Data used in the preparation of this article were obtained from the Parkinson’s Progression Markers Initiative (PPMI) database (Marek et al., 2018; www.ppmi-info.org/data). For up-to-date information on the study, visit www.ppmi-info.org. PPMI was conducted in accordance with the Declaration of Helsinki and the Good Clinical Practice (GCP) guidelines after approval of the local ethics committees of the participating sites. Informed consent was obtained from each subject.

At enrolment, PD patients had received a diagnosis no longer than 2years before and were not taking PD medication. DAT status was performed with Iodine-123-labeled ioflupane single-photon emission computerized tomography imaging for the DAT (Darcourt et al., 2010; DaTSCAN) at the first screening visit. Patients below age 50 were also excluded to avoid cases of early onset PD.

An age-, sex-, and education- matched healthy control (HC) group was free of neurological disease, no first-degree family member with PD and used no medication that might interfere with DAT imaging. For both patients and healthy controls, we included only participants with at least two scheduled visits. Data from unscheduled follow-ups outside the original yearly follow-ups were not included in this study. Following these criteria, the final sample size included in the present study were 349 PD patients with an average disease duration of 6.72months (SD=6.63) at baseline and 145 HCs. The data were downloaded from the PPMI database on 28 September 2020. At this point, individuals had undergone up to 5 annual follow-up visits.

At baseline and at each annual follow-up visit, a comprehensive neuropsychological battery was administered to all subjects by trained personnel as part of the PPMI testing procedure.

For the current study, the core five cognitive tests of the neuropsychological assessment in PPM were considered: (1) the WMS-III Letter-Number-Sequencing Test (Wechsler, 1997), (2) a Symbol Digit Modalities Test (Smith, 1973), (3) the Benton Judgement of Line Orientation Test (Benton et al., 1978), (4) a semantic fluency test (Gladsjo et al., 1999; animal naming), and (5) the Hopkins Verbal Learning Test-Revised (Benedict et al., 2010; alternate forms were used at follow-ups). We obtained the age-corrected, and where available, education-corrected, and normed cognitive scores (T and scale points) from the database and then transformed them into z-scores for each test. Further, where appropriate, cognitive domain composite scores were created based on the average of z-scores: An executive function (EF) composite included the average of z-scores from symbol digit modalities test, letter-number sequencing and semantic fluency, and memory composite included z-scores from immediate recall and delayed recall in the learning test. For the visuospatial domain, only the line orientation test was available. Patients were asked not to take their PD medication on the day of each annual study visit.

To identify early markers of different cognitive trajectories, a number of demographic, clinical variables and biomarkers (section “Genetic and CSF Biomarkers”) were used to compare patient sub-groups at baseline (section “Post-hoc Explorations”). Demographics included age, sex, and education in years. To evaluate motor impairment in the patients, The Movement Disorders Society Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) part III was used (Goetz et al., 2008). A mean tremor score and a mean postural instability and gait difficulties (PIGD) score were calculated from the respective items. Scores for a tremor dominant subtype (TD) and an akinetic-rigid subtype were then calculated according to the criteria: If mean tremor score/PIGD score≥1.15, OR if PIGD score=0 and Tremor score>0, then subject is TD. If ratio≤0.9, then subject is PIGD (Stebbins et al., 2013). As an indicator of disease stage/severity, the Hoehn and Yahr stage was included (Hoehn and Yahr, 1998). Part II total score of the MDS-UPDRS was used as an additional indicator of motor problems in daily living. Disease duration was calculated as the number of months between diagnosis and enrolment into the study.

Movement Disorders Society Unified Parkinson’s Disease Rating Scale score from part I was used as a general indicator of non-motor aspects of daily living associated with PD. In addition, cognitive impairment was assessed with the brief cognitive screening instrument montreal cognitive assessment (MOCA) (Nasreddine et al., 2005, adjusted for education); sleep disorder was assessed by the rapid eye movement (REM) Sleep Behavior Disorder Questionnaire (Stiasny-Kolster et al., 2007). Olfactory dysfunction with the University of Pennsylvania Smell Identification Test identified hyposmia as a score below the 15th percentile for the patient’s age and sex group (Jennings et al., 2014). Psychiatric features associated with PD were assessed with the short version of the Questionnaire for Impulse-Compulsive Disorder (QUIP; Weintraub et al., 2013), the 15-item Geriatric Depression Scale (GDS-15; Sheikh and Yesavage, 1986) and the 40-item psychiatric State-Trait Anxiety Inventory (STAI). The QUIP was used to indicate the total number of compulsive behaviors (max 7). The 40 items from the STAI were summed into a mean state and a mean trait anxiety score with a range of 20–80 and a higher score indicating greater anxiety (Spielberger et al., 1970).

For the current study, established cut-offs were used to separate normal from abnormal non-motor functions where available. A cut-off of 26 was used for the MOCA to indicate cognitive impairment vs. no cognitive impairment according to Nasreddine et al. (2005). The REM Sleep Behavior Disorder Questionnaire Score was dichotomized into REM sleep disorder (yes-no) using a cut-off of 5 (e.g., Nomura et al., 2011). For the GDS, prior research in PD patients has indicated that a cut-off score of <5 accurately distinguishes depressed from nondepressed patients in PD (Weintraub et al., 2006). Missing data for all questionnaires and clinical assessments were <1%.

APOE4 genotype and levels of CSF biomarkers were obtained as processed values from the PPMI database. CSF data were available for baseline and follow-up timepoints 1, 2, and 3 of the current study. Briefly, APOE genotyping followed the methods by Livak (1999). Taqman Assays were used per manufacturers protocol to genotype two non-synonymous single nucleotide polymorphisms (SNPs), rs429358 (APOE-C112R) and rs7412 (APOE-R158C), in each patient sample in order to distinguish between APOE ε2, ε3, and ε4 alleles. Patients were grouped into those having at least one ε4 allele vs. the rest. Genotype was missing for 37 patients (10.6%).

The measurements of α-synuclein, amyloidß42, total (t)-tau, and phosphorylated tau 181 (p-tau) levels in CSF are described in detail in Kang et al. (2016). CSF was collected by standard lumbar puncture. α-synuclein, amyloidß42, t-tau, and p-tau were measured by INNO-BIA AlzBio3 immunoassay (Innogenetics Inc.). α-synuclein was measured by enzyme-linked immunosorbent assay. Missing CSF data was 1.7% at baseline, 15.92% at follow-up 1, 17.57% at follow-up 2, and 34.84% at follow-up 3.

In order to investigate the association between cognitive decline and regional neuropathology, i.e., dopaminergic dysfunction, longitudinal DaTSCAN-data were included in the current study. DAT-data were quantified as Striatal Binding Ratios (SBR), a measure of DAT density computed as (target region/reference region)-1, with the target region being the DAT-rich striatum (consisting of caudate and putamen), and a reference region with negligible specific DAT binding, the occipital lobe. For the purposes of this study, the SBR values for right and left caudate and right and left putamen were downloaded from the database and averaged across hemispheres. In order to match the DaTSCAN data as closely as possible with the neuropsychological testing, DaTSCAN data were obtained from the screening visit (mean difference in months to baseline cognitive assessment=0.82, SD=1.06; available for 338 patients), and from the scans that were conducted in conjunction with annual follow-up visits 1 (mean diff=0.61months, SD=0.49; N=298), 2 (mean diff=0.62months, SD=0.58; N=294), and 4 (mean diff=0.58months, SD=0.62; N=249).

Fitting of a MLCLMM to the cognitive data suggested that longitudinal cognitive decline in the patient sample was best described by two latent groups (as indicated by the lowest BIC statistic (Table 2)). The two-class solution split the sample into one larger subgroup with 90% of the PD participants (Class 1. N=315) and one smaller subgroup, with 10% of the participants (Class 2. N=34, entropy=0.90). The worse-fitting three-and four-class solutions further divided the larger group but retained the smaller subgroup.

Plotting of the mean performance scores for each cognitive domain, timepoint, and both classes in Figures 1A–C shows that the smaller patient class (class 2, red) was characterized by performance deficits within 1SD of the larger class (class 1, blue) at baseline but then showed rapid linear decline over time.

Fitting linear mixed effects models with fixed effects of time, class, and time*class in the patients confirmed significant effects of class (i.e., a significant difference between patient classes at baseline) and a highly significant time*class interaction for each cognitive domain (i.e., a significant difference between classes in slopes, Table 3). HC were added (in grey) as a single control group after confirming with a separate multivariate latent class mixed model that cognitive decline across domains was best described as one group for HC (all multi-group solutions had a BIC>4777.33 (1-group), loglik=−2296.67, and 37 parameters).

Post-hoc comparisons between the large PD class (class 1) and HC showed that, on average, patients performed worse than HC in EF and memory but did not differ in performance on the visuospatial task (Figure 1; Table 3). Compared to the large performance decrements of patients in class 2, the performance deficit in class 1 patients was small, with an estimated difference of less than half a SD at baseline. Interactions between class and time were not significant (all ps>0.05). In summary, these analyses suggest longitudinal trajectories of cognitive change in PD are best described by two distinct latent classes: (1) A large group of patients that performed worse (compared to matched HC) on tests of EF and episodic memory but remained stable over 5years, suggesting subtle cognitive impairment is common in PD. (2) A small group of PD patients with rapid and pronounced loss of global cognitive functions across time.

For comparison, class differences in motor function (UPDRS score part III) were estimated over time (Figure 1D; Table 3). Here, both groups showed large differences to HC at baseline and over time. The rapid decliners also showed pronounced loss of motor functions over time, however they did not differ significantly from class 1 patients at baseline (p=0.08, cf. section Class Comparison on Baseline Demographics, Clinical Variables, and CSF Biomarkers).

At the first follow-up, 63.03% of patients in the large class (class 1) were on anti-PD medication, compared with 56.67% among the rapid decliners [χ² (df=1, N=349)=0.23, p=0.63]. At the second follow-up, the number of patients on medication was over 87% in both classes and over 90% at subsequent visits, suggesting differences in cognitive decline were not due to differences in medication.

Table 4 shows the number of follow-up visits for each class. The majority of patients in both classes participated in all five follow-up visits and there was no significant difference in the proportion of maximum follow-up visits (i.e., 5) between classes [χ² (df=1, N=349)=2.82, p=0.09], suggesting that study drop-out had no impact on the results.

In order to identify clinical predictors of rapid cognitive decline at baseline, post-hoc comparisons between the patient classes revealed that the rapid decliners were more likely than the rest of the patients to score below 26 on the MOCA, be mildly depressed, suffer from a REM sleep disorder, and have lower levels of CSF amyloidß42 level (Table 1) though these differences should be interpreted with caution as only the differences in amyloidß42 level reached significance after control for comparisons of 22 different variables.

In terms of the potential of these markers as stand-alone clinical screening tools at baseline, it should be noted that the majority of patients in both classes had a normal MOCA score of 26 or higher at baseline, suggesting that brief clinical screeners at baseline do not accurately identify rapid decliners. Prevalence of depression did not differ between that large patient class 1 and HC [χ² (df=1, N=349)=0.92, p=0.34], which indicates that a higher prevalence of depression may be specific to the rapid decliners and not PD in general. Nevertheless, only 10 out of 34 patients were considered mildly depressed, and no patient among the rapid decliners was judged to be moderately or severely depressed.

Amyloidß42 level was significantly lower in the rapid decliner at baseline whereas amyloidß42 level did not differ between the large patient group and HC [t(225.50)=−1.50, p=0.14].

Interestingly, neither demographics, such as age, gender, and education, nor motor symptoms, disease severity or disease duration were strongly predictive of rapid cognitive decline {all ps<0.05, though we note trend level differences in age [t(40.42)=−1.98, p=0.054] and UPDRS part III motor score [t(39.17)=1.99, p=0.054] between the patient classes}. Because of the large number of independent comparisons, trends of p>0.06 are not considered further.

Differences in regional neuropathology between the large PD class and the rapid decliners, estimated as DAT density in caudate and putamen, were compared between groups. DAT density estimates from SPECT scans obtained at the screening visit were available for 305 patients from the larger class and 33 patients from the class of rapid decliners and 293 and 28 patients, respectively, had at least one follow-up scan. Linear mixed effects models with DAT density as the outcome and class and time as predictors were fitted separately for caudate and putamen. For caudate DAT density, the results revealed a significant effect of time, class, and a trend for a significant time*class interaction, suggesting slightly accelerated loss of caudate DAT density for the rapid decliners. For putamen DAT density, a significant effect of time, class but no significant time*class interaction was found (Table 5). Figure 2 illustrates the main effects and interactions for both models. A linear model assessing the interaction between region (caudate/putamen) and class for baseline DAT level confirmed that a difference in DAT density between classes at baseline was pronounced for caudate relative to putamen (interaction Estimate=−0.22, SE=0.11, and p=0.05). Together, this suggests a highly significant DAT deficit for the rapid decliners that was marginally more pronounced for caudate both at baseline and in terms of loss over time. For comparison, the reduction of CSF biomarker levels over time did not differ between classes for myloidaß42 (Table 5), t-tau, or p-tau (ps>0.50). Of note, there was a significant interaction between class and time for α-synuclein (Estimate for time*class=67.36, SE=33.34, and p=0.04), but in the direction that α-synuclein increased at follow-up timepoints to the level of the comparison class (i.e., regression to the mean), suggesting that the marginal difference at baseline may be a false positive. Thus, unlike DAT, CSF biomarker change did not resemble the pattern of cognitive decline.

Within the group of rapid decliners, change in caudate DAT was associated with change in visuospatial ability (estimated with individually fitted slopes for each patient and each outcome, r=0.42, p=0.03, and N=28), but not with change in executive functions or memory (p>0.05).