| Summary: | The level of future impairment is difficult to predict with multiple sclerosis (MS). Previous MS
prognosis studies have mainly examined mortality or overall impairment and have used limited
analytic methods. The primary objectives of this study were to identify the predictors of later
functional impairment and to determine how predictable these impairments are. A secondary
objective was to compare prognosis modelling techniques.
A retrospective cohort study was conducted using patient data from the Vancouver MS clinic
(Canada). Some analyses used the level of impairment after 10-15 years of MS as an outcome
and involved binary and ordinal logistic regression (BLR, OLR) modelling. Other analyses used
the time from onset to a specific level of overall impairment (DSS 6) as the outcome and involved
proportional hazards PH) modelling. Candidate predictors were demographic (e.g., onset age)
or clinical (e.g., onset symptoms) in form. Model performance was assessed using both cross-validation
and data from a Dutch clinic in Groningen. Indicators of performance included
receiver-operator characteristic (ROC) curve analysis, Hosmer-Lemeshow chi-square analysis and
Somers’ Dxy correlation analysis.
Most of the previously recognized predictors were predictive of motor impairment. However,
the predictive value of patient characteristics often depended on the type of impairment being
predicted. Specific impairment types varied in their predictability. Moreover, differences in the
distribution of impairment frequently resulted in underestimates of risk in the Dutch cohort. In
terms of the patient characteristics identified as predictive, there was generally good agreement
amongst the different models. OLR models performed as well as BLR models.
While various predictors exist for all impairment types, the predictability of these impairments
is low or moderate. In terms of patient counselling, knowledge of predictors can be
communicated to patients about their prognosis. However, patients should also be informed that
improved knowledge about predictors of impairments does not translate into high predictability of
future function. Regarding the use of predictions in a patient care setting, only terms such as “low
risk” or “high risk” should be used. Improvements to the modelling strategy as well as a better
understanding of the underlying disease process should help to improve predictability.
|
|---|
