Information Transfer Rate (ITR)
The standard throughput metric for BCI communication, expressing how much information the system conveys per unit time by combining accuracy, the number of choices, and selection speed. The Wolpaw formula gives bits per selection as B = log2(K) + P·log2(P) + (1 − P)·log2((1 − P)/(K − 1)) for K equiprobable classes at accuracy P, and multiplies by selections per minute to get bits/min. It rewards more classes and higher accuracy and penalizes slow selection.
ITR is invaluable for comparing spellers but rests on strong assumptions: equiprobable, independent selections, a memoryless symmetric-error channel, and no benefit from error correction or language models — all violated by real systems. The Wolpaw form also assumes uniform confusion across the K − 1 wrong classes; when errors are structured, a confusion-matrix mutual-information (Nykopp) estimate is more honest. Report the definition, K, accuracy, and timing, because ITR values are not comparable across differing assumptions.
A 4-class SSVEP BCI at 95% accuracy with one selection every 2 s yields about 1.66 bits/selection × 30 selections/min ≈ 50 bits/min; dropping to 80% accuracy roughly halves the effective rate.
ITR combines accuracy, number of classes, and speed.
ITR counts raw channel information, not task-completed throughput. A predictive-text speller can produce far more characters per minute than its ITR suggests, so ITR undersells real communication aids while overselling lab tasks that have no error cost.