Posture-Dependent LFP–EMG Associations and Latent-State Conditioned Within-Recording Short-Term Prediction in Dystonia

Neural–muscle associations may vary with posture, complicating continuous decoding in dystonia. We examined posture-related associations and individualized predictions using postoperative local field potential (LFP) and electromyography (EMG) recordings. Of 18 patients, 16 contributed paired supine–seated data, and 15 contributed 98 recordings for prediction. A low-dimensional latent state inferred from causal LFP band-envelope histories was combined with current features to predict the mean EMG envelope over the subsequent 2 s. Raw EMG was divided chronologically into training, validation, and test partitions before independent target processing. The seated-to-supine participant-level EMG geometric mean ratio was 1.5687 (95% confidence interval, 1.2571–2.1971). Exploratory mixed-effects analyses identified selected posture-related power and power–EMG slope differences under model-based inference. Mean Pearson correlations in the principal analysis were 0.2856, 0.2631, and 0.3141 for subthalamic nucleus (STN), globus pallidus internus (GPi), and joint inputs, respectively. Descriptive model rankings varied across inputs and metrics. Neither connectivity-feature additions nor lead-time contrasts reached multiplicity-adjusted significance. These findings indicate modest within-recording EMG prediction using LFP alone. Generalization to independent recordings, strictly future-supported EMG targets, and clinical relevance require further validation.