Multiple Sclerosis Resource Centre

Welcome to the Multiple Sclerosis Resource Centre. This website is intended for international healthcare professionals with an interest in Multiple Sclerosis. By clicking the link below you are declaring and confirming that you are a healthcare professional

You are here

Continuous prediction of secondary progression in the individual course of multiple sclerosis

Multiple Sclerosis and Related Disorders



Prediction of the course of multiple sclerosis (MS) was traditionally based on features close to onset.


To evaluate predictors of the individual risk of secondary progression (SP) identified at any time during relapsing-remitting MS.


We analysed a database comprising an untreated MS incidence cohort (n=306) with five decades of follow-up. Data regarding predictors of all attacks (n=749) and demographics from patients (n=157) with at least one distinct second attack were included as covariates in a Poisson regression analysis with SP as outcome.


The average hazard function of transition to SPMS was 0.046 events per patient year, showing a maximum at age 33. Three covariates were significant predictors: age, a descriptor of the most recent relapse, and the interaction between the descriptor and time since the relapse. A hazard function termed “prediction score” estimated the risk of SP as number of transition events per patient year (range <0.01 to >0.15).


The insights gained from this study are that the risk of transition to SP varies over time in individual patients, that the risk of SP is linked to previous relapses, that predictors in the later stages of the course are more effective than the traditional onset predictors, and that the number of potential predictors can be reduced to a few (three in this study) essential items. This advanced simplification facilitates adaption of the “prediction score” to other (more recent, benign or treated) materials, and allows for compact web-based applications ( ).



  • Certain features of multiple sclerosis relapses predict impending progression.
  • A fundamental relationship is derived from a representative multiple sclerosis cohort.
  • A prediction score grades the yearly risk of secondary progression between 1% and 15%.
  • This clinically useful score is calculated from current age and recent clinical data.
  • It defines periods of increased disease activity and may prompt induction therapy.

Keywords: Multiple sclerosis, Secondary progression, Prognosis, Prediction, Continuous hazard function.

1. Introduction

Multiple sclerosis (MS) prognosis is particularly challenging due to the extreme variation in manifestations, which can include symptoms as subtle and fleeting as paraesthesia and as devastating as tetraparesis. Several natural history studies have reported that certain demographic and clinical features at onset or during short segments of the initial 2–5 years of MS are predictive of progression and disability development over the following decades (Confavreux et al, 2003, Eriksson et al, 2003, Runmarker and Andersen, 1993, Scalfari et al, 2010, and Weinshenker et al, 1989). However, the established prognostic values concerning long-term outcome are moderately effective at best, and the significance of the predictors is inconsistent between studies, particularly relating to attack frequency (Langer-Gould et al, 2006, Scalfari et al, 2013, and Yamout et al, 2010). One report suggested a decreasing predictive power during the three first 5-year periods ( Tremlett et al., 2009 ). The design of these studies was based on predictors restricted to gender, age at onset, and a set of characteristics of onset or early attacks – followed by long-term survival analyses. There is a gap in our knowledge on the predictive capacity of data obtained during the following decennia of disease. For some major diseases, such as myocardial infarction, risk factors can be obtained continuously during life. Specialized statistical methods were developed, which provide continuously updated risks in vitamin D deficiency and diabetes (Johansson et al, 2012 and Lind et al, 2010). The probability of a hip fracture was evaluated from present age, recent clinical data, and bone density assessment (Kanis et al, 2010, Kanis et al, 2009, and McCloskey et al, 2012).

In the present study, we applied similar methods of continuous prediction of secondary progression (SP) in multiple sclerosis. These statistical methods (hazard functions) are substantially different from the traditional survival functions or hazard ratios from early disease features. We analysed a database of a representative cohort including characteristics of successive relapses and subsequent events, such as SP and Expanded Disability Status Scale (EDSS) 6. In this study we chose SP as endpoint, in agreement with contemporary investigators (Scalfari et al, 2013 and Taylor, 2013). We searched for independent demographic and clinical factors that contributed to the risk of transition to SP. The aim was to simplify these complex relationships into a continuous individualised prediction based on repeated assessments expressed as a clinically and scientifically useful score.

2. Material and methods

2.1. Study population

The study used data from an untreated population-based incidence cohort of MS patients residing in Gothenburg at the predefined time of onset between January 1950 and December 1964 (Gothenburg Incidence Cohort, GIC;n=306). The cohort was closely followed by the research team at the Department of Neurology, Sahlgrenska University Hospital ( Skoog et al., 2012 ). Within this cohort, 255 patients had MS according to the Poser criteria ( Poser et al., 1983 ), of whom 202 had relapsing-remitting MS (RRMS) and 53 had primary progressive MS or undefined onset. Information on case identification procedures, conditions basic to MS ascertainment, and data on MS epidemiology in this cohort were previously reported (Runmarker and Andersen, 1993 and Skoog et al, 2012).

The present study included the subgroup of RRMS patients with at least one distinct second attack that confirmed the diagnosis of MS according to the Poser criteria (n=157; 102 women). Patients with single-attack progressive MS or a second attack in the year of onset of SP (n=45) were excluded ( Figure 1 ). The median time from the first attack to inclusion was 2 years. The mean age at inclusion was 33.5 years (SD, 10.1; range, 14.6–59.8 years).


Figure 1 Flowchart of the patient ascertainment procedure from the Gothenburg Incidence Cohort (GIC). The present study excluded “CIS-only”, primary progressive, and single-attack progressive patients, thus including GIC RRMS patients with at least one distinct relapse (a second attack) after the onset attack (n=157).

2.2. Definitions

Secondary progression was defined according to established criteria as continuous progression for at least one year without remission and detectable at time intervals of months or years ( Lublin and Reingold, 1996 ). This was determined retrospectively after one year of observation and the probable year of onset was retrospectively recorded ( Supplement 1 ). A relapse was defined as new MS-related symptoms appearing within a time-frame of weeks. Based on their previously reported predictive capacities, three dichotomous attack characteristics were used according to previous definitions: monofocal symptoms (yes/no), afferent symptoms (yes/no), and complete remission from the relevant attack (yes/no). The term afferent refers to lesions in afferent tracts from the skin, muscles, eyes or labyrinths. Afferent relapses included optic, sensory and vestibular symptoms with a documented absence of efferent symptoms such as central paresis. Minor associated efferent symptoms such as a Babinski sign and increased tendon reflexes were included. For instance, a parahypesthesia with hyperreflexia was recorded as afferent. Complete remission was defined as the absence of constant residual symptoms in the appropriate functional system, as evaluated one year after the acute phase of a relapse, including intermittent symptoms. A Babinski sign was not considered sufficient to indicate incomplete remission.

These variables have been previously defined (Amato and Ponziani, 2000, Eriksson et al, 2003, Levic et al, 1999, Runmarker and Andersen, 1993, Tedeholm et al, 2013, and Yamout et al, 2010) and variants of these predictors were extensively used in recent natural history studies (Amato and Ponziani, 2000, Levic et al, 1999, and Yamout et al, 2010).

2.3. Study design

We performed a Poisson regression analysis (Breslow and Day, 1987 and Preston, 2005) with transition to SP as outcome ( Supplement 2 ). Observation was terminated at onset of secondary progression, at censoring due to competing causes of death, other disabling diseases, migration or the end of follow-up. The following characteristics were tested as independent variables: age at onset attack; current age (continuous variable); gender; time from the second attack; number of previous attacks; the three above mentioned dichotomous characteristics of the onset attack and of the most recent attack; and time since the most recent attack (to current age). Variables remaining significant in the regression analysis yielded a continuous hazard function of SP, which we termed the “prediction score”.

3. Results

3.1. The study material

For the 157 patients, the total observation time from the first relapse (second attack) to end-point or censoring was 2568 years (median observation time from the first relapse to censoring or end-point was 11.5 years, range 0.7–56.7). A total of 118 patients converted to SP during the follow-up period. Of the 920 distinct attacks recorded in the GIC database, 749 had complete records of associated predictive characteristics and fulfilled the inclusion criteria for the present study ( Skoog et al., 2012 ). An average of 0.046 SP transition events occurred per patient-year. Supplementary Table S1 presents the 157 patients with the individual attack numbers. A Kaplan–Meier estimate of the survival function from the second attack illustrates the proportion of patients escaping SP ( Supplementary Figure S1 ).

3.2. Prediction from current age

Current age influenced the risk of transition to SP. The momentary risk (hazard function, Figure 2 ) showed a maximum at age 33, where it was estimated to 0.08 transition events (SP onset) per year.


Figure 2 The hazard function of transition to secondary progression with only patient age included in the model, with 95% confidence intervals. RRMS patients with at least one distinct relapse (a second attack) after the onset attack were included (n=157). The function forms a curve with a maximum value at approximately 33 years of age.

3.3. Prediction from current age with added clinical characteristics

Variables that did not reach significance as predictors in the Poisson regression analysis in the multivariable context included gender (p=0.0838), age at onset (p=0.4107), number of attacks (p=0.1630), and the three dichotomous (yes/no) onset attack characteristics ( Section 2 ): polyfocality (p=0.1380), afferent symptoms (p=0.6934) and complete remission (p=0.2027). The factors which remained significant in the analysis were clinical characteristics associated with a recent attack. They were, in addition to current age, independent and significant predictors of the current risk of secondary progression, including two of the dichotomous characteristics, namely afferent symptoms and complete remission, while polyfocality was not significant. The number (0, 1 or 2) of unfavourable “no” responses in these two attack characteristics was used to describe the severity or grade (0–2) of attack ( Table 1 ). An interaction term – the product of the attack grade and the interval between the most recent attack and current time – also remained independent and significantly predictive.

Table 1 Hazard function of progression, considering current relapses.

Variable β SE HR (95% CI) p
Constant β0: −11.5081 4.0138    
(a) Minimum (current age, 27) β1: 0.3167 0.1507 1.37 (1.02–1.84) 0.0356
(b) Maximum (current age – 27, 0) β2: −0.0199 0.0088 0.98 (0.96–1.00) 0.0232
(c) Attack grade (0, 1, or 2) β3: 0.7164 0.1467 2.05 (1.54–2.73) 0.0000
(d) Attack grade (0, 1, or 2) × time since last relapse β4: −0.0457 0.0158 0.96 (0.93–0.99) 0.0039

Minimum (x,y)=the smallest of the numbersxandy. Maximum (x,y)=the largest of the numbers (x,y).

The beta coefficients from the regression analysis provide the value of the prediction score (hazard function), expressing the current risk of conversion to SP: exp(β0+β1×aβ2×b+β3×c+β4×d)=exp(−11.5081 + 0.3167×current agea − 0.0199×current ageb+ 0.7164×type of attack − 0.0457×type of attack×time since last attack). Variablesa,b,c, anddare time dependent. Variablesaandbare a continuous age-dependent function that increases up to 27 years of age, earlier than in the hazard function for transition to SP based on age only (3:2), and then slowly decreases. The third component which mirrors the impact of successive attacks is the attack grade, which indicates the attack severity (number of unfavourable characteristics: 0, 1, or 2). It is updated at each attack. The fourth component. the variable (d), is the product of the attack grade (n=0, 1, or 2) and the time since the last attack. It mirrors the decreasing effect of an attack as time passes after the attack. The hazard ratio (HR) is the ratio of the momentary risk between two patients differing one unit with respect to the current variable (year of age or attack grade). The prediction score can be presented graphically, as shown in Figure 3 .

3.4. Prediction score

The prediction score was calculated as explained in Table 1 . When an attack occurred with efferent symptoms and no remission (defined in Section 2.2 ) during a low-risk period, the prediction score immediately increased. However, when an attack with afferent symptoms and complete remission occurred in a period with a high current baseline prediction score, it induced an immediate decrease in the prediction score. Representative graphic case reports with prediction scores are shown in Figure 3 . The prediction score calculation can be conveniently performed using an online wizard that prompts for the required items ( ). This requires the input of only a few items of raw data: current age (years), afferent symptoms of most recent attack (yes/no), remission of most recent attack (yes/no), and time (years) since that attack. The probability of SP during an interval is calculated from the area under the curve of the prediction score provided no new attacks occur.


Figure 3 Continuous prediction scores (hazard functions) throughout the course, and an illustrative diagram of the clinical course in six representative patients. Each relapse is represented by a peak, with the height describing the grade of the attack, defined as the number of unfavourable characteristics (0–2, Section 2.2 ) of each relapse, i.e., the grade of attack. Calculation of the prediction scores is described in Section 3 .

3.5. Identifying periods of high and low risk

The average yearly risk expressed as the prediction score, of transition to secondary progression was 0.046 events/year. There was a wide spectrum of individual risks of SP apparent from the number of patients who spent periods in the tiers with the highest and lowest risks ( Table 2 ). We compared the expected number of cases in different risk tiers with the observed number of cases. A chi-squared test showed that the fit of the model was very good (p=0.4811, Supplementary Table S2 ).

Table 2 SP onset events per patient year determined by the distribution of the prediction score.

Limit of prediction score Low risk periods High risk periods Whole cohort
<0.015 <0.02 <0.03 <0.04 >0.06 >0.10 >0.14 >0
(1) Expected number of SP onset events 2.0 4.5 17.1 34.6 56.3 19.8 10.3 118.3
(2) Patient years with indicated score 224.6 366.5 867.8 1369.6 619.7 144.1 63.4 2567.5
(3) Number of patients in this cohort who eventually fulfilled the condition 41 56 75 107 123 56 40 157
(4) SP onset events per patient year (=row 1/row 2) 0.009 0.012 0.020 0.025 0.091 0.137 0.162 0.046

This table demonstrates the capacity of the prediction score to define realistic periods with higher or lower risk of transition to SP. The basic observation is the “patient years” (row 2), periods of varying time imported from individual patients׳ courses, one patient often contributing to several periods. The point is that there is still a substantial number of patients in the periods (defined by the prediction score) when we leave aside the average 0.046 transition events to SP per patient year and approach the more extreme risks. Simply stated, the extreme high-risk period included 8% (10.3/118.3) of all transition events during only 63.4 patient-years, whereas the extreme low-risk periods included only 1.7% (2/118.3) of events during 224.6 patient-years.

4. Discussion

There is an urgent need for efficient predictors of long-term disability, which could enable treatments to be targeted to patients at high risk. Despite considerable efforts in long-term natural history studies (Confavreux et al, 2003, Runmarker and Andersen, 1993, Scalfari et al, 2010, Tremlett et al, 2006, and Weinshenker et al, 1989), only few predictors have been independently confirmed in another data set ( Bergamaschi et al., 2007 ), and there were only few practical applications of these studies ( Weinshenker et al., 1991 ).

4.1. Strategies for prediction

We recorded a moderate predictive effect of a series of demographic and clinical characteristics from the clinically isolated syndrome (CIS), from CDMS onset, or from a point in the course approximately 5 years after onset (Eriksson et al, 2003, Novakova et al, 2014, and Runmarker and Andersen, 1993). Although confirmed in studies of varying length (Scott et al, 2000 and Yamout et al, 2010), there is no general consensus as to the power of these predictors. It was suggested that the prognostic relevance of relapse frequency diminishes through the first 5-year periods, and one group reported an inverse effect of attacks after the first 2 years (Scalfari et al, 2010 and Tremlett et al, 2009). A web-based tool for personalised prediction of the long-term disease course – called the Evidence-Based Decision Support Tool in Multiple Sclerosis – provides long-term predictions based on the data from a large natural history cohort: it used disease course, age at first MS symptoms, number of attacks in the first two years, first inter attack interval and/or time to EDSS 3 as predictors. Transition to secondary progression or EDSS milestones was used as end-point. Its prediction had the same range of precision as prediction by neurology specialists, although it was more consistent (Daumer et al, 2007 and Galea et al, 2013). The present study is the first to show that a few clinical characteristics readily available throughout the disease course are strong and independent predictors. However this needs to be confirmed in another patient cohort. An unexpected observation was that mild attacks during periods with high baseline risk (evaluated from age and attack history) reduced the immediate risk of SP, a finding which may seem counterintuitive. Although the specific immunopathological background remains to be clarified, it is known that the recurrent process of inflammation with blood-brain barrier (BBB) damage also induces anti-inflammatory and reparative processes ( Wootla et al., 2012 ).

All hitherto reported prediction studies relied on a basic design where predictors were sampled at onset or during a period in the early stage followed by survival analysis producing hazard ratios. Recently, a more effective strategy was implemented where continuous prediction was based on repeated assessments. This method has been extensively applied for evaluation of the risk of hip fracture, and for the time relationship between a hyperglycaemia indicator and subsequent diabetic retinopathy (Johansson et al, 2012, Kanis et al, 2010, and Lind et al, 2010). The Fracture Risk Assessment Tool – a web-based algorithm that gives repeated estimates of the 10-year probability of major osteoporotic fractures – has achieved world-wide application ( Watts, 2011 ).

4.2. A representative cohort

The present analysis was unprejudiced concerning the timing of relative efficacy of early prediction, whether from onset, the early phase, or more recent clinical data. We found that the hazard function of the transition to SP had an age-related maximum at age 33, reproducing a previous finding ( Runmarker et al., 1994 ). Outcomes may be extremely variable depending on the mode of recruitment ( Andersen, 2008 ). However, there will be less recruitment bias in an incidence cohort, considered to be the most representative type of patient material (Kurland, 1994 and Sackett and Tugwell, 1991), fulfilling recently proposed requirements for historical cohorts ( Scalfari et al., 2013 ). We addressed the GIC database, an essentially untreated incidence cohort with 50 years of follow-up to SP or censoring mainly due to death from other diseases, which contains information on predictors associated with the majority of successive individual relapses. The reason for the trend towards shorter time to SP in this historical material, as compared with contemporary treated patients was recently analysed and partially attributed to the use of disease modifying drugs, DMD ( Tedeholm et al., 2013 ). The robust relationship between the characteristics of relapses and the risk of transition to SP found in the incidence cohort will, with due reason, not be significantly changed in today׳s core cases with typical clinically definite MS. It is necessary to have a historical database to obtain an informative range of follow-up.

4.3. A more relevant strategy of prediction

The prognosis based on classical predictors was far more powerful when repeatedly ascertained during the course, than from onset. After testing a number of conceivable predictors we found that a score based on only three characteristics had independent predictive capacity. These were: current individual age, a descriptor of the severity of the most recent attack, and the time elapsed since that attack. These three pieces of information were combined to determine the continuous hazard function for transition to SP, termed the “prediction score”. This function had a clinically relevant distribution scale with a significant proportion of patient time at the extremes of the scale of <0.01 or >0.15 transition events per patient year. Thus, the prognostic information beyond the average of 0.046 yearly SP events for the total present material is clinically meaningful, providing the momentary risk (events per time unit). From this, we have determined other quantities and functions, such as expected time to progression, survival functions, and estimation of the probability for events during a certain time (data not shown). Previous studies on prediction in MS, including our own, used the Cox proportional hazards model (Confavreux et al, 2003, Eriksson et al, 2003, and Scalfari et al, 2010). However, this model is too restrictive for an effective strategy, since it does not allow the predictive ability of variables to change with time, and it only provides hazard ratios rather than continuous hazard functions. Ratios are not very useful for advicing individual patients. The continuous hazard functions estimated by the Poisson regression models used here are responsive to both positive and negative prognostic indicators.

4.4. Secondary progression as an outcome parameter

We selected SP as an effective dependent (outcome) variable, with its main asset being that it is an intrinsic determinant of the subsequent disease course (Leray et al, 2007 and Scalfari et al, 2013). The fact that immunohistological changes occur with SP is a further argument for selecting it as an endpoint. The continued rate of SP was predictable from its initial phase, but unpredictable from the preceding RR disease course (Confavreux and Vukusic, 2006 and Eriksson et al, 2003). While the axonal pathology of RRMS may be stationary ( Benedetti et al., 2009 ), the pathology of SP includes slow expansion of pre-existing lesions which become the sites of axonal injury ( Lassmann et al., 2012 ). In the SP phase, lesions in the subpial cortical layer are abundant ( Kutzelnigg et al., 2005 ) and the inflammation tends to be compartmentalised behind the blood–brain barrier ( Serafini et al., 2004 ). The differentiation among four described histopathological types of MS is effaced in SP ( Franklin and Ffrench-Constant, 2008 ). However, this transition is probably not sudden, and whether it precedes the clinical transition remains unknown. Our present findings support the notion that high activity during a relatively short phase is indicative of the transition to the progressive phase.

4.5. Verification and modification in other materials

Our results are limited by the moderate size of the cohort, although the information content in the database is unique. In the absence of confirmation, the risk of data over-fitting cannot be excluded, even in a study such as this one with a small number of predictors ( Subramanian and Simon, 2013 ). However, our relatively large numbers of end-points (118) and attacks (749) reduce this risk. We hope our findings encourage further investigation and replication in an independent patient cohort. This would need to be defined in the same way as our original material, with two clinically defined relapses (including the onset attack). The prediction score may be slightly different, and this may actually provide additional information. Contemporary immunomodulatory therapy reduces inflammation and thereby attack frequency and severity. Therapy may modify the prediction score in three essentially different ways:

  • (i) An unchanged prediction score would indicate a direct and probably causal relationship between the reduction of inflammation in RRMS and subsequent neurodegeneration in SP. Then the effect on SP would be secondary to the effect on relapses.
  • (ii) An increased prediction score would indicate that SP occurs, regardless of therapy. Then, the effect on inflammation (and relapses) has not reduced the risk of SP.
  • (iii) A decreased prediction score would indicate that therapy has a direct effect against the SP process beyond its effects on inflammation in the RR phase. One of the theoretical factors that could mediate this change is the passage of drugs through the BBB.

4.6. Application of the prediction score

The following principle insights were gained from this study: (1) that periods of increased risk of SP occur, thus the data acquired during the course are more effective predictors than the traditional onset predictors and (2) the number of potential predictors can be reduced to a few core items. The prediction score wizard could help to evaluate the risk of SP in untreated patients and, possibly with some re-calculation of parameters in another material, also in DMD treated patients. It has not been excluded that near-significant factors (gender) in the present analysis should be included in the prediction score. There are indications that contemporary patients have a milder course than those in the incidence cohort although it would seem probable that the difference in risk of SP between the incidence cohort and contemporary patients is largely a result of the introduction of DMD ( Tedeholm et al., 2013 ).

In treating RRMS, one strategy is to start with induction therapy, while another is to wait until escalation or induction is indicated by certain criteria ( Rio et al., 2009 ). In the latter scenario, a modified prediction score may standardise the indication for induction therapy, defining periods of high risk of progressive deterioration. Furthermore, the prediction score can be used as a novel tool for assessing the statistical power of trials with SP as end-point, although this end-point has so far only been used in register studies. Using the prediction score, a trial will achieve the same statistical power (80%) with patients selected for a moderately high risk of progression (>0.1 onset events per year) as with twice the number of unselected patients ( Table 3 ).

Table 3 Calculation of number of patients required for a trial, with fictitious data.

Assumed reduction of risk of SP with treatment vs. placebo (%) Number of patients needed
Without selection Patients with prediction score>0.10
20 2×3234 2×1204
30 2×1340 2×498
40 2×704 2×259
50 2×474 2×152

End-point: secondary progression (SP). Treated and placebo groups are compared. The patients are followed for 2.5 years. Significance level, 0.05; two-sided test; power 80%. The groups are assumed to be of equal size.

5. Conclusions

The following principle insights were gained from this study:

  • the risk of transition to SP varies over time in individual patients,
  • the risk of SP is linked to previous relapses,
  • predictors in the later stages of the disease are more effective predictors than the traditional onset predictors,
  • the number of potential predictors can be reduced to a few (three in this study) essential items, allowing for compact web-based applications ( ).


We would like to thank the Swedish and Gothenburg Associations of Individuals with Neurological Disabilities (NHR) and the Bjo¨rnsson Foundation, Gothenburg, for their generous economic support. Basic resources were provided from ALF-LUA grants for Swedish University Hospital Research.

Conflict of interest statement

None declared.

Ethical approval

This study was approved by the Research Ethics Committee of Gothenburg, Sweden.


We are indebted to our co-workers in the Gothenburg Incidence Cohort study who contributed to the creation and follow-up of the cohort over many decades.

Appendix A. Supplementary information


Download file

Supplementary data


Download file

Supplementary data


Download file

Supplementary data


Download file

Supplementary data


Download file

Supplementary data


  • Amato and Ponziani, 2000 M.P. Amato, G. Ponziani. A prospective study on the prognosis of multiple sclerosis. Neurol Sci. 2000;21:S831-S838 Crossref
  • Andersen, 2008 O. Andersen. Natural history of multiple sclerosis. C. Raine, H.F. McFarland, R. Hohlfeld (Eds.) Multiple sclerosis: a comprehensive text (Saunders Elsevier, Edinburgh, 2008) 100-120
  • Benedetti et al., 2009 B. Benedetti, M. Rovaris, M.A. Rocca, D. Caputo, M. Zaffaroni, R. Capra, et al. In-vivo evidence for stable neuroaxonal damage in the brain of patients with benign multiple sclerosis. Mult Scler. 2009;15:789-794 Crossref
  • Bergamaschi et al., 2007 R. Bergamaschi, S. Quaglini, M. Trojano, M.P. Amato, E. Tavazzi, D. Paolicelli, et al. Early prediction of the long term evolution of multiple sclerosis: the Bayesian Risk Estimate for Multiple Sclerosis (BREMS) score. J Neurol, Neurosurg, Psychiatry. 2007;78:757-759
  • Breslow and Day, 1987 N.E. Breslow, N.E. Day. Statistical methods in cancer research. Volume II – the design and analysis of cohort studies. IARC Sci Publ. 1987;:1-406
  • Confavreux and Vukusic, 2006 C. Confavreux, S. Vukusic. Natural history of multiple sclerosis: a unifying concept. Brain. 2006;129:606-616 Crossref
  • Confavreux et al., 2003 C. Confavreux, S. Vukusic, P. Adeleine. Early clinical predictors and progression of irreversible disability in multiple sclerosis: an amnesic process. Brain. 2003;126:770-782 Crossref
  • Daumer et al., 2007 M. Daumer, A. Neuhaus, C. Lederer, M. Scholz, J.S. Wolinsky, M. Heiderhoff. Prognosis of the individual course of disease--steps in developing a decision support tool for Multiple Sclerosis. BMC Med Inform Decis Mak. 2007;7:11 Crossref
  • Eriksson et al., 2003 M. Eriksson, O. Andersen, B. Runmarker. Long-term follow up of patients with clinically isolated syndromes, relapsing-remitting and secondary progressive multiple sclerosis. Mult Scler (Houndmills, Basingstoke, England). 2003;9:260-274 Crossref
  • Franklin and Ffrench-Constant, 2008 R.J. Franklin, C. Ffrench-Constant. Remyelination in the CNS: from biology to therapy. Nat Rev Neurosci. 2008;9:839-855 Crossref
  • Galea et al., 2013 I. Galea, C. Lederer, A. Neuhaus, P.A. Muraro, A. Scalfari, N. Koch-Henriksen, et al. A web-based tool for personalized prediction of long-term disease course in patients with multiple sclerosis. Eur J Neurol. 2013;20:1107-1109 Crossref
  • Johansson et al., 2012 H. Johansson, A. Oden, J. Kanis, E. McCloskey, M. Lorentzon, O. Ljunggren, et al. Low serum vitamin D is associated with increased mortality in elderly men: MrOS Sweden. Osteoporos Int. 2012;23:991-999 Crossref
  • Kanis et al., 2010 J.A. Kanis, E.V. McCloskey, H. Johansson, A. Oden, O. Strom, F. Borgstrom. Development and use of FRAX in osteoporosis. Osteoporos Int. 2010;21(Suppl. 2):S407-S413
  • Kanis et al., 2009 J.A. Kanis, E.V. McCloskey, H. Johansson, O. Strom, F. Borgstrom, A. Oden. How to decide who to treat. Best Pract Res Clin Rheumatol. 2009;23:711-726 Crossref
  • Kurland, 1994 L.T. Kurland. The evolution of multiple sclerosis epidemiology. Ann Neurol. 1994;36(Suppl.):S2-S5 Crossref
  • Kutzelnigg et al., 2005 A. Kutzelnigg, C.F. Lucchinetti, C. Stadelmann, W. Bruck, H. Rauschka, M. Bergmann, et al. Cortical demyelination and diffuse white matter injury in multiple sclerosis. Brain. 2005;128:2705-2712 Crossref
  • Langer-Gould et al., 2006 A. Langer-Gould, R.A. Popat, S.M. Huang, K. Cobb, P. Fontoura, M.K. Gould, et al. Clinical and demographic predictors of long-term disability in patients with relapsing-remitting multiple sclerosis: a systematic review. Arch Neurol. 2006;63:1686-1691 Crossref
  • Lassmann et al., 2012 H. Lassmann, J. van Horssen, D. Mahad. Progressive multiple sclerosis: pathology and pathogenesis. Nat Rev Neurol. 2012;8:647-656 Crossref
  • Leray et al., 2007 E. Leray, S.P. Morrissey, J. Yaouanq, M. Coustans, E. Le Page, J. Chaperon, et al. Long-term survival of patients with multiple sclerosis in West France. Mult Scler (Houndmills, Basingstoke, England). 2007;13:865-874 Crossref
  • Levic et al., 1999 Z.M. Levic, I. Dujmovic, T. Pekmezovic, M. Jarebinski, J. Marinkovic, N. Stojsavljevic, et al. Prognostic factors for survival in multiple sclerosis. Mult Scler. 1999;5:171-178
  • Lind et al., 2010 M. Lind, A. Oden, M. Fahlen, B. Eliasson. The shape of the metabolic memory of HbA1c: re-analysing the DCCT with respect to time-dependent effects. Diabetologia. 2010;53:1093-1098 Crossref
  • Lublin and Reingold, 1996 F.D. Lublin, S.C. Reingold. Defining the clinical course of multiple sclerosis: results of an international survey. National Multiple Sclerosis Society (USA) Advisory Committee on Clinical Trials of New Agents in Multiple Sclerosis. Neurology. 1996;46:907-911 Crossref
  • McCloskey et al., 2012 E. McCloskey, H. Johansson, A. Oden, J.A. Kanis. Fracture risk assessment. Clin Biochem. 2012;45:887-893 Crossref
  • Novakova et al., 2014 L. Novakova, B. Skoog, B. Runmarker, S. Ekholm, S. Winblad, V. Lisovskaja, et al. Clinically isolated syndromes with no further disease activity suggestive of multiple sclerosis at the age of population life expectancy. Mult Scler (Houndmills, Basingstoke, England). 2014;20:496-500 Crossref
  • Poser et al., 1983 C.M. Poser, D.W. Paty, L. Scheinberg, W.I. McDonald, F.A. Davis, G.C. Ebers, et al. New diagnostic criteria for multiple sclerosis: guidelines for research protocols. Ann Neurol. 1983;13:227-231 Crossref
  • Preston, 2005 D.L. Preston. Poisson regression in epidemiology. Encyclopedia of biostatistics (, 2005)
  • Rio et al., 2009 J. Rio, M. Comabella, X. Montalban. Predicting responders to therapies for multiple sclerosis. Nat Rev Neurol. 2009;5:553-560 Crossref
  • Runmarker and Andersen, 1993 B. Runmarker, O. Andersen. Prognostic factors in a multiple sclerosis incidence cohort with twenty-five years of follow-up. Brain. 1993;116(Pt 1):117-134
  • Runmarker et al., 1994 B. Runmarker, C. Andersson, A. Oden, O. Andersen. Prediction of outcome in multiple sclerosis based on multivariate models. J Neurol. 1994;241:597-604 Crossref
  • Sackett and Tugwell, 1991 D.H.B. Sackett, P. Tugwell. Clinical epidemiology: a basic science for clinical medicine (Little, Brown Co., Boston, 1991)
  • Scalfari et al., 2013 A. Scalfari, V. Knappertz, G. Cutter, D.S. Goodin, R. Ashton, G.C. Ebers. Mortality in patients with multiple sclerosis. Neurology. 2013;81:184-192 Crossref
  • Scalfari et al., 2013 A. Scalfari, A. Neuhaus, M. Daumer, P.A. Muraro, G.C. Ebers. Onset of secondary progressive phase and long-term evolution of multiple sclerosis. J Neurol Neurosurg Psychiatry. 2014;85:67-75 Crossref
  • Scalfari et al., 2010 A. Scalfari, A. Neuhaus, A. Degenhardt, G.P. Rice, P.A. Muraro, M. Daumer, et al. The natural history of multiple sclerosis: a geographically based study 10: relapses and long-term disability. Brain. 2010;133:1914-1929 Crossref
  • Scott et al., 2000 T.F. Scott, C.J. Schramke, J. Novero, C. Chieffe. Short-term prognosis in early relapsing-remitting multiple sclerosis. Neurology. 2000;55:689-693 Crossref
  • Serafini et al., 2004 B. Serafini, B. Rosicarelli, R. Magliozzi, E. Stigliano, F. Aloisi. Detection of ectopic B-cell follicles with germinal centers in the meninges of patients with secondary progressive multiple sclerosis. Brain Pathol. 2004;14:164-174 Crossref
  • Skoog et al., 2012 B. Skoog, B. Runmarker, S. Winblad, S. Ekholm, O. Andersen. A representative cohort of patients with non-progressive multiple sclerosis at the age of normal life expectancy. Brain. 2012;135:900-911 Crossref
  • Subramanian and Simon, 2013 J. Subramanian, R. Simon. Overfitting in prediction models – is it a problem only in high dimensions?. Contemp Clin Trials. 2013;36:636-641 Crossref
  • Taylor, 2013 B.V. Taylor. Time to onset of secondary progression as an outcome in MS trials: a new paradigm?. J Neurol Neurosurg Psychiatry. 2014;85:3
  • Tedeholm et al., 2013 H. Tedeholm, J. Lycke, B. Skoog, V. Lisovskaja, J. Hillert, C. Dahle, et al. Time to secondary progression in patients with multiple sclerosis who were treated with first generation immunomodulating drugs. Mult Scler (Houndmills, Basingstoke, England). 2013;19:765-774 Crossref
  • Tremlett et al., 2006 H. Tremlett, D. Paty, V. Devonshire. Disability progression in multiple sclerosis is slower than previously reported. Neurology. 2006;66:172-177 Crossref
  • Tremlett et al., 2009 H. Tremlett, M. Yousefi, V. Devonshire, P. Rieckmann, Y. Zhao. Impact of multiple sclerosis relapses on progression diminishes with time. Neurology. 2009;73:1616-1623 Crossref
  • Watts, 2011 N.B. Watts. The Fracture Risk Assessment Tool (FRAX(R)): applications in clinical practice. J Womens Health (Larchmt). 2011;20:525-531 Crossref
  • Weinshenker et al., 1989 B.G. Weinshenker, B. Bass, G.P. Rice, J. Noseworthy, W. Carriere, J. Baskerville, et al. The natural history of multiple sclerosis: a geographically based study. 2. Predictive value of the early clinical course. Brain. 1989;112(Pt 6):1419-1428
  • Weinshenker et al., 1991 B.G. Weinshenker, G.P. Rice, J.H. Noseworthy, W. Carriere, J. Baskerville, G.C. Ebers. The natural history of multiple sclerosis: a geographically based study. 4. Applications to planning and interpretation of clinical therapeutic trials. Brain. 1991;114(Pt 2):1057-1067
  • Wootla et al., 2012 B. Wootla, A. Denic, A.E. Warrington, M. Rodriguez. Need for a paradigm shift in therapeutic approaches to CNS injury. Expert Rev Neurother. 2012;12:409-420 Crossref
  • Yamout et al., 2010 B. Yamout, S. Itani, A. Arabi, D. Hamzeh, S. Yaghi. Prognostic factors of multiple sclerosis in Lebanon. Int J Neurosci. 2010;120:206-210 Crossref


a University of Gothenburg, the Sahlgrenska Academy, Institute of Neuroscience and Physiology, Section of Clinical Neuroscience and Rehabilitation, Gothenburg, Sweden

b Department of Mathematical Sciences, Chalmers University of Technology, Gothenburg, Sweden

lowast Correspondence to: Department of Neurology, Blå Stråket 7, Sahlgrenska University Hospital, 413 45 Gothenburg, Sweden. Mobile: +46 706523390.