Fig. 5: Inspection of the best prediction model.

A Correlation of the predicted and observed reading speed difference, pre/post training from the best prediction model. The dashed line indicates the decision boundary (i.e., cases with >13.5% predicted reading speed increase are considered responders, established based on a sensitivity of 0.73 and specificity of 0.74; see Supplementary Fig. 2), and colors mark the true positive (Correctly predicted responders), true negative (Correctly predicted non-responders), false positive (Incorrectly predicted non-responders), and false negative cases (Incorrectly predicted responders). The black line indicates the overall correlation, and the gray area reflects the 95% confidence interval. B Feature relevance of all included features based on t values (i.e., signed median and standard deviations across 75 cross-validation runs including significance markers, red dashed line > 2 or < −2). LCM: lexical categorization model uncertainty effect, Lex: lexicality effect, Index: sequence index effect, OLD20: word-likeness effect based on the OLD20 measures⁴⁶, Frequency: word frequency effect based on SUBTLEX-DE measure⁴¹, Week: week of training effect (1 vs. 2), SLS: incoming reading speed effect (i.e., adult version of Salzburger Lesescreening⁷¹), Errors: Correct vs. incorrect lexical decisions in the training task, interaction sign “x'': interaction added at Level 2 during the feature selection, interaction sign “:'': interaction fitted with the random effect structure. C Applying the best-performing diagnostic model for categorization. Benefits in reading speed when no diagnostics applied vs. when diagnostics applied (i.e., if the training included only readers of which the model would predict learning success).