Understanding and detecting behaviours prior to a suicide attempt: A mixed-methods study

Abstract

Objective:

Prior research suggests there are observable behaviours preceding suicide attempts in public places. However, there are currently no ways to continually monitor such sites, limiting the potential to intervene. In this mixed-methods study, we examined the acceptability and feasibility of using an automated computer system to identify crisis behaviours.

Methods:

First, we conducted a large-scale acceptability survey to assess public perceptions on research using closed-circuit television and artificial intelligence for suicide prevention. Second, we identified crisis behaviours at a frequently used cliff location by manual structured analysis of closed-circuit television footage. Third, we configured a computer vision algorithm to identify crisis behaviours and evaluated its sensitivity and specificity using test footage.

Results:

Overall, attitudes were positive towards research using closed-circuit television and artificial intelligence for suicide prevention, including among those with lived experience. The second study revealed that there are identifiable behaviours, including repetitive pacing and an extended stay. Finally, the automated behaviour recognition algorithm was able to correctly identify 80% of acted crisis clips and correctly reject 90% of acted non-crisis clips.

Conclusion:

The results suggest that using computer vision to detect behaviours preceding suicide is feasible and well accepted by the community and may be a feasible method of initiating human contact during a crisis.

Introduction

Suicide ‘hotspots’ – frequently used locations that provide the means or privacy for suicide – can gain notoriety and thus attract more individuals. Approximately 30% of suicides occur in public locations (Owens et al., 2009; Too et al., 2016), and suicides at these locations can adversely affect bystanders and the local community (Pirkis et al., 2015).

One effort to prevent suicides at these locations has been to introduce closed-circuit television (CCTV) cameras to increase the possibility of intervention by a third party, such as emergency services (Pirkis et al., 2015). A recent review identified the use of CCTV in understanding and preventing suicide, including in public spaces (Onie et al., 2021). A key finding was that automated systems, to identify when an attempt is taking place, have been considered to address the need for continuous monitoring (see also Mishara et al., 2016). One such example is sending an alert when movement, other than a train, is detected on railway tracks (Mukherjee and Ghosh, 2017). However, even with these automated systems, often it is too late to intervene as an attempt may already be occurring at the time of detection. A new line of research seeks to identify behaviours preceding an attempt, thus allowing for earlier intervention. Thus far, two studies have investigated crisis behaviours on railways in the United Kingdom and Canada (Mackenzie et al., 2018; Mishara et al., 2016). The authors observed behaviours including pacing back-and-forth along the platform and to-and-from the platform edge, allowing trains go by, and switching between different stations and platforms. Thus far, crisis behaviours have only been studied at railways despite the wide range of settings used for suicide, and to date there is no automated way to detect these crisis behaviours. Computer vision algorithms using artificial intelligence (AI) is one possible approach to detect these crisis behaviours.

Any approach to automatically detect suicidal behaviours must be sensitive to, and respectful of, community attitudes to such an approach. Previous studies have assessed public acceptability of CCTV (Honess and Chairman, 1992) and AI in broader contexts (Lockey et al., 2020), with results showing people are generally favourable towards the use of either given sufficient governance and protections are in place, especially to protect privacy. However, to our knowledge, there are currently no data on the public acceptability of using CCTV or AI for suicide prevention.

To address these gaps, we conducted three studies examining automatic detection of behaviours preceding a suicide attempt. First, we conducted an acceptability study to understand the community’s attitudes towards this approach to suicide prevention. Second, we manually analysed CCTV footage from a frequently used Australian coastal cliffside location to identify behaviours preceding suicide in this new setting. Third, we developed AI algorithms to detect a selection of the identified behaviours and evaluated the algorithm’s ability to correctly identify an individual in crisis based on these behaviours. All three studies were approved by the University of New South Wales Human Research Ethics Committee (HC210256 and HC190663).

Methods

Study 1: acceptability study

In the first study, we assessed public attitudes towards the use of CCTV data for suicide prevention research purposes. Understanding these public perceptions is critical to ensure that this line of research is sensitive to potential concerns surrounding surveillance and use of public footage for suicide prevention.

We engaged a market research company to recruit a nationally representative sample of the Australian population. A sample size of 385 was required to ensure representation with a 5% margin of error. Participants were invited to complete an online questionnaire and were compensated for their participation through the company’s internal mechanisms.

Participants were first presented with a brief video explaining the purpose of the proposed research along with a definition of AI in this context. Participants were asked to respond to two statements: ‘It is ok to use the footage from CCTV cameras in public places for suicide prevention research’; and ‘It is ok to use machine learning or Artificial Intelligence [AI] to analyse the footage from publicly placed CCTV cameras for suicide prevention research’ on a 5-point Likert-type scale (from ‘strongly agree’ to ‘strongly disagree’). Given the nature of the study and by ethical requirement, participants were permitted to skip questions. We also conducted interviews with first responders and people with lived experience of suicide, the results of which will be reported elsewhere.

Study 2: identifying crisis behaviours

In the second study, we conducted a retrospective, structured analysis of existing CCTV recordings from a known frequently used location, to investigate the behaviours which may precede a suicide attempt at a coastal cliff location. This is a necessary step in understanding crisis behaviours at a broad range of locations other than the railways currently reported in the literature.

Setting

The frequently used location is a coastal cliff in Sydney, Australia. In 2010, the local council installed a series of suicide prevention interventions including CCTV cameras and updated fencing. The area studied is approximately 700 m in length with varying altitudes. Fencing runs along the entire cliffside.

Data

We retrospectively analysed 28 clips covering the period September 2019 to September 2020. This included 21 incident clips where individuals climbed the safety fence and either jumped or returned to safety following a police response, and 7 non-incident clips which contained routine behaviours. All clips were analysed using recordings from high-resolution infrared cameras to allow analysis of incidents that occurred in the evening. Clip duration ranged from 19 seconds to 64 minutes and 13 seconds, with a mean of 10 minutes and 44 seconds.

Analysis procedure

An initial coding framework was developed based on behaviours reported by Mackenzie et al. (2018) and iteratively updated after reviewing each video. Structured analysis of the CCTV data was conducted by three investigators – two suicide prevention researchers and a psychiatry registrar. Each investigator took free-text notes while watching each clip and recorded whether each behaviour in the framework was observed in the clip. Observed behaviours were discussed among the investigators until consensus was reached. In this paper, we only report behaviours that were observed in crisis clips but not observed in routine footage, i.e., if a behaviour was observed in both crisis and routine footage, it was deemed to be not associated with crisis and thus not reported.

Study 3: evaluation of an automated crisis detection system

In the third study, we explored whether a computer vision algorithm can detect a subset of crisis behaviours described above. In brief, an AI algorithm was evaluated, which is based on a modular computer vision pipeline that was developed specifically for this application. This pipeline detects the location of pedestrians and tracks their movement through the CCTV scene, detects their pose based on positions of their limbs, and recognises certain actions and behaviours based on the body movement. This specific pipeline was developed to detect behaviours preceding suicide. A range of existing deep learning-based computer vision algorithms were tested in various combinations, and a new skeleton-based action recognition model was proposed to find the optimal design for this setting and application. Full details of the algorithm selection are reported elsewhere (Li et al., 2022). The algorithm was trained using footage depicting a range of behaviours from this specific location.

Evaluation procedure

To evaluate the algorithm’s ability to detect behaviours associated with crisis, we developed a set of rules related to (a) behaviours, (b) duration of behaviours and (c) sequencing of behaviours. We focussed on clearly defined behaviours that involved greater movement or change in pose, as computer vision methods are more likely to detect these behaviours (e.g. crouching) rather than behaviours that encapsulate smaller gestures (e.g. psychomotor agitation). We also included behaviours that were made of smaller, clearly defined behaviours (e.g. extended stay which is standing or leaning for a period). For this study, we determined five rules:

1. Crouching in front of fence (detecting behaviours);

2. Leaning against the fence with head down (detecting behaviours);

3. Placing an object on the ground (detecting behaviours);

4. Standing or leaning in one location for an extended period (duration of behaviours);

5. Repeatedly transitioning between walking and standing/leaning (sequencing of behaviours).

To test the efficacy of this algorithm, we employed a method commonly used in computer science literature to test behaviour identification (Ludl et al., 2018, 2019), which is to simulate the behaviours of interest (acted by a team member on location) and determine whether the pipeline can detect behaviours of interest. Acted clips were used to allow the behaviours identified in study 2 to be recreated within a single camera scene, which reduces the complexity compared with identifying behaviours in different cameras with diverse angles and configurations, and tracking across multiple cameras as a person moves through the site.

Five clips were extracted which contained 12 individual instances of behaviours indicative of a crisis (pacing back and forth, visible agitation, leaning with head down, crouching in front of the fence and standing for extend periods of time), alongside 10 clips absent of crisis behaviours. In this evaluation, extended periods of time were classified as anything longer than 100 seconds, although this can be set to an arbitrary period. We confined analysis to a single, full colour camera to assess the algorithm at a location where fence climbing commonly occurs.

To assess algorithm performance, all recorded clips were passed through the pipeline, and the sensitivity and specificity were primarily computed at the level of individual video clips. Secondary analysis examined individual behaviours within the clips, including the number of correctly identified instances of individual behaviours along with a description of misclassifications.

Results

Study 1: acceptability study

A total of 1090 participants were recruited, 346 of whom identified as having a lived experience of suicide. As shown in Figure 1, 82.2% of respondents somewhat agreed or strongly agreed that it is acceptable to use footage from CCTV in public places for suicide prevention research. 6.6% of respondents somewhat or strongly disagreed with the use of footage. Agreement was slightly higher among those with lived experience (85.0% vs 81.9%).

Figure 1.

Respondents’ attitudes to the use of CCTV for suicide prevention research. Results are also disaggregated based on lived experience (LE) identification.

As shown in Figure 2, 72.0% of respondents somewhat or strongly agreed that the use of AI to analyse CCTV footage from public places for suicide prevention research was acceptable, and 9.6% somewhat or strongly disagreed. Agreement was similar in those with or without lived experience (73.3% vs 71.9%).

Figure 2.

Respondents’ attitudes to the use of AI to analyse CCTV footage for suicide prevention research. Results are also disaggregated based on lived experience (LE) identification.

Study 2: identifying crisis behaviours

Eleven crisis behaviours were identified, as described below:

• Visible psychomotor agitation. This is a group of behaviours including short erratic gestures of the hands, seemingly talking to oneself, and fidgeting. This occurred throughout an individual’s journey, but most often at the fence.

• Leaning with head down. Individuals would lean on the fence and look downwards (rather than out towards the sea).

• Crouching. Some individuals were observed crouching or squatting in front of the fence while holding onto the railing.

• Extended stay. We observed that individuals in crisis would often wait for extended periods of time, typically in the area near the fence. In one case, the individual walked to the fence, stood for 20–30 minutes, then sat down at a nearby bench for another 20–30 minutes and repeated back-and-forth multiple times (see also repetitive pacing, below). In another case, the individual would lean on the fence, and remain stationary for an extended period, in contrast with other individuals moving around them, often looking back to see whether others were still present and crossing over the fence once there were fewer persons nearby.

• Attempted climbing. This behaviour was observed when an individual attempted to climb or pull themselves over the fence – although it was unclear whether they were practising climbing or were unable to climb.

• Crying. In one case, an individual appeared to be crying – as inferred by hand gestures to their face.

• Repetitive pacing. Individuals in crisis would be observed pacing either along the fence line or from one location to another, with possible short breaks in between. For example, in one clip, the individual was sitting on a bench facing the fence. They would then stand up, walk behind the bench and crouch down while still facing the fence, stay for a few minutes and then return to sitting in the initial position. This cycle was repeated five times before climbing the fence.

• Location selection. In this behaviour, individuals would walk along the fence, stopping at multiple points, looking over the fence and often walking back along the same path. This behaviour is different from repetitive pacing, as location selection covers a much wider ground with a focus on the fence, whereas during repetitive pacing the individual has seemingly ‘chosen’ a location to climb and is pacing around the area.

• Intoxication. Some individuals were observed with an unsteady gait and walking path, suggesting intoxication. Some individuals were also carrying and drinking from a bottle or can, although due to the infrared footage it was not possible to determine if these were alcoholic beverages.

• Ignoring other individuals. One behaviour noted was that in incident clips, an individual would ignore other individuals where a response is typically observed. For example, typically as passers-by cross paths, a cursory glance between them is seen. However, for individuals in crisis, they would often keep their heads down.

• Placing belongings on the ground. Individuals were observed leaving behind their belongings. While placing belongings on a bench while sitting on it was a commonly observed routine behaviour, leaving possessions on the ground and walking away was observed in crisis incidents.

Study 3: evaluation of an automated crisis detection system

The algorithm correctly identified crisis behaviours in 4 of the 5 acted crisis clips (sensitivity = 0.80) and correctly rejected 9 of the 10 clips not containing acted crisis behaviours (specificity = 0.90). At the behaviours level, 8 of the 12 individual crisis behaviours were correctly identified and 4 were missed. Furthermore, three crisis behaviours were incorrectly identified within non-crisis clips.

The algorithm correctly identified cases of leaning with head down, crouching, staying at a location for a prolonged period as well as repetitive back-and-forth walking. However, in the one clip which was not correctly identified as containing crisis behaviours, the algorithm did not detect repetitive back-and-forth pacing as it associated the overall path with two people rather than a single individual. Furthermore, the algorithm failed to detect two instances of placing of an object on the ground and one instance of leaning with head down. For the one non-crisis clip that was misidentified as a crisis clip, there were two instances in which the individual was misidentified as crouching while sitting on the ground away from the fence and one instance where they were misidentified as leaning with head down while just leaning.

Discussion

Study 1: acceptability study

The survey showed broad public acceptance, across a nationally representative sample, towards the use of CCTV footage and AI for the purpose of suicide prevention research. To our knowledge, this is the first study that has assessed the acceptability of using CCTV and AI in suicide prevention research.

While these results are encouraging, it is noted that the survey and explanatory video were branded with the authors’ affiliation (Black Dog Institute) and it is unclear to what extent the responses may reflect the public’s level of trust in the Institute, and further work is required to understand whether the acceptability of the research depends on the organisations involved. Second, the survey questions asked specifically about acceptance in the research context, and further exploration is required to determine whether high levels of acceptability also apply to the implementation of the research findings.

The study continues with qualitative data collection through interviews with individuals bereaved by suicide and first responders at frequently used locations to further gain a nuanced understanding of how such research would be perceived by those that have been most impacted by people who have died by suicide or who respond to individuals in distress at suicide hotspots. Thus far, the responses suggest that there is good public acceptability for conducting research using CCTV and AI for suicide prevention.

Study 2: identifying crisis behaviours

In this study, we coded 21 video clips of incidents and 7 video clips with routine behaviour to ascertain whether there were behaviours that are indicative of crisis. Eleven behaviours associated with crisis were identified, suggesting that there are identifiable behaviours that precede a possible suicide attempt at the coastal cliff location.

It is also acknowledged that behaviours observed, in both incident and routine footage, may not be representative of all such behaviours. For example, while drinking from a can or bottle was observed in crisis footage, it was not observed in routine footage; however, drinking from a water bottle is not expected to be an indicator of distress. Individual behaviours may therefore have different sensitivities and specificities which can be used in combination to better understand the degree of risk – e.g., drinking with an unsteady gait may be distinct from drinking with a regular gait.

Previously, studies had only been conducted at metro stations where, due to differences in the setting, different behaviours may be observed (Mackenzie et al., 2018; Mishara et al., 2016). While both cliffs and metro stations require the crossing of a ‘boundary’ to access means, in metro stations the means (train) is not always available. Despite this, similar behaviours were observed across the two settings. For example, repetitive pacing has been reported along, and back and forth from, the yellow line at the edge of the platform (Mackenzie et al., 2018; Mishara et al., 2016). Individuals in crisis have been observed switching platforms, a form of location selection, as well as an extended stay in one place while individuals would let trains pass by. There were also location-specific behaviours that did not translate, e.g., attempted climbing as in the cliffside location was not observed in metro stations, as no barriers were present in the latter. Future studies should attempt to understand the underlying commonalities between these behaviours, potentially through their common cognitive processes through lived experience consultations.

Study 3: evaluation of an automated crisis detection system

The algorithm was able to correctly identify 80% of the crisis clips, correctly reject 90% of the non-crisis clips and correctly identify 8 of 12 individual instances of crisis behaviours (67%) across the clips. These results demonstrate that an automated computer algorithm can identify crisis behaviours at a hotspot setting.

It must be noted that these metrics were derived from a small number of video clips, and the evaluation was conducted using acted behaviour rather than real crisis clips. Further evaluation is needed to assess the performance using non-acted footage and across different settings. Furthermore, placing of an object on the ground and an instance of repetitive pacing were not detected, and more training footage may be necessary to refine the algorithm to detect these behaviours.

This study was conducted with colour cameras which require a sufficient degree of lighting (e.g. during daytime or at well-lit locations). Future work needs to consider the algorithm performance using thermal cameras to detect individuals in low light conditions at night. Nevertheless, some settings which have colour cameras are well-lit after dark, and this work may be directly useful in such settings.

This study inferred behaviours and actions based on the positions of individuals’ limbs; however, it is only possible to infer limb configurations when the person’s body is of a reasonable size within the frame. Therefore, an adaptive approach which continues to detect behaviours using location and trajectory information, even if the pose estimation is unreliable, may need to be considered.

General discussion

This study examined the potential for an automated computer vision system to identify behaviours preceding a suicide attempt. In the first study, we found positive public attitudes towards suicide prevention research using CCTV and AI. In the second study, we identified a set of 11 behaviours associated with crisis at a frequently used cliff location, with both distinct and similar behaviours from those previously identified at metro/underground railway settings. In the third study, we found promising results from an automated computer vision algorithm that can identify a subset of behaviours identified in study 2. Overall, the results give promise for a scalable intervention that is accepted by the public and may be able to identify individuals prior to a suicide attempt at frequently used locations. Furthermore, given the adaptability of the rules used in the automated system and the potential to acquire more data, the algorithm can be retrained and adapted to other settings.

While this is a promising finding, further research is needed prior to real-world implementation. Testing with naturalistic footage is needed to assess real-world sensitivity and specificity, as well as the timeliness of detecting behaviours. Furthermore, the performance across multiple cameras and views should be assessed. For example, locations may differ by the type of cameras installed (e.g. colour vs infrared) or their coverage (e.g. overlapping views vs sporadic placement). Further work is also required to understand how the algorithm performs in different conditions (e.g. time of day, and weather conditions).

Understanding the real-world performance will also inform the development of an appropriate intervention. The rate of false positive, false negatives, and the timeliness of true positive notifications will guide the intensity of any intervention, as there are potentially high, if asymmetric, costs associated with false positive and false negative misclassifications.

Regardless of the performance metrics achieved, such an implemented system would be inherently coupled to a human-centred response, such as alerting an operator to review the recent footage to confirm the identified behaviours, and to co-ordinate a response with an appropriate first responded. As such, this is a technology-supported approach for facilitating a human connection at the time of crisis. Overall, the current study gives promise that we may be able to detect and intervene earlier to interrupt suicide attempts at frequently used locations.