Abstract
Developmental dyslexia is classically divided into Boder's dysphonetic (DD), dyseidetic (DYD) and mixed (MD) subtypes, assigned by direct reading-and-spelling testing. Its computerised Italian adaptation, the Direct Test of Reading and Spelling (DTRS), assigns the subtype algorithmically from the reading quotient and the percentage of correctly spelt known and unknown words. Using the same source clinical cohort and the same stable and sparse regression classifier (SSRC) as our companion, we asked whether the subtypes form natural clusters or lie on a continuum, and whether reading-time (severity) or qualitative-error (route) information discriminates them. We analysed one DTRS assessment each from 520 children (378 DD, 41 DYD, 101 MD), deriving reading-time indices, including a severity/fluency index (SFI), and Boder's error taxonomies for reading and spelling. Subtype structure was probed with Gaussian-mixture clustering (adjusted Rand index against clinical labels) and Hartigan's dip test; discrimination with the SSRC and robust ROC under participant-level Monte-Carlo cross-validation, with subsampling stability intervals (5th-95th percentile over 300 iterations). Full-sample pairwise AUCs were additionally evaluated with percentile bootstrap 95% confidence intervals (10,000 subject-level resamples of the full-sample classifier scores) and label-permutation tests (10,000 permutations, one-sided and Bonferroni-corrected within each model). No feature space yielded label-aligned mixture solutions (adjusted Rand index ≈ 0); SFI was unimodal (dip p = 0.36). Reading-time indices separated the severity contrast (DD vs MD, stable AUC 0.73 [0.66-0.80]; p < 0.001) but were weak and non-significant for the route contrast (DYD vs DD, 0.58 [0.47-0.69]; p = 0.543, non-significant); reading qualitative errors showed a weak route signal (0.61 [0.50-0.73]; p = 0.062) that did not reach Bonferroni-corrected significance. Unknown-word spelling, the basis of the diagnostic algorithm, separated DD from DYD almost perfectly (0.94-0.98), but partialling out unknown-word accuracy collapsed this to 0.60: the separation is definitional. The severity contrast was likewise definitional: the reading quotient alone separated DD from MD at 0.86, and residualising the reading-time indices on the reading quotient within each training fold reduced the DD-MD AUC from 0.73 to 0.50. In the EEG subsample, no behavioural classifier separated DD from DYD significantly (TIME: 0.52 [0.31-0.75], p = 0.41; ERR: weak and non-significant 0.62 [0.41-0.81], p = 0.15), whereas qEEG cross-spectral connectivity did so at 0.85 in the published qEEG sample of 227 children (169 DD, 18 DYD, 40 MD; Chiarenza et al., 2026). No evidence for natural clusters emerged with the present methods: the subtypes behave as overlapping regions of graded continuous distributions. Both axes of Boder's scheme are threshold-defined - severity by the reading quotient, route by unknown-word spelling - and both are continuously distributed. Once the defining variable is removed, reading dynamics carry no residual severity signal (0.50) and independent behaviour carries only a faint route signal (0.58-0.61); the route distinction is clearest in brain connectivity, which predicts the behaviourally defined labels rather than independently validating a latent route construct. Together, the two papers advance a dimensional, mechanism-aware account of dyslexia subtyping.