Predictors of mortality in Parkinson's disease and the development of a prognostic model

Abstract Background Most previous prognostic studies in Parkinson's disease have used unrepresentative cohorts with selection biases resulting in major heterogeneity in estimates of mortality. Prognosis is best studied by long-term follow-up of community-based incident cohorts. Prognostic models can predict a patient's risk and improve clinical trial design, but none has been published for use in Parkinson's disease. We aimed to describe mortality in this disorder, identify independent prognostic factors, and develop a valid prognostic model. Methods All incident cases of Parkinson's disease in Aberdeen, UK (Nov 1, 2002, to April 30, 2004, and April 1, 2006, to March 31, 2009) identified with multiple, community-based ascertainment strategies were followed prospectively with death notification from the National Health Service central register. Mortality rates and age-adjusted and sex-adjusted standardised mortality ratios (SMRs) were calculated with indirect standardisation to regional mortality data. Weibull regression was used to create a prognostic model with clinical predictors measured at diagnosis and validated internally. Findings 198 incident cases of Parkinson's disease were identified. 91 of these 198 patients died after follow-up of up to 12 years (mean 6·1, SD 2·4). No losses to follow-up occurred. The mortality rate was 8·1 deaths per 100 person-years (95% CI 6·6–9·9) and was increased by 50% over those without Parkinson's disease (SMR 1·51, 95% CI 1·22–1·87). Independent baseline prognostic factors for increasing mortality were age (hazard ratio for 10 year increase 1·91, 1·43–2·54), male sex (1·84, 1·17–2·91), more bradykinesia (5 point increase 1·40, 1·14–1·72), more axial relative to limb features (1 unit increase in axial to limb ratio 1·81, 1·22–2·68), and comorbidity (1 unit increase in Charlson score 1·33, 1·14–1·56). A prognostic model incorporating these prognostic factors had good discrimination (Harrel's C-statistic=0·84) and calibration. An electronic predictive risk calculator was created. Interpretation This estimate of mortality is greater than many previous estimates, probably because we used a more representative cohort with a higher proportion of frail and elderly people than in most previous studies. Identification of prognostic factors and their combination in a prognostic model will allow individualised risk prediction, and facilitate stratified medicine and improved clinical trial design. Further work will include extern