Conference Agenda

Introduction

Beginning in 2015, a large-scale railway infrastructure program in Melbourne, Australia has resulted in level crossings being progressively removed. Our previous study conducted when 41 crossings had been removed showed there was a reduction in railway suicides close to removal sites. The aims of this current study were to determine whether: (1) railway suicides have continued to decline close to removal sites (now 88 sites); (2) there has been a reduction in railway suicide overall; and (3) there is evidence of substitution to other methods as the railway track has become restricted.

Methods

We classified Victorian Suicide Register data for the period January 2010 to June 2025 into three mutually exclusive categories: (1) railway suicides, (2) other non-residential suicides and (3) residential suicides. Each suicide was cross-joined to every level crossing removal site and flagged if they occurred within various buffer distances. We used Poisson regression models to estimate incidence rate ratios (IRRs) for pre- versus post-removal periods.

Results

Our dataset comprised 10,333 suicides, 5.1% occurred on railways, 20.5% in other non-residential locations and 74.4% in residential settings. We observed reductions in railway suicides of 54% within 500m of removal sites (IRR = 0.46), 49% within 1000m (IRR = 0.51), 36% within 2000m (IRR = 0.64) and 16% overall (IRR = 0.84). In contrast, other non-residential suicides showed no significant change at the 500–2000m buffers (IRRs 0.97–1.28) and a slight increase overall (IRR = 1.13). Residential suicides increased modestly, with IRRs ranging from 1.05 overall to 1.28–1.29.

Discussion

The reductions in railway suicides were greatest closest to removal sites, supporting the intervention effect. By contrast, the small increase in other non-residential suicides was only observed at distances of 5km or greater, suggesting it might be more indicative of a background rise in suicides rather than method substitution.

Death by suicide is a major societal issue, affecting railways in particular. However, the numbers of suicides have likely been underestimated, by being misclassified as accidents. In the year 2016, Sweden finalized the introduction of more extensive investigation procedures to resolve more deaths by suicide. Before year 2016, railway fatalities with undetermined suicidal intent were simply classified as accidents by default, unless forensic autopsies or police decided otherwise. During the years 2016–2023, all fatalities were evaluated by using a special investigation form and undetermined or drug use deaths were further investigated by a psychosocial investigator.

An expert group then classified these deaths as suicide or accident. We evaluated the effect of these altered procedures by interrupted time series (ITS) models using the official statistics concerning the percentage of deaths reported as suicides (%suicides). Over the entire study period, %suicides increased from ~75 % to ~87 %.

Using ITS models accounting for a roll-in implementation period (2012–2016), displayed the best fit compared to alternative competing models, as well as a significant increase of %suicides by at least ~4 % during the period after 2016 per se and ~10 % if including the roll-in period. The effect after 2016 was mediated by a decline in the official accident counts, but more detailed classification data suggested that this was rather due to a reduction of undetermined causes of deaths. We conclude that the more extensive classification procedures reduced the misclassification of suicides as accidents, whereby we recommend its continued usage.

AI-based video surveillance has been suggested for some time as a possible way to reduce suicides and suicide attempts in rail settings, but there has been no proper evaluation of its effects. In this study, we examine an automated camera system (AI-CCTV) used in the Stockholm Metro. The system reacts to certain predefined behaviours at stations and allows staff to act earlier in situations that may otherwise lead to a person-under-train (PUT) event.

We applied a Controlled Interrupted Time Series (CITS) design together with an ITS analysis, using data from 2010–2022. The models were adjusted for seasonality, long-term trends, and other factors that might influence the outcomes. Stations without AI-CCTV served as controls. The primary outcome was PUT due to suicidality (suicides and non-fatal suicide attempts). Secondary outcomes were suicide due to PUT (key secondary outcome), individuals safeguarded for suicide risk, PUT due to accidents, and cancelled train services linked to trespassing and PUT.

After the introduction of AI-CCTV, the patterns in the data suggest fewer PUT events related to suicides and non-fatal suicide attempts at the intervention stations. Similar tendencies were seen across the different modelling approaches, and we found no clear indication of incidents shifting to nearby stations. Detailed results will be presented at the conference.

We will also outline the next steps to improve the system's effectiveness. Based on current findings, the Stockholm Metro is preparing a pilot of a fully automated link between the video analysis system and direct alerts to train drivers for behaviours judged to be especially high-risk, aiming to shorten response times and strengthen the preventive effect.

Taken together, AI-CCTV seems to be a practical and scalable addition to suicide prevention in rail environments, particularly where larger infrastructural solutions such as platform screen doors are difficult to implement.

All urban transit systems (metro, subway, underground) that do not have physical barrier systems that impede accessing the rails, have suicide deaths and attempts. Installation of barriers that open when the train is in station effectively eliminate suicides and accidents by being hit by a train or touching high voltage rails. Although newly constructed stations often have barriers, most existing stations do not. The costs to install barriers in older stations is high. Some interventions, such as elevated rails so trains can pass over a person, and public service posters advertising suicide prevention helplines can prevent some suicide injuries and fatalities.

Mishara and colleagues (2016) analysed video recordings of all people who attempted suicide in the Montreal metro over 2 years to identify their pre-attempt behaviours. They identified behaviours unique to people who are going to attempt suicide and combinations of behaviours associated with high ris. These behaviours that did not occur or rarely occurred in metro users who do not attempt suicide.

Since humans cannot continually watch videos from over 1000 cameras to identify people at risk, we explored using computerized skeletal-based action recognition with deep learning architectures. We developed a skeleton-based spatio-temporal action recognition model (SSTAR) to recognize rapidly in real-time video streams identified pre-suicidal behaviours. SSTAR was tested and perfected with several human motion datasets, including Montreal metro videos. SSTAR demonstrated high performance across datasets, with high recognition accuracy and computational efficiency. Ongoing testing in several Montreal metro stations shows promise for alerting controllers to immediately slow trains entering the station and sending personnel to offer help to people. After extensive pilot testing, if confirmed to be effective, SSTAR will be implemented in all stations and its effectiveness in preventing suicides will be evaluated over several years, while continuing to improve its accuracy, efficiency and user interfaces.