Abstract (eng)
Real-time fMRI-based neurofeedback is an emerging scientific and clinical tool that allows for learning to self-regulate brain activity. It has been shown to modulate behaviour in healthy individuals, and further has demonstrated the capacity to improve clinical symptoms in various patient populations. However, the performance in self-regulating neural activity varies considerably across studies and individuals. Consistent learning curve patterns, such as steadily rising regulation performances across runs, are rare. Here, we investigate whether neurofeedback regulation performances across runs are merely random or follow a predictable pattern. This is achieved by applying machine-learning (L1-regularized Linear Regression & Randomized Trees) to predict the regulation performance of a training run based on previous training run performances. Additionally, we included subject- and study-specific characteristics such as age, sex, instructions, trained brain regions, and the length of regulation blocks in our machine-learning models to investigate how these factors affect performance. For assessing the relevance of each feature, we applied permutation-based feature importance analyses to our trained models. To obtain results that generalize across the field of real-time fMRI neurofeedback, our analyses was conducted on a large and heterogeneous real-time fMRI neurofeedback dataset of 197 participants from 11 different studies that included healthy participants as well as patients, different ROIs, and diverse experimental designs. We were able to predict regulation performance significantly better than chance level. However, with median R² values of up to 0.26 a considerable part of variance remains unexplained. For the predictions, previous regulation performances were the most crucial features. Overall, we found that performance in neurofeedback training is not random but to some degree predictable. These results might help to develop a better understanding of how self-regulation of brain activity with neurofeedback is accomplished, thus allowing for more effective clinical and scientific use of this promising method. Considering increased availability of suitable data in the context of the Open Science movement, our data-driven approach might become a promising avenue for advancing our understanding and the applicability of neurofeedback.