Smartphone-based Diagnostic Platform for Bovine Health Monitoring Using AI Image Analysis

Thesis Abstract

The rapid expansion of mobile technology presents a transformative opportunity for timely bovine disease detection, yet current veterinary diagnostics often rely on centralized facilities, causing delays and increased costs for producers. This study investigates a smartphone-based diagnostic platform that leverages AI image analysis to monitor bovine health in real time, aiming to shorten diagnostic turnaround and improve welfare outcomes. The objective is to develop and validate an AI-assisted mobile workflow capable of classifying key bovine health conditions (digital dermatitis, bovine respiratory disease indicators, lameness) from wound and gait/images captured in raised-density farm environments, and to evaluate its impact on early intervention decisions and farmer adoption. Specific objectives include (1) to design a mobile app integrated with a convolutional neural network (CNN) trained on a diverse image dataset of 5,000 labeled bovine health instances; (2) to assess diagnostic accuracy against veterinary ground truth using sensitivity, specificity, and area under the ROC curve; (3) to examine user acceptability and decision-making efficacy through a mixed-methods approach; and (4) to model cost-benefit implications for farm operations under varying prevalence scenarios. The study adopts a pragmatic, mixed-methods research design underpinned by the Technology Acceptance Model (TAM) and the Diffusion of Innovations theory to explain adoption patterns, complemented by the Clinical Decision Support framework to interpret diagnostic performance. The population comprises commercial dairy and beef farms within a defined temperate region, with a target sample of 60 farms contributing 600 animals for image-based evaluation and 120 veterinary case confirmations for external validation. Data collection instruments include a smartphone-based imaging protocol, a labeled image repository created with veterinary collaboration, in-app diagnostic outputs, structured farmer questionnaires, semi-structured interviews, and veterinary records. Validity and reliability are ensured through cross-validation of the CNN across an independent hold-out test set (20% of images), inter-rater agreement checks (Cohen’s kappa) on expert labels, and pilot testing of the app’s usability with 10 producers. Data analysis employs descriptive statistics for dataset characteristics, diagnostic performance metrics (sensitivity, specificity, precision, F1-score, AUROC), and inferential tests (McNemar’s test for paired proportions). A hierarchical mixed-effects model analyzes factors influencing diagnostic accuracy, while thematic analysis interprets qualitative interview data. Model specification includes a multi-class CNN with transfer learning from ImageNet, fine-tuned on bovine health imagery, and an auxiliary gradient boosting classifier integrating contextual features (breed, age, housing, season) to improve decision support. Expected findings indicate that the CNN-based classifier will achieve an AUROC above 0.90 for digital dermatitis and 0.85 for lameness-related cues under field conditions, with sensitivity and specificity each exceeding 0.80 in most scenarios. The integrated decision-support layer is anticipated to increase timely treatment initiation by 25–35% and reduce producer travel costs by 15–20% compared with standard practice, while maintaining acceptable user satisfaction scores (Likert scale ?4/5). The study will reveal nuanced barriers to adoption, including perceived reliability, data privacy concerns, and workflow integration challenges, with determinants such as farm size, previous ICT exposure, and perceived economic benefit significantly predicting intent to use. Contribution to knowledge includes (i) a rigorously validated, scalable smartphone AI pipeline for real-time bovine health assessment in farm environments, (ii) empirical evidence on diagnostic performance of mobile AI with field-derived data, and (iii) an integration framework linking veterinary clinical relevance with technology acceptance and diffusion processes. The main conclusion posits that smartphone-based AI image analysis can complement veterinary services by enabling earlier detection and more efficient resource allocation, provided robust validation, user-centered design, and clear data governance. Recommendations address standardization of imaging protocols, ongoing model updates with new field data, training programs for producers, and policy considerations for data ownership and interoperability with existing farm-management systems. Further research should explore longitudinal impact on herd-level welfare metrics and economic return across diverse geographical settings.