Data-Driven Models for Real-Time Prevention of Adverse Events in Robotic Needle Interventions




Journal Title

Journal ISSN

Volume Title



Robotic needle interventions enable surgeons to maneuver flexible needles into deep tissues with enhanced dexterity. The technology shows tremendous potential for therapeutic applications if critical events related to needle-tissue interaction mechanics are addressed during intra-operative procedure. Common events include needle buckling due to flexibility of the shaft, undesired needle deflections due to tissue inhomogeneity, and tissue displacements due to insertion force of the needle. Accurate and timely predictions of these adverse events is important to ensure patient safety. However, it is challenging to provide autonomous control for these events given the unknown dynamic nature of the tissues. The goal of this dissertation is to develop model-free methods to identify, predict and minimize adverse events in real time. The first topic of this dissertation focuses on identifying and detecting the listed three events using only sensor measurements. The algorithms use errors between the sensor readings and estimation models generated from the sensor data to track rapid changes in sensor patterns through scale independent metrics. These metrics enable classification of events based on intensity and behavior. Validation experiments in known gelatin tissues and biological tissues with unknown environments show generalizability. Early prediction of adverse events could improve procedural safety. The second topic of this dissertation focuses on developing computationally fast methods to forecast general time series data, as adverse events are characterized by time series based sensors. This novel forecasting technique is derived from a model-free adaptive control framework. The technique offers high computation efficiency and does not require any parameter tuning or training data, compared to existing statistical and learning models. Forecast accuracy is comparable with current state-of-the art statistical methods. Validations on realistic data sets show application to general time series for data prediction. Using sensor data forecasts and detection metrics from first topic can predict adverse events before occurrence. Preliminary validation experiments in gelatin and biological tissues show robustness to tissue environments. The low computation times of the combined prediction models show potential for integration with autonomous control routines. Dynamics of needle-tissue interactions are unknown and nonlinear. The last topic of the dissertation discusses model-free methods for needle steering control to prevent adverse events during insertions. The traditional model-free adaptive scheme is modified to improve robustness against unexpected changes in system behavior or process errors. Simulation results in known nonlinear systems show consistent stability and faster convergence under unexpected process errors. Experimental results in gelatin tissues show successful prevention of adverse events through prediction and steering control scheme. Results in this dissertation show potential applications to autonomous prediction and control of sensor-driven robotic systems that require safe operations.



Surgical instruments and apparatus, Surgical technology, Surgical robots, Machine learning