Risk factors of levodopa-induced dyskinesia in Parkinson’s disease: results from the PPMI cohort

Levodopa-induced dyskinesias (LID) negatively impact on the quality of life of patients with Parkinson’s disease (PD). We assessed the risk factors for LID in a cohort of de-novo PD patients enrolled in the Parkinson’s Progression Markers Initiative (PPMI). This retrospective cohort study included all PD patients enrolled in the PPMI cohort. Main outcome was the incidence rate of dyskinesia, defined as the first time the patient reported a non-zero score in the item “Time spent with dyskinesia” of the MDS-UPDRS part IV. Predictive value for LID development was assessed for clinical and demographical features, dopamine transporter imaging (DaTscan) pattern, cerebrospinal fluid (CSF) biomarkers (Aβ42, total tau, phosphorylated tau, total α synuclein) and genetic risk score for PD. Overall, data from 423 PD patients were analyzed. The cumulative incidence rate of LID was 27.4% (95% CI = 23.2–32.0%), with a mean onset time of 5.81 years from PD diagnosis. Multivariate Cox regression analysis showed several factors predicting LID development, including female gender (HR = 1.61, 95% CI = 1.05–2.47), being not completely functional independent as measured by the modified Schwab & England ADL scale (HR = 1.81, 95% CI = 0.98–3.38), higher MDS-UPDRS part III score (HR = 1.03, 95% CI = 1.00–1.05), postural instability gait disturbances or intermediate phenotypes (HR = 1.95, 95% CI = 1.28–2.96), higher DaTscan caudate asymmetry index (HR = 1.02, 95% CI = 1.00–1.03), higher polygenic genetic risk score (HR = 1.39, 95% CI = 1.08–1.78), and an anxiety trait (HR = 1.02, 95% CI = 1.00–1.04). In PD patients, cumulative levodopa exposure, female gender, severity of motor and functional impairment, non-tremor dominant clinical phenotype, genetic risk score, anxiety, and marked caudate asymmetric pattern at DaTscan at baseline represent independent risk factors for developing LID.


INTRODUCTION
Parkinson's disease (PD) is a neurodegenerative disorder characterized by motor and non-motor symptoms. 1 So far, levodopa still stands as the most effective symptomatic treatment for PD. 2 However, long-term dopamine repletion treatment may lead to motor fluctuations, such as wearing-off and dyskinesias. 3,4 Several factors participate in the development of motor fluctuations, including loss of dopaminergic neurons in the substantia nigra, changes in pre and post-synaptic striatal activity with chronic pulsatile stimulation of dopamine receptors, 5 and the daily dosage of levodopa. 6 Motor fluctuations highly impact on the quality of life of people with PD, representing a major criteria for eligibility to advanced treatments. 6 Observational studies have shown that more than 50% of PD patients treated with levodopa for more than 5 years develop levodopa-induced dyskinesia (LID). 7 Several risk factors for LID have been proposed, including levodopa dosage, 8 treatment duration, 9 female gender 10 and low body weight. 11 Other factors have been investigated as predisposing to LID, including neuroimaging findings, with conflicting results. [12][13][14][15] It is worth noticing that the available studies are poorly comparable, given the different methodological approaches and follow-up duration, with patients being rarely followed up ever since de novo stage. The Parkinson's Progression Markers Initiative (PPMI) is a large-scale international prospective observational study, started in 2010, designed to identify markers of disease progression in de novo PD patients. Clinical, neuroimaging and CSF/blood biomarkers are collected yearly. We wanted to define factors predictive of LID development already in de novo stage of PD.
Analysis of multiple risk factors for LID Multivariate Cox regression analysis has shown that a combination of multiple factors was fairly accurate in predicting the onset of dyskinesia (Concordance = 0.74, 95% CI 0.681 to 0.80). The Cox proportionality assumption was validated with chi-square test for Schoenfeld residuals (overall p-value = 0.383), and visual inspection of martingale residuals against individual covariates supports the hypothesis of linearity. The final model included female gender (HR = 1.63, 95% CI 1.06 to 2.50, Random survival forests for LID After fitting random survival forests with 1000 trees and missing data imputation, 12 variables were selected using minimal depth criterion leading to an out-of-bag error of 24%. The final model was consistent with the results of multivariate Cox regression in terms of the selected variables. Using random survival forests, also CSF biomarkers entered in the panel of predictive factors (Table 2) for LID development, with α-syn ranking 6th according to the importance of all variables (Fig. 2).

DISCUSSION
LID negatively affects the quality of life of patients with PD. Despite extensive research, conflicting results have been gathered regarding modifiable and non-modifiable risk factors for LID development, 16 with few studies assessing de novo patients. In this study, evaluating data of de novo PD patients included in the PPMI cohort, we identified a set of seven independent risk factors for LID development that can be taken into consideration already in the very early phase of the disease. First of all, we confirm that female gender represents a crucial non-modifiable predictive factor for LID, 10 independently from body weight and genetic factors. 10,11,17 Second, our results underline that cumulative exposure to levodopa is positively associated with the development of LID, in line with the results deriving from other large cohorts of PD patients. 18,19 Noteworthy, it is well known that levodopa accelerates the loss of nigrostriatal dopamine nerve terminals, a key pathophysiologic element in the development of dyskinesia. Dyskinesia in Parkinson's disease are associated with changes in long term neuroadaptation and neuronal synaptic plasticity which in turn are linked to dopamine transporters and receptors density, respectively at a presynaptic and a postsynaptic level. 6 Based on this concept, both positron emission tomography (PET) and single photon emission computed tomography (SPECT) were used to assess changes in neurotransmitter pathways involved in dyskinesia and, subsequently, to identify imaging biomarkers for LID development. Notably, lower dopamine transporter activity in the putamen evaluated by 18 F-FP-CIT-PET in de-novo PD patients has been described as significant predictor of LID. 13 Our results indicate that dopamine deficits in the contralateral putamen in de-novo PD patients is an independent predictor of a shorter time to levodopa initiation. Besides, both putamen asymmetry and caudate asymmetry indices evaluated by [ 123 I]FP-CIT-SPECT significantly correlated with the development of dyskinesia, with the caudate asymmetry entering in the multivariate model. These findings reflect previous evidences indicating a positive relationship between the striatal asymmetric index and the magnitude of response to levodopa. 20 Thus, PD patients with higher striatal asymmetric index show an increase of both response to levodopa and susceptibility to dyskinesia.
The probability of developing dyskinesia in Parkinson's disease is also influenced by the initial clinical phenotype. Our findings showed that tremor-dominant (TD) phenotype is at lower risk of LID compared to PIGD or Intermediate. Previous studies have already demonstrated that tremor-dominant manifestation at disease onset is associated with a reduced risk of LID compared to rigid-akinetic (RA) phenotype. 21 The reason of a lower risk of dyskinesia among TD patients may lay in different patterns of nigrostriatal denervation, morphologic lesions of the basal ganglia subregions and pathophysiological mechanisms between different phenotypes. 5 Furthermore, TD patients usually show lower striatal dopamine depletion compared to RA patients on [ 123 I]FP-CIT-SPECT, 22 suggesting a less pronounced predisposition to LID. Nigrostriatal dopamine depletion is one of the main prerequisite for developing dyskinesia and dopaminergic denervation is increased by disease severity. Therefore, it is not surprising that disease severity at baseline represents an important predictive factor for LID. Our findings indicate that UPDRS Part III score in denovo patients is a significant clinical biomarker to predict dyskinesia. However, it is not fully consistent with data from STRIDE-PD trial which showed UPDRS Part II score as risk factor for dyskinesia, whereas UPDRS Part III did not correlate with dyskinesia development. 19 Such contrasting results may be due to the fact that UPDRS motor score does not always reflect precisely the status of presynaptic dopamine denervation evaluated by PET or SPECT. 23 The relationship between the severity of disease and the susceptibility to dyskinesia is also supported by the significant inverse correlation we found between ADL Scale score and risk of LID. Thus, patients with higher disease severity testified by greater impairment on daily living activities are more prone to develop dyskinesia. Genetic susceptibility to dyskinesia is of great interest, with conflicting results deriving from several studies focusing on single-nucleotide polymorphisms (SNPs) of different genes. PPMI polygenic risk score was developed to explain the risk of idiopathic PD onset so far and it does not include several genes that are known to increase the risk of developing LID. 24 However, to the best of our   knowledge, for the first time in this study we highlighted a correlation between cumulative effect of known genetic risk variants of PD and LID development. Such results, if corroborated by extensive genome-wide association studies (GWAS), might represent a focus for future research. Interestingly, we also found that anxiety was associated to an increased risk of dyskinesia. This could be linked to the relationship between dopaminergic dysfunction and neuropsychiatric symptoms in early PD. 25 Although CSF biomarkers did not explain LID onset in the Cox regression, we found that α-syn ranked 6th, according to the variable importance criterion, in the random survival forests model. This suggest that further investigations are warranted for exploring the role of CSF α-syn in predicting LID, with special attention to pay for the observation in larger cohorts or the measurement of oligomeric and phosphorylated forms.
There are limitations in our data, mainly due to the evaluation of dyskinesia. Even if specific scales have been validated to assess dyskinesia in PD, 26 they were not used for evaluating qualitative and quantitative aspects of dyskinesia in the PPMI. Therefore, our results neither identified predictive factors for different types of dyskinetic complication in PD (chorea versus dystonia) nor found a possible explanation for the severity of dyskinesia.
In summary, our findings indicate that data deriving from a large cohort of de-novo PD patients monitored longitudinally are useful in understanding the composite aspects involved in the progression of disease. Our results highlight the role of several factors in determining dyskinesia, thus providing useful information for future design of both biomarker studies and randomized clinical trials.

Study design
Overall, 423 de novo PD participants were enrolled in the PPMI study between January 2011 and December 2012. Data were obtained from the PPMI database accessed 30 December 2017.

Participants
The inclusion criteria for entering PPMI were: (i) age >30; (ii) presence of at least two parkinsonian signs such as bradykinesia, rigidity and resting tremor or have an asymmetric resting tremor, or asymmetric bradykinesia; (iii) having received the diagnosis not earlier than two years before enrollment; (iv) documented reduced striatal 123-I Ioflupane dopamine transporter (DatScan, GE Healthcare, Arlington Heights, IL) imaging binding consistent with PD; (v) no ongoing symptomatic therapy. Each PPMI participant received extensive assessment of motor and non-motor features.

Standard protocol approvals, registrations, and consents
Each participating PPMI site received approval from an ethical standards committee on human experimentation before study initiation. Written informed consent for research was obtained from all individuals participating in the study.

Baseline features
With the aim of defining predictive factors od LID development among de novo patients, we have investigated several variables available at baseline, including: (i) demographics (age, gender, family history, disease duration, education years); (ii) bradykinesia, rigidity and tremor were assessed according to UK Parkinson's Disease Society Brain Bank Criteria (iii) motor features (International Parkinson's disease and Movement Disorder Society-Unified Parkinson's Disease Rating Scale (MDS-UPDRS)-Part II and III, total tremor score, postural instability-gait disturbance (PIGD) score, tremor/PIGD motor phenotype, Schwab-England activities of daily living (ADL) score, bradykinesia, tremor, and (iv) age/education adjusted Montreal Cognitive Assessment (MoCA); (v) non-motor manifestations (MDS-UPDRS-Part I), olfactory dysfunction via University of Pennsylvania Smell Identification Test (UPSIT) score, autonomic dysfunction via Scale for Outcomes in Parkinson's disease-Autonomic (SCOPA-AUT) total score, REM sleep behaviour disorder (RBD) via RBD screening questionnaire (RBDSQ) score, sleep disturbances via Epworth Sleepiness Score (ESS), anxiety via State-Trait Anxiety Inventory scores (STAI), depression via Geriatric Depression Scale (GDS-15); (vi) genetic risk score including 28 independent risk variants for PD that have been selected according to the results of a meta-analysis of PD genome-wide association studies, 24 also including p. N370S in GBA and p.G2019S in LRRK2 27,28 ; (vii) CSF biomarkers (CSF amyloid-β1-42, total (t)-tau, and phosphorylated tau (P-tau181) and asynuclein); (viii) dopamine transporter imaging striatal-binding ratios (DATscan) (single-photon emission computed tomography (SPECT) with the DAT tracer 123I-ioflupane at baseline, with striatal-binding ratio calculated for left and right putamen separately using the occipital lobe as reference, obtaining ipsilateral, contralateral, mean measurements and asymmetry indices). The asymmetry index for caudate and putamen was calculated, according to PPMI indications for deriving variables, as the difference between left and right divided by the mean value.

Levodopa exposure
Levodopa equivalent daily dose (LEDD) was reported for each participant reporting the initiation of dopaminergic therapy. We included total LEDD as the cumulative exposure to all dopaminergic drugs, as well as levodopa LEDD as the cumulative exposure to levodopa until the onset of dyskinesia (observed events) or the study exit (censored events).

Levodopa-induced dyskinesia
The primary outcome was the incidence of LID, defined as the first time the patient reported a positive score in the item "Time spent with dyskinesia" of the MDS-UPDRS part IV.

Statistical analysis
Statistical analyses were performed using R software version 3.4. Continuous variables were described by means and standard deviations, while categorical ones were reported as count and percentages. Predictors of LID onset were assessed using multivariate Cox proportional hazards regression models. In the absence of a conversion event, data were censored at the most recent clinic visit. We have first assessed the role of each risk factor in conjunction with the total LEDD. In order to avoid rejection of potentially important variables due to uncontrolled confounders, a p-value lower than 0.20 was used as screening criterion to consider the risk factor as candidate for the multivariate analysis. Backward elimination based on the Akaike's information criterion was used to select a final model. Hazard proportionality was assessed through analysis of scaled Schoenfeld residuals whereas martingale residuals were plotted against continuous covariates to detect nonlinearity. As a complementary approach, we have applied random survival forests, 29 a machine learning technique to determine important variables for the prediction of individual survival times. Variable selection was carried out using a conservative approach based on minimal depth criterion. Significance level of 5% was assumed for all the analyses.