Related Experiment Video
Updated: Jul 6, 2026
Implementation of a Real-Time Psychosis Risk Detection and Alerting System Based on Electronic Health Records using CogStack
Published on: May 15, 2020
Reduction in suicides and suicide attempts following implementation of AI-based video surveillance in the Stockholm
Johan Fredin-Knutzén1,2, Gergö Hadlaczky3,4, Anna-Lena Andersson3
1National Centre for Suicide Research and Prevention of Mental Ill-Health (NASP), Karolinska Institutet, Stockholm, S-171 77, Sweden. johan.fredin.2@ki.se.
Background:
Over 700,000 suicides occurring globally each year are a major public health issue. Railways and metros provide a lethal means of suicide, mainly occurring by persons under train (PUT) events. Restriction of means has been shown to be effective in reducing suicides and is increasingly being prioritized in railway settings, e.g., using physical barriers. Here, we instead investigated the changes in suicidal behavior on metro platforms following the implementation of an AI-based CCTV detection system.
Methods:
We used longitudinal data about PUT due to suicidality in the Swedish metro system in Stockholm (2010-2025). A controlled interrupted time series (CITS) analysis, as well as uncontrolled analyses, were used to test if an AI-CCTV implementation in Q4 2021 (at 14 stations) was associated with decreased rates of PUT due to suicidality, compared to the other 86 stations as controls. We also evaluated secondary outcomes (e.g., suicide deaths and train-traffic cancellations). Sensitivity analyses assessed the robustness of the primary model. A separate exploratory analysis examined an extended post-period, which included multiple heterogeneous system-wide exposures.
Results:
Rates of PUT due to suicidality were lower after AI-CCTV implementation, in analyses with (IRR = 0.27, p < 0.05) or without (IRR = 0.41, p < 0.05) controls. Secondary outcomes showed consistent point estimates, e.g. death by suicide after PUT (IRR = 0.3), safeguarded individuals (IRR = 0.8) and less cancelled train-kilometers in the metro system. The robustness of the changes in PUT and death by suicide outcomes were confirmed by Bayesian sensitivity analyses using weak priors.
Conclusions:
Implementation of AI-CCTV as described herein was associated with lower rates of PUT due to suicidality at metro stations, as well as changes in the same direction for e.g. deaths by suicide and less cancelled train-kilometers. This preliminary study of AI-CCTV in the metro system provides a specific example of how such an intervention may support suicide prevention in the metro system, at least in the short-term.
