Preprocessing and feature extraction co-optimization for low-power EEG-based neurodegenerative disease detection: a benchmark on STM32 embedded platforms

Early differential diagnosis of Alzheimer’s disease and Frontotemporal Dementia (FTD) using electroencephalography (EEG) enables low-cost and non-invasive neurological monitoring. This study presents a lightweight TinyML framework for embedded EEG classification on resource-constrained microcontrollers. Twelve preprocessing techniques and six machine learning classifiers were evaluated on two public OpenNeuro datasets containing 88 participants. All pipelines were implemented on STM32 NUCLEO-64 and STM32WB55 platforms using INT8 quantization with CMSIS-NN. Results show that Sparse Decomposition combined with XGBoost achieved the best performance, reaching 94.46% accuracy and 92.1% sensitivity while requiring only 45 KB SRAM and low execution time, demonstrating the feasibility of wearable EEG diagnostic systems.