Prediction of nanobubble-assisted focused ultrasound-induced blood-brain barrier opening with machine learning

Abstract

Novel approaches for predicting the outcomes of blood-brain barrier (BBB) opening with focused ultrasound (FUS) and microbubbles are highly desired. This study aims to explore machine learning-based methods for reliably predicting the FUS-induced BBB opening efficacy and safety. Methods: Sixteen female rats were used in this study. An acoustic feedback-controlled FUS system (f0: 0.5MHz) was used for the BBB opening with the infusion of custom-made nanobubbles/Definity. Evans Blue was injected for the BBB opening efficacy verification and the brain tissue was harvested for the safety assessment. Acoustic emissions were recorded, preprocessed and fed into three machine learning models for BBB opening outcomes prediction. Conventional stable and inertial cavitation dose were also calculated. Results: Among the tested machine learning models, a modified Support Vector Data Description (mSVDD) model achieved the best performance in the BBB opening efficacy and safety prediction with an accuracy of 85.0+/-16.6% and 62.5+/-12.8%, respectively. Conventional stable and inertial cavitation dose-based prediction has a prediction accuracy of 80.0% in efficacy and 34.3% in safety, respectively. The mSVDD model trained with the overall bubble response (0-2 MHz) performed better than that trained with the ultra-harmonic bubble response (0.7-0.8 MHz) in both efficacy prediction (85.0+/-16.6% vs 76.0+/-8.0%, p=0.04) and safety prediction (62.5+/-12.8% vs 55.0+/-10.7%, p>0.05). It is also found that the mSVDD model trained with nanobubble data cannot be directly applied to Definity. Conclusion: Our investigations demonstrated that it is feasible to achieve a reliable prediction of FUS-induced BBB opening outcomes with machine learning and acoustic signals from stimulated nanobubbles. This study provided a new approach for the prediction of FUS-BBB opening outcomes with a clinical translation potential.