Hybrid Quantum-Classical Model That Combines Spatial-Temporal Electroencephalography and Digitized Counterdiabatic Quantum Features for Motor Imagery Classification

To develop a quantum feature-enhanced algorithm for motor imagery (MI) classification using electroencephalographic (EEG) waveforms.

Data were obtained from a published, deidentified data set. Fifty participants were available for model training and one participant's data was reserved for external evaluation. Each participant underwent 600 MI task cues over 3 sessions associated with 4-second, 58-lead EEG waveforms. Our approach used a spatial-temporal deep learning architecture to reduce the raw waveforms to a 24-dimension feature set for digitized counterdiabatic quantum feature extraction. The final classifier combined the deep learning prediction with the quantum features to provide an overall MI prediction weighting the quantum features more heavily when the spatial-temporal deep learning model was uncertain.

All 51 participants were right-handed and 34 (66.7%) were male. The mean (SD) age was 23 (2.2) years with a range of 17 to 30 years of age. The evaluation participant was a 22-year-old man. The classification accuracy in the evaluation participant was 88.8% (533 of 600 MI trials; 95% CI, 86.0% to 91.2%) with an area under the receiver operating characteristic curve of 0.962 (95% CI, 0.948 to 0.976). The classification accuracy for the entire set of 51 participants was 89.8% (27,478 of 30,590; 95% CI, 87.5% to 92.2%) with an area under the receiver operating characteristic curve of 0.970 (95% CI, 0.956 to 0.980). Model performance improved with age (Spearman ρ = 0.35, P=.012).

A combination of deep learning feature extraction enriched with quantum features resulted in robust classification of MI EEG waveforms, setting a foundation for further study of quantum computing in health care.