Amyotrophic lateral sclerosis (ALS) is a progressive and usually fatal neurodegenerative disease. Survival from diagnosis varies considerably. Several prognostic factors are known, including site of onset (bulbar or limb), age at symptom onset, delay from onset to diagnosis and the use of riluzole and non-invasive ventilation (NIV). Clinicians and patients would benefit from a practical way of using these factors to provide an individualised prognosis.

575 consecutive patients with incident ALS from a population-based registry in South-East England register for ALS (SEALS) were studied. Their survival was modelled as a two-step process: the time from diagnosis to respiratory muscle involvement, followed by the time from respiratory involvement to death. The effects of predictor variables were assessed separately for each time interval.

Younger age at symptom onset, longer delay from onset to diagnosis and riluzole use were associated with slower progression to respiratory involvement, and NIV use was associated with lower mortality after respiratory involvement, each with a clinically significant effect size. Riluzole may have a greater effect in younger patients and those with longer delay to diagnosis. A patient's survival time has a roughly 50% chance of falling between half and twice the predicted median.

A simple and clinically applicable graphical method of predicting an individual patient's survival from diagnosis is presented. The model should be validated in an independent cohort, and extended to include other important prognostic factors.

Amyotrophic lateral sclerosis (ALS, also known as motor neuron disease) is a degenerative disease of the nervous system which is progressive, incurable and almost always fatal. Around 1 in 400 people will develop the condition, usually aged between 50 and 70 years.

The rate at which the condition progresses varies greatly, as does the length of time from symptom onset to death, from a few months to more than 10 years.

Three medical treatments have also been shown to prolong survival: non-invasive ventilation (NIV) when respiratory support is necessary, gastrostomy feeding when dysphagia is present, and riluzole. As a drug treatment which has been shown to prolong survival in a neurodegenerative disease, riluzole is of great theoretical as well as practical interest, but it is not known whether the survival improvement occurs throughout the disease course or only at certain stages.

A practical way of combining the factors to produce a good prediction for a particular individual would be of great use to the clinician and the affected person. While there are many multivariate survival analyses in the literature, most would require the clinician to reconstruct the mathematical model (and in many cases the necessary information is not given). Further, most use a Cox's proportional hazards model; in a similar study of mortality in ALS,

To develop such a model, one would need a large population-based cohort of patients with newly diagnosed ALS, with consistently recorded data on the known prognostic factors and length of survival, along with a robust statistical technique requiring a minimum of arbitrary assumptions and informed by understanding of the biology of the disease. The South-East England register for ALS (SEALS) registry provides such a data set, ascertaining all cases of ALS in a population of some 3 million in the South-East of England. Full details of the registry and case ascertainment are available elsewhere.

The dominant cause of death in ALS is respiratory muscle weakness, causing respiratory failure or infection: estimates range from 65% to 89%.

The criteria for inclusion in the SEALS registry are described elsewhere,

We included in the model the most significant and easily assessed factors that have been found to be associated with survival, namely site of onset (bulbar, upper limb or lower limb), age at symptom onset, gender, time from onset to diagnosis (diagnostic delay) and treatment with riluzole. Riluzole use was analysed as a binary variable: if a patient had ever taken riluzole, it was coded as 1, otherwise 0. This method gives the closest result parallel with the intention-to-treat method of analysis used in clinical trials. Insufficient information was available on cognitive impairment, respiratory muscle weakness, genetic status, use of NIV and gastrostomy feeding. Analyses relating to survival in patients receiving NIV were performed using data from patients attending the Motor Nerve Clinic at King's College Hospital, London (

NIV data set: descriptive statistics

Total cases | n=128 | |
---|---|---|

Site of onset | Bulbar 32 (25%) | Limb 96 (75%) |

Gender | Female 37 (29%) | Male 91 (71%) |

Age at symptom onset (years) | Median 57.0 | IQR 50–64 |

Diagnostic delay (months) | Median 10.3 | IQR 6.0–15.0 |

Riluzole use | Yes 32 (25%) | No 96 (75%) |

Survival from NIV initiation (months) | Median 5.7 | IQR 2.1–11.3 |

NIV, non-invasive ventilation.

The technical details of the model are given below, but the reasoning behind it can be summarised as follows. We observed that patients with ALS using NIV have a constant risk of death, namely 20% per month in bulbar-onset cases and 10% per month with limb onset; this does not change significantly over time. Noting that most deaths in ALS occur in the context of respiratory muscle weakness, we inferred that this death rate would apply from the onset of respiratory muscle involvement, and that the rate would also be constant (albeit higher) in patients with respiratory muscle weakness who were not using NIV—we used data from the largest suitable study

Most patients do not have respiratory muscle involvement at diagnosis, but develop it later on after a variable length of time. We found that it fitted the observed survival data very well if we assumed that the logarithm of the time from diagnosis to respiratory muscle involvement was normally distributed. Using this model, we could ask whether the predictor variables (age at symptom onset, diagnostic delay, riluzole use) were associated with progression to respiratory muscle weakness, or separately with mortality after respiratory muscle involvement.

Our data did not include the date of onset of respiratory involvement. We therefore examined Kaplan-Meier plots of the NIV data set, which suggested that mortality in patients on NIV (all of whom have respiratory involvement) is well modelled by an exponential distribution, that is, a constant mortality rate, at least until survival falls below 20% (

Constant-hazard fitted post-non-invasive-ventilation (post-NIV) survival curves in the clinic data set.

For patients with respiratory involvement but not using NIV, these mortality rates needed to be adjusted for the survival benefit of NIV. We were not able to estimate this survival benefit using our data, and estimates in the literature vary widely. The largest study is reported by Bourke

Reanalysis of data from Bourke

Between the four modelled events (symptom onset, diagnosis, respiratory involvement and death) lie three time intervals, which were modelled separately. As our goal was to create a clinically useful model, we chose to model survival from diagnosis, as it would be at least as useful as predicting from onset and would allow us to use the time from symptom onset to diagnosis (diagnostic delay) as a predictor variable.

The effects of predictor variables on survival can be modelled in two ways: as affecting the chance of death occurring at any particular time (a Cox's proportional hazards model) or as affecting the length of time until death (an accelerated failure time (AFT) model). In the case of time from diagnosis to respiratory involvement, we were dealing with the progression of a condition rather than the stochastic occurrence of an event, and so an AFT model was appropriate.

To maximise the statistical power, we chose to use a parametric model with an a priori choice of modelling distribution. The distribution of times from diagnosis to respiratory involvement could not be visualised directly, but only in the context of the combined model, so the choice of model distribution was arbitrary among those amenable to AFT analysis. A Weibull distribution was tried first, resulting in a poor fit; a log-normal distribution gave a much better fit. To avoid overfitting, the log-normal distribution was chosen without trying other distributions.

The overall survival time distribution was derived as the convolution of the log-normal distribution of time from diagnosis to respiratory involvement with the cumulative exponential distribution of time from respiratory involvement to death.

The combined model was then fitted to the survival data. Two parameters were allowed to vary in fitting the model: the log-scale mean of the log-normal distribution of time from diagnosis to respiratory involvement, and its SD. The mean was the only parameter which was allowed to depend on the predictor variables of age at symptom onset, diagnostic delay (log transformed) and riluzole use. The model was fitted by maximum-likelihood estimation, using the flexsurv package in R (C Jackson. Flexsurv: Flexible parametric survival models. R package version 0.2. 2013. http://cran.r-project.org/package=flexsurv).

Fitting the model to each group of patients according to onset site (bulbar, upper limb or lower limb) gave substantially closer fits than the data set as a whole. This was due to a greater spread (log-scale SD) of the log-normal distribution in the limb-onset groups than in the bulbar-onset group. The two limb-onset groups were practically identical in all analyses, so we pooled the limb-onset groups but analysed the bulbar-onset group separately.

All analyses were conducted and figures produced using R 2.15.2 (R Core Team. R: A language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing, 2012. ISBN 3-900051-07-0.

A total of 575 patients were included. Very few cases were excluded due to missing dates of birth, diagnosis or onset or missing data on predictor variables; patients who remained alive at the time of analysis were included, treating survival as censored, but in fact the date of death was known in 97% of cases (

SEALS data set: descriptive statistics

Total cases | n=575 | |
---|---|---|

Site of onset | Bulbar 189 (33%) | Limb 386 (67%) |

Gender | Female 239 (42%) | Male 336 (58%) |

Age at symptom onset (years) | Median 61.5 | IQR 54–69 |

Diagnostic delay (months) | Median 11.1 | IQR 7.0–19.0 |

Riluzole use | Yes 260 (45%) | No 315 (55%) |

Survival from diagnosis (months) | Median 17.7 | IQR 8.3–29.9 |

SEALS, South-East England register for amyotrophic lateral sclerosis.

Pairwise associations between the variables

Bulbar onset | Female gender | Age at symptom onset | Riluzole | Log diagnostic delay | |
---|---|---|---|---|---|

Bulbar onset | |||||

Female gender | 1.4 to 2.8 | 0.73 | 0.05 | ||

Age at symptom onset | 3.9 to 7.8 | 0.5 to 4.5 | −0.06 | ||

Riluzole | 0.5 to 1.0 | 0.5 to 1.0 | −8.0 to −4.2 | −0.07 | |

Log diagnostic delay | −0.4 to −0.1 | −0.1 to 0.2 | −0.1 to 0.02 | −0.2 to 0.1 |

The figures in the upper triangle represent the strength of association between each pair of variables. Where both variables are binary (onset site, gender, riluzole), the OR is given; where one is continuous (age, delay) and one binary, the difference in means is given; where both are continuous, Pearson's r is given. In each case, the 95% CI is given in the corresponding cell in the lower triangle. Figures in

Raw Kaplan-Meier survival curves for the bulbar-onset and limb-onset groups.

Fitted survival curves.

No significant associations were found between mortality in the NIV data set and age at symptom onset, diagnostic delay, gender or riluzole use (

Effects of covariates on survival in patients on NIV: univariate analyses

Variable | Comparison | Survival benefit estimate | Survival benefit 95% CI | p Value |
---|---|---|---|---|

Age at symptom onset | Across IQR (69 cf. 54) | 1.02 | 0.79 to 1.31 | 0.89 |

Log diagnostic delay | Across IQR (19 m versus 7 m) | 1.14 | 0.91 to 1.43 | 0.24 |

Gender | Male cf. female | 1.40 | 0.94 to 2.07 | 0.09 |

Riluzole | Yes cf. no | 1.38 | 0.92 to 2.08 | 0.12 |

NIV, non-invasive ventilation.

Age, diagnostic delay and riluzole use were each independently associated with slower progression from diagnosis to respiratory involvement. This held true in univariate (

Effects of covariates on time from diagnosis to respiratory involvement: univariate analyses

Variable | Onset-site group | Coefficient | 95% CI | p Value |
---|---|---|---|---|

Age at symptom onset (years) | Bulbar | −0.0235 | −0.030 to −0.017 | 1×10^{−12} |

Limb | −0.0253 | −0.031 to −0.020 | 1×10^{−18} | |

Diagnostic delay (months) | Bulbar | 0.226 | 0.11 to 0.35 | 3×10^{−4} |

Limb | 0.166 | 0.07 to 0.26 | 9×10^{−4} | |

Riluzole use | Bulbar | 0.328 | 0.15 to 0.51 | 3×10^{−4} |

Limb | 0.408 | 0.22 to 0.60 | 2×10^{−5} | |

Male gender | Bulbar | −0.068 | −0.25 to 0.12 | 0.5 |

Limb | 0.007 | −0.18 to 0.19 | 0.9 |

Effects of covariates on time from diagnosis to respiratory involvement: multivariate analysis

Variable | Onset-site group | Coefficient | 95% CI (main analysis) | 95% CI (80% sample) | p Value (main analysis) |
---|---|---|---|---|---|

Baseline mean | Bulbar | 4.43 | 2.7 to 7.3 | 3.5 to 5.9 | |

Log scale | Limb | 4.86 | 3.3 to 7.1 | 3.9 to 6.2 | |

Baseline SD | Bulbar | 0.97 | 0.85 to 1.11 | 0.89 to 1.03 | |

Log scale | Limb | 1.32 | 1.19 to 1.48 | 1.24 to 1.37 | |

Age at symptom onset (years) | Bulbar | −0.0185 | −0.025 to −0.012 | −0.022 to −0.016 | 2×10^{−9} |

Limb | −0.0223 | −0.028 to −0.017 | −0.026 to −0.019 | 9×10^{−16} | |

Diagnostic delay (months) | Bulbar | 0.130 | 0.039 to 0.220 | 0.079 to 0.174 | 0.005 |

Limb | 0.140 | 0.063 to 0.215 | 0.094 to 0.180 | 0.0003 | |

Riluzole use | Bulbar | 0.228 | 0.077 to 0.380 | 0.16 to 0.29 | 0.003 |

Limb | 0.205 | 0.057 to 0.353 | 0.12 to 0.29 | 0.006 |

Effects of predictors on time to respiratory involvement.

Riluzole appears to have a greater benefit, even in ratio terms, in younger patients and in those with longer diagnostic delay (ie, in patients with longer predicted survival;

To predict the survival time of a patient at diagnosis, we suggest the following procedure. If the patient already has respiratory involvement, their mortality is given simply by 20% per month if there is bulbar involvement, or 10% per month if there is not; if the patient is not using NIV, these figures increase to 54% and 31%, respectively. For patients without respiratory involvement,

For example, if

Eighty per cent of our cases had an age at symptom onset between 45 and 75, and 80% had a diagnostic delay between 5 and 30 months. The model should be interpreted with caution for cases lying outside these ranges.

Although AFT models are considered relatively robust in general, we wished to test our model in detail.

To examine the sensitivity of the model to changes in the data, the analyses were rerun many times, removing 20% of cases at random each time. For post-NIV mortality, 95% of the 500 resulting estimates lay between 17.1% and 24.4% (bulbar onset) or 8.6% and 10.8% (limb onset) monthly mortality. The results for the combined model are shown in

The figures for the survival benefit of NIV (3.5×), and for the proportion of our cases who received NIV (80%), were derived from sources external to the study. Varying these figures within ranges of 2× to 6× and 50–90%, respectively, the coefficients for age at symptom onset, diagnostic delay and riluzole all remained within 22% of the values from the main analysis.

The choice of a log-normal distribution to model time from diagnosis to respiratory involvement was arbitrary. However, the fit remained good when the data were stratified by the predictor variables, and the distribution also fitted well to the data derived from the patients attending the Motor Nerve Clinic at King's College Hospital, London (of which the NIV data set used in the present study was a subset;

Survival curves fitted to the KCH Motor Nerve Clinic data set.

We present a straightforward graphical method of predicting a patient's survival time from ALS diagnosis, taking into account the known prognostic factors of onset site, age at symptom onset, diagnostic delay, riluzole use, NIV use and respiratory muscle involvement. The model may predict a median survival of as much as 4 years or as little as 8 months, and even less in patients with respiratory involvement. The true survival time falls between half and twice the median in about half of the patients. Pending validation of the model, we propose this as a clinically useful tool.

The present study improves on the existing prognostic models in several important ways. We give an explicit, straightforward method which can be applied in the clinic in moments, without specialised tests or detailed calculation. It provides an individualised estimate, not merely a broad prognostic category, along with a guide to the uncertainty in this estimate which is easy to calculate and interpret. Using a large population-based cohort avoids biases associated with referral to and treatment in a specialist clinic; we avoid making the proportional hazards assumption, found to be invalid in a similar study;

Riluzole use was associated with slower progression to respiratory involvement in this study. Riluzole is known to prolong overall survival in ALS, but its effects on particular stages of the condition have not previously been reported. While our data do not rule out a continuing effect of riluzole on mortality after respiratory involvement, this finding emphasises the importance of starting riluzole treatment as soon as possible after diagnosis. Our analysis also suggests that riluzole may be especially beneficial for younger patients and those with longer diagnostic delay.

The association of bulbar onset with shorter median survival is accounted for by higher mortality after respiratory involvement, a higher average age at symptom onset and less use of riluzole in the bulbar-onset group. Patients with bulbar weakness are at higher risk of aspiration and respiratory tract infection, a common mode of death; they are also less likely to tolerate NIV. While the present analysis related to bulbar involvement at onset, it is clinically plausible that the effect on mortality would be seen in any patients who have bulbar weakness at the time of respiratory involvement. We have not found evidence that patients with bulbar-onset progress to respiratory involvement sooner, after allowing for age and riluzole use.

The effects of known prognostic factors were confirmed in our model, with similar effect sizes: for example, a recent study

Like any model, the present study represents a compromise between simplicity and robustness on the one hand, and the faithful representation of reality on the other.

There is evidence that patients with upper motor neuron features alone

In constructing the model, we assumed that the patients did not have respiratory muscle involvement at diagnosis. This biases our model slightly towards predicting shorter survival in those without respiratory involvement. The basic assumption that respiratory involvement always occurs before death will also have exceptions. The availability and use of gastrostomy and NIV may vary between centres, as well as other aspects of care relevant to survival.

The model is useful for predicting survival at the time of diagnosis, and once respiratory involvement has occurred; the prediction of survival at other times during a patient's illness is not addressed by the model.

The present study is not a clinical trial, and therefore the effect of riluzole use should be interpreted cautiously as ‘those who receive riluzole survive longer’ rather than as ‘riluzole prolongs survival’, though both are consistent with the data.

A weakness of this study is that the population studied has a lower median age of onset than is typical for population registers. This may reflect underascertainment of older age groups, which would be consistent with the lifetime risk curves previously reported.

Our model makes quantitative statements about the effects of prognostic factors on the distribution of survival time in ALS, and suggests the specific testable hypotheses that riluzole treatment may be more beneficial in patients who do not yet have respiratory involvement, in younger patients and in those with longer diagnostic delay. Analysis of an independent population-based cohort is necessary to investigate these findings and predictions.

Instead of using mortality data, the speed of progression of ALS can also be assessed using the time from clinical onset to the time of involvement of the second site. We would predict that age, diagnostic delay and riluzole use would be associated with this, but that gender and site of onset would not have clinically significant independent effects.

JAK and AA-C were involved in the conception, design of the study and data interpretation. NK, AK, PNL, SM, CES, MT, AA-C contributed to the acquisition of the data. JAK analysed the data. All authors were involved in critically revising the manuscript and approving the final version to be submitted, and agreed to be accountable for all aspects of the work. The corresponding author had full access to all the data in the study and has final responsibility for the decision to submit for publication.

This work was funded by the UK Medical Research Council and Economic and Social Research Council under the aegis of the EU Joint Programme—Neurodegenerative Disease research project. AAC and CES receive salary support from the National Institute for Health Research (NIHR) Biomedical Research Centre and Dementia Biomedical Research Unit at South London and Maudsley NHS Trust and King's College London. The work leading up to this publication was funded by the European Community's Health Seventh Framework Programme (FP7/2007–2013; grant agreement number 259867). The funding sources had no role in writing the manuscript or deciding to submit it for publication.

PNL reports grants from MND Association Funding to support MND Clinic from 1995 to 2010, during the conduct of the study. CES reports consultancy and/or collaborative relationships with Glaxo Smith Kline, Vertex Pharmaceuticals, Chronos Therapeutics and Eli Lilly; he receives no personal gain and has never had any shares of any sort. AAC reports grants from EU FP7, grants from EU JPND—MRC and ESRC, grants from NIHR BRC in Mental Health and Dementia BRU, during the conduct of the study. All other authors declare no conflicts of interest.

The study was approved by the local ethics committees of King's College Hospital (15/LO/0810) and South London and Maudsley (222/02) NHS Foundation Trusts.

Not commissioned; externally peer reviewed.