1233 - Predictive Motion Management Using Surface Electromyography (sEMG)

Purpose/Objective(s):

Current motion management systems are reactive in that they detect patient displacement but require 200-400 ms to respond. This limitation prevents anticipatory intervention during breath-holds, forcing patients to their physiological limits before the system can respond, and fundamentally prevents ultra-fast delivery techniques like FLASH-RT where total treatment time is shorter than system response latency. We investigated whether non-invasive surface electromyography (sEMG) can overcome this limitation by predicting movement 200-500 ms before physical displacement, exploiting the electromechanical delay between muscle depolarization and force generation.

Materials/Methods:

We developed a machine learning framework to predict motion onset using two publicly available sEMG datasets: voluntary upper-limb gestures (10 subjects) and involuntary respiratory cycles (8 subjects). Ground truth was obtained from kinematic sensors and spirometry. We extracted signal envelopes and computed features at lead times from 50 to 500 ms before motion onset. Linear Discriminant Analysis classified whether movement would occur within each prediction window. Performance was evaluated using leave-one-subject-out cross-validation and quantified by area under the Receiver Operating Characteristic (ROC) curve (AUC). Additionally, computational latency was measured to assess real-time feasibility.

Results:

sEMG successfully predicted motion before physical onset; For voluntary upper-limb movement, we achieved AUC of 0.86 ± 0.03 at 50 ms lead time, maintaining robust performance (AUC > 0.77) up to 500 ms. For respiratory motion, optimal prediction occurred at 300 ms lead time (AUC = 0.79 ± 0.16) and robust performance (AUC> 0.70) was maintained up to 800 ms. Total inference latency remained under 6 ms, enabling real-time beam control. These prediction windows provide sufficient temporal margin for system response, enabling the beam to stop before motion begins. Additionally, respiratory sEMG yielded accurate respiratory rate estimation (MAE = 2.3 ± 0.7 breaths/min when compared to the spirometer) and heart rate extraction (mean ~88 bpm), indicating multi-modal real-time physiology monitoring potential. Prediction results below.

Conclusion:

This proof-of-concept demonstrates that sEMG-based prediction can anticipate motion with sufficient lead time to overcome hardware response latencies, essentially shifting the motion management paradigm from reactive detection to pro-active detection. Future work will validate this approach using radiation therapy-specific training data and hardware-integrated closed-loop systems.

Lead Time (ms)	Upper-limb AUC (mean +/- SD)	Upper-limb Accuracy	Respiratory AUC (mean +/- SD)	Respiratory Accuracy
50	0.859 +/- 0.04	77.0%	0.752 +/- 0.20	69.1%
100	0.800 +/- 0.04	72.0%	0.768 +/- 0.19	72.1%
200	0.741 +/- 0.05	67.2%	0.782 +/- 0.17	70.9%
300	0.750 +/- 0.05	68.3%	0.786 +/- 0.19	72.6%
500	0.773 +/- 0.05	69.8%	0.759 +/- 0.16	70.4%