Smartphone-based AI Diagnostics for Bovine Health Monitoring under Farm Conditions

Thesis Abstract

This study addresses the critical need for rapid, scalable, and cost-effective bovine health monitoring in farm settings through a smartphone-based artificial intelligence (AI) diagnostic platform. Delays in detecting common health issues such as mastitis, bovine respiratory disease, lameness, and metabolic disorders often lead to reduced productivity, increased veterinary costs, and elevated animal welfare concerns. The aim is to develop, validate, and deploy an AI-assisted diagnostic workflow that leverages smartphone-acquired images, audio cues, and wearable sensor data to identify early signs of illness in dairy cattle under real-world farm conditions. Specific objectives include (1) designing a mobile data-collection protocol that integrates multimodal inputs (facial/udder imagery, gait and posture videos, vocalizations, and body temperature via connected thermometry), (2) constructing and validating machine learning models—convolutional neural networks (CNNs) for image/video data, recurrent neural networks (RNNs) for time-series sensor data, and ensemble fusion for decision making—(3) evaluating model performance against gold-standard veterinary diagnoses, (4) assessing user acceptability and workflow integration among farm personnel via mixed-methods evaluation, and (5) performing a cost-benefit analysis to determine adoption viability. The methodological framework combines a pragmatic mixed-methods approach with a diagnostic accuracy study design. The population comprises lactating Holstein dairy cows across five commercial farms in a temperate agricultural region, with a target sample of 1,200 cow-days for imaging and sensor data collection. A stratified sampling strategy ensures representation of key health states (healthy, subclinical, and clinically ill). Data collection instruments include a smartphone app for image, video, and audio capture; Bluetooth-enabled wearables measuring rumination, activity, and body temperature; calibrated infrared thermometers; and veterinary clinical records for gold-standard labeling. Data preprocessing employs standardized augmentation, noise reduction, and temporal alignment to synchronize multimodal streams. Validity and reliability are established through cross-validation of models, test-retest reliability of measurements, and inter-rater reliability for veterinary diagnoses. Analytical methods comprise a hierarchical modeling pipeline CNN architectures (e.g., EfficientNet) for static imagery, 3D-CNNs or Transformer-based video models for dynamic footage, LSTM or Temporal Convolutional Networks for time-series sensor data, and late fusion to produce an aggregated health risk score. Model evaluation uses sensitivity, specificity, area under the receiver operating characteristic curve (AUC-ROC), and F1 scores, with bootstrap95% confidence intervals. Calibration assessments (reliability diagrams, Brier score) and decision-curve analysis will quantify clinical utility. The theoretical underpinning draws on the Technology Acceptance Model (TAM) to interpret user adoption, and the Diffusion of Innovations (DOI) framework to explore factors affecting farm-level integration. A qualitative component based on thematic analysis of semi-structured interviews with farmers and veterinarians elucidates workflow barriers and facilitators. Expected findings include high diagnostic performance for several conditions subclinical mastitis (AUC-0.90–0.95, sensitivity ~0.88, specificity ~0.92), early respiratory disease indicators (AUC 0.85–0.90), and locomotion abnormalities associated with lameness (precision-recall balance improving with multimodal fusion). Multimodal models are anticipated to outperform unimodal counterparts, with late fusion yielding the most robust performance. The study also anticipates that user perceived usefulness and ease of use will correlate positively with adoption intentions, moderated by perceived reliability and integration depth into existing farm management systems. The contribution to knowledge includes demonstrating the feasibility and accuracy of on-device, AI-driven bovine health diagnostics under commercial farm conditions, providing a replicable multimodal data-collection and modeling framework, and offering empirical evidence on economic viability for ICT-enabled veterinary surveillance. The study concludes that smartphone-based AI diagnostics can significantly enhance early disease detection, enabling timely interventions, reducing antibiotic use, and improving animal welfare and farm profitability. Recommendations include standardizing data governance and privacy protocols, integrating the diagnostic outputs with existing farm management information systems, developing tiered alert levels aligned with veterinary thresholds, and conducting longitudinal validation across diverse breeds, management practices, and ecological regions to generalize the approach.