Progression of motor and nonmotor features of Parkinson's disease and their response to treatment


Professor Nick Holford, Department of Pharmacology & Clinical Pharmacology, University of Auckland, 85 Park Road, Private Bag 92019, Auckland, New Zealand. Tel.: +64 9 923 6730. Fax: +64 9 373 7090. E-mail:



• There is marked variability among patients with Parkinson's disease (PD) in the rate of progression of status (severity) assessed with a global functional score (Unified Parkinson's Disease Rating Scale; UPDRS). It has been hypothesized that there are distinct PD subtypes with different rates of progression.

• Previous studies attempted to quantify the rates of progression for tremor-dominant and postural instability and gait disorder (PIGD)-dominant subtypes using only baseline clinical features.


• We used a nonlinear mixed effects modelling approach to describe the time course of the four cardinal features of PD before and during anti-parkinsonian treatment.

• Tremor-dominant and PIGD-dominant subtypes appear to be different stages of the disease rather than persistent attributes in individual patients and do not explain variability in progression among patients. Tremor progresses more slowly than other cardinal features with and without drug treatment. Postural instability and gait disorder is much less sensitive to the symptomatic effects of levodopa than the other cardinal features.

• We have extended the finding that anti-parkinsonian treatments have symptomatic and disease-modifying effects on overall function and demonstrate similar effects on each of the four cardinal features of PD.


(i) To describe the progression of the cardinal features of Parkinson's disease (PD); (ii) to investigate whether baseline PD subtypes explain disease progression; and (iii) to quantify the symptomatic and disease-modifying effects of anti-parkinsonian treatments.


Data were available for 795 PD subjects, initially untreated, followed for up to 8 years. Cardinal features [tremor, rigidity, bradykinesia, and postural instability and gait disorder (PIGD)] were derived from the total unified Parkinson's disease rating scale (total UPDRS), cognitive status from the mini-mental status exam score (MMSE) and depression status from the Hamilton depression scale (HAM-D). Analysis was performed using a nonlinear mixed effects approach with an asymptotic model for natural disease progression. Treatment effects (i.e. symptomatic and disease modifying) were evaluated by describing changes in the natural history model parameters.


Tremor progressed more slowly (half-time of 3.9 years) than all other motor features (half-time 2–3 years). The MMSE progression was negligible, while HAM-D progressed with a half-time of 5 years. Levodopa had marked symptomatic effects on all features, but low potency for effect on PIGD (ED50 of 1237 mg day−1 compared with 7–24 mg day−1 for other motor and nonmotor features). Other anti-parkinsonian treatments had much smaller symptomatic effects. All treatments had disease-modifying effects on the cardinal features of PD. Baseline PD subtypes only explained small differences in disease progression.


This analysis indicates that tremor progresses more slowly than other cardinal features and that PIGD is less treatment responsive in early PD patients. There was no evidence of baseline PD subtypes as a clinically useful predictor of disease progression rate. Anti-parkinsonian treatments have symptomatic and disease-modifying effects on all major features of PD.


The rate of progression of Parkinson's disease (PD), as well as the array of parkinsonian signs and symptoms, differs widely among individual patients [1]. This observation raises the question of whether there are distinct subtypes of Parkinson's disease with different rates of progression. The question about the existence of subtypes is important for the care of patients, in that subtypes may dictate different treatment approaches, as well as improve counselling of patients about prognosis. Furthermore, recognition of subtypes may allow recruiting or stratifying subjects in clinical trials to reduce variability and improve the efficiency of trials. We have investigated the proposal that there may be distinct disease subtypes [2] by describing the rates of progression of four cardinal PD features [tremor, rigidity, bradykinesia, and postural instability and gait disorder (PIGD)], and the effects of treatment on the time course of these features.

Questions about rates of progression have been difficult to answer because of the confounding effects of treatment. Nevertheless, there is evidence that the cardinal features of PD progress at different rates. Louis et al. studied patients on treatment and found that tremor, measured with the motor portion of the Unified Parkinson's Disease Rating Scale (UPDRS), did not significantly progress with time, whereas the other features progressed at different annual rates (total motor UPDRS 1.5%, rigidity 2%, bradykinesia 2.1% and PIGD 3.1% per year) over 8 years follow-up [3]. Jankovic and Kapadia observed a 77% faster progression rate of total UPDRS in patients who were PIGD dominant at baseline than in those who were tremor dominant at baseline after an average of 6 years follow-up [4]. Furthermore, greater disability and more depression were observed in patients who were PIGD dominant than in those who were tremor dominant for the Deprenyl and Tocopherol Antioxidative Therapy of Parkinsonism (DATATOP) cohort at baseline [2].

Disease progression rates have typically been described as the difference between the first and last clinic visits and expressed as an annual rate of change [5]. As the majority of PD studies are observational, with unbalanced designs, variable dosing patterns and treatment combinations, common statistical methods have not been able to distinguish natural progression from treatment effects. The shape of the time course of the four cardinal features in treated and untreated PD patients has therefore remained largely unknown. Holford et al. described the time course of treatment effects on total UPDRS progression by studying the DATATOP cohort [6]. This analysis suggested evidence of symptomatic and disease-modifying effects of levodopa, selegiline, bromocriptine and pergolide. The model developed for the DATATOP cohort was externally validated by successfully using clinical trial simulation to predict the results of the ELLDOPA trial [7]. As shown in these previous analyses, the modelling method handled the observational nature of the data, separated the symptomatic and disease-modifying effects of anti-parkinsonian drugs during the course of PD, and described the time courses of PD status in subjects before and after anti-parkinsonian medications were initiated.

We have extended this earlier work, which used total UPDRS (UPDRST) as a marker for disease status, by applying the same methodology to characterize the progression of the UPDRS subscales [i.e. the cardinal features plus activities of daily living (ADL)] and nonmotor features (i.e. cognitive function and depression) in the DATATOP cohort. We used pharmacodynamic models to quantify the influence of various anti-parkinsonian treatments in terms of symptomatic or disease-modifying effects on the parameters of the disease progress models. Lastly, we investigated whether the subtype classification of each patient at baseline (i.e. tremor dominant or PIGD dominant) could account for between-patient differences in disease progression rates.


Data source

The DATATOP study enrolled 800 patients. Five of these patients dropped out very early in the clinical trial. The DATATOP cohort is 795 patients who were followed up over a period of nearly 8 years. Treatment information, patient demographics and clinical assessments were obtained from the DATATOP cohort. The average follow-up time was 4.97 years, with a maximum of 7.73 years. As characteristics of the DATATOP trial design and patients have been published [2, 8], only the relevant details are described in the present analysis. Briefly, the trial started with 800 treatment-naïve subjects who were randomized to placebo, tocopherol, selegiline or both drugs and continued for up to 2 years. Selegiline delayed the need for symptomatic anti-parkinsonian therapy, while tocopherol was found to be inactive and, as a result, all patients were switched to selegiline. As PD disability emerged or worsened during the subsequent follow-up period, symptomatic treatments, including levodopa (median daily dose of 300 mg), pergolide (median daily dose of 1.25 mg) and bromocriptine (median daily dose of 15 mg) were added and adjusted as clinically indicated. Only one dose regimen of selegiline (10 mg day−1) was used. Approximately 50% of patients were receiving a combination of levodopa and selegiline at some time during the trial.

In the present analysis, clinical markers for the four motor features and disability were measurements derived from total UPDRS (range 0–176) and UPDRS subscales. The tremor subscale was the sum of 0 (none) to 4 (severe) scores for resting tremor in five body parts and for postural and action tremor in the hands (range 0–28). The rigidity subscale was the sum of rigidity scores of 0 (none) to 4 (severe) for limbs and neck (range 0–20). The bradykinesia subscale was the sum of 0 (none) to 4 (severe) scores for bradykinesia for right and left finger tapping, opening and closing the hands, pronation and supination of the arms, tapping the feet and for global bradykinesia (range 0–36). The PIGD subscale was the sum of falling, freezing, walking, gait and postural stability (range 0–20). The ADL subscale was the total of UPDRS part II (range 0–52). Markers for nonmotor features were the mini-mental state exam (MMSE; range 0–30) score for cognitive impairment and Hamilton rating scale for depression (HAM-D; range 0–30). When the patients were receiving symptomatic therapies (levodopa, bromocriptine or pergolide), the evaluation was performed without regard to the ‘on’ state (i.e. the patient felt that they were experiencing beneficial effects from treatment).

All disease status variables were modelled as continuous time-varying scores. As each of these variables is determined by a combination of random-effect parameters, there is no simple transformation to constrain the predicted value within the range of each scale. Even if such a transformation were available, we have no reason to suppose that it would more correctly describe the distribution, as the boundaries of a range are approached. For simulation purposes, we discarded any prediction lying outside its clinical range.

Repeated disease status measurements collected at intervals of 3–6 months were available for 795 of 800 patients. The average number of observations and average score of each sign at the beginning and end of DATATOP are shown in Table 1. Note that all features have an increasing value for worsening status except for MMSE, which decreases when the status is worsening. None of the patients was diagnosed with dementia or depression at baseline.

Table 1.  Summary of observed Unified Parkinson's Disease Rating Scale (UPDRS) subscales and nonmotor Parkinson's disease (PD) features
PD featuresAverage no. of observations per patient (maximum)Observed mean score ± SD (range)
First observationsLast observations
  1. ADL, activities of daily living; HAM-D, Hamilton rating scale for depression; MMSE, mini-mental state exam; PIGD, postural instability and gait disorder.

UPDRS subscales   
 Tremor19 (30)3 ± 2 (0–15)3 ± 3 (0–21)
 Rigidity19 (30)4 ± 3 (0–18)5 ± 3 (0–20)
 Bradykinesia19 (30)7 ± 4 (0–21)10 ± 6 (0–36)
 PIGD19 (30)2 ± 1(0–8)3 ± 3 (0–19)
 ADL19 (30)7 ± 4 (0–21)11 ± 6 (0–44)
 HAM-D14 (22)3 ± 3 (0–15)3 ± 3 (0–30)
 MMSE10 (16)29 ± 1 (23–30)28 ± 2 (11–30)

Disease progress models

The specifics of disease progress models to describe the time course of total UPDRS for the DATATOP cohort have been described by Holford et al. [6]. Consequently, only model-building details useful for the comprehension of the present work are presented here. Both linear and nonlinear models of progression were evaluated.

Linear progression  Patient's disease status (S) at any time (S(t)) was estimated from the patient's baseline status (S0) and the rate of progression (α; Equation 1). The treatment effects were modelled as functions that offset the disease progress curve, Eoffset(t) and/or changed its slope, EDM(t) (Equation 2).


Nonlinear progression  The Gompertz model has been used previously to describe the approach of Parkinson's disease status to an asymptote when little further progression is discernible [6]. The parameters describing the Gompertz model are the progression time constant (Tprog) and steady-state status (Sss; Equation 3).


In the presence of treatment, similar to the linear model, symptomatic effect would be defined by a nonzero Eoffset(t). A disease-modifying effect occurs if treatment changes either Tprog (Equation 4) or Sss (Equation 5), described by a nonzero EDM(t).


The progression time constant (Tprog) is related to a second-order rate constant. We define the half-time of progression when the status is halfway between the baseline and the asymptotic disease status (Equation 6).


Dominant motor subtypes

The ratio of tremor score to PIGD score was computed at baseline (tremor/PIGD ratio). This ratio has been used to describe motor subtypes known as tremor dominant and PIGD dominant [2]. When baseline PIGD was zero, the tremor/PIGD ratio could not be computed, so the value was assumed to be 20, the highest ratio observed in the cohort.

Sixty-two per cent (n= 499) of patients were tremor-dominant subtype and 32% (n= 253) were PIGD-dominant subtype at study entry (baseline subtype); the remaining 6% of patients were in an indeterminate group. The baseline subtype effect on disease progression was quantified using the tremor-dominant subtype as a reference value. The indeterminate group was omitted from this analysis, because our main interest was the relative rate of progression between PIGD- and tremor-dominant baseline subtypes.

To account for differences in the time course of disease progression for the baseline subtype, the Tprog parameter was modified as shown in Equation 7, where FPIGD is the fractional change for patients with the PIGD baseline subtype.


Pharmacodynamic models

Symptomatic effects  A sigmoid Emax model was used to describe the symptomatic effect for levodopa (Eoffset,LD(t)). The time course of Emax (i.e. Emax(t)) was described by an exponential increase starting from Emax0, which was indistinguishable from a value of 0, so it was assumed to be zero [6, 9]. The value of Emax(t) reached an asymptote of BEML with a half-life of TEML. At steady-state, BEML is equal to Emax(t).


The effect is assumed to be proportional to the concentration in a hypothetical effect compartment, Ce (Equation 8). The effect compartment lags behind the average steady-state plasma drug concentration (assumed to be proportional to dose rate) with an equilibrium half-life TEQL and therefore accounts for the delay in attainment of the symptomatic effect after a change in dose rate. As no drug concentrations were measured in the DATATOP trial, the concentrations in the effect compartment at equilibrium were assumed to be proportional to daily doses. The ED50 is the dose relative to the median levodopa dose rate (i.e. 300 mg day−1) at which 50% of Emax(t) is observed.

Based on previous model-building results [6], a linear model with no delay (Equation 9) was used to describe the symptomatic effects for selegiline (SG), bromocriptine (BC) and pergolide (PG), where BOD, BOB and BOP are parameters describing the effect on Eoffset(t) produced by each treatment at the median daily dose. The linear model assumed no delay in symptomatic effects; therefore, values of Ce for these three treatments are predicted from the ratio of the daily dose to the median daily dose. The symptomatic effects are reported as the percentage change relative to the baseline disease status, S0.


Disease-modifying effects  Treatments may also exhibit disease-modifying effects by changing either Tprog or Sss. Equation 10 shows the disease-modifying effect of levodopa, ELD(t), where KPLT is a parameter that relates CeLD(t) to the disease-modifying effect of levodopa. There is no delay in CeLD(t), the effect compartment concentration of levodopa, which is predicted from the ratio of the daily dose to the median daily dose.


In Equation 11, KPDT is the disease-modifying parameter for levodopa and pergolide. There is no delay in effect compartment concentrations, which are predicted from the ratio of the daily dose to the median daily dose. An interaction parameter, FLXDT, was estimated when levodopa and selegiline were given in combination (Equation 11).

For the purpose of comparison among PD features, the disease-modifying effects on Sss are reported as percentage change relative to the natural history steady-state disease status, and effects on Tprog are reported as percentage change relative to Tprog of the natural history progression.

Dropout model

A dropout model was included to simulate subjects who did not complete 8 years follow-up due to death or other reasons [10]. Details of dropout model development are described in an accompanying report [11]. The dropout model described the distributions of time to dropout or death using hazard functions involving UPDRS and selegiline treatment. The model predicted that subjects with worsening disease status (i.e. high UPDRS) were more likely to drop out of the trial or die, whereas those who remained on selegiline treatment were more likely to stay in the trial.

Random effects

The distribution of individual parameter values (e.g. the ith individual) was estimated with the random individual difference being proportional to the magnitude of the population parameters (e.g. Equation 12). This allows for either worsening or improved disease progress in individual responses. A normal distribution was assumed for η, with a mean of 0 and a variance of ω2 (OMEGA).


The correlation between individual patient parameters was investigated by estimating the covariance of values of η. Some parameters could not have negative values because of their biological meanings (e.g. ED50, S0, Tprog or Sss), and therefore an exponential model was used for the variability (e.g. Equation 13).


The residual variance model was described as an additive random error, ε, that is normally distributed with a mean of 0 and a variance of σ2 (Equation 14).


The estimates of population parameter variability (PPV) and residual unidentified variability (RUV) are calculated from the square root of the NONMEM estimate of the variance of the random effect [13].

Parameter estimation

A mixed effects modelling approach was used to describe the time course of disease progression and response to treatment [12]. Data analysis was performed using NONMEM (version VI, release 1.3) [13] with the first-order conditional estimation method. NM-TRAN abbreviated code for the model has been published [6].

Model selection

Selection of disease-modifying effect models was primarily based on changes in the NONMEM objective function value (OFV) using the likelihood ratio test with a nominal Type 1 error of 0.005. The degrees of freedom for the likelihood ratio test were based on the difference in the number of all parameters in the model, including the elements of OMEGA.

Model evaluations

Owing to rounding errors, the NONMEM covariance step could not be completed for computation of asymptotic standard errors. NONMEM 7 with the UNCOND option was also not successful. The standard error of parameter estimates was assessed by a nonparametric bootstrap method. Owing to long individual runtimes (e.g. 5 days), the number of bootstraps was limited to 20.

To assess the adequacy of model structure, simulation-based diagnostics were used, with plots of individual observed and predicted time courses of disease status. Model variability components were assessed using a visual predictive check (VPC) based on simulations from the final parameter estimates [14]. Random samples were drawn from the multivariate distributions of η and ε to simulate 100 replications of the data set. Both simulated and observed disease status values were normalized to the typical population predicted values (i.e. PRED-corrected) for each nominal time interval, as described by Bergstrand et al. [15]. This PRED-correction was done to account for the unique dosing pattern for each individual. Visual predictive checks were created for all disease status variables using a combined death and nondeath dropout event model. These dropout estimation models used the empirical Bayes predictions of total UPDRS score as an explanatory variable. There is a technical challenge to create a dropout simulation model for nontotal UPDRS variables. The total UPDRS value for each simulated subject is expected to be highly correlated with other simulated disease status variables. However, this simulation-specific total UPDRS value cannot be predicted by using either observed or empirical Bayes predictions of total UPDRS from the original data set. A parametric model including correlations between disease status variables would require more parameters to be estimated than is currently possible with NONMEM, even for total UPDRS paired with only one other variable. A partial solution to the problem was obtained by assuming that the simulated total UPDRS value was perfectly correlated with the simulated disease status variable value. The total UPDRS hazard model parameters (death and nondeath) were adjusted by the ratio of the population baseline (S0) for each variable to the total UPDRS baseline in order to implement this assumption. For the simulated data replicates, any simulated total UPDRS scores after simulated dropout or death were considered missing. Median and 90% intervals of PRED-corrected observed and simulated total UPDRS scores were plotted to evaluate the model predictions.

As the PRED-corrected VPC relies on the treatments used in the original data, it cannot be used to simulate future trials. We therefore also tested the behaviour of an adaptive dosing algorithm based on dose and treatment changes triggered by the simulated total UPDRS value. The algorithm was based on plausible clinical practice, because there was no defined treatment change protocol for the DATATOP cohort. The adaptive dosing algorithm followed the two DATATOP randomizations for selegiline at study entry and at 5 years. All patients received selegiline from year 2 to year 5. Levodopa was started (250 mg day−1) if the total UPDRS score was greater than 35 and increased in steps of 50 mg day−1 at each 3 months visit if the score was greater than 35, up to a maximal daily dose of 800 mg day−1. If both selegiline and levodopa were being used and the levodopa dose was greater than 300 mg day−1, then a dopamine agonist was added (randomly) as either bromocriptine or pergolide. Bromocriptine was added starting at 7.5 mg day−1 and increased by 1.25 mg day−1 at the following visits up to a maximum of 60 mg day−1. Pergolide was started at a dose of 1.25 mg day−1 and increased by 0.25 mg day−1 at the following visits up to 6 mg day−1.


Model performance

Individual predictions of UPDRS subscales and nonmotor features followed the observed values as illustrated for two illustrative subjects (ID nos 4 and 10; Figure 1). These two subjects showed the typical patterns of adaptive dosing in the trial. At the start of the trial, subject 4 was randomized to receive placebo and subject 10 was randomized to receive selegiline. As the trial progressed, the disease status of subject 4 worsened, which prompted the clinician to start levodopa. Subject 10 also received pergolide in addition to levodopa and selegiline later in the trial.

Figure 1.

Observed time courses (open triangles) of total Unified Parkinson's Disease Rating Scale (UPDRS) and subscales and nonmotor features for two individuals to illustrate the symptomatic and disease-modifying effects of anti-parkinsonian treatments. Individual predicted scores are shown as continuous lines; treatment dosages are normalized to 300 mg day−1 for levodopa (dotted lines), 10 mg day−1 selegiline (short dashed-dotted lines) and 1.25 mg day−1 pergolide (long dashed-dotted lines). Subject ID no. 4 was tremor-dominant subtype [tremor/postural instability and gait disorder (PIGD) ratio of 3.5] and subject ID no. 10 was indeterminate subtype at study entry (ratio of 1.3). Both converted to PIGD-dominant subtype at the end of the study (ratio of 0.5 for both)

UPDRS subscales

Natural disease progress of motor function  Consistent with previous results for total UPDRS in the same patients [6], the Gompertz asymptotic model provided a better fit than the linear model for progression of the subscales (Table 2). Bootstrap averages and standard errors of the estimates are shown in Tables 3 and 4. We have used the relative standard error (RSE) of the parameters to decide which parameters are adequately defined (RSE ≤ 40%). Parameters with RSE > 40% are not reliably estimated. Only parameters with RSE ≤ 40% are shown in Table 5.

Table 2.  NONMEM Objective Function Values (OFVs) for linear and nonlinear models with disease-modifying effects on parameters α, Tprog or Sss
PD featuresLinear modelGompertz models
  • *

    OFV of best-fitted models. ADL, activities of daily living; HAM-D, Hamilton rating scale for depression; MMSE, mini-mental state exam; PD, Parkinson's disease; PIGD, postural instability and gait disorder; UPDRS, Unified Parkinson's Disease Rating Scale.

UPDRS subscales   
 Tremor29 93630 02829 697*
 Rigidity34 19834 04733 832*
 Bradykinesia48 93348 50148 397*
 PIGD17 03516 66016 558*
 ADL46 27546 100*47 109
 HAM-D31 97732 05631 022*
 MMSE13 012*16 99915 621
Table 3.  Bootstrap average and Relative Standard Errors (RSEs) for motor-related structural model parameters
ParameterTotal UPDRSADLBradykinesiaPIGDRigidityTremor
  • *

    ED 50 estimates are relative to the median daily dose (300 mg) of levodopa. ADL, activities of daily living; PIGD, postural instability and gait disorder; UPDRS, Unified Parkinson's Disease Rating Scale.

Natural disease progression parameters
 Baseline statusS021.72.36%6.402.23%5.382.65%1.232.30%2.903.50%2.423.13%
 Steady-state statusSss10912.5%27.212.5%25.58.26%18.726.9%12.22.96%18.329.6%
 Progression timeTprog13619.2%31.319.4%33.211.3%30.729.2%16.34.04%40.440.8%
Disease-modifying effect parameters
 Levodopa + selegilineFLXDT6.2216.1%5.5526.3%0.7116.73%0.35117.0%0.6584.88%0.31418.4%
Symptomatic effect parameters
Table 4.  Bootstrap average and RSEs for motor related Population Parameter Variability (PPV) and Residual Unidentified Variability (RUV) parameters
ParameterTotal UPDRSADLBradykinesiaPIGDRigidityTremor
  • *

    PPV was jointly estimated for levodopa, selegiline, bromocriptine and pergolide parameters of disease-modifying effects.

  • †PPV was jointly estimated for selegiline, bromocriptine and pergolide parameters of symptomatic effects. ADL, activities of daily living; PIGD, postural instability and gait disorder; RSE, relative standard errors; UPDRS, Unified Parkinson's Disease Rating Scale.

PPV of natural disease progression parameters
 Baseline statusPPV S00.4793.8%0.4952.69%0.6483.15%0.6463.41%0.7722.89%0.6483.46%
 Steady-state statusPPV Sss0.74315.0%0.69220.3%0.4066.56%0.85418.8%0.3659.07%0.50819.1%
 Progression timePPV TP0.97912.0%0.86917.1%0.7897.29%1.0012.4%0.8466.57%1.1121.4%
PPV of disease modifying effect parameters
 Levodopa + selegilinePPV FLXDT0.84313.6%0.87017.2%0.092562.0%0.71415.4%0.25322.6%0.47119.7%
 Other drugsPPV BOFF*0.95430.1%1.1717.2%0.437112%2.0544.2%1.5179.3%1.6530.8%
PPV of symptomatic effect parameters
 LevodopaPPV BEML0.654.6%0.7054.01%0.6298.26%1.0310.9%0.0085198.8%0.82719.3%
PPV ED500.39676.6%0.53668.1%0.51748.8%0.5952.4%1.1353.7%0.6948.0%
PPV TEML0.97732.4%1.1013.3%1.1432.8%0.73634.9%1.4928.4%0.87525.9%
PPV TEQL0.92220.4%0.69956.2%0.94213.0%1.1420.2%2.0616.5%0.96938.2%
PPV Hill0.931926.5%0.91827.0%0.74238.2%0.99546.2%0.0971208%1.0225.1%
 Other drugsPPV KPROG0.91012.8%1.0623.2%1.5324.6%1.1914.3%1.5913.4%1.2314.0%
Correlation parameters
 S0 with SssR120.12241.4%0.21457.1%0.38619.6%0.233101%0.25134.8%0.060258%
 S0 with TprogR13-0.014476%0.057341%0.030199%0.111229%-0.14653.7%-0.21557.0%
 Sss with TprogR230.8245.7%0.8209.48%0.58813.0%0.7976.28%0.64914.6%0.66128.4%
 Residual variabilityRUV SD5.521.9%2.111.59%2.312.31%0.8291.62%1.481.42%1.301.47%
Table 5.  Comparison of treatment effects at median daily doses
Treatment effectsTotal UPDRSTremorRigidityBradykinesiaPIGDADLMMSEHAM-D
  • *

    Percentage change relative to baseline S0: Levodopa 100/(1 + (C50L/1)^(–Hill))/S0; others; e.g. 100 ×BOD/S0.

  • Percentage change relative to parameter estimates in untreated state, e.g. (exp(KPLT) – 1) × 100. Tprog for UPDRS and ADL (positive percentages mean slowing of disease progress); Sss for tremor, rigidity, bradykinesia, PIGD and HAM-D (negative percentages mean beneficial disease-modifying effect); slope for MMSE (negative percentages mean slowing of disease progress).

  • ‡RSE of parameter determining treatment effect >40%. ADL, activities of daily living; HAM-D, Hamilton rating scale for depression; MMSE, mini-mental state exam; PIGD, postural instability and gait disorder; UPDRS, Unified Parkinson's Disease Rating Scale.

Symptomatic effects*        
Disease-modifying effects        
 Levodopa + selegiline2685%–88%–58%–53%–89%2428%706%–96%

Owing to the relatively slow rate of progression, tremor is predicted to reach the asymptotic disease state of 18.3 units after approximately 16 years. Although the mean natural progression curve of tremor appears to increase linearly during the 8 years of observations, the nonlinear model fit was superior to the linear model fit, with a decrease in objective function of 239 for one additional parameter (P < 0.0001; Table 2). Our models suggest that the natural progression of other subscales would reach their asymptotes (Sss) in 8–12 years. Most patients would have approached the asymptotic steady-state status for ADL, rigidity and bradykinesia at the end of 8 years, whereas PIGD scores would continue to rise at a slower rate after that time.

Baseline subtypes and disease progression  Using PIGD- relative to tremor-dominant baseline subtype as a covariate to Tprog significantly improved the fit for all motor features but did not affect the ADL subscale. The PIGD-dominant subtype group progressed at a faster rate compared with the tremor-dominant subtype group for tremor (8% faster; change OFV = 8), bradykinesia (20% faster; change OFV = 12), PIGD (4% faster; change OFV = 11) and rigidity subscales (13% faster; change OFV = 26) and total UPDRS (17% faster; change OFV = 11). Despite these statistically significant differences in progression rates, the use of PIGD- relative to tremor-dominant baseline subtype explained a negligible part of the variability for Tprog (relative reduction in PPV ranged from 0.7 to 10%).

As the PIGD subscale progressed faster (half-time 3 years) than did tremor (half-time 3.9 years), many patients with the tremor-dominant subtype at baseline switched to PIGD-dominant subtype during the follow-up period (Figure 5).

Symptomatic effects of anti-parkinsonian medications on motor function  The sigmoid Emax pharmacodynamic model best described the levodopa symptomatic effects. The mean daily dose that produced 50% of levodopa symptomatic effect (ED50) was similar for all features (7–24 mg day−1) except for PIGD (1237 mg day−1). For the purpose of comparison across features with different maximal scores, Table 5 shows treatment symptomatic effects as percentages of offset relative to S0 at the median daily dose of each drug.

Disease-modifying effects on motor function  Two models of disease-modifying effects were examined: the first tested the possibility that drug treatments reduced the steady-state disease status, Sss; the second tested the possibility that treatments slowed the time to attain steady state through an effect on the progression time constant, Tprog. Based on the objective function values (Table 2), models with disease-modifying effects on Sss best described the disease progress for tremor, rigidity, bradykinesia and PIGD; a model with disease-modifying effect on Tprog best described the progress of ADL. The disease-modifying effects are expressed as the percentage change in Sss or Tprog in the treated state relative to the untreated state (Table 5).

Nonmotor features

Natural disease progression of mood and cognitive function  The Gompertz model described the natural progression of HAM-D. A linear model adequately described the progression of MMSE. Bootstrap averages and standard errors of the nonmotor parameter estimates are shown in Tables 6 and 7. The HAM-D disease status had a score of 1.92 units at study entry and a steady-state score of 10.7 units. The linear model estimated the baseline MMSE score of 28.8 and a rate of decline of 0.0055 units (0.02% of baseline) per year.

Table 6.  Bootstrap average and RSEs for nonmotor structural model parameters
  • *

    Slope for MMSE. HAM-D, Hamilton rating scale for depression; MMSE, mini-mental state exam; RSE, relative standard errors.

Natural disease progression parameters     
 Baseline statusS01.923.50%28.80.096%
 Steady-state statusSss10.73.16%–0.00546*11.0%
 Progression timeTprog30.36.31%..
Disease-modifying effect parameters     
 Levodopa + selegilineFLXDT0.3553.50%1.230.450%
Symptomatic effect parameters     
Table 7.  Bootstrap average and RSEs for nonmotor PPV and RUV parameters
  • *

    PPV was jointly estimated for levodopa, selegiline, bromocriptine and pergolide parameters of disease-modifying effects.

  • †PPV was jointly estimated for selegiline, bromocriptine and pergolide parameters of symptomatic effects. HAM-D, Hamilton rating scale for depression; MMSE, mini-mental state exam; PPV, population parameter variability; RSE, relative standard errors; RUV, residual unidentified variability.

PPV of natural disease progression parameters     
 Baseline statusPPV S00.8133.35%0.03385.49%
 Steady-state statusPPV RSS0.5755.91%2.1716.8%
 Progression timePPV TPROG2.038.02%
PPV of disease-modifying effect parameters     
 Levodopa + selegilinePPV FLXDT0.7203.01%0.7544.35%
 Other drugsPPV BOFF*0.8916.83%1.831.91%
PPV of symptomatic effect parameters     
 LevodopaPPV BEML0.6256.69%1.671.83%
PPV ED501.071.95%1.123.72%
PPV TEML2.016.11%1.161.11%
PPV TEQL1.125.26%0.561.45%
PPV Hill2.183.90%2.921.97%
 Other drugsPPV KPROG1.116.36%1.574.76%
Correlation parameters     
 S0 with SssR12–0.060734.6%–0.9274.05%
 S0 with TprogR13–0.049639.9%
 Sss with TprogR230.31530.1%
 Residual variabilityRUV SD2.002.26%1.152.88%

Symptomatic effects and disease-modifying effects on mood and cognitive function  Levodopa had a similar potency for MMSE and HAM-D effects (ED50 of 23.5 and 20.4 mg day−1, respectively). Levodopa and selegiline produced a beneficial offset (symptomatic effect) on HAM-D, while bromocriptine and pergolide worsened HAM-D (Table 5). Levodopa worsened MMSE slightly, while selegiline produced a minor improvement.

The disease-modifying effects of treatments were best described by changes in Sss for HAM-D and linear slope, α, for MMSE. Levodopa and selegiline treatments produced large improvements (reductions) in Sss of HAM-D (Table 5). Steady-state disease status (Sss) was reduced to almost zero when both drugs were given together.

Unlike the motor features and depression, MMSE rating scores decrease with worsening disease status. The disease-modifying effects of treatments were described by changes in the linear slope of the rate of progression. Levodopa and selegiline both increased the rate of progression, whereas pergolide halved the rate and bromocriptine had almost no effect (Table 5). It should be noted that the baseline rate of progression was small, so that apparently large relative changes, expressed as percentage changes, produced by treatment are not large in absolute terms. When all disease-modifying effects on the slope were removed from the model for MMSE, the model fit worsened significantly (OFV increased by 29 units, P < 0.0001).

Visual predictive checks

The PRED-corrected VPCs for each of the eight disease status variables show the scatter of observed values and the fifth and 95th percentiles for the observations and predictions (Figures 2 and 3).

Figure 2.

Visual predictive check (VPC) for total UPDRS showing observations and observed percentiles (a), prediction-corrected (PRED-corrected) VPC without dropouts (b) and with dropouts (c), and simulated adaptive treatment with dropouts (uncorrected; d). Median (continuous lines) and 90% prediction intervals (dashed lines) of the observed (continuous lines with filled circles) and simulated data are shown. The 90% confidence intervals for the median and the prediction intervals are shown in the plot as shaded bands

Figure 3.

Figure 3.

Visual predictive checks for other disease status variables (lines and symbols similar to Figure 2)

Figure 3.

Figure 3.

Visual predictive checks for other disease status variables (lines and symbols similar to Figure 2)

Figure 3.

Figure 3.

Visual predictive checks for other disease status variables (lines and symbols similar to Figure 2)

The overall distribution of total UPDRS observations is shown in Figure 2a. The behaviour of the PRED-corrected VPC with and without simulated death or dropout illustrates why it is important to include a model for missing patient data (Figure 2b). The PRED-corrected VPC with simulated missing patients shows a good agreement with the observed time course of UPDRS (Figure 2c). The adaptive dosing algorithm with simulated missing patients (Figure 2d) was almost as good as the empirical PRED-corrected VPC, which confirmed the suitability of the algorithm for simulating future clinical trials.


This analysis is the first to describe with a quantitative model the time course of PD cardinal features (tremor, rigidity, bradykinesia and PIGD), disability and nonmotor features (depression and cognitive impairment) in a large cohort of PD subjects starting with early untreated PD and continuing into later, more severe stages. Unlike other analyses of PD progression [3, 4], we modelled the progression using longitudinal clinical assessments of these features (Figure 4) instead of the difference in scores between first and last visits. In addition, we have quantified the anti-parkinsonian treatment responsiveness for each feature in terms of symptomatic and disease-modifying effects.

Figure 4.

Predicted natural disease progress (continuous lines) and symptomatic and disease-modifying effects starting at year 2 of levodopa (dotted lines) and combination of levodopa and selegiline (dashed lines) on disease progress curves (y-axes are scaled to maximal possible score of the respective features)

UPDRS subscales

We used an empirical asymptotic function, the Gompertz model, to describe the progression of UPDRS subscales, in which the disease status progressed at different rates before it reached an asymptote. The asymptotic model was chosen in part because the clinical measurements are based on a scoring system (UPDRS) that has a maximal value, but also because of a clinical impression that the motor features eventually approach a stable state. A nonlinear rate of progression of bradykinesia has been reported for patients with idiopathic PD [16]. Nonlinear rate of progression has also been described by a prospective longitudinal study using serial [18F]-fluorodopa positron emission tomography (PET) as a marker for neurodegeneration in 31 PD patients [17]. Louis et al. [3] noted evidence of a nonlinear rate of progression, by observing that patients with shorter disease duration progressed faster than those with longer disease duration. Consistent with these studies, our model-based analysis suggested that a nonlinear function best described the progression of UPDRS and its subscales.

We found that the progression half-times of rigidity (2.2 years), bradykinesia (2.2 years) and PIGD (3.1 years) are comparable to total UPDRS (2.1 years) [6], whereas that of tremor is slower (3.9 years). Tremor does indeed progress, despite the common clinical impression that tremor does not progress after the initial stages of the disease. This difference in rate of progression suggests that the pathophysiology for tremor may be different from the other motor features, an idea supported by the lack of correlation between fluorodopa measures of dopamine terminal loss in PD and tremor, unlike the good correlations for rigidity and bradykinesia with PET markers of dopamine terminal loss [17–19]. Although the methodology is different, our findings are consistent with Louis et al., who used generalized linear models to describe 237 PD patients followed annually for a mean of 3 years [3]. They found that tremor did not appear to progress, but the other three motor signs progressed at the same rate of 2–3% per year. As all patients received anti-parkinsonian treatments in that observational study, they were not able to describe the progression of untreated PD as we have attempted to do. In another observational study in advanced PD patients, the linear rate of progression of total UPDRS (1.43% per year) was faster than that of ADL derived from UPDRS part II (0.56% per year) in the ‘on’ state [4]. Likewise, this study was unable to correct for anti-parkinsonian treatment effects. In contrast, we did not find the rates of progression of total UPDRS and ADL to be appreciably different.

According to Jankovic and Kapadia, patients who were PIGD dominant at baseline progressed more quickly (77% greater annual slope of total UPDRS) than patients who were tremor dominant at baseline. That study [4] suggested that the baseline dominant subtype may explain some of the differences in rates of progression of PD. We also found that the half-time of UPDRS progression differed significantly between the two subtypes, but found that Tprog for total UPDRS was only 18% faster in the PIGD-dominant group. As the PIGD subscale progressed faster than the tremor subscale, many patients with the tremor-dominant subtype at baseline converted to PIGD-dominant subtype during the trial (Figure 5). This is consistent with the finding of Alves et al. that a large proportion of patients progressed from tremor-dominant to PIGD-dominant subtype over 8 years of follow-up [20]. These observations that many tremor-dominant patients convert to PIGD dominant (and not vice versa) suggest that these subtypes are not distinct biological entities but reflect different stages of the disease.

Figure 5.

Median ratio of tremor/PIGD over 8 years, illustrating the transition from tremor-dominant to PIGD-dominant subtypes in the study cohort

Response of motor features to anti-parkinsonian treatments

The motor subscales responded to anti-parkinsonian treatments to different degrees in terms of both symptomatic and disease-modifying effects. Similar to previous results for total UPDRS [6], we observed a slow onset of response over years due to both a time-dependent change in Emax and a slow build up of drug in the hypothetical effect compartment. The predicted time to approach 90% of Emax for levodopa symptomatic response was relatively faster for the PIGD subscale (2 years) than for the other subscales (8–12 years). The slow onset of levodopa response has been shown in a previous pharmacokinetic–pharmacodynamic study using a different end-point of response (finger tapping) [21].

The maximal (Emax) motor symptomatic benefit of 300 mg day−1 levodopa (median dose in this cohort) was smallest for tremor (6% decrease relative to baseline disease status). Sensitivity to the symptomatic effect of levodopa was almost 85 times less for PIGD than for other motor features. Nevertheless, symptomatic responsiveness of PIGD to levodopa was greater when compared with other subscales when expressed relative to the baseline status, but this is in part a reflection of the small numerical value of the baseline (1.23 units).

We found that symptomatic benefits of the other anti-parkinsonian medications were small relative to levodopa, especially for total UPDRS. Selegiline had small symptomatic effects by itself, but these were consistent across all motor features. This is compatible with other reports of monoamine oxidase type B inhibitors in PD subjects not receiving other anti-parkinsonian medications [22, 23]. The disease-modifying effects of treatments were mainly evident as changes in steady-state disease status but, for total UPDRS and ADL, changes in the progression half-time provided a better description. The beneficial disease-modifying effects of levodopa and selegiline were consistently seen for all features. The combination of selegiline and levodopa had a greater disease-modifying effect than expected from their individual disease-modifying effects; that is, there was a beneficial interaction between the two drugs.

Nonmotor features

A nonlinear model best described the progression of depression as measured by HAM-D. The progression half-time to steady-state asymptote was fairly slow for HAM-D (4.8 years). Given that only 4% of HAM-D observations were greater than 10 units and the steady-state asymptote was 11 units, most patients in this cohort did not develop depression at the end of 8 years. Symptomatic treatment effects had little influence on the HAM-D score, although this was the greatest percentage influence of selegiline, which may be attributable to its monoamine oxidase B inhibitory effect. However, disease-modifying effects of levodopa and selegiline attributable to lowering the asymptotic steady-state HAM-D score were clearly evident. The results suggested that anti-parkinsonian treatments may reduce the prevalence or severity of depression as reflected in HAM-D. As the DATATOP trial excluded patients at baseline with MMSE < 23 and with other major coexisting medical or neurological diseases, the range of MMSE score was limited (85% of observations were between 28 and 30). The progression of MMSE was linear, unlike other status measures. The rate of MMSE decline was very slow and gave little support for the idea that PD is associated with an acceleration of a decline in cognitive function, consistent with another report of this cohort that found low rates of cognitive impairment [24]. Treatments appeared to accelerate the rate of decline of MMSE. However, the natural history of MMSE progression is so slight that these effects would not be considered clinically significant.

Model evaluation

We have shown that failure to account properly for dropout (Figure 3), especially after the first year of follow-up, will over-predict the observed percentiles of the VPC. Some motor function status variables are clearly well correlated with the predictive ability of total UPDRS to describe dropout hazard (ADL and tremor) and have acceptable predictions in the VPC. Other motor variables (bradykinesia, rigidity and PIGD) are not predicted well in later years, which suggests that they are not well correlated with total UPDRS and patient dropout is not simulated correctly. The nonmotor status variables (HAM-D and MMSE) show little variation with time, with good prediction of the percentiles.


The robustness and limitations of a model-based approach to study disease progress have been discussed in publications describing the total UPDRS score [6, 25]. The modelling method makes several assumptions. Firstly, we assumed empirical nonlinear or linear functions for the natural history time course. Secondly, we assumed an empirical delay model to describe the symptomatic effects of levodopa, which involved both an increase in efficacy (Emax) with time as well as a slow equilibration of the effect with each change in dose rate. Finally, we assumed that empirical functions describe the effect of treatments on the disease progression model parameters. Overall, the model, with these assumptions, produces a good description of the time course of PD features and the response to treatment. The same assumptions were made for the model describing total UPDRS and have been successfully validated for short-term predictions by comparison with the results of the ELLDOPA study [7]. In the spirit of Sheiner's definitions, this work is more in the learning than confirming category [26]. Examining the disease progression and response to therapy in other long-term clinical trial databases will be a further test of the model and the assumptions.

Some parameters are clearly better described than others, so that quantitative conclusions about each individual parameter will vary a lot. We do not make any strong quantitative statements about any particular parameter. The numerical values should be regarded as indicative of the magnitude of differences in rate of progression and drug effects for comparative purposes.

Owing to the selected characteristics of the original DATATOP cohort, there are limitations in extrapolating our results to a more general PD population. As our models showed, cognitive function and mood did not worsen to any clinically important degree over the study duration. The lack of major change over such a long period of follow-up has not been previously reported. Our results are based on the DATATOP cohort that was initially selected because they were at an early stage of the disease. This may not be the case in an unselected population and more advanced PD patients.

Despite these limitations, we have a very high degree of confidence in the qualitative conclusions about the DATATOP cohort based on the full model structure and the set of parameters taken together. This confidence comes from individual subject predictions, internal evaluation using VPCs and a successful external evaluation of the UPDRS model predictions [7].


With a model-based approach, we have described the time course of five Parkinson's disease features (tremor, rigidity, bradykinesia, PIGD and ADL) and two nonmotor features (cognitive impairment and depression) in the DATATOP cohort of PD subjects. The progression of tremor is clearly slower than other disease status subscales, which suggests a distinct pathophysiological process for this feature. We have also shown differential sensitivity of PIGD compared with the other UPDRS subscales to anti-parkinsonian treatments. Of greater long-term importance, we have extended our previous observation that treatments, especially the combination of levodopa and selegiline, slow the functional progression of all cardinal features of PD.

Competing Interests

N. Holford has received consulting fees from Novartis. Novartis produces anti-Parkinsonian medication. There are no other competing interests to declare.


This work was supported by a grant from the Michael J. Fox Foundation. We are grateful to the Parkinson Study Group for providing access to the DATATOP cohort. We wish to acknowledge the continuing support and encouragement of the Parkinson Study Group and Dr Arthur Watts, University of Rochester, for help in providing the data for the DATATOP cohort.