Hybrid Edge-Cloud Asymmetric Analytics for Portable Multimodal BCI Biosensors

Area of Science:

• Biomedical Engineering
• Machine Learning
• Wearable Technology

Background:

• Portable biosensor hardware enables continuous multimodal physiological data acquisition.
• The analytical layer for real-time inference remains a bottleneck for practical embedded systems.

Purpose of the Study:

• To present the machine-learning layer of the Real-time Cognitive Grid for real-time physiological-state classification.
• To establish an asymmetric edge-cloud workflow for biosensor data analysis.
• To demonstrate analytical consistency across diverse public datasets and hardware implementations.

Main Methods:

• Developed a 17-feature schema (EMG, EEG, HRV, EOG, EOG) for edge and cloud tiers.
• Utilized an Arduino Nano RP2040 Connect for edge inference (LDA) and cloud for pretraining/refinement (Random Forest).
• Evaluated performance on five public repositories (WESAD, DEAP, PAMAP2, EMG Gestures, EEGMMIDB) and a hardware branch using GroupKFold cross-validation.

Main Results:

• Multimodal fusion improved macro-F1 by up to 0.141 in WESAD and 0.109 in PAMAP2.
• The edge LDA model achieved 0.9435 macro-F1 and 0.9470 accuracy.
• Cloud Random Forest reached 0.8792 macro-F1 and 0.8799 accuracy; EMG features dominated importance analysis.

Conclusions:

• A compact multimodal biosensor assembly can perform context-aware interpretation with minimal user intervention.
• The analytical workflow is coherent across heterogeneous benchmarks, hardware refinement, and microcontroller deployment.
• Cross-session bench feasibility is established for future multi-subject wearable validation.