Pith. sign in

REVIEW 1 major objections 1 minor 24 references

An interpretable framework assesses suicide risk in metro stations from surveillance video by accumulating evidence through person tracking, activity recognition, platform segmentation, and trajectory heatmaps instead of inferring intent.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.3

2026-06-30 16:41 UTC pith:J7COEMFW

load-bearing objection The paper outlines a multi-component pipeline for suicide risk scoring from metro video but the 83.2% ROC-AUC rests on an undefined labeling process for the positive cases, so the number cannot be interpreted yet. the 1 major comments →

arxiv 2605.22904 v2 pith:J7COEMFW submitted 2026-05-21 cs.CV cs.AI

Suicide Risk Assessment from AI-powered Video Surveillance: An Interpretable Framework for Prevention in Metro Stations

Safwen Naimi , Wassim Bouachir , Guillaume-Alexandre Bilodeau , Brian Mishara This is my paper
classification cs.CV cs.AI
keywords suicide risk assessmentvideo surveillancemetro stationsperson trackingactivity recognitionsemantic segmentationrisk heatmap modelinginterpretable framework
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper formalizes suicide risk assessment as a distinct task in metro station surveillance video. It presents an interpretable pipeline that jointly processes person tracking, activity recognition, semantic segmentation of the platform, and trajectory-driven risk heatmap modeling to build risk scores from accumulated behavioral cues over time. The approach is tested on real data, reaching 83.2 percent ROC-AUC. A sympathetic reader would care because the method aims to support timely intervention in suicide prevention by handling the spatial context and temporal dynamics that isolated subtasks miss.

Core claim

The authors introduce the first interpretable framework for Suicide Risk Assessment in metro stations that assesses risk from accumulated evidence by incorporating person tracking, activity recognition, semantic segmentation of the platform, and trajectory-driven risk heatmap modeling, achieving 83.2 percent ROC-AUC on real surveillance data rather than focusing on isolated subtasks or attempting to infer intent directly.

What carries the argument

The interpretable pipeline that integrates person tracking, activity recognition, semantic segmentation of the platform, and trajectory-driven risk heatmap modeling to accumulate behavioral evidence over time and produce risk scores.

Load-bearing premise

The integrated pipeline of tracking, recognition, segmentation, and heatmap modeling can reliably assess suicide risk from video without direct intent inference, and the reported ROC-AUC indicates practical utility for prevention.

What would settle it

A controlled test on new metro surveillance footage in which the framework produces risk scores that do not correlate with actual recorded incidents or intervention outcomes at rates above random chance.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

1 major / 1 minor

Summary. The paper formalizes the task of Suicide Risk Assessment (SRA) from metro station surveillance videos as distinct from isolated subtasks or direct intent inference. It proposes an interpretable framework that accumulates evidence via person tracking, activity recognition, semantic segmentation of the platform, and trajectory-driven risk heatmap modeling, and reports benchmarking a complete operational pipeline that achieves 83.2% ROC-AUC on real surveillance data.

Significance. If the validation details were supplied and the metric shown to be reliable, the work would be significant as one of the first end-to-end interpretable pipelines for this socially important application in computer vision. It correctly emphasizes accumulated behavioral and contextual cues over direct intent modeling and opens directions for AI for social good. The absence of any dataset construction, labeling protocol, or baseline comparison currently prevents assessing whether the result advances the state of the art.

major comments (1)
  1. [Abstract and Evaluation] Abstract (and the evaluation section it summarizes): the central claim that the pipeline 'achieves 83.2% ROC-AUC on real surveillance data' is load-bearing for the contribution, yet the manuscript supplies no information on dataset size, how ground-truth positive and negative suicide-risk events are defined independently of the model (e.g., incident logs, expert annotation criteria, temporal windows), train/test splits, cross-validation, baselines, or error analysis. Without an independent labeling protocol the reported AUC cannot be interpreted as evidence that the accumulated-evidence formulation measures the intended construct.
minor comments (1)
  1. [Abstract] The abstract states the framework is 'the first' without referencing prior related work on behavior analysis or risk modeling in surveillance; a brief related-work paragraph would improve context.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback and for acknowledging the potential significance of formalizing suicide risk assessment as a distinct task. We address the major comment on evaluation details below.

read point-by-point responses

  1. Referee: [Abstract and Evaluation] Abstract (and the evaluation section it summarizes): the central claim that the pipeline 'achieves 83.2% ROC-AUC on real surveillance data' is load-bearing for the contribution, yet the manuscript supplies no information on dataset size, how ground-truth positive and negative suicide-risk events are defined independently of the model (e.g., incident logs, expert annotation criteria, temporal windows), train/test splits, cross-validation, baselines, or error analysis. Without an independent labeling protocol the reported AUC cannot be interpreted as evidence that the accumulated-evidence formulation measures the intended construct.

    Authors: We agree that the manuscript currently provides insufficient details on the dataset and evaluation protocol. In the revised manuscript we will add a dedicated subsection describing: the size and characteristics of the real surveillance dataset; the independent ground-truth labeling protocol (including use of incident logs, expert annotation criteria, and temporal windows for positive/negative events); the train/test splits and any cross-validation procedure; relevant baselines; and error analysis. These additions will allow readers to properly interpret the 83.2% ROC-AUC in the context of the accumulated-evidence formulation. revision: yes

Circularity Check

0 steps flagged

No derivation chain or equations; framework is descriptive pipeline without self-referential reduction

full rationale

The paper describes an operational pipeline combining person tracking, activity recognition, semantic segmentation, and trajectory risk heatmaps to produce a risk score, benchmarked at 83.2% ROC-AUC. No equations, formal derivations, fitted parameters renamed as predictions, or self-citations appear in the provided text. The central claim is an engineering integration validated on real data rather than a mathematical result that reduces to its inputs by construction. Absence of any load-bearing derivation steps means the analysis is self-contained with no circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No specific free parameters, axioms, or invented entities are identifiable from the abstract alone.

pith-pipeline@v0.9.1-grok · 5744 in / 1157 out tokens · 49594 ms · 2026-06-30T16:41:34.890151+00:00 · methodology

0 comments
read the original abstract

Understanding and monitoring human behavior in metro stations play an important role in supporting suicide prevention efforts, where early identification of high-risk situations can enable timely intervention. This requires assessing suicide risk from a surveillance video by jointly reasoning about the behavior of each passenger, his/her spatial context, and temporal dynamics. However, this assessment using videos captured by surveillance cameras is challenging, as it demands accurate perception of human motion, understanding of platform geometry, and aggregation of heterogeneous behavioral cues over time. In this work, we formalize the task of Suicide Risk Assessment (SRA) in metro stations and introduce the first interpretable framework that addresses this challenge. Unlike approaches that focus on isolated subtasks or attempt to infer intent directly, our formulation assesses suicide risk from accumulated evidence by incorporating person tracking, activity recognition, semantic segmentation of the platform, and trajectory-driven risk heatmap modeling. By formalizing SRA as a distinct task and benchmarking a complete operational pipeline achieving 83.2% ROC-AUC on real surveillance data, this work highlights the complexity of suicide risk assessment and opens new directions for research on interpretable AI systems for social good.

Figures

Figures reproduced from arXiv: 2605.22904 by Brian Mishara, Guillaume-Alexandre Bilodeau, Safwen Naimi, Wassim Bouachir.

Figure 1
Figure 1. Figure 1: Architecture overview of our proposed SRA-Framework. It takes a video as input and outputs the complete platform state and a suicide risk assessment for each person. different areas of the platform may warrant preventive atten￾tion. It is influenced not only by what actions are performed, but also by where they occur and how long they persist. This formulation is consistent with operational safety practice… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the Platform Semantics Modeling Process: (a–b) Boundary and Offset Lines Estimation, and (c) Final Zone Partitioning. Wt, providing high-level action cues for subsequent risk as￾sessment. In our SRA-Framework, we employ SSTAR pre￾trained on the ARMM dataset [Naimi et al., 2025], a skeleton￾based spatio-temporal action recognition model specifically designed for surveillance scenarios. In our ca… view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of Individual At-Risk Heatmaps and their ag [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Feature importance of the eight risk indicators used for [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative Results. Output predictions of two frames from surveillance video streams. high risk score (R0=0.98), driven by a combination of a pro￾longed crossing of the yellow line (11 seconds) and elevated position risk score along the position risk heatmap. Our sys￾tem also successfully assigned lower suicide risk scores for control individuals (R2=0.12 and R3=0.14). In the second video illustrated in F… view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

24 extracted references · 2 canonical work pages · 1 internal anchor

  1. [1]

    Rushi Kumar

    [Anjali and Kumar, 2024] T Anjali and B. Rushi Kumar. Spatial analysis of multivariate factors influencing suicide hotspots in urban tamil nadu.Journal of Affective Disor- ders Reports,

    2024

  2. [2]

    Humans in the loop, lives on the line: Ai in high-risk decision making

    [Balamuruganet al., 2025 ] Merlin Balamurugan, Kumar Shanmugasamy, and Saranya Balaguru. Humans in the loop, lives on the line: Ai in high-risk decision making. European Journal of Computer Science and Information Technology,

    2025

  3. [3]

    Intelligent video surveillance for real-time detection of suicide attempts.Pattern Recog- nit

    [Bouachiret al., 2018 ] Wassim Bouachir, Rafik Gouiaa, Bo Li, and Rita Noumeir. Intelligent video surveillance for real-time detection of suicide attempts.Pattern Recog- nit. Lett., 110:1–7,

    2018

  4. [4]

    Using deep learning model integration to build a smart railway traffic safety monitoring system.Scientific Re- ports, 15,

    [Changet al., 2025 ] Chin-Chieh Chang, Kai-Hsiang Huang, Tsz-Kin Lau, Chung-Fah Huang, and Chun-Hsiung Wang. Using deep learning model integration to build a smart railway traffic safety monitoring system.Scientific Re- ports, 15,

    2025

  5. [5]

    Xgboost: A scalable tree boosting system.Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining,

    [Chen and Guestrin, 2016] Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system.Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining,

    2016

  6. [6]

    The role of new technologies to prevent suicide in adolescence: A systematic review of the literature.Medicina, 57,

    [Forteet al., 2021 ] Alberto Forte, Giuseppe Sarli, Lorenzo Polidori, David Lester, and Maurizio Pompili. The role of new technologies to prevent suicide in adolescence: A systematic review of the literature.Medicina, 57,

    2021

  7. [7]

    Osgood, Benjamin Yeoh, Eva Graham, and Heather Orpana

    [Gari´epyet al., 2024 ] Genevi`eve Gari ´epy, Rifat Zahan, Nathaniel D. Osgood, Benjamin Yeoh, Eva Graham, and Heather Orpana. Dynamic simulation models of suicide and suicide-related behaviors: Systematic review.JMIR Public Health and Surveillance, 10,

    2024

  8. [8]

    Boring, and Anahita Khojandi

    [Gurselet al., 2024 ] Ezgi Gursel, Mahboubeh Madadi, Jamie Baalis Coble, Vivek Agarwal, Vaibhav Yadav, Ronald L. Boring, and Anahita Khojandi. The role of ai in detecting and mitigating human errors in safety-critical industries: A review.Reliab. Eng. Syst. Saf., 256:110682,

    2024

  9. [9]

    Girshick

    [Heet al., 2017 ] Kaiming He, Georgia Gkioxari, Piotr Doll´ar, and Ross B. Girshick. Mask r-cnn

    2017

  10. [10]

    Hightower and Mor- gan J

    [Hightower and Grant, 2024] Heath H. Hightower and Mor- gan J. Grant. Proposing an individual-in-contexts model for reimagining suicide screening, assessment, and inter- vention in black communities.Journal of Human Services,

    2024

  11. [11]

    Ultralytics yolov8 https://github.com/ultralytics/ultralytics

    [Jocheret al., 2023 ] Glenn Jocher, Ayush Chaura- sia, and Jing Qiu. Ultralytics yolov8 https://github.com/ultralytics/ultralytics

    2023

  12. [12]

    Laliberte, Brandon Roth, Beau Edwards, and Jason I

    [Laliberteet al., 2021 ] Avery Z. Laliberte, Brandon Roth, Beau Edwards, and Jason I. Chen. Suicide screening and risk assessment in the emergency department: Case review of a suicide attempt survivor.Journal of emergency nurs- ing,

    2021

  13. [13]

    Behaviours preceding suicides at railway and underground locations: a multimethodological quali- tative approach.BMJ Open, 8,

    [Mackenzieet al., 2018 ] Jay-Marie Mackenzie, Jo Borrill, Emily Hawkins, Bob Fields, Ian Kruger, Ian Noonan, and Lisa Marzano. Behaviours preceding suicides at railway and underground locations: a multimethodological quali- tative approach.BMJ Open, 8,

    2018

  14. [14]

    Mishara, C ´ecile Bardon, and Serge Dupont

    [Misharaet al., 2016 ] Brian L. Mishara, C ´ecile Bardon, and Serge Dupont. Can cctv identify people in public transit stations who are at risk of attempting suicide? an analysis of cctv video recordings of attempters and a comparative investigation.BMC Public Health, 16,

    2016

  15. [15]

    [Naimiet al., 2025 ] Safwen Naimi, Wassim Bouachir, Guillaume-Alexandre Bilodeau, and Brian L. Mishara. Sstar: Skeleton-based spatio-temporal action recognition for intelligent video surveillance and suicide prevention in metro stations.2025 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW), pages 765–775,

    2025

  16. [16]

    Koudys, and Anthony C

    [Parsapooret al., 2023 ] Mahboobeh Parsapoor, Jacob W. Koudys, and Anthony C. Ruocco. Suicide risk detec- tion using artificial intelligence: the promise of creating a benchmark dataset for research on the detection of suicide risk.Frontiers in Psychiatry, 14,

    2023

  17. [17]

    [R˚adbo and Andersson, 2012] Helena R ˚adbo and Ragnar Andersson. Patterns of suicide and other trespassing fatal- ities on state-owned railways in greater stockholm; impli- cations for prevention.International Journal of Environ- mental Research and Public Health, 9:772 – 780,

    2012

  18. [18]

    Information and com- munication technology use in suicide prevention: Scoping review.Journal of Medical Internet Research, 23,

    [Rassyet al., 2021 ] Jessica Rassy, C ´ecile Bardon, Luc Dar- gis, Louis-Philippe Cˆot´e, Laurent Corth´esy-Blondin, Carl- Maria M ¨orch, and R ´eal Labelle. Information and com- munication technology use in suicide prevention: Scoping review.Journal of Medical Internet Research, 23,

    2021

  19. [19]

    YOLOv3: An Incremental Improvement

    [Redmon and Farhadi, 2018] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement.ArXiv, abs/1804.02767,

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  20. [20]

    Main characteristics of train-pedestrian fatalities on finnish rail- roads.Accident; analysis and prevention, 45:61–6,

    [Silla and Luoma, 2012] Anne Silla and Juha Luoma. Main characteristics of train-pedestrian fatalities on finnish rail- roads.Accident; analysis and prevention, 45:61–6,

    2012

  21. [21]

    Evaluating xgboost for balanced and im- balanced data: Application to fraud detection,

    [Velardeet al., 2023 ] Gissel Velarde, Anindya Sudhir, San- jay Deshmane, Anuj Deshmunkh, Khushboo Sharma, and Vaibhav Joshi. Evaluating xgboost for balanced and im- balanced data: Application to fraud detection,

    2023

  22. [22]

    Deep high-resolution representation learn- ing for visual recognition.IEEE Transactions on Pattern Analysis and Machine Intelligence, 43:3349–3364,

    [Wanget al., 2019 ] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, Wenyu Liu, and Bin Xiao. Deep high-resolution representation learn- ing for visual recognition.IEEE Transactions on Pattern Analysis and Machine Intelligence, 43:3349–3364,

    2019

  23. [23]

    Bytetrack: Multi-object tracking by associating every detection box.arXiv preprint arXiv:2110.06864, 2021

    [Zhanget al., 2021 ] Yifu Zhang, Pei Sun, Yi Jiang, Dong- dong Yu, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xing- gang Wang. Bytetrack: Multi-object tracking by associat- ing every detection box.ArXiv, abs/2110.06864,

    work page arXiv 2021

  24. [24]

    The research on recognition of passenger abnormal behavior in urban rail transit stations based on conditional generative adversarial networks.Alexandria Engineering Journal, 2025

    [Zuoet al., 2025 ] Jing Zuo, Zexin Li, Ming He, and Qiaoli Yang. The research on recognition of passenger abnormal behavior in urban rail transit stations based on conditional generative adversarial networks.Alexandria Engineering Journal, 2025

    2025