AI-assisted post-mortem classification of vehicle trauma in free-roaming cats using transformer-based large language models

Abstract

Veterinary forensic science is a specialized discipline that investigates non-natural animal deaths and injuries. Determining the cause of death in free-roaming cats often presents challenges due to underdeveloped pattern recognition methods and the intricate interplay of events and contributing factors. The integration of natural language processing into forensic investigations may offer potential for the computerized automation of document analysis, with the goal of supporting investigational decision-making. The objective of this study was to develop an AI framework leveraging a transformer-based large language model (LLM) to classify veterinary forensic autopsy reports derived from 271 feline autopsies and to identify the causes of death. We first identified discriminative lexical patterns between vehicular trauma (VT; n = 111) and non-VT cases (n = 160) using Term Frequency-Inverse Document Frequency analysis. The best-performing model, based on Clinical-Longformer, achieved an accuracy of 0.945, precision of 0.885, recall of 1.000, F1 score of 0.939, and AUC-ROC of 0.950. Model interpretability using Integrated Gradients revealed that VT cases were characterized by terms associated with traumatic physical injuries, whereas non-VT cases exhibited a broader range of diagnostic descriptions. These findings demonstrate the feasibility of applying LLMs to veterinary forensic pathological texts and underscore the importance of domain-specific token interpretation for model explainability.

Introduction

Free-roaming cats (FRCs) are at high risk of exposure to traffic-related accidents, such as hit-by-car, as well as human-inflicted violence (Bruce et al. 2019; Tan et al. 2020). Both scenarios often require forensic investigations to be performed, particularly in cases where the public or law enforcement suspect animal abuse and the possibility of future legal proceedings (Loss et al. 2013; Watson 2018). The field of veterinary forensic science is gaining prominence amid growing concerns for animal welfare, increasing societal awareness, and a rising demand for integrated medical expertise (Cooper and Cooper 2008; Ottinger et al. 2014; Doukas 2022; Aradilla et al. 2025). For example, a 2016 survey of board-certified veterinary pathologists felt that their training has not adequately prepared them to handle forensic cases (McEwen and McDonough 2016).

Veterinary forensic science encompasses a multi-disciplinary approach, drawing on pathology, entomology, analytic chemistry, toxicology, and genetics (Bagsby et al. 2024). Veterinary forensic pathology is a specialised domain that plays a vital role in the professional examination of animal deaths and injuries, especially in cases involving trauma, poisoning, disease, or other life-threatening events (Brownlie and Munro 2016; Stern and Sula 2018). When investigating the unexplained deaths of FRCs, it is essential to distinguish between accidental causes and intentional harm (Kihara et al. 2024; Gilbert et al. 2026). However, differentiating between injuries caused by vehicular trauma (VT) and those resulting from human-inflicted violence and in some instances, predation can be particularly challenging (Intarapanich et al. 2016; Gilbert et al. 2026). This difficulty arises from the overlapping characteristics of blunt force trauma associated with motor vehicle collisions and physical assaults by humans (O'Donovan et al. 2022). A significant compounding challenge is the frequent lack of critical information from the scene, such as physical evidence, eyewitness accounts, or surveillance video, which are often vital for accurate forensic interpretation (Gilbert et al. 2026). In some cases, the complexity increases when both types of trauma are present; for example, when an animal is physically abused and subsequently struck by a vehicle, or vice versa (Intarapanich et al. 2016; O'Donovan et al. 2022). One author (AWS) has investigated several cases where the animals were intentionally struck by a vehicle, and further investigations revealed the intentional nature of the incidents.

Forensic reports are primarily text-based narrative documents summarising gross findings from the autopsy, supplemented by radiographic imaging, histopathological assessments, and ancillary analytic tests, when necessary (Bolliger et al. 2008; Schweitzer et al. 2014; Dedouit et al. 2025; Golomingi et al. 2025). The unstructured nature of forensic reports poses challenges for large-scale data analysis, retrospective review, on-demand document retrieval, statistical evaluation, and the identification of key forensic features in specific cases (Golomingi et al. 2025; Skipanes et al. 2025). In the digital era, scientific documentation is increasingly being transformed into computation-friendly formats, revolutionising many scientific domains (Amiraslani and Dragovich 2022; Gordon et al. 2022). Within forensic pathology, the standardisation of report formats and consistent documentation practices offers significant benefits (Ong and Milne 2009; de Boer et al. 2022; Van Den Bogaert et al. 2025). These include improved quality control, enhanced global collaboration, and better outcomes for stakeholders (Meilia et al. 2021; O'Keefe et al. 2025). Standardisation also helps mitigate potential conflicts, strengthens legal protections, and reduces the risk of harm to both human and animal victims (Meilia et al. 2021). However, narrative text remains difficult to parse automatically, not only in forensic science but also across medicine, due to semantic ambiguity and lack of fielded data, which hinders interoperability with computational models (Awaysheh et al. 2018).

Artificial intelligence (AI) technology is transforming virtually every discipline and aspect of society through digital innovation. One area of significant advancement is AI-driven text analysis, which has progressed rapidly through natural language processing (NLP), particularly with transformer-based large language models (LLMs) such as BERT and Longformer. These models have demonstrated strong potential in extracting meaningful patterns from clinical texts in human medicine (Devlin et al. 2019; Lyu et al. 2022; Li et al. 2023). In this study, we applied NLP techniques to analyse veterinary forensic autopsy reports and employed Clinical-Longformer, a domain-adapted transformer model optimised for medical text. To improve the interpretability of model predictions within sensitive forensic contexts, we incorporated explainable AI methods, specifically Integrated Gradients.

Here, we demonstrate a novel AI-driven workflow capable of distinguishing VT and non-VT fatalities in FRCs, providing a foundation for enhanced forensic applications and broader adoption of AI in veterinary forensic investigations. Importantly, our model was designed to evaluate the potential of AI models as a decision‑support tool for veterinary forensic pathologists, helping to highlight possibly overlooked features in narrative reports. This study also offers insights into how such as model might function as a quality control mechanism, assist in triaging cases for further investigation, and enable retrospective classification to improve case management and support research efforts.

Methods

Case selection

Deceased FRCs were submitted to the ‘A Cat Has No Name’ (ACHNN) programme at the University of Florida (UF) College of Veterinary Medicine, United States, from August 2020 to December 2024. Briefly, the ACHNN programme is a free service dedicated to the postmortem examination of FRCs and all forensic autopsies were performed by a board-certified veterinary pathologist. The forensic autopsy reports of cats were analysed to determine the cause of death, which was classified as either hit-by-car (i.e. VT) or non-VT. VT cases included animals confirmed as hit-by-car by the submitter (i.e. traumatic event was observed) and animals with autopsy findings suggestive of hit-by-car, based on our standard diagnostic criteria. These criteria were primarily derived from gross findings, including characteristic trauma patterns such as frayed claws, fractures consistent with vehicular impact, and soft tissue injuries. When available, radiographic and histologic analyses were reviewed to support the diagnosis, but classification did not rely exclusively on these ancillary tests. Non-VT cases encompassed a wide range of injuries and natural diseases, including bite and fall injuries, bacterial infections, and poisoning. Each report included the following fields: Submitter Name, Case Number, Submitter Agency, Patient Name, Gross Findings, Gross Diagnoses, and Comments. For rigorous modelling, we also used autopsy cases of deceased FRCs submitted to the Animal Disease Diagnostic Division of the Animal and Plant Quarantine Agency (APQA) in the Republic of Korea from 2019 to 2025. The original reports, written in Korean, were translated into English and reformatted to align with the structure of UF’s ACHNN reports. This study did not require approval from the Institutional Animal Care and Use Committee as all samples were collected postmortem and none of the animals were euthanized for purposes related to this study.

Data preparation and pre-processing

For this study, we extracted the Comments section to identify distinctive keywords for each class, and utilised the Gross Diagnoses section as input for cause-of-death classification (i.e. VT vs non-VT). This approach generated two datasets: (1) annotated necropsy comments (final_all_comment.xlsx) and (2) gross diagnoses (diagnoses.xlsx). Each entry was labelled as either VT-related (Label 1) or non-VT (Label 0). Texts were lemmatised and tokenized using Natural Language Toolkit (NLTK), and stopwords were removed.

The dataset comprises 271 forensic investigation reports, including 111 cases categorised as VT cases (74 from the ACHNN programme and 37 from APQA) and 160 cases as non-VT cases (88 from the ACHNN programme and 72 from APQA). The cause of death was labelled as 1 for cases confirmed as VT, and 0 for non-VT causes. Following the preprocessing approach described in the previous study (Li et al. 2023), the Comments text was lowercased and concatenated into a single paragraph without line breaks or additional whitespace. The Gross Diagnoses text was used as-is without additional preprocessing. The dataset was split into training, validation, and test sets in a 60:20:20 ratio.

Term Frequency-Inverse Document Frequency (TF-IDF) analysis

To identify discriminative terms between the two groups, TF-IDF analysis was applied to both the Comments and Gross Diagnoses datasets. The preprocessing steps include lowercasing, word tokenization, alphabet filtering, lemmatization, and removal of English stop words. TF-IDF scores were computed based on unigrams. For each term, the mean TF-IDF score across all documents was calculated. The top 30 words with the highest average TF-IDF scores were then visualised as word clouds separately for each group.

Tokenizer and model initialisation

We employed the Clinical-Longformer model (yikuan8/Clinical-Longformer) for binary classification. The Gross Diagnoses texts were tokenized using the corresponding tokenizer with a maximum input length of 512 tokens, applying truncation and padding as needed. Tokenization was performed separately for the training, validation, and test sets. To ensure compatibility with the model input requirements, we computed the average and maximum token lengths for each split. The average token lengths were approximately 280 for training, 179 for validation, and 310 for testing, with no instance exceeding the 512-token limit.

Hyperparameter optimisation

To optimise model performance, we conducted hyperparameter tuning using the Optuna framework integrated with the Hugging Face Trainer API. A total of 81 hyperparameter combinations were evaluated, with a categorical search space defined for learning rate ([1e-5, 3e-5, 5e-5]), batch size ([8, 16, 32]), number of training epochs ([5, 10, 20]), and weight decay ([0.0, 0.001, 0.01]). The optimisation objective was the weighted F1-score on the validation set. For each trial, the model was trained and evaluated on the predefined training and validation splits, which internally employs the AdamW optimiser and the BCEWithLogitsLoss function by default for binary classification. Training arguments were kept constant across all trials, including a fixed random seed (42) to ensure reproducibility. The decision threshold was optimised using F1-score maximisation and Youden’s J statistic. The best-performing hyperparameter configuration was recorded for subsequent training and analysis.

Training and validation procedure

We fine-tuned the Clinical-Longformer model for binary classification of cause of death using the Gross Diagnoses text (Li et al. 2023). The model was trained with a batch size of 8, learning rate of 2e-5, weight decay of 0.01, and up to 10 training epochs. Early stopping with a patience of 3 was applied to prevent overfitting. Model performance on the training and validation sets was evaluated using accuracy as the primary metric. The best model was selected based on the lowest validation loss, and a fixed random seed of 42 was used to ensure reproducibility.

Model evaluation

Model performance was evaluated on the hold-out test set using the optimal classification threshold of 0.50. For evaluation, we reported accuracy, precision, recall, and F1-score, all computed with weighted averaging to account for class imbalance. Additionally, the area under the receiver operating characteristic (ROC) curve (AUC-ROC) was calculated based on the predicted probabilities. A confusion matrix and ROC curve were generated to visualise model performance. All misclassified samples (false positives and false negatives), including predicted labels and probabilities, were saved for further error analysis.

Interpretability analysis using Integrated Gradients (IG)

To enhance model interpretability, we applied the IG method to the test set (Lyu et al. 2022; Wang et al. 2024). The IG attribution scores were computed for each token by integrating gradients with respect to the model embedding layer over 20 steps. For each class (Label = 0 and Label = 1), we calculated the average IG scores for all tokens across the correctly classified samples. Tokens corresponding to special characters or stopwords were filtered out to focus on semantically meaningful words. The top 20 tokens with the highest average IG scores were selected for each classification group and visualised using word clouds to highlight the most influential features contributing to model predictions. The attributions were aggregated per token, and the top 3 most influential tokens were colour-coded in red based on their IG scores. An HTML table was constructed to display the true label, predicted label, and highlighted tokens to illustrate which words the model considered most decisive in its prediction. This visualisation enables qualitative analysis of model decision-making behaviour, facilitating the identification of class-discriminative lexical cues.

Computational environment

All experiments were conducted using the PyTorch 2.6.0 framework with CUDA 12.4. The models were trained on an NVIDIA A100 GPU via the UF HiPerGator cluster environment. Tokenization and training were implemented using transformers 4.48.3 and datasets 3.3.2. Other dependencies include scikit-learn 1.6.1, captum 0.8.0, and Python 3.10.16.

Results

TF-IDF analysis reveals distinct lexical patterns in feline forensic autopsy reports

To assess the readability and lexical structure of the feline forensic reports, we extracted key terms to determine whether the text features were interpretable to humans, particularly in the context of determining the cause of death related to traffic accidents. For this analysis, terms were independently extracted from two major sections of the reports, Comments and Gross Diagnoses, with a representative example of each institutional report provided in Supplementary Figures S1 and S2. First, our TF-IDF analysis revealed distinct lexical patterns between VT and non-VT cases in the Comments section. Specifically, our initial analysis showed that reports classified as VT group frequently included high-scoring terms such as “trauma,” “fracture,” and “vehicular,” reflecting language commonly used to describe hit-by-car incidents (Supplementary Figure S3A). In contrast, non-VT cases were characterised by terms such as “testing,” “feline,” “urine,” “wound,” and “barbiturate,” indicating a broader range of general terms and investigative purposes (Supplementary Figure S3B). When we excluded general words common to both groups, such as “cat,” “feline,” “testing,” and “multiple,” the distinct patterns remained, further highlighting trauma-related words (Figure 1).

Figure 1.

TF-IDF analysis of the “Comments” section in veterinary forensic reports on FRCs. WordCloud visualises the top 30 most frequent commentary terms for VT cases (A) and non-VT cases (B).

Next, in the Gross Diagnoses section, the TF-IDF scores revealed differential patterns in pathology-related terms between the two groups. VT cases were associated with terms such as “fracture,” “skull,” “laceration,” “claw,” “rib,” and “bone,” while non-VT cases featured terms including “wound,” “flea,” “intestine,” and “effusion,” and “hyperaemia” (Supplementary Figure S4). When we excluded general words commonly found in both groups, such as “acute,” “haemorrhage,” “multifocal,” “focal,” “associated,” “moderate,” “severe,” “right,” “left,” “multiple,” and “mild,” the patterns became more distinct, highlighting trauma-related diagnoses (Figure 2).

Figure 2.

TF-IDF analysis for “Gross diagnoses” section in veterinary forensic reports on FRCs. WordCloud visualises the top 30 most frequent pathology-related terms for VT cases (A) and non-VT cases (B).

These findings indicate that the forensic autopsy reports used in this study were documented in a consistent format and contained identifiable lexical patterns associated with the causes of death in FRCs. The results suggest that lexical features within these reports can serve as effective discriminative signals for cause-of-death classification. Also, they highlight the utility of TF-IDF analysis in generating descriptive statistics and uncovering class-specific terminology.

Clinical-Longformer enables the classification of forensic autopsy reports by cause of death in cats

Subsequently, to establish AI-driven classification models for determining causes of death in cats, the Clinical-Longformer model was fine-tuned to distinguish between VT and non-VT cases using textual descriptions from the Gross Diagnoses section of forensic reports (dataset splits are shown in Figure 3A. Training accuracy increased steadily from 0.858 to 1.000 by the fifth epoch, indicating effective learning through early stopping and hyperparameter tuning (Figure 3B). Concurrently, training loss decreased from 0.445 to 0.003. Validation accuracy exhibited minor fluctuations but trended upward, ranging from 0.778 and 0.889 by the fifth epoch (Figure 3C). Validation loss showed greater variability, starting at 0.507 and decreasing to below 0.375. Based on these trends, the model checkpoint from epoch 2 was selected for testing. On the 20% hold-out test dataset, the model demonstrated strong and well-balanced performance, achieving an accuracy of 0.946, precision of 0.885, recall of 1.000, and F1-score of 0.939 (Figure 3D). The confusion matrix revealed balanced classification across both classes, with zero false negative and three false positives among the 55 test sets (23 VT and 32 non-VT cases) (Figure 3E). This performance was obtained using an optimised decision threshold of 0.50. The discriminative ability of the model was further supported by an AUC-ROC of 0.950 (Figure 3F). These results demonstrate that Clinical-Longformer can be effectively trained on structured forensic text data to distinguish between VT and non-VT causes of death in cats. Our model also showed strong and well-balanced performance on independent hold-out test datasets. Such robust performance across datasets obtained from two institutions indicates the potential of the model for generalisation, advancing AI-driven applications in veterinary forensic pathology.

Figure 3.

Test performance evaluation of the Clinical-Longformer model for classifying VT and non-VT feline forensic cases. Panels show test performance metrics (A), confusion matrix (B), and AUC-ROC curve (C).

IG tool enhances model interpretability

To interpret the model predictions, we applied IG technique to the test set. From this analysis, the top 30 tokens with the highest average IG attribution scores were extracted for each class, revealing distinct lexical patterns with some overlap (Figure 4). In VT cases, the most influential tokens were associated with traumatic physical injuries, including “fracture,” “injury,” and “trauma.” Other prominent tokens included “laceration,” “frontal,” “spinal,” “skull,” “brain,” and “bone,” all suggestive of damage and diagnostic observations related to the skeleton. While “fracture” was also common in non-VT cases, they featured terms such as “tooth,” “pubis,” “mandible,” and “avulsion,” indicating a wide range of trauma sources or conditions potentially resulting from falls or other accidents. The model highlighted “fracture” as a key token for both classes, but the attribution scores differed, suggesting that the interpretation of the model is influenced by contextual and co-occurring terms.

Figure 4.

IG analysis of token importance in model predictions. WordCloud visualises the top 20 influential tokens contributing to predictions for VT cases (A) and non-VT cases (B).

Token-level visualisation highlights discriminative features between VT and non-VT feline cases

To further enhance the model interpretability, we generated token-level IG visualisations for individual test samples. Word importance was computed using IG data, and the top three most influential tokens were highlighted in red to provide insight into the reasoning for model predictions (Supplementary Table S1). Among the 52 correctly predicted cases, the mean predicted probability was 0.951 for VT cases (SD = 0.017) and 0.048 for non-VT cases (SD = 0.026), indicating strong confidence in distinguishing between the two groups (Figure 5). Three non-VT cases were misclassified as VT, with predicted probabilities of 0.925, 0.961, and 0.962, respectively. All three cases involved blunt force trauma, and IG visualisations highlighted terms such as “fractures,” “age,” “acute,” and “tooth,” which may have contributed to their misclassification as VT (Table 1). These findings provide human-interpretable insights into the decision-making process and demonstrate that the model attends to features associated with established forensic indicators, supporting the pathological plausibility of its predictions.

Figure 5.

IG-based comparison of VT and non-VT cases with identification of misclassified cases. Plots show predicted probabilities for test data and highlight incorrectly classified cases.

Table 1.

Interpretation of word importance results for incorrectly predicted cases using IG analysis.

*Text highlighted in red indicates highly influential features for the model decision.

Discussion

This study demonstrates the utility of transformer-based LLMs in classifying veterinary forensic diagnoses. The Clinical-Longformer model, originally on human clinical text, achieved strong performance on veterinary autopsy reports with domain-specific fine-tuning. Our analyses using TF-IDF and IG identified biologically meaningful terms, enhancing the interpretability of the model. The ability to spotlight key terms within text-based reports may assist pathologists and forensic investigators in reviewing complex cases and discerning patterns across large datasets.

A minimum standards document exists for performing veterinary forensic autopsies, outlining procedures typically applied in the forensics realm, such as DNA analysis, morphology, tyre tracks, fingerprints, human death investigations, and crime scene investigations (Ottinger et al. 2014; AAFS 2026). However, these standards apply only to forensic cases, and most laboratories, pathologists, and general practitioners do not adhere to them in routine practice (Aradilla et al. 2025). In addition, no standardised postmortem protocol currently exists for FRCs worldwide. It is partly due to the lack of systemic survey data and the practical challenges of establishing standards based on the limited cases examined by trained veterinary pathologists (McDonough and McEwen 2016). Financial constraints further limit the investigation of FRC deaths (Ottinger et al. 2014): most funding comes from government or community programmes and it is primarily allocated to spay/neuter procedures, with minimal resources devoted to severely ill or deceased cats. These financial and practical challenges impede substantial progress in determining causes of death in FRCs, leaving the growing societal demand largely unmet. Our study demonstrating AI-powered analysis of FRC autopsy reports offers valuable resources and reference points that support the establishment of practical guidelines for standardising procedures and documenting autopsy findings.

This study confirms that incorporating human-interpretable terms in documents is beneficial for enhancing the utility and interpretability of language models. Word extraction and text mining should prioritise human interpretability, particularly during the initial phases of data preprocessing. Our model was developed within the specialised field of veterinary forensic science and a unique domain within veterinary pathology. In this context, autopsy findings and gross diagnoses are recorded by specially trained veterinary pathologists for forensic investigations, necessitating a consistently structured and standardised textual format. The datasets used in this study were derived from a uniform documentation system across two veterinary forensic laboratories (UF and APQA). Despite the inherent variability and complexity of forensic cases, our multi-institutional study highlights the potential for model generalisability. Another critical area of focus involves necropsies or autopsies performed by general veterinary practitioners in clinical settings. In-depth examinations of the kinds conducted by specialised forensic laboratories in this study are rarely feasible in general practice. In other words, while numerous FRC cases are submitted to local animal clinics, only a small fraction are referred to forensic laboratories for detailed investigation. AI-assisted platforms could help bridge this gap by supporting case interpretation in settings without access to specialised expertise, particularly given that many FRCs may never be evaluated by a board-certified pathologist. In this context, AI-based tools may be integrated into routine workflows as complementary decision-support systems. For example, following completion of a postmortem examination, AI models could serve as a secondary review to compare interpretations and identify potential discrepancies requiring further evaluation. Such tools may also assist non-specialists by providing preliminary assessments and flagging complex cases for referral to forensic pathologists. These applications highlight the potential of AI to improve diagnostic consistency, support clinical decision-making, and expand access to forensic expertise.

A key limitation of this study is that the Clinical-Longformer model learns from token-level features rather than complete words, which can disrupt typographic and linguistic structures such as punctuation, spacing, syllables, and morphemes. As a result, the model may produce word fragments that are not readily interpretable without specialised domain knowledge. This presents challenges for readers outside the field, particularly when interpreting complex medical or domain-specific terminology. For instance, in our study, the IG explainability tool identified tokens like “age” and “hemorrh” (from “haemorrhage”) as influential features. Without the appropriate anatomical or pathological context, these fragments are insufficient to reconstruct the full terms and understand their significance. To mitigate token fragmentation and domain dependency, several practical strategies may be considered, including standardised vocabularies, text normalisation during preprocessing, and domain-specific model training. In addition, structured reporting formats and ontology-based annotation frameworks may improve consistency and alignment between textual inputs and clinically meaningful concepts, thereby enhancing interpretability. Hence, the development and application of such models require active involvement of domain experts, veterinary forensic pathologists in this case, to ensure interpretability and relevance. Their expertise is also important for identifying ambiguous or overlapping terms that may call for further contextual modelling.

Another limitation to acknowledge is that trauma in FRCs can manifest in diverse patterns, and the models developed in this study do not fully account for this variability. For example, traumatic injuries as a result of predation in a kitten may present as circular puncture wounds accompanied by subcutaneous soft tissue tearing. In contrast, projectile trauma may also produce circular external wounds but typically involves track formation with minimal soft tissue disruption. Since both types of trauma, as well as VT, can result in fractures, careful differentiation is required. Other trauma categories, such as gunshot wounds, injuries caused by air gun pellets, and additional predation scenarios, should also be considered in forensic evaluations. Moreover, cases involving animal poisoning, particularly those caused by ethylene glycol or anticoagulants, should be explored, as FRCs are at risk for such injuries. While experienced pathologists may readily recognise these conditions, they can be challenging for practitioners or pathologists who are not routinely engaged in forensic casework. In this regard, future efforts to enhance AI-supported tools may assist veterinarians in interpreting injury patterns, identifying underlying causes, and refining differential diagnoses. Expanding datasets to include a broader spectrum of trauma and non-trauma conditions, along with rigorous external validation using independent datasets, will be beneficial for improving model robustness and better reflecting the complexity of real-world forensic cases. The present study was designed as an initial feasibility assessment using a relatively focused dataset. We anticipate that incorporation of additional case types, larger multi-institutional datasets, and systematic external validation will further strengthen the clinical and forensic applicability of AI-based approaches.

Furthermore, it is important to note that, in the specific context of FRC forensic cases, trauma is commonly encountered, and it is essential not only to identify its presence but also to determine the underlying cause of the injuries. Submitters or reporters of FRCs often seek confirmation of intentional harm, placing forensic investigators under a legal obligation to remain neutral in interpreting examination results. It is therefore imperative to secure scientific evidence that can both support and refute such allegations. The implementation of AI-assisted systems in this domain must be guided by rigorous ethical and legal considerations, including the establishment of appropriate rules and regulations to govern their development and real-world application.

In conclusion, this study demonstrates that transformer-based LLMs, when paired with text analysis and explainability tools, offer a powerful framework for automating and interpreting veterinary forensic diagnostics. Although our AI models are still in the early stages of development, once fully developed, they may also offer timely, localised insights for general practitioners and other stakeholders. Future work should explore domain-specific pretraining and multi-label classification to enhance model performance and maximise the utility of AI across a broader range of investigational categories.

Supplementary Material

Supple_Tableclean.xlsx

Supple_Tableclean.xlsx

Supple_Figs.docx

Supple_Figs.docx

Funding Statement

This work was partly supported by the UF Foundation ‘A Cat Has No Name’ fund and the UF AI Initiative. This work was also supported by a grant (I-1543069-2024-26-01) from the Animal and Plant Quarantine Agency (APQA), Ministry of Agriculture, Food and Rural Affairs, Republic of Korea.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The original contributions presented in this study are included in the manuscript and the supplementary data. Further inquiries can be directed to the corresponding author. The data supporting the conclusions of this article may be made available by the corresponding author upon request, subject to institutional guidelines and a data use agreement.

Supplementary material

Supplemental data for this article can be accessed at https://doi.org/10.1080/01652176.2026.2705200.