Proof-of-Concept Validation in Healthy Participants of an sEMG-Controlled 4-DOF Upper Limb Exoskeleton for Post-Stroke Rehabilitation

Abstract

Two improvements are needed in upper-limb exoskeletons (ULEs) for rehabilitation after stroke: a mechanism for the shoulder joint that will ensure interaction forces are minimized throughout the entire joint’s range of motion and an easy-to-use interface that detects neural intent. These have previously been addressed separately; here, we validate them together in one integrated device, the first closed-loop 4-DOF ULE with a surface electromyography (sEMG)-based classification interface. The objective is to test, in healthy participants, whether an anatomically aligned shoulder mechanism preserves sEMG signal quality during physical coupling while a lightweight classifier recognizes movement intent in real time. The mechanism uses three revolute glenohumeral joints whose axes intersect at the shoulder center of rotation (COR), giving the basic shoulder degrees of freedom (DOF) with minimal acromion interference, plus a proximal joint emulating the sternoclavicular joint to expand the workspace; an elbow joint completes the kinematic chain. The controller identifies four activities of daily living (ADL), namely shoulder abduction (SAB), pick-up cup (PUC), push-forward cup (PFC), and rest arm (RA). These activities are recognized by classifying a 72-dimensional feature vector (6 features × 12 channels) at 16 Hz, using a 125 ms sliding time window and a linear support vector machine (SVM). In a closed-loop study with 15 healthy volunteers, recognition accuracy was 92.4%, 89.7%, and 85.1% for SVM, artificial neural network (ANN), and logistic regression, respectively; SVM recall was 96% (SAB) and 98% (RA). End-to-end latency was 148 ± 19 ms, below the accepted 200 ms threshold. Participants rated the system “Good” (SUS 71.3 ± 8.4) with only slight discomfort (NRS 1.8 ± 0.7). Misclassification was dominated by the kinematically adjacent PUC–PFC pair. As validation used healthy participants, these findings establish technical feasibility as a precursor to future clinical trials in stroke patients. ABSTRAK: Eksoskeleton anggota atas (upper limb exoskeletons, ULE) untuk pemulihan pasca-strok memerlukan dua kemajuan: mekanisme bahu yang mengekalkan daya interaksi rendah sepanjang julat pergerakan, dan antara muka intuitif yang mengecam niat neuro-otot tanpa beban mental tambahan. Kedua-duanya sebelum ini ditangani secara berasingan; di sini kami mengesahkan gabungannya dalam satu peranti bersepadu — ULE 4-DOF gelung tertutup pertama dengan antara muka pengelasan sEMG. Objektif kajian ialah menguji, dengan peserta sihat, sama ada mekanisme bahu yang sejajar secara anatomi dapat mengekalkan kualiti isyarat sEMG semasa gandingan fizikal, sementara pengelas ringan mengecam niat pergerakan secara masa nyata. Mekanisme ini menggunakan tiga sendi glenohumeral berputar yang paksinya bersilang pada pusat putaran (center of rotation, COR) bahu dengan gangguan minimum pada akromion, serta satu sendi proksimal yang meniru sendi sternoklavikular untuk memperluas ruang kerja; satu sendi siku melengkapkan rantaian kinematik. Pengawal tersebut mengenal pasti empat jenis aktiviti kehidupan harian (Activities of Daily Living, ADL), iaitu abduksi bahu (Shoulder Abduction, SAB), mengambil cawan (Pick-Up Cup, PUC), menolak cawan ke hadapan (Push Forward Cup, PFC), dan merehatkan lengan (Rest Arm, RA). Aktiviti-aktiviti ini dikelaskan melalui vektor ciri berdimensi 72 (6 ciri × 12 saluran) pada kadar 16 Hz, menggunakan tetingkap masa gelongsor 125 ms dan Mesin Vektor Sokongan linear (linear SVM). Dalam eksperimen gelung tertutup dengan 15 subjek sihat, ketepatan pengecaman ialah 92.4%, 89.7%, dan 85.1% masing-masing bagi SVM, rangkaian neural buatan (ANN), dan regresi logistik; kepekaan (recall) SVM ialah 96% (SAB) dan 98% (RA). Kependaman hujung-ke-hujung ialah 148 ± 19 ms, di bawah ambang 200 ms yang boleh diterima. Peserta menilai sistem ini “Baik” (SUS 71.3 ± 8.4) dengan sedikit ketidakselesaan sahaja (NRS 1.8 ± 0.7). Salah pengelasan didominasi oleh pasangan PUC–PFC yang berdekatan secara kinematik. Oleh kerana pengesahan dijalankan ke atas peserta sihat, dapatan ini merupakan langkah awal kepada ujian klinikal dengan pesakit strok.