Predicting Parkinson disease in the community using a nonmotor risk score

Parkinson Disease (PD) is the second most common neurodegenerative disorder among elderly [1]. At present, no treatment can effectively modify disease progression in patients with PD. This may be due to the advanced stage of pathology that PD patients already have at the time of clinical diagnosis [2]. The identification of persons from the general population who are at high risk of PD might open the door to earlier diagnosis, and possibly enable early symptomatic treatment. Equally important, it would enable the selection of persons who, possibly after additional refined screening, can be enrolled in neuroprotective trials.

In the most recent comprehensive meta-analysis of nonmotor features and risk factors for PD to date, several variables were determined to affect the risk of PD [3]. Subsequently, a cohort study in the United Kingdom was initiated (PREDICT-PD) to assess the validity of a risk score based on 11 of these variables to prospectively predict PD. In cross-sectional analyses at baseline, the PREDICT-PD risk score was associated with several proxies for PD [4]. However, the prospective usefulness of the PREDICT-PD risk score for PD remains unclear.

In this study, we investigated the prospective prognostic value of the PREDICT-PD risk score in an independent, population-based sample with 20 years of follow-up.

The most prevalent nonmotor risk factors were coffee and alcohol use, while constipation and a family history of PD were the least prevalent (Table 1). During follow-up (87,321 person-years, median 16.1 years), 110 individuals had incident PD (age-adjusted incidence rate 1.4 per 1000 person-years) of whom 56 were men and 54 were women. In total, 3713 persons died, and 1021 were diagnosed with incident dementia while at risk of parkinsonism. In our population, the only risk factors that were independently associated with incident PD were current smoking, former smoking and depression (Table 1). As shown in Online Resource 2, women had effect estimates of laxative use, family history, hypertension, NSAID use, CCB use, and alcohol for incident PD that were direction-consistent with the meta-analysis, whereas men had opposite estimates. Furthermore, we observed a significant association between family history and incident PD in women, but not in men.

Predicted 20-year risk of PD ranged from 0.7 to 18.8 % in model I (median 2.2 %), and from 0.5 to 22.5 % (median 2.2 %) in model II (Fig. 1). During follow-up, persons in the highest PREDICT-PD risk score tertile consistently had the highest cumulative hazard of incident PD (Online Resource 3). The PREDICT-PD risk score was independently associated with incident PD and yielded a small, non-significant improvement in discrimination of incident PD beyond age and sex (Table 2; ΔC = 0.018 [−0.005; 0.041]). Compared to model I, model II slightly improved overall classification of PD risk.

The association between the PREDICT-PD risk score and incident PD was strongly modified by sex (p = 0.004). Stratified analyses showed that the risk score was associated with incident PD independently of age in women but not in men. In line with this, risk prediction of PD based solely on age was more accurate in men than in women, but this difference faded after application of the PREDICT-PD risk score. Classification of PD risk was improved by model II in women, but not in men.

After restriction of follow-up to 5 years, discrimination and risk classification of incident PD did not significantly improve (ΔC = 0.008 [−0.022; 0.037] and risk classification (NRI = 0.012 [−0.091; 0.145]) from model I to II. Similarly, after restriction of follow-up to 10 years, prediction did not improve (ΔC = 0.013 [−0.011; 0.038] and NRI = 0.031 [−0.069; 0.140]).

In this prospective, population-based sample with 20 years of follow-up, we found that the PREDICT-PD risk score yielded a small, non-significant improvement in overall discrimination and classification of incident PD. This was due to improvement of PD risk prediction in women to the level of men.

At present, there are no validated methods to identify persons at high risk for PD from the general population so that they can be monitored for onset of symptoms or enrolled in neuroprotective trials. The recently proposed PREDICT-PD risk score was based on a meta-analysis of early nonmotor features and risk factors [3, 4]. Strengths of our study were its prospective design and inclusion of community-dwelling individuals irrespective of PD risk. Compared to cross-sectional case–control data, such as from the multi-center Parkinson’s Progression Marker Initiative [14], prospective community-based studies such as the Rotterdam Study have the advantage that all participants (i.e., both PD future cases and controls) were included and followed up using the same methodology, presumably ensuring a realistic estimate of the risk of incident PD in the general population. Further strengths include long duration of follow-up for PD (median 16.1 years) and standardized assessment of PD diagnosis. In addition, our sample was completely independent of discovery samples used for relative risk estimates in the meta-analysis [3].

Limitations included lack of data on erectile dysfunction as well as the assessment of head trauma and depression using a single question. In addition, we used laxative medication as a proxy for constipation, which likely caused a severe underestimate of the true prevalence of constipation, since many people who suffer from constipation do not use drugs and change their dietary and lifestyle habits. In our sample, only a small proportion of male participants who did not develop incident PD and not a single male incident PD patient used laxatives at baseline, suggesting that our underestimate may have been larger in men than in women. If we would have had complete information on these factors, the PREDICT-PD risk score may have improved PD prediction significantly in our population. Furthermore, we lacked histologic confirmation of PD diagnosis, which may have introduced non-differential misclassification of PD cases. Also, we may have been underpowered to detect a small significant improvement in PD prediction, especially in the middle-long term (i.e., 5 years) The estimates used in the PREDICT-PD risk score were mostly based on studies that did not assess the majority of variables simultaneously, and the estimates were not sex-specific. In our sample, only 3 risk factors were independently associated with incident PD (current smoking, former smoking and depression), which may indicate that the meta-analyzed estimates were inflated due to limitations of the meta-analysis, such as publication bias, a substantial degree of selection in some discovery samples, or insufficient adjustment for covariates [3]. Alternatively, we may have been underpowered to detect significant associations with PD for separate risk factors, and limitations in our assessment methods may have led to underestimates of true associations. Future collaboration across cohort studies who have prospectively assessed (nearly) all risk variables in the score will probably increase the accuracy of risk estimates. Similarly, while we observed clear sex differences in associations between risk factors and incident PD, most of the sex-specific associations in our sample were non-significant. Collaborative studies may distinguish true sex differences from limitations in assessment methods that may have worse in men (e.g., laxative use).

Future studies can further build on the PREDICT-PD risk score by focusing on three other key aspects. First, some relatively common nonmotor risk factors for PD were not yet part of the risk score, such as impaired olfactory function [15]. Recently, dedicated olfactory function testing was shown to distinguish patients with a PD diagnosis from controls with very high accuracy [14]. Although the long-term prospective predictive value of olfactory testing for PD in the community has not yet been demonstrated empirically, a previous study showed that impaired olfaction is associated with PD up to 4 years before clinical diagnosis [15]. Therefore, inclusion of prospective measures of olfactory function in the risk score may further improve prediction of PD in the community. Second, while the Rotterdam Study comprises a suburban-based study population with only few farmers, discrimination and classification accuracy in other communities may be improved by inclusion of data on rural living and farming occupation. Third, motor features were not included in the risk score. Even in the absence of objective signs on routine screening, prediagnostic PD patients have subjective parkinsonian complaints more frequently than controls [16], and tremor is the most common presentation of PD patients in primary care practice 10 years before clinical diagnosis [17]. The advancement of dedicated motor screening tests might not only lead to reliably detection of PD in select subgroups of very high-risk persons (e.g., RBD-patients [18]), but potentially also in community-dwelling persons.

In conclusion, the PREDICT-PD risk score is a small step forward towards predicting incident PD in the community, in particular in women, but there is still a clear need for improvement.