P080 A Representation Learning approach captures clinical effects of slow subthalamic beta activity drifts
Salvatore Falciglia*1,2, Laura Caffi1,2,3,4, Claudio Baiata3,4, Chiara Palmisano3,4, Ioannis U. Isaias3,4, Alberto Mazzoni1,2
1The BioRobotics Institute, Scuola Superiore Sant’Anna, Pisa, Italy
2Department of Excellence in Robotics and AI, Scuola Superiore Sant’Anna, Pisa, Italy
3University Hospital Würzburg and Julius Maximilian University of Würzburg, Würzburg, Germany
4Parkinson Institute Milan, ASST G. Pini-CTO, Milan, Italy
*Email: salvatore.falciglia@santannapisa.it
Introduction
Deep brain stimulation (DBS) of the subthalamic nucleus (STN) is a mainstay treatment for drug-resistant Parkinson's disease (PD) [1]. Adaptive DBS (aDBS) dynamically adjusts stimulation according to the beta ([12-30] Hz) power of the STN local field potentials (LFPs) to match the patient's clinical status [2]. Today, aDBS control depends on accurate determination of pathological beta power thresholds [3]. Notably, in the days/months timescale, STN beta power shows irregular temporal drifts affecting the long-term efficacy of the aDBS treatment. Here we aim at characterizing these drifts and their clinical effects with a multimodal study, integrating neural and non-neural data streams.
Methods
We conducted home monitoring of patients with PD, focusing on periods of rest and gait activity. Multimodal data were collected, including STN LFPs from chronically implanted DBS electrodes, wearable inertial sensor recordings, and patient-reported diaries. A low-dimensional feature space was derived by integrating the acquired signals through Representation Learning techniques [4]. Leveraging LAURA, our transformer-based framework for predicting the long-term evolution of subthalamic beta power under aDBS therapy [5], we present a multimodal approach where neural data are paired with kinematic data and labelled according to the patient’s clinical status during the monitored activity.
Results
We observed that STN beta power distributions show large irregular non-linear fluctuations over several days. Consequently, patients spend a significant portion of time in suboptimal stimulation states. A fully informative description of the STN LFPs dynamics is achieved by integrating neural, kinematics, and clinical data into a low-dimensional feature-based representation. Latent patterns of STN activity correlate with clinical outcomes as well as motor and non-motor daily activities, necessitating further explainability within the same low-dimensional space. This might support clinically effective recalibration of aDBS parameters on a daily basis.
Discussion
Our study advances the understanding of slow timescales of pathological activity in PD patients implanted with DBS. We developed a comprehensive deep learning framework that integrates neural data with longitudinal clinical information, enabling a more precise characterization of patient status. This will enable personalized control strategies for stimulation parameters (Fig. 1) and enhance the clinician-in-the-loop paradigm by improving patient status assessment and automating aspects of neuromodulation to prevent suboptimal stimulations due to beta power drifts. Ultimately, this work paves the way for novel long-term neuromodulation strategies with potential applications to neurological disorders beyond PD [6].
Figure 1. Block diagram of aDBS as a closed-loop control system. The control loop operates on two separate timescales. In the short-term, the modulation changes with fluctuations in beta power (solid box). In the long-term, the parameters of the fast aDBS algorithm are updated based on the expected drifts of daily beta distributions combined with the neurologist’s clinical assessments (dashed box).
Acknowledgements
The authors declare that financial support was received for the research. The European Union - Next-Generation EU - NRRP M6C2 - Investment 2.1: projects IMAD23ALM MAD, Fit4MedRob, and BRIEF. Fondazione Pezzoli per la Malattia di Parkinson. The Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) Project-ID 424778381 - TRR 295.
References
1.https://doi.org/10.1007/s00221-020-05834-7
2.https://doi.org/10.1088/1741-2552/ac3267
3.https://doi.org/10.3390/bioengineering11100990
4.https://doi.org/10.1109/TPAMI.2013.50
5.https://doi.org/10.1101/2024.11.25.24317759
6.https://doi.org/10.3389/fnhum.2024.1320806
