Threshold-Free Neural Network Models for Swim Bout Detection of Larval Zebrafish

Accurate detection of swim bouts is essential for decoding motor behavior in larval zebrafish, yet conventional threshold-based methods rely on subjective cutoffs and per-experiment tuning. To address this limitation, we developed threshold-free deep learning models operating on tail-curvature time series: an offline model for post hoc analysis and an online model for real-time detection. Both improved precision over threshold-based methods, particularly for low-amplitude bouts. Using the offline model, we show that a head-fixed, closed-loop preparation largely preserves naturalistic swimming kinematics. Repeated trials reduced bout frequency and increased interbout interval, whereas genetic background primarily affects bout duration and maximum amplitude; notably, casper larvae lacked the late-trial decline observed in wildtype and nacre. We further show that detector choice can alter inferred behavioral outcomes. Together, these results establish a high-throughput framework for robust and reproducible analysis of zebrafish swimming behavior across experimental conditions.