relsa-severity-assessment

Use the RELSA procedure: normalize each variable to the animal's baseline, declare rising variables as turned, build a reference set from the highest-burden group, and run scripts/relsa_score.py. It outputs one score per animal per time point with per-variable weights.

Fit an ARIMA model to each animal's RELSA trajectory using scripts/forecast_relsa.py, which forecasts the endpoint score with a 95% prediction interval. Report RMSE, PICP, and MPIW together, and act on the upper interval bound since ARIMA cannot foresee abrupt deterioration.

The scoring and KDE threshold scripts need numpy, pandas, and scipy. Forecasting additionally requires statsmodels >= 0.14, and matplotlib is needed only for figures. Python 3.10 or later is required; no network access is needed.

A variable whose worsening direction is rising but is not declared as turned contributes exactly zero, because deviations in the wrong direction are floored at zero. Declare rising variables like clinical scores and biomarkers with --turned, and check the reference model table for plausible max reached values.

No. KDE-derived attention and danger zones are model-specific candidate thresholds on the RELSA scale, not regulatory severity gradings. EU categories such as mild, moderate, and severe are assigned prospectively by a separate process, and thresholds are sensitive to bandwidth choice.

Avoid relying on ARIMA when deterioration is abrupt, since the model assumes stationarity and linearity and cannot predict a cliff from a smooth prior trajectory. For sudden changes, consider Bayesian online changepoint detection or Markov switching models instead.