Motor trajectories in early-treated pediatric SMA
Spinal muscular atrophy used to be defined by what children lost. Newborn screening and modern therapy changed the question: these children do well, so what still separates one outcome from another?
Read this firstNot a predictor, on purpose
With thirteen patients, a model that returns a number for an individual child would overfit, and it would look convincing while doing it. So this is characterisation: how the cohort moves over time, which endpoint carries the signal, and a reference band a clinician can read a new child against. The individual predictor is the multi-centre sequel, and we say so rather than shipping one early.
The study runs with Neurologia Pediatrica at Policlinico “Giovanni XXIII” in Bari, with Dr. Valentina Fioretti as clinical lead.
CohortA screening-detected population
Puglia screens newborns for SMA. This cohort is therefore largely presymptomatic: first assessment at roughly 3 to 18 days old, treatment at 10 to 23 days, diagnosis by screening rather than by clinical signs. Motor function is tracked on CHOP-INTEND in infancy and HFMSE later, across two treatment arms, with a short oral bridge before gene therapy in a few cases.
That context is the reason almost everyone recovers, and it is the first thing any reader should know before seeing the results. It also fixes the external validity: these findings describe exactly today’s real-world scenario in the region, not a historical untreated population.
FindingOne factor separates outcome
SMN2 copy number is the single robust signal, and it holds in both treatment arms, so it is not an artefact of one drug.
Children with two copies recover but tend to plateau below the ceiling on HFMSE, sometimes with late functional loss; three and four copies reach the ceiling and hold it. A Bayesian two-group estimate puts the difference at +16 HFMSE points, 95% credible interval [+6, +26], with a 99.8% posterior probability that the higher-copy group sits above the lower. A literature-informed prior, borrowing published levels from the presymptomatic trials rather than their patient data, returns the same +16.
What the infant scale hides
On CHOP-INTEND nearly everyone reaches the 64-point ceiling within months, which reads as uniform success. The spread only reappears on HFMSE. And the baseline score, the number most readily to hand at diagnosis, does not predict the long-term endpoint at all (rho 0.08). The prognosis rides on the genotype, not on how the child scored on day one. That corrected an earlier overstatement of our own.
FindingThe same gradient, in a separate endpoint
The pre-specified primary endpoint is age at independent walking, recovered objectively from dated functional flags in the treatment registry rather than from recall. The SMN2 gradient shows up again, in a measure with no scale ceiling at all:
| Independent walking | Median age | Reference: NURTURE |
|---|---|---|
| 2 copies | 24.4 months | 20.4 months |
| 3-4 copies | 16.4 months | 12.3 months |
Same direction, similar magnitude, our real-world cohort slightly later than the trial. Head control arrives early and is copy-independent; the gradient emerges at the later, harder milestones. Several ages are interval-censored upper bounds, widened by visit spacing, and we report them as such.
MethodWhere we refused to overreach
- Ninety assessments are not ninety patients. They are repeated measures nested in thirteen children. A mixed-effects model puts the ICC at 0.39, a design effect of about 3.3: the information content of roughly 27 independent observations, and for a between-patient factor still just thirteen.
- Formal significance is not there, and we say so. Under correct clustering the interaction with the long-term endpoint is directionally right but not significant (p 0.20). The descriptive pattern is clear; repeated measures tighten the longitudinal picture, they do not manufacture patients.
- No multivariable model. Thirteen patients affords roughly one predictor, not six. We report ranked univariate associations instead, and flag the moderate correlations for gestational age and age at first assessment as almost certainly confounded: exactly the false signals a fitted model would seize on.
- Clusters tested for stability. Consensus clustering over 1000 subsamples separates the structures that survive every resample from the pair that co-clusters only because their follow-up is a month long. The fragile one is reported as an artefact, not a phenotype.
DataGovernance
This study handles real identifiable health data of minors with a rare disease, under GDPR and Italian medical-secrecy rules. Nothing identifiable is committed to the analysis repository: patients are keyed by opaque sequential study IDs, the linking key lives outside the repository as a secret, and free-text clinical narrative is never ingested into the modelling tables.
This page carries no patient-level content of any kind: no study IDs, no per-child figures, no cohort plots. A rare disease, one region and a cohort this small is not a combination to publish points from.
NextWhere this goes
A WHO-anchored parent questionnaire on milestone achievement and regression is ethics-approved and live, feeding the walking endpoint. The reference bands recompute from the cohort as it grows.
The individual trajectory predictor is a multi-centre job: more children, and real spread in timing, gestation and presentation, which a uniformly early single-centre cohort cannot provide.
Talk to us about this work