Real-world chronic recordings from implantable adaptive deep brain stimulation systems for Parkinson's disease motor state classification.

Introduction: Precise identification of motor states in Parkinson's disease (PD) is critical for adaptive deep brain stimulation (aDBS) therapies. Recent developments in implantable neurostimulators now support continuous neural recordings, enabling long-term monitoring of PD-related neural dynamics under real-world conditions. However, the real potential of such deep brain continuous recordings in chronic home-based conditions has not been explored yet. In this work, we introduce a multimodal classification framework that integrates subthalamic nucleus local field potentials (STN-LFPs), wearable sensors, and patient-reported diaries to distinguish between three key motor states: Methods: The framework was initially validated on data collected from three patients over an average of 30 days (totaling 2,136 h of LFP recordings), and then applied to a larger dataset from thirteen patients recorded over an average of 14 days (totaling 4,440 h of LFP recordings). Feature extraction was performed in the time, frequency and time-frequency domains, after applying principal component analysis (PCA) to decrease dimensionality, which preserved 95% of the variance of the data and reduced computational complexity. Results: The multilayer perceptron (MLP) classifier using time-frequency domain features achieved the highest F1-score among the models tested. For the three validation patients, the MLP reached F1-scores of 81.1% with wearable sensor data and 94.3% with diary-derived data. When evaluated across all thirteen patients, it maintained a strong F1-score of 93.1% using diary time-frequency features. Discussion: These results demonstrate that chronic LFP recordings available 24/7 from implantable aDBS devices enable robust motor state classification, thus supporting the personalization and optimization of aDBS systems for real-life use.