Company
Office
© 2026 Graphity Limited. All rights reserved.
A time series is a sequence of measurements taken over time – for example, an animal patient's heart rate recorded every hour, daily blood test results, or continuous telemetry from a veterinary ICU monitor. Time series analysis uses statistical and machine learning models to find patterns, predict future values, detect anomalies, or classify the patient's state (e.g., "stable" vs "deteriorating"). In a triage system, this provides a dynamic view of the patient's condition, complementing static inputs like symptoms or lab results from a single time point.
Deterioration is a process, not an event. A rising heart rate or falling blood pressure often develops over hours. Time series models can detect subtle trends that single measurements miss – for example, a heart rate creeping upward for six hours while each reading is still within normal limits.
Species‑specific baselines. A dog's normal temperature differs from a cat's. Time series models can be trained on species‑specific data and adapt to individual baselines – enabling personalised monitoring.
Forecasting (prediction): Given past vital signs, what will the next value be? A large forecast error signals an anomaly. Forecasts can anticipate future instability, such as a predicted drop in blood pressure.
Anomaly detection: Identify unusual patterns – a sudden spike, prolonged flatline, or gradual drift. Anomalies can trigger alerts for arrhythmia, apnoea, or decompensated shock.
Change point detection: Pinpoint the moment when a patient's state changes significantly (e.g., from stable to deteriorating). This helps evaluate treatment effects.
Classification of entire sequences: Given a 24‑hour recording, classify whether the patient is in pain, has sepsis, or is recovering well – turning a long time series into a diagnostic label.
Traditional statistical models (e.g., ARIMA, exponential smoothing): These are well‑understood and work well for regular, univariate time series like hourly temperature. They are interpretable and require little data, but struggle with complex, multivariate, or irregularly sampled sequences.
Long short‑term memory (LSTM) networks: A type of recurrent neural network (RNN) specifically designed to remember information over long periods. LSTMs maintain an internal "memory" that updates with each new measurement – allowing them to link a low heart rate 12 hours ago to a current fever. They excel at detecting subtle trends and anomalies in streams of vital signs, and can handle multiple input signals (e.g., heart rate, respiratory rate, temperature) simultaneously.
In practice, LSTMs are a go‑to choice for continuous monitoring in veterinary ICUs because they balance predictive power with reasonable training data requirements. The model's ability to learn from long sequences makes it particularly good at early warning – detecting deterioration before clinical signs become obvious.
The choice of model ultimately depends on the task (forecasting vs anomaly detection), sampling frequency, available labelled data, and need for interpretability. For many triage applications, LSTMs or simpler baseline models provide reliable performance.
A practical time series monitoring pipeline for veterinary triage:
Because time series models run continuously, the system can detect deterioration minutes or hours before it becomes clinically obvious – enabling pre‑emptive interventions such as moving the patient to intensive care or alerting the veterinary team.
Time series analysis – a key component of modern veterinary triage systems.
