Reading pressure scatterplots and agreement plots AI
The reference-versus-prediction view plots the same held-out pressure pairs used for scoring. The diagonal marks exact agreement. Absolute coordinates include large differences in mean pressure between people; a visually tight overall scatter need not mean the model tracks small within-person changes well.
Calibration-relative coordinates subtract each person’s early reference-pressure average from both reference and prediction. Using the same offset on both axes preserves errors and makes deviations from calibration easier to inspect. It does not recalibrate the model on test pressure.
Error agreement
A Bland–Altman-style view plots signed error against the average of prediction and reference. Its central line is pooled bias; the other lines are bias ±1.96 sample error SD for the displayed points. These are descriptive limits of agreement, not confidence intervals, guaranteed coverage or a validated repeated-measures agreement analysis. Repeated, overlapping windows are dependent.
Colors identify people. Selecting one participant reveals reference and predicted time series, which can expose an offset, a missed pressure change or a large individual failure. Figure selection changes the displayed subset; the main accuracy table remains a full-cohort summary. Predictions are not clipped.
See prediction-error definitions and calibration. The classic method is described by Bland and Altman (1986).