Predicting Sitting After Nusinersen

This single-study explainer examines a machine-learning survival-prediction analysis of sitting after nusinersen treatment in non-sitter patients with spinal muscular atrophy. The investigators drew participants from SMA Reach UK and the Italian Telethon Network, included children treated with nusinersen under 4 years of age, and focused on whether participants subsequently achieved sitting and on predicted time-to-sitting. The work models heterogeneity within this treated cohort; it does not compare nusinersen with an untreated group. (pmid:42900957)

Evidence

The study included 124 non-sitter spinal muscular atrophy participants, each with a median of 7 assessments. Participants were randomly allocated into training and testing datasets in an 80/20% split. The source describes an average follow-up time of 2.6 years, with an interquartile range of 1.5 to 4.1 years. Thus, the reported evaluation was an internal training-and-testing exercise within patients from the two named clinical networks rather than an external validation in a separate dataset. (pmid:42900957)

The investigators considered tree-based, conditional-inference-tree-based, and regularised regression-based approaches for survival outcomes. They selected oblique random forests. Candidate features came from items in the reported CHOP-INTEND and HINE-2 motor-function assessments, respiratory status, and swallowing status, with mutual information used for feature selection. For participants who initiated zolgensma or risdiplam, follow-up was censored 4 months after the last nusinersen dose or at first non-nusinersen dosing, whichever occurred earlier. This handling defines how later treatment changes entered the time-to-sitting analysis. (pmid:42900957)

Sitting was observed in 62 participants, representing 50% of the cohort, at a median age of 2.4 years. The selected model had a testing concordance index of 0.74. That metric reports the model’s discrimination in the held-out testing portion of this dataset; it is not a direct measure of clinical benefit, nor does it establish accuracy in another care setting or population. The study’s outcome was the ability and timing to sit among treated non-sitters, rather than a broad measure of overall disease progression. (pmid:42900957)

Feeding support status was reported as the most influential factor for predicted time-to-sitting. Participants requiring tube feeding at treatment initiation had a predicted median time-to-sitting 4 months later than those without tube feeding. In a related baseline comparison considered in isolation, feeding support and swallowing problems were associated with 4- and 3-month delays in median time-to-sitting, respectively. These are modeled and observed associations within the study cohort, with feeding and swallowing measures considered alongside motor, respiratory, and other recorded features. (pmid:42900957)

The report also identifies specific motor-function features, including the ability to kick in supine, as strongly associated with a higher likelihood of sitting. In individual predictions, a feeding tube had substantial influence on predicted time-to-sitting among participants with intermediate and weak baseline function. The authors interpreted feeding support as potentially representing underlying baseline disease severity, rather than treating support itself as the mechanism determining sitting. This distinction is important because the measured predictor may stand in for severity-related characteristics not fully captured by the model inputs. (pmid:42900957)

Analysis — Prediction and association

The central pattern is one of stratification within a treated, non-sitter spinal muscular atrophy cohort: baseline bulbar-related measures and selected motor features differentiated modeled prospects for sitting and modeled timing to sitting. The testing concordance index of 0.74 indicates that the chosen oblique random-forest model had reported discriminatory ability in its internal testing dataset. However, the train/test allocation evaluates a prediction process, not the effect of nusinersen or of feeding support. Because every participant in the analyzed cohort received nusinersen, the design contains no untreated comparator for attributing observed sitting outcomes to treatment. (pmid:42900957)

The observed pattern is consistent with the reported design and selection process: children entering these two networks as non-sitters and treated under 4 years were assessed repeatedly, and the model used baseline and longitudinally available clinical features to predict a survival-type outcome. Feeding support and swallowing problems may identify participants with more severe baseline disease, as the authors note. Consequently, their association with later predicted sitting should not be read as evidence that these characteristics cause the outcome pattern. The analysis is most appropriately interpreted as an exploratory, internally evaluated prognostic model for this defined cohort. (pmid:42900957)

Limitations

The observational prediction design cannot establish causation. The source explicitly notes that its methodology cannot determine whether an individual feature causes a lower likelihood of sitting, and describes latent disease severity as a plausible common factor related both to feeding support and sitting. The cohort also has no untreated comparator, so the reported results compare modeled outcomes across characteristics among nusinersen-treated participants rather than across treatment strategies. What was measured was sitting and time-to-sitting, with clinical features including motor-function items, respiratory status, swallowing status, and feeding support. (pmid:42900957)

Selection and generalizability require caution. Participants came from SMA Reach UK and the Italian Telethon Network and met the study’s inclusion criteria, so patterns may reflect this selected clinical cohort and its recorded assessments. The authors further note that feeding support status and swallowing difficulties may be susceptible to country-level variability between the UK and Italy. Such variability could affect comparability of these predictors across settings, even while the outcome modeling is internally evaluated. (pmid:42900957)

Internal validation is also limited by the sequence of multiple imputation before both the validation split and cross-validation, which the authors state may create information leakage and bias. They therefore characterize all internal validation results as exploratory and state that external validation in a new dataset is imperative. The reported concordance index and predictor rankings should consequently remain scoped to this internal evaluation rather than assumed to transfer unchanged to other spinal muscular atrophy populations or data-collection systems. (pmid:42900957)