The Irish population-based cohort represented 61.6 % of patients diagnosed with ALS in the Republic of Ireland during the set study period (244/396). Forty patients were subsequently excluded, and the remaining 204 patients were included in the final study. Reasons for non-capture included death prior to contact (n = 95) and patients declining participation (n = 47) or not responding to invitation (n = 10). Reasons for exclusion included history of major hemispheric stroke (n = 9), alcohol dependence syndrome (n = 6), pre-morbid learning disability (n = 1), major psychiatric disorder (n = 3), atypical disease course suggestive of variant (n = 4), severe active epilepsy (n = 1), patients being too sick to participate adequately in the study (n = 12), patient not fully informed of diagnosis (n = 1) and co-morbid Alzheimer’s disease at baseline (n = 3).
Recruited ALS patients displayed no significant differences with regard to age, sex distribution, or site of onset when compared to patients who were diagnosed in the same period but did not participate in the study, although non-participants experienced a shorter median survival time (p < 0.0001).
At time of analysis (May 2014), 177 of the 204 patients in the Irish cohort were deceased (86.8 %). Median survival time from symptom onset of the deceased patients was 32 months (range 7–126, interquartile range = 21). Among patients who remained alive (n = 27) the median follow-up time was 47 months from study enrolment, and median follow-up time measured from symptom onset was 75 months (range 51–114).
There were no significant differences in baseline characteristic between the two Irish sub-cohorts (Training and Test groups, see Table 1).
Table 1 Baseline characteristics of the two Irish sub-cohorts (the training and test sets) and the Italian validation cohort
Data from the Training sub-cohort were used to identify significant predictors of survival time. Univariate survival analysis was carried out for the following clinical variables: age at symptom onset, gender, site of disease onset (spinal-onset versus non-spinal or bulbar/respiratory onset), diagnostic category as per the El Escorial (possible, probable or definite), the ALSFRS-R slope (48-ALSFRS-R/disease duration at time assessment), the presence (versus absence) of family history of ALS and/or frontotemporal lobar degeneration in a 1st or 2nd degree relative, and the presence (versus absence) of executive dysfunction on cognitive testing.
Factors associated with significantly worse prognosis on univariate analyses in the Training sub-cohort (n = 117) were (1) Bulbar or respiratory (i.e. non-spinal) onset of disease with a median time of 30 months (95 % CI 26.9–33.1, SE 1.6) compared to 36 months in patients with spinal-onset disease (95 % CI 30.9–41.4, SE 2.6, p = 0.032); (2) higher ALSFRS-R slope (indicating faster functional decline), HR 2.6, 95 % CI 1.9–3.5, SE 0.15, p < 0.0001; (3) and the presence of executive dysfunction (median survival = 27 months, 95 % CI 19.9–34.1, SE 3.6) as opposed to absence of executive dysfunction (median survival time 37 months, 95 % CI 28.2–45.8, SE = 4.5, p < 0.000).
Although patients with older age at symptom onset and female patients tended to have shorter survival, the effect did not reach statistical significance in either case (p = 0.094 and p = 0.064, respectively). Similarly, a positive family history for ALS and/or FTLD and El Escorial diagnostic category at diagnosis had no significant effect on survival on univariate analyses (p = 0.972 and p = 0.109, respectively).
Proportional hazards Cox regression was used to build a multivariate model that included site of disease onset, ALSFRS-R slope, and executive dysfunction (n = 117). The survival effect of all the three factors persisted on multivariate analyses: (1) non-spinal onset of disease, HR = 1.7 (95 % CI 1.12–2.63, SE 0.22, p = 0.012); (2) ALSFRS-R slope: HR = 2.8 (95 % 2.00–3.81, SE = 0.166, p < 0.0001); and executive dysfunction: HR = 2.11 (95 % 1.37–3.28, SE = 0.233, p = 0.001). Internal validation of the model was carried out using boot-strapping techniques. Based on the results of 1000 randomly generated samples, the robustness of the three-parameter model was confirmed.
Based on these results a simple prognostic index, named the ALS Prognostic Index (or API), was generated (Fig. 1) with possible scores ranging from zero to six (higher scores indicating worse predicted prognosis). The figure also shows how patients were then divided using the total API score into three risk groups, termed the ALS risk groups.
The index and prognostic risk group classification procedure were applied to the Irish Training set. The index and classification were also applied to the Irish Test sub-cohort (after excluding one patient for missing data precluding full classification, n = 86) and, for external validation purposes, to the Italian cohort (n = 122).
As shown in Table 2, in all three cohorts the ALS risk groups predicted survival time (log-rank test p < 0.0001 in all three cohorts) with no overlap of the 95 % confidence intervals (Kaplan–Meier survival plots for validation cohorts shown in Fig. 2).
Table 2 This table summarises the Kaplan–Meier estimated median survival time for the three ALS risk groups in the Irish training and test cohorts and the Italian cohort
To investigate the utility of the ALS risk group classification in predicting risk of (1) poor prognosis, defined as death within 25 months of symptom onset and (2) good prognosis, defined as survival time of at least 50 months post-symptom onset, we included only patients who were either deceased at time of analyses or whose follow-up time measured from symptom onset was at least 50 months (all Irish patients and 91 Italian patients). In all three cohorts, the API risk group was a reliable predictor of both poor and good prognosis (Fig. 3a, b, Chi-square test p < 0.0001 in all cases). In the validation cohorts, classifying a patient into a high-risk group was associated with a positive predictive value for poor prognosis of 73.3–85.7 % and a negative predictive value for having good prognosis was 93.3–100 %. Conversely, the low-risk group was associated with a positive predictive value for good prognosis of 59.1–60.1 % and negative predictive value for bad prognosis of 100 %.
As the ALSFRS-R slope was the strongest predictor of survival in the model, we investigated the utility of a classification system based on this measure only (ALSFRS-R slope <0.025 points/month, 0.25–0.49 points/month, 0.50–0.99 points/month, and ≥1 points/months). As shown in Table 3 and Fig. 4, although this model was useful in the Irish validation cohort with only minor overlap of survival times, it was poor predictor of survival in the Italian cohort (external validation cohort).
Table 3 This table summarises the Kaplan–Meier estimated median survival time for the patients in the Irish test cohorts and the Italian cohort classified by ALSFRS-R slope only
Risk groups and genetic status
We investigated the relationship of the risk group allocation and the C9orf72 pathogenic hexanucleotide repeat in both populations. Genetic screening for common ALS genetic mutation was undertaken in 197 Irish patients (96.6 % of the cohort). TARDP gene and FUS gene mutations were identified in one patient each (0.5 % of cohort in each case) and 19 patients carried the C9orf72 hexanucleotide repeat expansion (9.3 %).
Carriers of the C9orf72 repeat expansion represented 10.6 % of the medium-risk group and 13.2 % of the high-risk groups compared to 4.4 % of the low-risk group, although the difference did not reach statistical significance.
In the Italian cohort genetic status was available in all 122 cases, with TARDP gene mutation identified in 5 cases, FUS and optineurin in 1 case each. Three patients carried the C9orf72 repeat expansion (2.6 %) with two cases in the Medium Risk and 1 case in the High risk, representing 2.8 and 5.6 % of each group, respectively, and no C9orf72 positive cases in the low-risk group.