decoder models. Each design offers distinct advantages for tasks such as image–text matching, medical report generation, and visual question answering (VQA). 3.1 Encoder-Only Architectures Encoder-only architectures form the backbone of early multimodal models and are particularly effective for representation learning and retrieval tasks. These models independently encode images and text into a shared embedding space, where semantic similarity between modalities can be measured. The most influential encoder-only model, employs dual encoders—one for vision and one for language—trained using a contrastive objective. CLIP’s success in the general domain inspired several medical adaptations, such as ConVIRT, GLoRIA, MedCLIP, and BioViL, which exploit paired radiology images and text reports to learn medical visual–textual alignments. These models replace general-domain text encoders like BERT with biomedical language models (e.g., BioBERT, ClinicalBERT) and fine-tune the vision encoders (e.g., ResNet, ViT) on domain-specific data. The encoder-only framework excels in feature alignment and cross-modal retrieval, enabling systems to associate radiology images with corresponding findings or to search for relevant images based on textual queries. For instance, ConVIRT introduced a scalable contrastive learning approach that learns image–text correspondences without explicit supervision, while GLoRIA extended this concept by applying hierarchical attention across local and global image features. Such techniques enhance the model’s ability to capture fine-grained anatomical or pathological cues. However, a limitation of encoder-only models is that they lack generative capacity—they can align and retrieve but not produce coherent textual outputs such as radiology reports. Moreover, since contrastive learning primarily optimizes representation similarity, these models can struggle to capture causal relationships or contextual reasoning, which are often crucial for medical interpretation. 3.2 Decoder-Only Architectures Decoder-only architectures, inspired by autoregressive language models like GPT, have recently been adapted for multimodal applications. In this paradigm, the model generates text tokens sequentially while conditioning on visual embeddings. Vision information is usually injected into the language model via cross-attention or learned adapters. Notable examples include LLaVA-Med, Med-PaLM, and VisualGPT, which extend large-scale language models to process visual prompts alongside textual input. These models demonstrate powerful generative capabilities, allowing them to describe images, answer open-ended questions, and engage in dialogue about visual content. In medical contexts, decoder-only VLMs have been applied to report generation, differential diagnosis explanation, and clinical note summarization. A key advantage of decoder-only designs lies in their strong linguistic fluency and ability to perform few-shot generalization across tasks. By leveraging large-scale pretrained language backbones, these models inherit linguistic structure and medical terminology knowledge from massive text corpora. For instance, Med-PaLM adapts the PaLM language model using multimodal 77

instruction tuning, allowing it to understand medical images and respond to clinician queries in natural language. Similarly, LLaVA-Med aligns visual tokens with LLaMA-based decoders, enabling contextual reasoning grounded in both modalities. Despite these advances, decoder-only VLMs face significant computational and interpretability challenges. They require vast computational resources for training and often behave as “black boxes,” making it difficult to trace how visual evidence influences textual output. Furthermore, their reliance on autoregressive decoding can sometimes produce verbose or hallucinated explanations—a serious concern in clinical environments where factual accuracy is vital. 3.3 Encoder–Decoder Architectures The encoder–decoder paradigm combines the representational power of dual encoders with the generative flexibility of decoders, offering a more balanced and interpretable framework for medical applications. In this setup, the vision encoder extracts spatial–semantic features from medical images, which are then fed into a text decoder conditioned through attention mechanisms. Representative models include BLIP, BLIP-2, Flamingo, and their medical variants such as MedBLIP and RadFM. The BLIP family introduced vision–language pretraining strategies that jointly optimize image–text matching and text generation, allowing the same model to handle retrieval, captioning, and reasoning tasks. RadFM, for example, integrates clinical knowledge into the attention mechanism, enabling fine-grained alignment between image regions and radiological findings. Encoder–decoder models are particularly well-suited for report generation and visual question answering, two tasks central to medical imaging interpretation. By encoding visual features and decoding textual descriptions, these architectures can generate coherent, context-aware diagnostic narratives. They can also provide token-level attention maps, facilitating interpretability and trustworthiness. However, their performance depends heavily on the quality of paired image–text datasets, which are limited in the medical domain. Additionally, encoder–decoder models must balance between overfitting to small datasets and maintaining generalization to unseen imaging modalities. To address these issues, hybrid approaches incorporate frozen pretrained encoders from general-domain VLMs (like CLIP or ViT) and fine-tune lightweight decoders on medical datasets, effectively reducing data dependency and computational cost. 3.4 Multimodal Fusion and Training Strategies Beyond architectural topology, VLMs differ in how they fuse visual and textual representations. Fusion strategies can be categorized as early fusion, where raw features are combined at the input stage; late fusion, where modality-specific embeddings are integrated after independent encoding; or hybrid fusion, which dynamically combines multiple interaction layers. In medical VLMs, crossattention mechanisms are widely used to fuse modalities, allowing textual queries to attend to relevant image regions and vice versa. Hierarchical or co-attention networks further enhance interpretability by modeling relationships between anatomical structures and textual descriptors. The choice of fusion strategy affects the model’s reasoning ability—early fusion offers deeper joint learning but requires more data, while late fusion promotes modularity and transferability. Training strategies play an equally crucial role. Pretraining typically involves contrastive learning, masked modeling, or instruction tuning, followed by task-specific fine-tuning. Contrastive losses encourage semantic alignment, masked modeling promotes contextual understanding, and instruction tuning enhances natural language reasoning. Recently, instruction-based VLMs have 78

gained traction, where models are trained with multimodal prompts designed to mimic clinical dialogues. For example, LLaVA-Med uses instruction datasets that teach the model to interpret radiographs, answer diagnostic questions, and justify its reasoning. Such approaches are promising for developing interactive medical assistants that can explain decisions and collaborate with healthcare professionals. 4 MEDICAL MULTIMODAL DATASETS AND EVALUATION METRICS This section reports the most widely used multimodal datasets that pair medical images with textual information, including radiology reports and question–answer annotations. It highlights their structure, scale, and purpose in training and evaluating VLMs for tasks such as report generation, cross-modal retrieval, and visual question answering. 4.1 Overview of Medical Multimodal Datasets The rapid evolution of VLMs in medicine has been closely tied to the availability of highquality multimodal datasets that pair medical images with corresponding textual information such as diagnostic reports, annotations, or question–answer pairs. Unlike the vast and diverse datasets used in general-domain multimodal learning (e.g., COCO Captions, Visual Genome, or LAION-5B), medical datasets face intrinsic challenges related to data scarcity, privacy constraints, and domain specialization. Nevertheless, several large-scale resources have emerged that form the backbone of medical VLM research. Among the most influential are MIMIC-CXR, CheXpert, Open-I (IU X-Ray), PathVQA, and VQA-RAD—each serving as a cornerstone for distinct multimodal tasks such as medical report generation, image–text retrieval, and visual question answering (VQA). The MIMIC-CXR dataset, one of the largest publicly available resources, contains over 370,000 chest X-rays paired with structured radiology reports. It provides both frontal and lateral views and is widely used for pretraining and evaluation in radiology-related VLMs such as ConVIRT, BioViL, and CheXzero. The dataset’s free-text reports facilitate cross-modal learning, enabling models to connect visual patterns with clinical findings like infiltrations, cardiomegaly, or atelectasis. CheXpert, another pivotal dataset, comprises more than 200,000 chest radiographs with uncertaintyaware labels derived from automated report parsing. Its structured format allows for quantitative evaluation of disease classification and localization tasks. IU X-Ray (Open-I), although smaller in scale, remains valuable for benchmarking report generation models due to its concise image–report pairs and detailed sentence-level annotations. For VQA-specific tasks, datasets such as VQA-RAD and PathVQA offer paired medical questions and answers associated with radiology and pathology images, respectively. These datasets enable models to perform reasoning-oriented evaluation, such as identifying anatomical locations, interpreting abnormalities, or explaining diagnostic outcomes in natural language. Despite these achievements, several limitations persist. Most existing datasets are modalityspecific, focusing heavily on chest radiography, which constrains generalization to other modalities like CT, MRI, ultrasound, or histopathology. Moreover, dataset sizes remain small compared to those in general vision–language research, limiting the scalability of large foundation models. Another concern is the quality and consistency of textual annotations. Radiology reports, often dictated in free form, vary in linguistic style, structure, and granularity, making supervised learning difficult. Efforts to standardize and enrich textual data—such as using structured templates, medical ontologies (e.g., 79