JMIR Publications

ABSTRACT

Background:

In highly secure psychiatric wards, surveillance cameras are used to prevent incidents such as self-harm. However, round-the-clock monitoring remains challenging due to limited staffing and the difficulty of detecting covert behaviors. To address these issues, the introduction of automated monitoring systems has been proposed, yet their effectiveness relies heavily on the availability of large-scale self-harm datasets—resources that are notoriously hard to obtain. Gathering reliable self-harm data is crucial for bolstering prevention and monitoring capabilities, but the collection and management of sensitive patient footage raise ethical and legal concerns that complicate the process. Even in wards equipped with continuous surveillance, acquiring high-quality, comprehensive data proves difficult. This scarcity of data impedes the development of robust automated monitoring systems and, ultimately, slows progress in enhancing safety measures within mental health care settings.

Objective:

This study aims to address the lack of self-harm datasets by introducing a publicly available dataset and evaluating its effectiveness in training action recognition models for automated monitoring in psychiatric wards, assessing their applicability to real-world settings.

Methods:

A self-harm dataset was created, consisting of 1,120 reenactment videos recorded in a controlled studio environment, along with 118 real-world videos collected from secure psychiatric wards over a 10-month period. These datasets were used to evaluate state-of-the-art action recognition models, including Temporal Segment Network, SlowFast, TimeSformer, UniformerV2, and Vision Transformer (ViT). The models were tested on both the controlled and real-world datasets to examine their generalization performance.

Results:

Experimental results show that ViT-Base, pre-trained on Kinetics 400 using Video Masked AutoEncoderV2, achieved the highest F1 score of 0.78 on the studio dataset. However, all models performed significantly worse on the real-world dataset, with ViT-Base achieving only 0.65 accuracy, highlighting the challenge of generalizing to real-world scenarios.

Conclusions:

By making this dataset publicly available, this study aims to promote further research in self-harm detection and enhance automated monitoring in secure medical environments. Future work will expand the dataset and explore methods to bridge the gap between controlled and real-world conditions, providing a foundation for robust patient safety solutions.