Tattoo-Based Emergency Patient Identification: From Internal Research to Public Deployment

Tattoo-Based Emergency Patient Identification: From Internal Research to Public Deployment

Abstract #

We describe the public release of a tattoo-based emergency patient identification framework whose conceptual roots trace to OTG-bot — a UNDP-grant-winning civic technology project developed in 2021 for Ukrainian territorial communities. That project received a $10,000 USD grant from the United Nations Development Programme at the Hack Locals 2.0 hackathon and included an automated missing-person lookup feature in a Telegram-based community service bot. Its pilot city, Popasna, was destroyed in 2022. The infrastructure is gone; the research line is not. The present work formalises the identification methodology as a computer vision module within the ScanLab medical AI platform — automated feature extraction from tattoo photographs, quantitative similarity matching, and a GDPR-compliant voluntary registry. The emergence of government-backed tattoo identification services in conflict zones in early 2026, addressing over 90,000 missing persons cases in Ukraine alone, confirms that the problem identified in 2021 remains unsolved at scale and that the technical approach refined since is viable.

1. Background and Motivation #

In 2024, while building ScanLab’s diagnostic platform, we identified a recurring problem in emergency medicine datasets: a non-trivial fraction of admitted patients could not be reliably identified from standard intake procedures. Peth (2003) documented this at 1–2% of emergency department volume; in conflict-adjacent or mass-casualty settings the figure is substantially higher. Fingerprint infrastructure requires enrolled databases. DNA identification requires laboratory processing time measured in days. Dental records require prior documentation.

Tattoos require only a camera and a searchable record. They are permanent, visible, and distinctive. The forensic literature had established their utility in individual casework — the FBI’s Next Generation Identification system includes a tattoo module since 2015 (Maltoni et al., 2022) — but no open clinical framework existed for integrating tattoo-based identification into a general-purpose medical AI platform. We built one. Its origin, however, is not 2024 — it is 2021.

In 2021, our team competed in Hack Locals 2.0, co-organised by Ukraine’s Ministry of Digital Transformation and the United Nations Development Programme (UNDP). We built OTG-bot: a Telegram-based community service router that allowed residents to submit requests and had them automatically dispatched to the appropriate city department. One module handled missing or absent persons, routing reports to the relevant municipal service. The hackathon planned three winning grants. The jury awarded four. OTG-bot received a $10,000 USD UNDP grant as the additional winner (UNDP Ukraine, 2021).

The pilot deployment was in Popasna, Luhansk Oblast. In spring 2022, Popasna was entirely destroyed in combat operations. The municipal infrastructure OTG-bot connected to no longer exists. The grant-funded codebase survived as open-source research on this platform. The missing-person identification module — structured intake, routing, searchable records — is the direct conceptual ancestor of the system described here.

2. Prior Art and Related Systems #

2.1 Forensic Systems #

The academic literature on automated tattoo recognition predates our work by a decade. Cao and Jain (2019) demonstrated deep learning retrieval on the NIST Tatt-C benchmark. Han et al. (2018) applied CNNs to tattoo detection and attribute classification. These systems assume large labelled datasets and GPU inference — appropriate for law enforcement data centres, not hospital-deployed diagnostic platforms on modest server hardware.

Our 2024 design chose a non-deep-learning approach: classical image analysis — k-means colour clustering, edge density, texture statistics — producing a compact feature vector for CPU-based similarity matching. The trade-off is precision: the system does not achieve forensic-grade certainty. It achieves clinical-grade utility, surfacing candidate matches for human verification. This constraint is explicit in the design and mirrors acknowledged limitations in comparable government deployments: final identification is always a human judgement.

2.2 Government Deployments #

In early 2026, government institutions began deploying manual tattoo identification registries at national scale in Eastern Europe. These systems confirm that the operational model is sound — independent parties arrived at similar schema decisions (sequential record numbering, category/body-location taxonomy, multi-tattoo linking per individual). They also illustrate the ceiling of manual approaches: they depend on sketch artists, are not searchable by photograph, and produce no similarity score.

The critical differentiators of our system: automated feature extraction from photographs (no sketch artist required), image-to-image similarity matching with a quantitative confidence score, REST API access for clinical integration, images not retained after analysis, and a GDPR Article 9-compliant consent architecture. These were specified in 2024 and have not required revision. Contact: contact@stabilarity.com.

3. Research Provenance: From OTG-bot to ScanLab #

The intellectual lineage of this work is worth documenting precisely, because it illustrates how civic technology research survives the destruction of its deployment context. OTG-bot was not a research prototype — it was a production deployment, used by real residents of Popasna to contact municipal services. The missing-person module processed structured intake: name, last known location, physical description, and — where available — identifying features including tattoos and distinguishing marks. That last field was free-text in 2021. The question we asked in 2024: what would it take to make that field searchable by image?

The answer required three components: a feature extraction pipeline representing a tattoo photograph as a compact comparable vector; a registry schema storing voluntary records with consent and access controls; and an API layer integrating with clinical workflows. All three are described in Section 4. The OTG-bot codebase, archived as open research on this platform, contains the original intake schema from which the registry design was derived.

The UNDP grant was awarded under the criterion of civic security innovation. The jury’s decision to create a fourth grant outside the planned three was documented in the UNDP Ukraine press release (UNDP Ukraine, 2021). We cite this not as credential but as evidence of external review: the missing-person identification use case was evaluated by an independent jury in 2021 and found to represent a meaningful contribution to community security. This paper is, in part, a technical account of what happened to that contribution after the pilot city ceased to exist.

4. Technical Architecture #

4.1 Feature Extraction Pipeline #

The pipeline processes input images through five stages. First, normalisation to a standard 256×256 RGB representation with contrast enhancement. Second, dominant colour profiles via iterative k-means clustering over the pixel colour space, yielding five representative colours with proportional coverage. Third, edge density computed by Laplacian filter — edge pixels as proportion of total image area, a proxy for visual complexity and marking style. Fourth, texture complexity as the standard deviation of grayscale pixel values. Fifth, style classification derived from edge density: above 0.35 = linework; below 0.15 = shading or watercolour; intermediate = mixed.

These five feature types are serialised into a compact JSON representation and hashed to produce the visual fingerprint — a SHA-256 hash of the normalised feature vector providing a deterministic identifier for exact duplicate detection, while the feature vector itself enables similarity matching.

4.2 Registry and Matching #

The registry stores records with: sketch_number, category, type, body_location, body_part, feature_vector, fingerprint, contact_reference, and linked_sketches for persons with multiple registered markings. Similarity matching computes cosine distance between the query feature vector and all registered vectors, returning candidates above a configurable threshold with similarity scores. The system presents these as candidates for human review, not as identifications.

4.3 API Endpoints #

ScanLab v2.2.0 exposes three new endpoints: POST /v2/tattoo/analyze — image upload returning extracted feature profile and registry matches above threshold; GET /v2/tattoo/registry/browse — filtered registry metadata access without image data; GET /v2/tattoo/registry/count — aggregate statistics. A fourth endpoint, POST /v2/tattoo/register, is restricted to authorised administrators handling voluntary registration with explicit consent documentation. Images are not persisted after extraction.

5. Legal and Ethical Framework #

Tattoo photographs constitute biometric data under GDPR Article 4(14) when used for identification. Processing is subject to the Article 9 prohibition on special category data, with the applicable exception being Article 9(2)(c): processing necessary to protect vital interests where the data subject is physically or legally incapable of giving consent — the exact scenario of an unconscious emergency patient. For voluntary registry submissions, Article 9(2)(a) explicit consent applies.

All records in the registry are voluntary and consent-documented. Images are not retained. Access to the registration endpoint is administratively controlled. These safeguards were designed in 2024 and remain unchanged in the public release.

6. Validation and Limitations #

Internal validation over 2024–2025 on a synthetic test set of 47 image pairs confirmed the expected precision-recall trade-off: at similarity threshold 0.75, the system achieves 91% precision and 68% recall. High precision is the design target — false positives (wrong identification) are a more serious error than false negatives (missed match). Recall improves as the registry grows.

The principal limitation is scope: classical feature extraction performs well on distinctive, high-contrast markings and poorly on small, faded, or heavily stylised tattoos. A CNN-based retrieval backbone would improve forensic-grade performance — documented as future work. A second limitation is registry sparsity: with three demonstration records in the public registry, clinical value is currently illustrative.

7. Discussion and Conclusions #

OTG-bot was built for Popasna. Popasna no longer exists. The problem it addressed — routing missing-person reports to the people who could act on them — exists at a scale in 2026 that could not have been anticipated in 2021. The 90,000 missing persons figure is not an abstraction; it is the direct continuation of the problem that earned a UNDP jury grant that year.

We release this system publicly now because the combination of original proof of civic utility, two years of technical refinement, and independent government deployments constitutes sufficient evidence that the approach is sound and the need is real. The ScanLab implementation is open to clinical and municipal partners for integration. We hope to renew OTG-bot — rebuilt for the communities that need it now. Registry participation and integration enquiries: contact@stabilarity.com.

flowchart TD
    A[Tattoo Photograph Input
REST API Upload] --> B[Feature Extraction Pipeline]
    B --> C1[Dominant Colour
Analysis]
    B --> C2[Edge Density
Detection]
    B --> C3[Texture Complexity
Measurement]
    B --> C4[Style Classification
Traditional/Fine-line/etc.]
    C1 & C2 & C3 & C4 --> D[Feature Vector
Composite Score]
    D --> E[GDPR-Compliant Registry
Matching Engine]
    E --> F[Ranked Similarity Results
Quantitative Score 0–1]
    F --> G[REST API Response
Image NOT retained]

flowchart LR
    A[OTG-bot 2021
Manual Intake Form] -->|feature fields| B[Identifying Features Schema]
    B --> C[Tattoo / Scar / Mark fields]
    C -->|2024 automation| D[ScanLab CV Module]
    D -->|2026 validation| E[Government Deployment
90,000+ missing persons]
    E -->|confirms schema design| F[Sequential ID Numbers]
    E -->|confirms schema design| G[Body-Location Taxonomy]
    E -->|confirms schema design| H[Linked Records per Individual]
    D --> F & G & H

sequenceDiagram
    participant Operator
    participant API as ScanLab API
    participant CV as CV Engine
    participant Registry as GDPR Registry

    Operator->>API: POST /identify {photo, metadata}
    API->>CV: Extract feature vector
    CV->>CV: Colour + Edge + Texture + Style analysis
    CV-->>API: Feature vector [256-dim]
    API->>Registry: Query similarity (cosine distance)
    Registry-->>API: Top-N matches with scores
    API-->>Operator: {matches, scores, record_links}
    Note over API: Image discarded — not retained