Machine learning discriminates bacterial, fungal, and viral infections using temporal blood analyte dynamics in bottlenose dolphins (Tursiops truncatus).

OBJECTIVE: To determine underlying infectious disease as bacterial, fungal, or viral in origin from longitudinal assessment of biochemistry and hematology in bottlenose dolphins (Tursiops truncatus). METHODS: Blood samples, including up to 63 blood analytes, obtained from 1995 through 2025 from professional-care dolphins with 33 confirmed disease episodes (11 bacterial, 10 fungal, and 12 viral) were included in this retrospective, longitudinal, observational study. Random forest models were trained on 5 defined temporal phases of infection, selecting 34 blood analytes as key features in predicting the most underlying pathogen from routine blood results. Model parameters included the analyte's absolute value in addition to temporal slope (change in analyte over time), clinical ratios, and statistical aggregates to create a classification model to predict the most likely underlying pathogen from routine blood results. RESULTS: 860 blood samples were included from 31 dolphins. The model achieved 75.8% accuracy in pathogen classification (25 of 33 episodes), with disease-specific performance of 80% for fungal, 75% for viral, and 72.7% for bacterial. Temporal slope features, particularly eosinophil and total WBC count rate of change, were selected in 96.8% of validation folds. CONCLUSIONS: Temporal shifts in disease-specific biomarkers can provide superior diagnostic information compared to single-time-point measures or static threshold ranges. CLINICAL RELEVANCE: Personalized medicine involving longitudinal blood sample monitoring in marine mammals provides the opportunity to detect subtle shifts in blood analyte levels over time specific to that individual. This can be used to discriminate between underlying pathogens, identifying the likelihood of bacterial, fungal, and viral infections.