Time-to-Focus: Prosodic Milestones as Survival Events with Calibrated Brain-Penetration Covariates

When children learn to speak, they gradually acquire the ability to use their voice to highlight information — stretching the word "blue" in "I want the blue ball" to correct a listener who picked up the red one. This skill, called prosodic focus, matures slowly and unevenly, and in children with hearing loss or neurodegenerative disease it can be delayed by years or lost entirely. Developmental science currently studies it with cross-sectional snapshots: compare 4-year-olds to 5-year-olds to adults, and ask who has reached the adult pattern. This proposal instead treats the moment a child first produces adult-like focus marking as a survival event — the same kind of "time until something happens" outcome that medicine uses for death or disease onset — because many children simply have not yet reached the milestone when a study ends, and current methods either discard them or mislabel them as failures. The third ingredient enters as a covariate: for children whose speech development depends on drugs or gene therapies reaching the brain, we ask whether a computationally predicted, uncertainty-calibrated blood-brain barrier penetration score — an estimate of whether a molecule crosses the vessel wall that isolates brain tissue — changes the hazard of reaching or preserving speech milestones.

The two-sentence version: we rebuild the analysis of prosodic-focus acquisition around proportional hazards models (the statistical workhorse of clinical time-to-event data), where the "event" is crossing a quantitative acoustic threshold and the time axis is auditory experience rather than birth age; and we replace the crude binary "crosses the barrier / doesn't" drug labels used in existing dementia survival studies with continuous, confidence-weighted penetration predictions from molecular machine learning, then test whether that replacement improves risk estimates for speech and neurological outcomes.

Core Claim

Two literatures sit close to this idea but neither makes the move. Clinical epidemiology has already linked barrier penetration to time-to-dementia: Lower Risk of Alzheimer's Disease and Related Dementias associated with Angiotensin Receptor II Blockers (ARBs) treatment for individuals with Hypertension in High-volume Claims Data uses Cox proportional hazards models — where the instantaneous event rate is written as a baseline rate $h_0(t)$ multiplied by $\exp(\beta^T X)$, so covariates scale risk rather than determine it absolutely — and separates brain-permeable from non-permeable blood pressure drugs with a binary label. That study motivates our covariate idea but is limited to a two-valued penetration class with no pharmacokinetic confirmation and no behavioral endpoints. Separately, Prosodic Focus Acquisition in French Early Cochlear Implanted Children shows that deaf children implanted before 18 months gradually acquire adult-like contrastive duration lengthening, reaching it around age 7 (hearing age roughly 6 years, 2 months) — but analyzes this with mixed-effects regression on age strata of 2–3 children, excluding the 4-year-olds entirely for task compliance. That study supplies our endpoint but not the time-to-event machinery, because attrition and "not yet" outcomes were handled by exclusion rather than by censoring.

Our claim is that both gaps close at once. Defining the event as an acoustic threshold crossing lets censored children contribute partial information; and making the treatment covariate a calibrated penetration score rather than a binary class connects molecular computation to behavioral trajectory. The biological plausibility of that link is documented: in AAV-mediated ARSA replacement for the treatment of Metachromatic Leukodystrophy, a barrier-crossing engineered viral capsid (AAV.GMU01) restored enzyme activity across the brain and prevented auditory threshold shifts in a mouse model of a pediatric disease whose symptoms include progressive speech and hearing loss. Barrier delivery is the rate-limiting step for preserving the auditory substrate on which prosody is built. We do not claim deep non-linearity is the point — on standard survival benchmarks the deep Cox network DeepSurv gains only modestly over linear Cox (concordance index 0.643 versus 0.630 on METABRIC, from Deep Survival: A Deep Cox Proportional Hazards Network). The value here is the event definition, the censoring-correct time axis, and honest covariate calibration.

Research Program

The program has three components. First, we formalize the milestone event. Following the French cochlear-implant paradigm, each child completes an elicitation game producing broad focus (BF), contrastive narrow focus on noun or adjective (FN/FA), and corrective narrow focus (CFN/CFA) utterances; we compute a focal-lengthening ratio $\rho = \text{mean duration of focused words} / \text{mean duration of the same words under broad focus}$. A session counts as an acquisition event when $\rho$ exceeds the lower bound of the adult control distribution for the corrective condition. The hazard model is then

$$h(t \mid X) = h_0(t)\,\exp\big(g(X;\theta)\big),$$

where $t$ is hearing age (time since auditory access, whether implantation or birth), $h_0$ is an unspecified baseline left free, and $g$ maps covariates — duration of auditory deprivation, cue-availability (cochlear implants transmit pitch poorly, forcing a duration-only pathway, consistent with Acoustic correlates of contrastive stress in German children, where vowel duration is the only stable cue for 4-to-6-year-olds), and respiratory capacity (children sustain 5–7 syllables per breath versus adult 14–16, a constraint shown to matter in Predicting Prosodic Boundaries for Children's Texts) — to a relative risk score. Children left-truncated at study entry and right-censored at last session are retained, not dropped.

Second, we build the penetration covariate. We reuse the uncertainty machinery of Blood-brain barrier penetration prediction enhanced by uncertainty estimation: graph neural network predictions of barrier penetration, trained on the refined MoleculeNet-derived M-data (1,937 molecules) and validated on the independent clinical S-data (527 molecules), filtered by combined Shannon-entropy and MC-dropout confidence so that low-confidence scores are shrunk toward neutral rather than propagated. This is essential because the same paper showed that uncalibrated deep models collapse to simple-descriptor parity on independent clinical data, and the whole point of our pipeline is to avoid injecting that false confidence into hazard estimates.

Third, we confront non-proportionality head-on. The constant-hazard-ratio assumption is known to fail in exactly our two regimes: prosodic discourse intonation undergoes a sharp transition near age 7 (A Longitudinal Study of Children's Intonation in Narrative Speech), and barrier-disruption effects after hemorrhage show an acute peak within 3 weeks followed by chronic elevation over months (Blood–brain barrier opening as a predictor of epilepsy and mortality after subarachnoid haemorrhage). We therefore build in time-varying coefficients and compare against existing non-proportional machinery from A Federated Cox Model with Non-Proportional Hazards, which supplies the estimation technique we adapt rather than a competing claim.

Proposed Experiments

Experiment 1 re-analyzes the existing French cochlear-implant cohort under the milestone-hazard framing. With only 10 children this cannot yield publishable hazard estimates by itself; instead we use it to calibrate a parametric simulation (event-time distributions matched to the observed cross-sectional attainment and the age-7 asymptote), run power analyses, and publish the event-definition protocol. No reported result found on Lacuna for a hazard ratio or event time of prosodic-focus mastery — that absence is precisely the gap.

Experiment 2 collects the longitudinal cohort the design needs: roughly 50 early-implanted children and 30 typically developing controls, quarterly sessions over three years, using the BF/FN/FA/CFN/CFA elicitation plus the leveled story-reading task from Predicting Prosodic Boundaries for Children's Texts (54 stories, 8,662 words, 21 annotators, Fleiss' Kappa 0.68). Automated boundary segmentation must approach the reported LightGBM performance (boundary detection F1 of 0.89, forbidden-pause detection F1 of 0.87) before we trust it for scoring; if it falls short, human annotation backs it up. Outcomes are the estimated acquisition curve with censoring-correct confidence bands, and a concordance index for predicting which children master focus early versus late.

Experiment 3 tests the penetration covariate on the pharmacological side. We assemble a retrospective cohort patterned on the ARB dementia design, with endpoints of time-to-cognitive-or-communicative decline, and run three covariate arms on identical data: binary barrier-crossing labels (the existing approach), raw continuous penetration predictions, and uncertainty-weighted continuous scores. The penetration model is evaluated first on held-out benchmarks — the reference standard is HimNet's ROC-AUC of 0.954 on scaffold-split MoleculeNet BBBP from Learning Hierarchical Interaction for Accurate Molecular Property Prediction, and the calibration target is GROVER-BBBp's above-99% accuracy on predictions with uncertainty below 0.1 from the Tong et al. paper. The decisive comparison is concordance across the three arms.

Experiment 4 runs non-proportionality diagnostics (Schoenfeld-type residual tests, time-varying coefficient fits) on both the prosody hazard and any barrier-related hazard, reporting where proportionality fails and how much time-varying terms recover.

Baselines

The naive reference point is what the field does now: age-stratified group means with mixed-effects regression, as in the French cochlear-implant study, which reports that adult-like corrective-focus lengthening emerges by age 7 (hearing age about 6 years, 2 months) but reports no numeric coefficient or event-time estimate we can cite — no reported result found on Lacuna for the underlying regression coefficients. This baseline answers "what proportion of children at each age look adult-like," and our method beats it if it reproduces that age-7 pattern while additionally producing unbiased acquisition curves for the children the stratified design excluded, validated by simulation with known ground truth.

The strongest existing competitor is linear Cox proportional hazards with binary barrier-penetration drug labels — the exact design of the ARB dementia study, extended with our prosodic covariates. On the machinery benchmark METABRIC, Deep Survival reports linear Cox at a concordance index of 0.630 and DeepSurv at 0.643, so we expect the deep versus linear gap to be small and do not rest the proposal on it. We run both on the same splits with matched tuning budgets (grid search over covariate sets for the linear model; matched epoch and dropout budgets for the deep model, per the DeepSurv recipe), and the criterion for beating the baseline is a concordance-index improvement whose bootstrap confidence interval excludes zero, plus better calibration of predicted versus observed acquisition curves. A second fixed-window competitor is binary logistic regression on a 3-week horizon, which achieved an AUC of 0.866 for early death in the subarachnoid-hemorrhage cohort; our time-to-event version must at least match its fixed-window AUC while handling the heterogeneous follow-up durations that logistic regression cannot.

The key ablation isolates calibration: the identical hazard model, same data and budget, run once with raw graph-network penetration scores and once with uncertainty-weighted scores (the moderated-Z filtering of the Tong et al. pipeline). If the raw scores win, the covariate-calibration ingredient is unsupported and we will report that; if the weighted scores win, calibration is doing the work we claim. A second ablation drops all acoustic covariates to isolate how much the prosodic trajectory itself contributes beyond demographics and deprivation history.

Why not just…?

Why not just use mixed-effects logistic regression, as the field already does? Because the failure mode it creates is visible in the source study itself: 4-year-olds who could not complete the task were excluded, and 2–3 children per age stratum forced coarse grouping. A child observed at 40 months of hearing age who has not yet acquired focus is not a failure — she is a censored observation whose eventual acquisition time is unknown but informative. Logistic regression must either drop her or count her as a non-acquirer, biasing the estimated acquisition rate, a limitation documented for binary endpoints generally in the survival literature compiled here. We are honest that survival models alone do not fix a sample of ten; that is why Experiment 1 is a power analysis and Experiment 2 is new collection, not a re-labeling exercise.

Why not just keep binary drug categories, as the ARB dementia study already does? This is the most serious objection, because the binary design is simple, defensible, and already published. The counter-evidence is Blood-brain barrier penetration prediction enhanced by uncertainty estimation: uncalibrated deep penetration predictions performed no better than simple physicochemical descriptors on independent clinical data (their "DL hype check"), so naively upgrading the covariate to continuous scores would inject overconfident noise into hazard estimates — exactly the failure the uncertainty filter exists to prevent. Experiment 3 settles the question empirically and symmetrically: if binary labels match or beat calibrated continuous scores in concordance, the simpler design wins and we will say so; the proposal's value then rests on the milestone-hazard framing alone. The ARB study also lacked any pharmacokinetic confirmation of brain exposure, a gap the continuous score directly addresses, though we cannot manufacture the missing unbound-brain-concentration datasets — large public ones do not exist yet.

Why not just rely on text, punctuation, or standard clinical endpoints instead of acoustic prosody? Punctuation misses nearly 30% of natural prosodic pauses in the children's story-reading corpus, and the syntactic text is literally identical across the broad, contrastive, and corrective focus conditions in the French paradigm — only the acoustics distinguish them, so no text-based measure can encode the event at all. The production–function dissociation (Prosodic marking of continuation versus completion in children's narratives) — children who can insert pauses but place them inappropriately — means milestone definitions must be functional, which requires listening. Standard clinical endpoints like dementia diagnosis, used in the ARB and atherogenic-index studies, arrive years after the speech-motor changes that precede them, which is the whole argument for the earlier, cheaper acoustic signal.

Decisive Checks

The proposal stands or falls on three questions. Does the milestone-hazard model recover the documented age-7 acquisition pattern while correctly handling simulated censoring that breaks the stratified baseline? Does the uncertainty-weighted continuous penetration covariate beat binary drug labels in concordance on the retrospective cohort, with the calibration ablation confirming which ingredient carries the gain? And where proportionality fails — near the age-7 developmental transition and across the acute-to-chronic barrier-disruption window — do time-varying coefficients measurably recover fit? A clear no on the second question would reframe the work as a developmental-speech methodology contribution with the penetration covariate left as a negative result; a no on the first would end it.