Multivariate respiratory motion prediction.

Area of Science:

• Medical Physics
• Robotics
• Biomedical Engineering

Background:

• Extracranial robotic radiotherapy relies on tracking surrogates to compensate for tumor motion.
• System-specific time delays necessitate time series prediction of external optical surrogates.

Purpose of the Study:

• To evaluate if incorporating an accelerometer, strain belt, and flow sensor enhances prediction accuracy.
• To adapt and compare four prediction algorithms (nLMS, wLMS, SVR, RVM) for multivariate inputs.
• To propose and utilize a sequential forward selection (SFS) method for optimal sensor selection.

Main Methods:

• Four prediction algorithms (nLMS, wLMS, SVR, RVM) were adapted for multivariate sensor inputs.
• A sequential forward selection (SFS) method was used to identify the most relevant and least redundant sensors.
• Data from 18 subjects were collected in two phases (M1: long-term trends, M2: breathing artifacts).

Main Results:

• The clinically used univariate normalized least mean squares (nLMS) algorithm showed susceptibility to outliers.
• For irregular breathing (M2), the root mean square error (RMSE) decreased from 0.66 mm (univariate nLMS) to 0.46 mm (multivariate RVM).
• Further optimization using the full test set showed potential for RVM to reach 0.42 mm RMSE, with accelerometer and strain belt being frequently selected sensors.

Conclusions:

• Multivariate models, particularly Relevance Vector Machines (RVM), significantly improve tumor motion prediction accuracy compared to univariate methods.
• Integrating additional sensors like accelerometers and strain belts into clinical setups can enhance motion compensation precision in robotic radiotherapy.
• Further research into sensor selection methods is warranted to fully leverage multivariate approaches for improved treatment accuracy.