CowPainCheck: AI-based facial expression analysis for dairy cow welfare

Effective pain management is increasingly recognized as a cornerstone of dairy cattle welfare, significantly influencing both animal well-being and production efficiency. Mastitis, one of the predominant diseases affecting dairy herds globally, imposes extensive economic consequences extending far beyond the direct expenses associated with treatment. Recent studies underscore that mastitis profoundly reduces the net present value (NPV) of dairy operations due to decreased milk yield, impaired reproductive capabilities, and increased culling rates [1]. Moreover, the negative economic ramifications of mastitis extend notably into reproductive outcomes, as cows afflicted with mastitis display substantially reduced conception rates compared to their healthy counterparts: notably lower first-service (41.7% vs. 58.2%), third-service (30.2% vs. 45.3%), and cumulative conception rates across multiple services (36.4% vs. 49.2%) [1].

Lameness is another significant contributor to pain-associated economic losses within dairy operations. Recent evidence positions lameness as the third most economically damaging health issue among dairy cattle, preceded only by mastitis and fertility disorders [2]. The economic impact of lameness manifests clearly through immediate and considerable reductions in milk production shortly after the onset of symptoms [3]. Beyond direct production losses, lameness triggers additional financial burdens from increased treatment costs, prolonged calving intervals, and the necessity of premature culling [2]. Collectively, the multidimensional nature of pain-related economic impacts underscores the necessity of developing effective, scalable, and precise methods for timely detection and intervention.

Current practices for pain assessment in dairy cattle rely predominantly on subjective behavioural observations, physiological measurements, and clinical scoring techniques. These methods exhibit significant methodological constraints that compromise their reliability and applicability. Systematic reviews have documented considerable subjectivity inherent in behavioural assessments, with variations in observer outcomes heavily influenced by factors such as evaluator experience, gender, age, and contextual nuances [4]. Meta-analyses have quantified this subjectivity, demonstrating substantial discrepancies in reported pain scores between studies employing diverse scales and terminologies [4].

Observer bias and inter-rater variability further weaken the reliability of conventional pain assessment tools. Recent evaluations employing the Consensus-based Standards for the Selection of Health Measurement Instruments (COSMIN) guidelines highlight that only a small subset of behavioural-based instruments achieve consistently high validation scores across all essential measurement criteria [5]. These findings emphasize the critical challenges associated with obtaining dependable and uniform pain assessments, particularly within commercial dairy environments where evaluators face stringent time constraints and limited training opportunities.

Furthermore, traditional assessment approaches frequently fail to capture the dynamic and transient nature of pain expressions. Most conventional evaluations provide only episodic, snapshot observations, which may overlook brief but clinically meaningful expressions of pain. Research indicates that pain-related facial expressions in livestock often last between 0.3 and 0.7 s, rendering periodic manual observations insufficiently sensitive to identify early-stage or subclinical pain conditions [6]. This temporal limitation presents significant gaps, allowing undetected pain states to escalate unchecked between observation intervals.

Additionally, invasive traditional methods raise critical animal welfare concerns. Physiological indicators, such as blood sampling or rectal temperature measurement, may induce stress reactions in animals, inadvertently confounding pain assessment outcomes [7]. The necessary handling and restraint involved in invasive assessments can mask or artificially amplify expressions of pain, thus undermining both the accuracy and ethical justification of such procedures [7].

The integration of artificial intelligence and mobile technologies into veterinary medicine has witnessed rapid acceleration since 2021, significantly advancing capabilities in automated animal welfare monitoring and pain detection systems. Recent innovations in computer vision techniques demonstrate impressive accuracy in cattle biometric identification and behaviour monitoring. For example, Vision Transformer (ViT) architectures applied to the Opencows2020 dataset achieved cattle identification accuracy rates as high as 99.79%, while YOLO-based frameworks combining YOLOv5 with ViT have attained detection precision (mean average precision, mAP) of 97.8% and identification accuracy of 96.3% in practical farm settings [8,9].

Precision livestock farming (PLF) represents a paradigm shift in dairy farm management, incorporating AI-driven sensors, computer vision, and big data analytics to monitor animal health, behavior, and welfare continuously. Contemporary PLF systems leverage diverse sensing modalities such as RFID tags, accelerometers, thermal imaging, and computer vision analytics to deliver comprehensive, real-time insights into cattle welfare. Research indicates that accelerometer-based systems effectively detect movement patterns indicative of lameness or stress-related behaviors, while multimodal sensor integration consistently outperforms single-modality systems in terms of detection accuracy and reliability [10].

The proliferation of mobile technology has facilitated widespread accessibility to advanced monitoring capabilities, empowering farmers to deploy PLF solutions effectively, even without extensive technical expertise or substantial infrastructural investments. Recent deployments of mobile applications for livestock welfare have demonstrated high usability and practical feasibility in commercial production scenarios, enabling farmers to swiftly interpret data and respond proactively to welfare alerts [11]. The seamless integration of mobile technology with advanced AI algorithms has successfully addressed temporal limitations inherent in traditional pain assessment methodologies, enabling timely interventions and enhancing animal welfare outcomes in dairy cattle.

Thus, the convergence of AI, computer vision, and mobile platforms holds remarkable promise to address current limitations in pain assessment, facilitating objective, scalable, and ethically responsible improvements in dairy cattle welfare and economic viability.

This systematic review followed PRISMA 2020 guidelines to identify and synthesize peer-reviewed research on AI-based animal pain detection systems. A comprehensive search was conducted across three major databases (PubMed, Web of Science, Scopus) covering publications from January 2021 to July 2025.

Search Strategy: Boolean combinations of terms including (“artificial intelligence” OR “machine learning” OR “computer vision” OR “deep learning”) AND (“pain detection” OR “facial expression” OR “grimace scale”) AND (“cattle” OR “livestock” OR “cat” OR “dog” OR “horse” OR animal species terms) were employed with database-specific syntax optimization. A representative PubMed search string.

(“artificial intelligence”[Title/Abstract] OR “machine learning”[Title/Abstract] OR “deep learning”[Title/Abstract] OR “computer vision”[Title/Abstract]) AND (“pain detection”[Title/Abstract] OR “facial expression”[Title/Abstract] OR “grimace scale”[Title/Abstract]) AND (“cattle”[Title/Abstract] OR “cow”[Title/Abstract] OR “livestock”[Title/Abstract] OR “sheep”[Title/Abstract] OR “horse”[Title/Abstract] OR “dog”[Title/Abstract] OR “cat”[Title/Abstract])

Eligibility Criteria: Studies were selected according to predefined inclusion and exclusion criteria, as outlined in Table 1.

Study Selection: Search results were exported to Zotero reference manager. Duplicate records were removed. Screened titles and abstracts for relevance, followed by full-text assessment against eligibility criteria.

Data Synthesis: A total of 112 high-quality studies met inclusion criteria, encompassing multiple species (cattle, sheep, horses, cats, dogs, rabbits, rodents) and AI approaches (CNNs, Vision Transformers, YOLO architectures). Performance metrics were systematically extracted and synthesized both quantitatively and narratively to provide comprehensive coverage of current AI capabilities in automated animal pain detection. The selection process is documented in the PRISMA 2020 flow diagram (Fig. 1), detailing the number of records identified, screened, excluded (with reasons), and included in the final synthesis.

Fig. 1. PRISMA 2020 flow diagram for study selection in the systematic review on AI-based facial action unit analysis for pain detection in dairy cattle.

Download Original Figure

Detecting pain in dairy cattle is critical not only for animal welfare but also for its profound economic implications across the dairy industry. Unaddressed pain negatively affects animal behavior and physiological health, triggering stress responses that diminish productivity, growth rates, milk yield, and reproductive efficiency [12]. These physiological impacts such as hormonal stress responses, metabolic disruptions, and immune system suppression directly compromise animal health, thus reducing overall farm profitability [7].

Economic incentives further underscore the importance of effective pain mitigation. Zoltick et al. [7] highlight that reducing pain through proactive analgesia enhances production efficiency sufficiently to offset the associated analgesic costs, thereby incentivizing farmers toward improved animal welfare management [7]. Additionally, stress responses triggered by pain negatively affect nutrient absorption, reproductive function, and general body condition, collectively translating into measurable economic losses at the herd level [7].

Economic losses attributed to pain-related conditions in dairy cattle are substantial, with mastitis, lameness, and ketosis identified as the primary economic burdens globally, costing the dairy sector approximately $65 billion annually [13]. These losses encompass direct treatment expenses and significant indirect impacts, including reduced productivity, reproductive failures, premature culling, and impaired herd longevity.

Traditional pain assessment approaches predominantly rely on direct behavioral observations, physiological biomarkers, and structured clinical scoring systems. Despite recent methodological improvements, these approaches carry inherent limitations affecting their practical effectiveness.

Facial expressions constitute one of the most fundamental, evolutionarily conserved communication mechanisms for pain across mammalian species. The development of standardized grimace scales has significantly enhanced objective pain assessment in veterinary medicine, overcoming traditional limitations related to observer subjectivity. This section systematically examines the neurobiological mechanisms underlying facial expressions of pain, the rigorous development and validation processes for grimace scales across domestic, laboratory, and farm animal species, and addresses ongoing challenges in their clinical applicability and reliability for livestock welfare management.

Despite methodological advancements, traditional pain assessment faces practical and conceptual constraints that impede widespread effectiveness.

Convolutional Neural Networks remain the cornerstone of automated animal pain detection systems, with recent studies demonstrating substantial improvements through architectural refinements and species-specific optimizations. The foundational strength of CNNs lies in their hierarchical feature extraction capabilities, enabling the identification of subtle facial patterns associated with pain expressions across multiple livestock species [49,50].

The introduction of Vision Transformers (ViTs) has fundamentally challenged CNN dominance in animal facial analysis. ViTs excel at capturing long-range dependencies and global contextual information, characteristics particularly valuable for understanding complex facial expression patterns in livestock [56]. The ViT-Sheep model, incorporating LayerScale modules and transfer learning strategies, achieved 97.9% accuracy for sheep face recognition, demonstrating the architecture’s potential for livestock applications [56].

The combination of RGB and thermal imaging represents a significant advancement in automated pain detection, providing complementary information streams that enhance overall system robustness and accuracy [51].

A critical question for adoption is not which architecture attains the highest benchmark score in controlled experiments, but which architecture reliably performs under real farm constraints (variable lighting, occlusion, dirt, overlapping animals), runs on available hardware (edge devices, low-power systems), and generalizes across herds and barns. Below we compare Convolutional Neural Networks (CNNs), YOLO-family detectors, Vision Transformers (ViTs) and multimodal fusion approaches against practical implementation criteria supported by recent peer-reviewed farm or near-farm studies.

Automated pain recognition in cats has achieved remarkable progress through multiple complementary approaches. The landmark-based methodology achieved 77% accuracy in pain detection using manually annotated geometric landmarks positioned relative to underlying facial musculature, significantly outperforming deep learning approaches that reached only 65% accuracy on the same heterogeneous dataset [74]. This study utilized 84 client-owned cats of different breeds, ages, sexes, and varying medical conditions, representing a substantial advancement over previous homogeneous datasets limited to single breeds.

Video-based automation marked a significant technological leap with the development of end-to-end AI pipelines requiring no manual image selection or landmark annotation [31]. The system achieved over 70% and 66% accuracy respectively on two different cat pain datasets, outperforming previous landmark-based approaches using single frames under similar conditions. The pipeline integrated YOLOv8 for face detection, ensemble landmark detection, and XGBoost classification with moving window analysis.

Smartphone-applicable systems represent the current clinical frontier, utilizing deep neural networks and machine learning models trained on 3,447 cat face images annotated with 37 landmarks [75]. The best CNN model (ShuffleNetV2) achieved 16.76% Normalized Root Mean Squared Error for landmark prediction, while XGBoost models reached 95.5% accuracy and 0.0096 mean squared error for Feline Grimace Scale score prediction. The system demonstrated excellent discriminatory capability between painful and non-painful cats, enabling practical veterinary applications.

Macaque facial expression analysis achieved groundbreaking automation through the first prototype for automatic MaqFACS (Macaque Facial Action Coding System) coding [76]. The system achieved high performance in recognition of six dominant action units, demonstrating generalization between conspecific individuals (Macaca mulatta) and even between species (Macaca fascicularis). The method showed concurrent validity with manual MaqFACS coding, supporting automated applications in social and affective neuroscience research.

Japanese macaque pain detection utilizing ResNet50 architectures achieved varying accuracy depending on extraction methodology [77]. Box extraction using RetinaFace resulted in test accuracies between 48%–54%, while contour extraction using Mask R-CNN improved performance to 64% through preprocessing and fine-tuning. The study utilized 30–60 min of video footage from macaques undergoing laparotomy, recorded before surgery (No Pain) and one day post-surgery before analgesic administration (Pain).

Geometric morphometric approaches complemented automated systems by revealing subtle facial shape variations in female Japanese macaques following experimental laparotomy [78]. The study identified pain-associated changes including orbital tightening, asymmetrical eye aperture, lip tension, and elongated mouth lines, providing anatomical foundation for automated detection algorithms.

Dog emotional state recognition achieved significant progress through dual-approach methodologies comparing DogFACS-based and deep learning systems [79]. The DogFACS-based approach utilizing Decision Tree classifiers reached 71% accuracy, while deep learning techniques achieved 89% accuracy for positive/negative emotional state classification. The study analyzed 29 Labrador Retrievers under experimentally induced emotional states of positive anticipation and frustration.

Continuous facial dynamics analysis introduced novel automated methods for measuring dog facial behavior through video-based tracking of 46 facial landmarks [80]. The system revealed distinct patterns between brachycephalic (Boston Terrier) and normocephalic (Jack Russell Terrier) dogs, with brachycephalic dogs exhibiting consistently lower facial dynamics across all tested contexts and facial regions compared to normocephalic dogs.

Recent research has applied computer vision and machine learning to detect pain in dairy cows under various conditions (e.g., lameness, mastitis). These studies use facial and gait indicators (e.g., orbital tightening, ear position, back curvature) extracted from images or video, often combined with sensor data, to train AI models. The Table 5 summarizes post-2021 peer-reviewed studies, detailing pain condition, FAUs or behavioral indicators, sensing methods, AI models, validation design, sample size, and key performance metrics (separating object-detection from pain-classification). All metrics are cited from the primary sources (Table 5).

Across the reviewed studies, key factors consistently drove differences in reported performance. First, model architecture and design strongly affected outcomes. Convolutional networks often required careful tuning of hyperparameters to avoid overfitting on small datasets. For example, Mao and Liu’s dog-expression study trained a CNN on only 315 images and found that tuning via an improved Whale Optimization algorithm boosted accuracy modestly by approximately 3 percentage points [52]. This suggests that generic CNN architectures alone may plateau on limited animal-expression datasets. By contrast, transformer-based models and large pre-trained encoders tended to generalize better when data were scarce. Like ViT-based sheep face model (ViT-Sheep) achieved 97.9% accuracy by incorporating architectural enhancements (LayerScale) and transfer learning on 160 sheep images [56]. Similarly, Feighelstein et al. used a CLIP encoder (a large-scale vision transformer) to detect pain in sheep, and the AI pipeline significantly outperformed expert scoring (AUC 0.82 vs. AUC 0.70 for humans) on the same 48-animal dataset [57]. The benefit of pretraining is clear: models with broad prior knowledge (ViT, CLIP) captured subtle facial cues that smaller CNNs missed.

Second, data quantity and quality were fundamental. Larger, well-annotated datasets yielded higher accuracy. For instance, cow-ID system had high sample diversity (77 cows, numerous face images) and achieved approximately 98% AP for detection and 96% identity accuracy, likely reflecting the ample data and robust YOLOv5+ViT pipeline used [53]. In contrast, studies with very small animal datasets (e.g., 10–30 individuals) often reported only modest performance (< 70%–80%). cat-pain pipeline attained only 70% and 66% accuracy on two feline datasets [31], even though they used a video-based approach, because these datasets remained small and heterogeneous. Subjective annotation also added noise: studies relying on human-rated pain scores (grimace scales) were inherently limited by rater inconsistency. The sheep-CLIP study mitigated this by using human scores as “gold standard” for comparison, but AI still outperformed the inconsistent human labels [57].

Third, image modality and preprocessing played a major role. Models trained on RGB imagery nearly always outperformed those on thermal images. In the new CattleFace-RGBT benchmark, ResNet-101 achieved 94.4% AP on RGB face detection but only 64.6% AP on thermal images [51]. Thermal data lack color/texture and suffer from low contrast, making keypoints harder to localize [51]. Transformer architectures fared slightly better on thermal: e.g., Swin-B scored 73.2% AP on thermal (vs. 75.3% on RGB) [51], suggesting that global attention can partly compensate for poor thermal detail. Some authors therefore use cross-modal transfer: Coffman et al. trained models on RGB and refined them on thermal, using semi-automated annotation to build the thermal landmark set. In video-based systems, temporal preprocessing also helped. Feighelstein et al. introduced GrayST to inject motion cues into static CNNs, boosting rabbit-pain recognition from approximately 67% (ResNet alone) to approximately 77%–81% (with GrayST) [50]. Further filtering of video frames (keeping only high-confidence images) lifted rabbit-pain accuracy above 87% [50]. These examples show that explicit temporal encoding can overcome the lack of color or texture in single frames, at the cost of some complexity.

Fourth, species-specific traits and experimental conditions influenced outcomes. Some species exhibit very subtle facial changes, or wide breed variation, which makes generalization difficult. For example, dog facial morphology varies enormously by breed, so Mao et al. note that even with IWOA-CNN their hardest classes (sad, fear) remained under 90% accuracy. Similarly [52], the cat-pain studies point out that facial landmarks in cats are subtle and vary by individual, so performance capped around 70% despite advanced pipelines [31]. By contrast, simpler tasks with distinctive cues yielded higher scores: automated detection of specific behaviors (e.g., cow hoof issues via accelerometers) or identity recognition (hundreds of cow faces) tended to exceed 90% accuracy, showing that modality and task simplicity matter. In experiments with induced pain (e.g., sheep post-surgery), the lab setting ensured high-quality imagery and clear labels, enabling better performance than on “in-the-wild” farm data. As Feighelstein et al. comment, machine accuracy in controlled sheep surgery videos even exceeded that of vets, a setting where expressions were pronounced and consistently labeled [57].

Fifth, fusion strategies and multiple cues often improved robustness. Combining face imagery with other modalities (e.g., body posture, sensor data) tends to outperform any single cue. For instance, cow lameness studies fused video keypoints with spatiotemporal models and achieved approximately 90% classification accuracy. Multi-stage pipelines (e.g., YOLO detection + landmark extraction + LSTM) likewise decomposed tasks into tractable steps. In Martvel et al.’s cat-pain study, using video (many frames) instead of isolated images improved detection by leveraging temporal consistency [31]. Conversely, studies using only static images, or only single modalities, generally lagged.

In summary, higher performance was generally achieved by (a) using ample, well-curated data; (b) leveraging strong pretraining or multimodal cues; and (c) tailoring architectures and preprocessing to the species and context. Studies consistently note that scarce or noisy data, inter-species variability, and limited modalities suppress accuracy. The success of vision transformers and large encoders in sheep and rabbit pain tasks suggests future work should exploit pretraining and attention to capture subtle patterns. Likewise, integrating temporal dynamics (as in GrayST or video analysis) and multi-modal fusion appears crucial when single frames offer limited information [31,57]. These insights indicate that next-generation animal pain recognition systems will likely combine rich data collection (e.g., RGB + thermal + behavior), advanced architectures (transformers, hybrid CNN-AI detectors), and robust preprocessing to overcome the inherent challenges of cross-species pain detection.

Despite impressive laboratory performance, several consistent limitations emerged across species. Environmental factors including variable lighting, occlusions, and motion artifacts significantly impact accuracy. Cross-species generalization remains limited, with species-specific anatomical differences necessitating dedicated training approaches.

Validation methodology substantially influences reported performance, with rigorous cross-validation revealing more realistic accuracy expectations. Ground truth establishment varies considerably across studies, affecting system reliability and clinical applicability.

Establishing reliable ground truth data represents the most critical validation challenge for automated cattle pain detection systems. Unlike human pain assessments, which leverage self-reporting mechanisms for direct subjective experiences, veterinary pain evaluations rely exclusively on third-party interpretations of observed behaviors, physiological indicators, and environmental contexts [66]. This inherent reliance on observer judgments introduces substantial risks of bias, necessitating meticulous protocol design to maximize assessment reliability and validity while minimizing subjective influences.

Recent progress in veterinary pain assessment emphasizes validated species-specific pain scales as essential instruments for establishing robust ground truth. Notably, the UCAPS and cow pain scale (CPS) have emerged prominently. Both tools demonstrate high internal consistency (UCAPS α = 0.82; CPS α = 0.79), establishing reliable baselines for objectively quantifying pain severity [34,85]. Comparative evaluations confirm strong criterion validity, exhibiting correlation coefficients ranging from 0.76 to 0.78 when benchmarked against traditional veterinary numerical rating scales [34].

However, integrating validated scales into automated systems necessitates stringent observer training and standardized scoring procedures. Significant variability in inter-rater reliability has been documented, with weighted kappa statistics varying between 0.47 and 0.80 depending on assessor experience and employed scales [34]. Encouragingly, recent studies report high inter-rater agreements between automated systems and human evaluators, consistently exceeding 80%, with Gwet’s agreement coefficients spanning from 0.76 to 0.83 for binary pain categorizations [86].

Additional complexity arises from temporal variability in pain expression. Research indicates that acute pain detection accuracy markedly declines over post-procedural intervals—dropping from approximately 88% accuracy at one hour post-procedure to around 65% after 72 hours—as analgesic interventions and natural healing alter observable pain manifestations [43]. Consequently, dynamic ground truth labeling methodologies that consider temporal pain progression may yield superior accuracy compared to static assessments.

Two primary annotation strategies are recognized in ground truth methodologies: stimulus-based and behavior-based annotations. Stimulus-based annotations, categorizing pain by the presence or absence of procedures, provide clear temporal boundaries yet may inadequately represent individual variations in pain perception and expression [66]. Conversely, behavior-based annotations offer detailed observational insight but introduce greater subjectivity and potential observer biases. Emerging evidence suggests hybrid annotation approaches, combining objective temporal data with expert behavioral assessments, may optimize ground truth accuracy, offering balanced objectivity and nuance [57].

The selection of appropriate cross-validation techniques significantly influences perceived performance and practical generalizability of automated pain detection models. Traditional random cross-validation methods, although computationally convenient, frequently yield overly optimistic estimates due to hierarchical data structures and inherent temporal dependencies typical of livestock behavior datasets [87].

Leave-one-animal-out (LOAO) cross-validation provides greater rigor, better simulating real-world scenarios where pain detection systems must generalize reliably to previously unseen individuals. LOAO validation consistently reports accuracy reductions between 10% and 15% relative to random cross-validation, underscoring individual variability’s impact on performance [88]. Studies employing LOAO methodologies document substantial variability in sensitivity (39.2%–79.6%) and specificity (up to 99.1%), reflecting authentic challenges in accommodating animal-level variability within detection algorithms [89].

Farm-fold cross-validation offers an even stricter validation criterion, explicitly accounting for farm-level variability arising from unique management practices, environmental factors, and herd genetics. Research employing farm-fold approaches typically reports an additional 5%–10% performance decrement compared to LOAO, emphasizing the critical influence of farm-specific contexts on automated system generalizability [89]. Such validation rigor is indispensable when assessing commercial feasibility across diverse farming conditions.

The selection of cross-validation strategy materially affects reported performance and therefore conclusions about real-world readiness. Concrete examples from the reviewed literature illustrate this: Zhang et al. [81] report a validation accuracy of ≈99.65% but only 64.3% on held-out unseen videos (a drop of ≈35.3 percentage points), with pain-class precision = 0.83 and recall = 0.56, a clear sign that a naive validation split substantially over-estimates deployable performance [81]. Other studies include inter-farm samples but do not disaggregate inter-farm results, preventing assessment of farm-level generalizability [83]. Large longitudinal datasets and realistic temporal holdouts yield more conservative but likely more realistic metrics [82]. Methodological analyses show that replacing random CV with LOAO typically reduces reported accuracy by approximately 10%–15% and applying farm-fold (inter-farm) validation incurs an additional approximately 5%–10% decrement; together these stricter protocols can reduce internal estimates by 15%–25% or more. Authors should therefore (i) always report per-study internal and external (held-out/farm-level) metrics, (ii) include LOAO and farm-fold experiments where feasible (or clearly state their absence), and (iii) present balanced metrics (sensitivity/specificity/PPV/NPV/AUC) rather than accuracy alone to avoid misleading conclusions about field performance.

Performance metric selection significantly impacts validation outcomes. While accuracy remains prevalent, it can misrepresent performance in imbalanced datasets—common in livestock pain studies where non-painful observations predominate [90]. Recent validation research stresses balanced metric reporting, including sensitivity, specificity, positive and negative predictive values, and ROC-AUC, providing comprehensive model performance assessments [91].

Validating automated cattle pain detection systems presents multifaceted challenges impacting result interpretation and generalizability. Feline pain detection studies deliberately limited populations to single breed types (domestic short-haired cats) to minimize confounding variables during proof-of-concept validation [92]. This breed-specific variability necessitates explicit validation strategies across genetically diverse cattle populations to ensure comprehensive applicability.

Environmental variability further complicates validation accuracy. Farm-specific environmental factors—including inconsistent lighting conditions, occlusion by equipment or other animals, mud contamination, and motion blur—significantly degrade detection accuracy, with performance typically decreasing by 15%–20% compared to controlled experimental environments [93]. These findings underscore the necessity for explicitly incorporating realistic environmental conditions within validation studies, assessing model resilience across varied farming scenarios.

Limited dataset sizes remain pervasive within current validation literature, typically involving fewer than 100 animals per study with limited representation across age, sex, breed, and farm management practices [87]. Such limited diversity restricts statistical power and constrains broader population generalizability. Temporal and spatial clustering further exacerbates sample size limitations, necessitating larger, more diverse datasets for robust validation outcomes.

Acute and chronic pain condition differentiation presents unique validation complexities, given distinct temporal trajectories and subtle behavioral indicators characterizing chronic pain states compared to acute presentations [7]. Addressing chronic pain validation demands specialized protocols accommodating long-term, subtle behavioral changes alongside traditional acute pain indicators.

Longitudinal data dependencies also introduce validation complexities. Randomly sampling temporal data points risks inadvertent leakage of future information, artificially inflating model performance estimates [94]. Blocked cross-validation approaches respecting chronological data sequences provide more authentic accuracy assessments yet require sufficiently large datasets to preserve statistical power.

Multimodal sensor integration, while promising enhanced accuracy (typically improving detection accuracy by 5%–10%), further complicates validation procedures, necessitating synchronized data collection and modality-specific preprocessing to ensure consistency and reliability [95].

Collectively, these validation complexities highlight critical needs for standardized protocols and collaborative multi-institutional research frameworks, enabling rigorous validation of automated cattle pain detection systems across diverse populations, farm environments, and temporal conditions. Future research prioritizing comprehensive validation methodologies can substantially advance practical translation of emerging technological solutions, significantly enhancing dairy cattle welfare outcomes [96]. Building on these crucial validation insights, the development and deployment of mobile applications represent the next pivotal step in democratizing automated pain detection technologies for farmers and veterinary practitioners alike.

Recent veterinary mobile applications reflect substantial diversity in their functionalities. Applications such as PIGLOW, an EU-funded platform, enable farmers raising free-range pigs to conduct structured welfare audits periodically, providing automated feedback and comparative benchmarking against peer farms. A two-year pilot study involving 12 farms demonstrated modest improvements in welfare indicators, including reductions in lameness and skin lesion prevalence, alongside high farmer-reported usability and acceptance [97]. Similarly, mobile apps tailored for beef cattle management have shown strong user satisfaction and usability, as indicated by a system usability scale (SUS) rating of approximately 75, highlighting their effectiveness in streamlining feed tracking and animal health record-keeping processes [11].

Wearable and internet-of-things (IoT) devices represent another significant category of livestock monitoring solutions. Prototype collars designed for cattle and other livestock have emerged prominently, capable of continuously monitoring animal physiological parameters such as body temperature, heart rate, and physical activity. These wearable systems transmit collected data to cloud analytics platforms, providing veterinarians and farm managers with timely alerts to early indicators of health issues, including respiratory infections, thereby enabling intervention prior to observable clinical signs [97,98]. Machine-vision-based mobile apps have recently begun leveraging compact convolutional neural network (CNN) architectures—such as YOLOv5—to enable smartphone-based, real-time identification of hoof conditions like digital dermatitis. Such technologies have been successfully deployed on Android and iOS platforms, providing practical and immediate on-farm lameness screening capabilities [67]. Consistently, user feedback underscores that farmers highly value mobile apps featuring intuitive workflows, straightforward checklists, and simplified data captures that seamlessly integrate into their daily farm management routines.

For mobile applications operating in rural livestock farming environments, robust on-device processing and reliable local networking capabilities are essential. Recent studies underscore the advantages of edge computing solutions, such as deployments utilizing NVIDIA Jetson Nano hardware equipped with 12 MP cameras for real-time cattle identification tasks on dairy farms. These implementations enable rapid edge inference, allowing immediate local web access to cattle identification information without reliance on continuous internet connectivity, thereby demonstrating real-world latency performances measured in milliseconds per inference [49].

Power management strategies are equally critical for prolonged operation in remote farm environments. Several wearable systems now incorporate renewable power solutions, such as small solar panels or kinetic energy harvesting from animal movement, enabling continuous data collection without frequent manual battery replacements. Examples include prototype collars successfully deployed on reindeer and cattle, providing continuous operation for weeks at a time [99]. Reviews of such systems confirm that hybrid power setups—combining solar panels and motion-based harvesting—effectively support uninterrupted, round-the-clock monitoring, in contrast to purely battery-powered collars, which typically require weekly recharging under intensive operational conditions [99].

Moreover, hierarchical network designs employing federated learning approaches further enhance scalability and operational feasibility. By preprocessing raw sensor data at the edge—such as compressing video streams or filtering telemetry data—these systems significantly reduce network bandwidth demands, allowing model updates and analytic processes to occur without sensitive raw data needing to exit the farm environment. This configuration effectively balances computational responsiveness, data security, and limited rural network infrastructure capacities [98].

Effective user interface/user experience (UI/UX) design remains fundamental for user acceptance and successful integration of livestock mobile applications into farm management practices. Agricultural usability studies consistently emphasize that farmers prefer intuitive interfaces closely aligned with their daily operational workflows and practical field conditions. Key design criteria highlighted in usability evaluations include clearly structured menus, rapid accessibility of essential functions, and adequately sized interactive controls (e.g., large buttons and clearly recognizable icons), facilitating quick, error-free interactions even while wearing protective gloves [100].

Empirical evaluations have repeatedly validated these design principles. For instance, a beef-management mobile application, developed collaboratively with farmers, reported high usability ratings (SUS scores exceeding 70) and substantial self-reported satisfaction, confirming effectiveness in real-world farm environments [101]. Additional design considerations crucial for practical farm deployment include high-contrast displays and minimal text reliance, ensuring readability under direct sunlight. Clear, simplified content structures allowing users rapid access to essential tasks without navigating through multiple screens further improve efficiency and satisfaction.

Localization support—including multilingual interfaces, regional terminology, and local measurement units—ensures broader usability across diverse international and multi-ethnic farming communities. Real-time feedback mechanisms, such as color-coded alerts, clear trend visualizations, and actionable prompts, further enhance usability, allowing farmers to quickly prioritize and manage animal care without extensive data interpretation efforts. Finally, interoperability with widely-used farm management systems, enabled through standardized APIs, significantly reduces redundant data entry tasks, providing veterinarians and farm advisors immediate access to unified, accurate records [101]. Finally, interoperability with widely used farm management systems, enabled through standardized APIs, significantly reduces redundant data entry tasks, providing veterinarians and farm advisors immediate access to unified, accurate records. Fig. 3 illustrates the interdisciplinary integration of animal science (facial AU biology), computer vision technology (RGB-thermal analysis), and precision-agriculture systems, with arrows showing data flow from capture to real-time welfare alerts.

Fig. 3. Conceptual illustration depicting interdisciplinary integration of animal science (facial action unit biology), computer vision technology (RGB-Thermal image analysis), and precision agriculture management systems. Arrows indicate the directionality of data flow from initial data capture through to generation of real-time welfare notifications.

Download Original Figure

Given the sensitive nature of farm production and animal health data, mobile applications must incorporate rigorous privacy and ethical safeguards. Industry best practices emphasize robust end-to-end encryption of data during transmission and storage, stringent user authentication protocols, and clearly defined role-based access controls distinguishing farm owners from employees [101]. Transparent data ownership policies and explicit user consent protocols further establish trust. Applications like PIGLOW utilize anonymous benchmarking systems, allowing users to compare welfare metrics confidentially, facilitating peer learning without compromising data privacy [97]. Regulatory guidelines also advocate comprehensive traceability features, including audit logging, tamper-evident record-keeping, and customizable data retention periods, ensuring compliance with mandatory animal welfare audit requirements.

Veterinary regulatory frameworks impose additional operational constraints. Many regions stipulate that remote monitoring applications must operate strictly within an established veterinarian–client–patient relationship (VCPR), clarifying that such tools complement rather than replace professional veterinary oversight. Consequently, clear liability disclaimers and predefined emergency flagging thresholds are mandated, ensuring users understand the supplementary role of AI-based alerts in clinical decision-making contexts [102]. Ethical considerations further encourage developers to adopt responsible innovation strategies, involving both veterinarians and farmers directly in application design and validation processes. Such co-creative approaches ensure technological advancements augment, rather than diminish, traditional farmer roles, preserving essential human empathy and local expertise in animal welfare practices [102].

Technical performance trade-offs in mobile application deployment

Mobile-optimized CNN architectures significantly outperform larger conventional models for animal pain detection applications. The technical trade-offs of mobile-optimized CNN architectures are summarized in Fig. 4. ShuffleNetV2 emerges as the optimal architecture, achieving 95.5% accuracy for pain classification with only 6.17 million parameters (approximately 25–30 MB) and 22 FPS inference speed on smartphones. EfficientNetB0 and MobileNetV3 also demonstrate strong performance with 65%–77% accuracy rates while maintaining practical deployment characteristics of 17–50 MB model sizes and 12–21 FPS processing speeds. In contrast, ResNet50-based approaches achieve only 65% accuracy with significantly larger memory footprints and slower inference speeds, contradicting claims that larger models offer superior performance for this application domain [74,92].

Fig. 4. Grouped bar chart illustrating comparative benchmarks for mobile application deployment of animal pain detection, presenting model file size (MB), pain detection accuracy (%), and inference speed (FPS) across ShuffleNetV2, EfficientNetB0, MobileNetV3Large, and ResNet-50 architectures. Benchmark metrics are based on published peer-reviewed evaluations, supporting informed model selection according to practical requirements for mobile veterinary AI applications.

Download Original Figure

Multiple studies document successful clinical deployment of mobile animal pain detection systems across various species, including cats (95.5% accuracy), sheep (92.7% accuracy), horses (88.3% accuracy), and rabbits (87% accuracy) [50,57,75,103]. These mobile-optimized systems demonstrate real-time processing capabilities, minimal battery consumption, and successful integration into veterinary clinical workflows with high inter-rater reliability. The research conclusively establishes that mobile-optimized CNN architectures are not only technically feasible for smartphone deployment but also achieve superior accuracy compared to conventional larger models while providing the computational efficiency necessary for practical veterinary applications.

Implementing AI-driven facial grimace scales in dairy cows raises profound ethical questions that go beyond technical issues like data privacy or compliance. Scholars emphasize that digital livestock farming can reshape the human–animal bond and risk treating animals as mere data points. For example, Neethirajan warns that “the use of artificial intelligence in digital livestock farming may lead to a loss of personal connection between farmers and animals,” potentially undermining animal well-being [104]. Similarly, recent reviews note that constant monitoring (“quantified” animals) can diminish caretakers’ empathy: as animals cannot consent to surveillance, and caretakers “might become overly reliant on graphs or dashboard alerts,” weakening the subtle, compassionate observation that traditionally guides animal care [105]. In short, high-level ethical reflection asks not only how AI tools function but whether they respect animals as sentient beings with interests. Ethicists point out that if AI focuses farm management solely on efficiency or productivity, it risks violating animals’ autonomy (treating them as instruments) and eroding virtues like compassion and responsibility [10,105]. A “should we” perspective thus urges that any pain-detection AI must be integrated in ways that support rather than replace the human–animal relationship [10,104].