Boundary Density Likelihood Improves Event Detection in Sleep Studies

A novel approach for training models in event detection, known as Boundary Density Likelihood (BDL), optimizes the final output directly instead of focusing on intermediate segmentation. In a predefined five-fold nested sleep study, BDL-Hard improved the pooled out-of-fold mean average precision (mAP) from 0.586 to 0.705, an increase of 11.9 percentage points (95% interval [10.8, 13.0]), and enhanced strict one-minute average precision from 0.071 to 0.286 (4.0x; +21.5 points). This method demonstrated improvements across all outer folds, and a separate rerun confirmed the results. BDL allocates one unit of target mass per annotated event, maintaining this mass through smoothing and temporal downsampling, while employing a Poisson objective for expected event mass estimation. The research can be found on arXiv under identifier 2408.12792, categorized as 'replace'. The paper details the method and findings, emphasizing event detection in lengthy recordings, relevant to sleep analysis and other time-series applications.