Predictive Mortality and Anomaly Detection Models

Predictive mortality models use historical water quality, feeding, and mortality data to flag batches or systems at elevated near-term mortality risk before a visible die-off begins, typically by learning what combinations of conditions, a specific dissolved oxygen trajectory, a feeding-rate drop, a temperature swing, historically preceded mortality spikes in that farm's own past data, rather than relying on a single fixed threshold that ignores how conditions are trending or interacting. Anomaly detection is a related but distinct technique: rather than predicting a specific known risk pattern, it flags when current readings deviate unusually from a system's own established normal range, catching problems a rule-based threshold system was never explicitly configured to look for, such as an unusual feeding-camera response pattern that precedes a disease outbreak days before mortality itself rises. Model quality depends entirely on having enough clean, well-tagged historical mortality and root-cause data to learn from, which is precisely why consistent mortality event logging and tagging discipline earlier in a farm's data pipeline directly determines whether predictive features built later actually work or produce unreliable, low-confidence alerts that staff learn to ignore. Deployed responsibly, these models are framed as an early-warning decision aid for experienced staff to investigate, not an autonomous diagnosis, since false positives and false negatives both carry real operational cost if the model's confidence level is not communicated honestly alongside its alert.