Parkinson’s disease (PD) is the second most common neurodegenerative disorder (Nussbaum and Ellis, 2003). It is thought to affect roughly 1% of the population over 65 years old and 5% of the population over 85 years old (Wood-Kaczmar et al., 2006).

A reliable diagnostic test for PD is not yet available. The diagnosis of PD is often based on clinical symptoms. Current criteria for PD diagnosis include the presence of bradykinesia as well as resting tremor and/or rigidity (De Lau and Breteler, 2006). However, the gold standard for diagnosis is still pathological confirmation post-mortem (Belle et al., 2017). When diagnosis is investigated at autopsy, it is thought between 10 and 20% of previously clinically diagnosed cases were misdiagnosed (De Lau and Breteler, 2006). Even though clinical diagnosis of PD is based on motor symptoms, many individuals with PD also suffer from non-motor symptoms including dribbling, constipation and anxiety (Poewe, 2008).

Loss of dopaminergic neurons in the substantia nigra is considered the main cause of PD (Reeve et al., 2014), which is associated with loss of dopaminergic terminals in the forebrain, reduced dopamine release and thus, the onset of motor symptoms. The current gold standard of treatment for PD is Levodopa, which is converted to dopamine within the brain, which restores (at least partially) the levels of dopamine in the striatum (LeWitt, 2008). However, this effect is only transient, as it does not stop the progression of disease, as dopaminergic neurons continue to die. Neurodegeneration of dopaminergic neurons has been associated with the presence of aggregated alpha-synuclein in the form of Lewy bodies (Ruipérez et al., 2010).

Prasinezumab is a humanized immunoglobulin G1 monoclonal antibody designed to selectively bind aggregated alpha-synuclein at the C-terminus. PASADENA is a Phase II, randomized, double blind, placebo controlled trial that investigates the efficacy and safety of prasinezumab in early PD (Pagano et al., 2021). Here, the placebo arm from PASADENA was used to investigate the progression of the disease in individuals with early PD and to identify prognostic factors that can predict PD progression.

We aim to produce a model, to identify baseline characteristics, which can predict disease progression in individuals with PD. These predictors might also aid the identification of PD patients who are at risk of degenerating more rapidly. These patients are the ideal population to enroll into clinical trials to enable an earlier and faster detection of a treatment effect, if one exists, as, these patients may be more likely to respond to treatments aimed at slowing disease progression.

Demographic and baseline characteristics of the 76 PD participants in the PASADENA trial (training dataset) and the 139 clinically matched PD participants in the PPMI cohort (test dataset) are presented in Table 1. The t-test was used to investigate the difference in means of the continuous variables and the z-test was used to investigate the difference in proportions for the two binary variables. The p-values of these tests are shown in Table 1. The two datasets had a similar: mean age, proportion of men and mean MDS-UPDRS Part I, II, III and total scores. They did, however, differ in the time from PD diagnosis (p = 0.0063) and the Hoehn and Yahr stage (p = 0.0002). PASADENA had a larger mean time since PD diagnosis (9.55 months), compared to PPMI (6.97 months) and it had a larger proportion of Hoehn and Yahr stage II (77.63%), than the PPMI database (50.36%). This could affect the accuracy of the Hoehn and Yahr Stage as a predictor of progression.

The clinically selected model 1, which includes variables considered by PD experts to be the best predictors, is displayed in Table 2. As expected, age had a positive coefficient, with older patients declining faster than younger patients. In addition, both Dat-SPECT variables had negative coefficients, with lower Dat-SPECT values (i.e., more neurodegeneration) associated with fast progression and hence they are more likely to progress. The predictive values of each predictor are displayed in Table 2, which indicates that Ipsilateral putamen Dat-SPECT SBR was the most important variable in predicting the progression of PD patients.

Multiple clinical models were considered, focusing on the variables which were ranked most important and MDS-UPDRS part III as a covariate. The model which performed best, clinically selected model 2, is shown in Table 3. Out of the models explored it produced the largest NPV, PPV and TPV.

The first data driven model is displayed in Table 4. As expected, Hoehn and Yahr had a positive coefficient, with higher stages associated with fast progression. Again, Dat-SPECT SBRs in the ipsilateral putamen had a negative coefficient. MDS-UPDRS part II had a positive coefficient, with higher scores associated with fast progression.

The predictive accuracy of the variables within the data-driven model were further compared using the permutation importance of each variable. Their variable importance is also displayed in Table 4, which shows the two most important variables were MDS-UPDRS part II and Ipsilateral putamen Dat-SPECT. Several models were investigated, focusing on the variables which were ranked most important including MDS-UPDRS part III. The model which performed best in terms of NPV, PPV and TPV, the second data driven model, is shown in Table 5.

Table 6 displays the predictive accuracy of the two clinically selected models and the two data driven models, in predicting the progression of the 139 PD participants in the PPMI dataset. Table 6 shows the second data driven model had a high predictive accuracy, producing the joint largest NPV = 0.45 and the largest PPV = 0.75. However, the second clinically selected model also had a high predictive accuracy. It produced the joint largest NPV = 0.45 and largest TPV = 0.62 and sensitivity = 0.71. The first data driven model produced the largest specificity = 0.67. Both the clinically selected models had very similar small Brier scores, 0.2395 vs. 0.2363. Figure 3 displays the ROC curve and AUC for each model, where the two data driven models gave the joint largest AUC.

Our finding suggests that Hoehn and Yahr stage and Ipsilateral putamen Dat-SPECT predict progression of motor signs in early PD. These results were derived by using four models to predict PD progression at 12 months evaluating a sample of 76 individuals with early PD (training set from the PASADENA placebo arm) and validating them in a sample of 139 clinically matched PD individuals (validation set from the PPMI database).

Data driven models gave a larger specificity than the clinically selected models, due to correctly predicting a larger proportion of the actual non-progressors, 0.67 and 0.63 vs. 0.40 and 0.44. However, the data driven models generated a lower sensitivity, as they also incorrectly predicted large quantities of progressors, 0.51 and 0.59 vs. 0.69 and 0.71. All 4 models produced quite similar non-progressor predictive values (predicted non-progressors who will actually not progress), 40%, 45, 42, and 45%. Furthermore, the data driven models had larger progressor predictive values (predicted progressors who will actually progress), 74 and 75% vs. 68 and 71%. Additionally, the two clinically selected models generated better (lower) Brier scores than the data driven models: 0.2395 and 0.2363 vs. 0.2752 and 0.2497. Lastly, the two data driven models produced larger AUC than the two clinically selected models: 0.62 and 0.62 vs. 0.58 and 0.59.

The predictive strength of each model depends on the relative power of each predictor variable. However, the second clinically selected model and both data driven models performed well for different measures of predictive accuracy. The second clinically selected model yielded the joint largest NPV, largest sensitivity, largest TPV and the lowest Brier score. However, the first data driven model gave the largest specificity and joint highest AUC and the second data driven model produced the joint best NPV, the best PPV and joint highest AUC. To decide which model performed best, we must choose which prediction measure we believe to be the most important. In this situation we found the mean of the PPV, NPV, and TPV for each model, which produced 0.56, 0.59, 0.57, and 0.6, respectively. The second data driven model appears to be the most well rounded model taking into account these three predictive values, and indeed when sensitivity, specificity and AUC are also included, the second data driven model still produced the largest mean of these 6 performance measures. Therefore, we conclude the second data driven model, which included the variables: Hoehn and Yahr stage, Ipsilateral putamen Dat-SPECT and MDS-UPDRS Part III as a covariate, is the optimal prediction model.

We note that there may be other baseline variables which have an effect on the progression of PD patients, which have not been explored. Another limitation of the models, is the small PASADENA dataset used to train said models and the differences between the training and test datasets. The two datasets had differing mean time from diagnosis and proportions of PD patients in Hoehn and Yahr Stage I and II. Furthermore, the follow-up time did not match between the two datasets. In PASADENA, progression was defined 52 weeks after baseline, whereas in PPMI, progression was defined at a patient’s visit closest to 12 months after baseline, where the visit could have been anywhere between 10 and 14 months after baseline. These differences could also affect the predictive accuracy of our models and levels of validation.

Finally, there were missing values in both the training and test datasets. The one missing value in the training dataset will not have a large effect on the accuracy of the prediction models. However, the 36 patients of the 139 in the PPMI dataset (25.9%) who did not have an Ipsilateral putamen or Contralateral putamen Dat-SPECT value could have affected the accuracy of our prediction models if the missing values had not been imputed correctly.

Further studies are needed to validate these results in an independent dataset and evaluate other baseline characteristics not investigated here. In addition, these prediction models could be further explored by investigating different cut off values for determining the probability of becoming a progressor or not. Here, a cut-off of 0.5 was used, thus, any patient with a predicted probability of progressing above 0.5, was predicted to progress. However, other cut-off values could be considered toward improving the above model’s predictive accuracy.

In conclusion, four models predicting motor sign progression in early PD patients are presented. They were trained on a sample of 76 PD patients who had not started symptomatic dopaminergic therapy assigned to the placebo arm of the phase II PASADENA study. These models were validated in 139 demographically and clinically matched PD patients from the PPMI database. Baseline Hoehn and Yahr stage and Ipsilateral putamen Dat-SPECT SBR best predicted motor sign progression in early PD. Further studies are needed to confirm the validity of these predictors to identify individuals with early PD with faster progression phenotype.

Qualified researchers may request access to individual patient-level data through the clinical study data request platform (https://vivli.org/). Further details on Roche’s criteria for eligible studies are available at https://vivli.org/members/ourmembers/. For further details on Roche’s Global Policy on the Sharing of Clinical Information and how to request access to related clinical study documents, see https://www.roche.com/research_and_development/who_we_are_how_we_work/clinical_trials/our_commitment_to_data_sharing.htm. Data used in the preparation of this article were obtained from the Parkinson’s Progression Markers Initiative (PPMI) database (www.ppmi-info.org/access-data-specimens/download-data). For up-to-date information on the study, visit ppmi-info.org. Further inquiries can be directed to the corresponding author/s.

Participants were identified for potential recruitment using site-specific recruitment plans prior to consenting to take part in this study. Recruitment materials for participants had received Institutional Review Board or Ethics Committee approval prior to use. The following Institutional Review Boards ruled on ethics of the PASADENA study: Ethikkommission der Medizinischen Universität Innnsbruck, Innsbruck, Austria; Comité de Protection des Personnes (CPP) Ouest IV, Nantes, France; Ethikkommission der Universität Leipzig and Geschäftsstelle der Ethikkommission an der medizinischen Fakultät der Universität Leipzig, Leipzig, Germany; Ethikkommission der Fakultät für Medizin der Technischen Universität München, München, Germany; Ethikkommission der Universität Ulm (Oberer Eselsberg), Ulm, Germany; Landesamt für Gesundheit und Soziales Berlin and Geschäftsstelle der Ethik-Kommission des Landes Berlin, Berlin, Germany; Ethikkommission des FB Medizin der Philipps-Universität Marburg, Marburg, Germany; Ethikkommission an der Medizinischen Fakultät der Eberhard-Karls-Universität und am Universitätsklinikum Tübingen, Tübingen, Germany; Ethikkommission an der Med. Fakultät der HHU Düsseldorf, Düsseldorf, Germany; Ethikkommission der LÄK Hessen, Frankfurt, Germany; CEIm Hospital Universitari Vall d’Hebron, Barcelona, Spain; Copernicus Group Independent Review Board, Puyallup, WA, United States; Western Institutional Review Board, Puyallup, WA, United States; The University of Kansas Medical Center Human Research Protection Program, Kansas City, KS, United States; Oregon Health & Science University Independent Review Board, Portland, OR, United States; Northwestern University Institutional Review Board, Chicago, IL, United States; Spectrum Health Human Research Protection Program, Grand Rapids, MI, United States; The University of Vermont Committees on Human Subjects, Burlington, VT, United States; Beth Israel Deaconess Medical Center Committee on Clinical Investigations, New Procedures and New Forms of Therapy, Boston, MA, United States; Vanderbilt Human Research Protection Program Health, Boston, MA, United States; Vanderbilt Human Research Protection Program Health, Nashville, TN, United States; University of Maryland, Baltimore Institutional Review Board, Baltimore, MD, United States; University of Southern California Institutional Review Board, Los Angeles, CA, United States; Columbia University Medical Center Institutional Review Board, New York, NY, United States; University of Southern California San Francisco Institutional Review Board, San Francisco, CA, United States; University of Pennsylvania Institutional Review Board, Philadelphia, PA, United States; and HCA – HealthOne Institutional Review Board, Denver, CO, United States. All Institutional Review Boards gave ethical approval of the study.

GP, KT, and JA-C designed the study. HJ and JA-C were involved in data collection and analyzing the data. All authors contributed to data interpretation, revised and gave input on the article, and approved the final manuscript.

HJ was a paid intern of F. Hoffmann-La Roche Ltd. JA-C, KT, and GP were full-time employees and shareholders of F. Hoffmann-La Roche Ltd.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.