Abstract
The development of effective automated systems to prevent patient self-harm in psychiatric wards is severely hampered by a scarcity of realistic training data. To address this critical gap, this study introduces a new public dataset of 1120 videos simulating cutting actions in a controlled studio environment, alongside a validation set of 118 real-world videos from secure wards that include more diverse behaviors such as picking and scratching. We conducted a comprehensive benchmark of state-of-the-art action recognition models, including both convolution-based and transformer-based architectures, to evaluate their performance and generalizability from simulated to real-world conditions. Our results reveal a significant "sim-to-real" gap: while the top-performing model, VideoMAEv2, achieved a high F1 score of 0.65 on the simulated data using 7-fold LOAO cross-validation, its performance degraded to a mean F1 score of 0.61 on the real-world data. This performance drop is attributed to the models' inability to generalize from the uniform, simulated actions to the diverse and often occluded behaviors observed in authentic clinical settings. By providing a foundational dataset, a systematic benchmark, and a qualitative analysis of model failure points, this study quantitatively demonstrates the limitations of current approaches. Our findings underscore the urgent need for more diverse data and advanced approaches to develop robust technologies that can enhance patient safety.