Contributions
Abstract: P324
Type: Poster
Abstract Category: Clinical aspects of MS - 4 Natural course
Background: More precise methods to estimate prognosis for individual MS patients are needed. Model building, selection and external validation are key steps. However, small sample sizes, inadequate data standardization, short follow-up and lack of quantitative metrics have limited progress.
Objectives: Use advanced modelling and data mining techniques to assess baseline prognostic factors for MS disease progression in a pooled sample of RCT placebo arms.
Methods: Four RCTs of RRMS were combined in an integrated clinical trial database (IDB). Studies included AFFIRM (natalizumab registration trial), DEFINE and CONFIRM (dimethyl fumarate registration trials) and ADVANCE (peginterferon beta-1a registration trial). IDB data were used to build prognostic models for outcomes over 2 years of follow-up. LASSO and ridge regression, elastic nets, support vector machines (SVM) and unconditional and conditional random forests (RF) were applied to model time to disease progression confirmed at 24 weeks, in which the combined endpoint was defined as progression on at least one of EDSS, MSFC-4 (T25FW, 9HPT, PASAT, LCLA). Sensitivity analyses for different definitions of the combined endpoint and separate assessment of EDSS progression and MSFC-4 progression alone were also carried out. Baseline factors tested were: age, gender, ethnicity, number of prior relapses, disease duration, time since pre-study relapse, prior treatment, EDSS, MSFC-4, Gd+ lesion presence, T1 and T2 lesion volume, brain volume, SF-36 PCS and MCS. Survival c-index was used to benchmark model performance.
Results: A total of 434 subjects (27.4%) out of 1582 had progression in the combined endpoint over 2 years. PASAT, SF-36 PCS and LCLA were the most important factors selected by the different competing algorithms, yet without consistent ranking of these 3 factors for prognostic importance. Model prediction performance was relatively poor (c-indices< 0.60) across all models including standard Cox regression and other definitions of progression.
Conclusions: Disagreement in prognostic factors ranking obtained by more powerful statistical tools confirmed the relatively poor prediction performance of baseline factors in modelling disease progression. The performance of the selected modeling approaches, including traditional regression methods, makes it important to explore alternative predictors, or use dynamic prognostic models which account for predictor changes during follow-up.
Supported by: Biogen, Inc.
Disclosure:
Massimiliano Copetti received consulting fees from Biogen, Teva and Eisai.
Andrea Fontana and Francesca Bovis have nothing to disclose.
Maria Pia Sormani received consulting fees from Biogen, Novartis, Genzyme, Roche, Teva, GeNeuro, Merck Serono and Medday.
Ulrich Freudensprung, Carl de Moor, Robert Hyde and Fabio Pellegrini are employees of, and stockholders in, Biogen.
Abstract: P324
Type: Poster
Abstract Category: Clinical aspects of MS - 4 Natural course
Background: More precise methods to estimate prognosis for individual MS patients are needed. Model building, selection and external validation are key steps. However, small sample sizes, inadequate data standardization, short follow-up and lack of quantitative metrics have limited progress.
Objectives: Use advanced modelling and data mining techniques to assess baseline prognostic factors for MS disease progression in a pooled sample of RCT placebo arms.
Methods: Four RCTs of RRMS were combined in an integrated clinical trial database (IDB). Studies included AFFIRM (natalizumab registration trial), DEFINE and CONFIRM (dimethyl fumarate registration trials) and ADVANCE (peginterferon beta-1a registration trial). IDB data were used to build prognostic models for outcomes over 2 years of follow-up. LASSO and ridge regression, elastic nets, support vector machines (SVM) and unconditional and conditional random forests (RF) were applied to model time to disease progression confirmed at 24 weeks, in which the combined endpoint was defined as progression on at least one of EDSS, MSFC-4 (T25FW, 9HPT, PASAT, LCLA). Sensitivity analyses for different definitions of the combined endpoint and separate assessment of EDSS progression and MSFC-4 progression alone were also carried out. Baseline factors tested were: age, gender, ethnicity, number of prior relapses, disease duration, time since pre-study relapse, prior treatment, EDSS, MSFC-4, Gd+ lesion presence, T1 and T2 lesion volume, brain volume, SF-36 PCS and MCS. Survival c-index was used to benchmark model performance.
Results: A total of 434 subjects (27.4%) out of 1582 had progression in the combined endpoint over 2 years. PASAT, SF-36 PCS and LCLA were the most important factors selected by the different competing algorithms, yet without consistent ranking of these 3 factors for prognostic importance. Model prediction performance was relatively poor (c-indices< 0.60) across all models including standard Cox regression and other definitions of progression.
Conclusions: Disagreement in prognostic factors ranking obtained by more powerful statistical tools confirmed the relatively poor prediction performance of baseline factors in modelling disease progression. The performance of the selected modeling approaches, including traditional regression methods, makes it important to explore alternative predictors, or use dynamic prognostic models which account for predictor changes during follow-up.
Supported by: Biogen, Inc.
Disclosure:
Massimiliano Copetti received consulting fees from Biogen, Teva and Eisai.
Andrea Fontana and Francesca Bovis have nothing to disclose.
Maria Pia Sormani received consulting fees from Biogen, Novartis, Genzyme, Roche, Teva, GeNeuro, Merck Serono and Medday.
Ulrich Freudensprung, Carl de Moor, Robert Hyde and Fabio Pellegrini are employees of, and stockholders in, Biogen.