Identification of Parkinson's disease subtypes with distinct brain atrophy progression and its association with clinical progression

Abstract Background Parkinson's disease (PD) patients suffer from progressive gray matter volume (GMV) loss, but whether distinct patterns of atrophy progression exist within PD are still unclear. Objective This study aims to identify PD subtypes with different rates of GMV loss and assess their association with clinical progression. Methods This study included 107 PD patients (mean age: 60.06 ± 9.98 years, 70.09% male) with baseline and ≥ 3-year follow-up structural MRI scans. A linear mixed-effects model was employed to assess the rates of regional GMV loss. Hierarchical cluster analysis was conducted to explore potential subtypes based on individual rates of GMV loss. Clinical score changes were then compared across these subtypes. Results Two PD subtypes were identified based on brain atrophy rates. Subtype 1 (n = 63) showed moderate atrophy, notably in the prefrontal and lateral temporal lobes, while Subtype 2 (n = 44) had faster atrophy across the brain, particularly in the lateral temporal region. Furthermore, subtype 2 exhibited faster deterioration in non-motor (MDS-UPDRS-Part Ⅰ, β = 1.26 ± 0.18, P = 0.016) and motor (MDS-UPDRS-Part Ⅱ, β = 1.34 ± 0.20, P = 0.017) symptoms, autonomic dysfunction (SCOPA-AUT, β = 1.15 ± 0.22, P = 0.043), memory (HVLT-Retention, β = −0.02 ± 0.01, P = 0.016) and depression (GDS, β = 0.26 ± 0.083, P = 0.019) compared to subtype 1. Conclusion The study has identified two PD subtypes with distinct patterns of atrophy progression and clinical progression, which may have implications for developing personalized treatment strategies.


Introduction
Parkinson's disease (PD) is the second most pr e v alent neur odegener ativ e disease, affecting an estimated 6.1 million people worldwide in 2016 alone and having a large effect on society (Feigin et al., 2019 ).PD is c har acterized by a r ange of motor and non-motor symptoms, including typical motor symptoms such as bradykinesia, rigidity, and tremor, as well as typical non-motor symptoms such as olfactory d ysfunction, cogniti ve impairment, psychiatric symptoms , sleep disorders , and autonomic dysfunction (Kalia & Lang, 2015 ;Thenganatt & Janko vic , 2014 ;Wang et al., 2020 ).The clinical manifestations and prognosis of PD vary widely, suggesting that it may not be a singular entity (Thenganatt & Janko vic , 2014 ).Identifying PD subtypes is crucial for understanding the underlying pathophysiological mechanisms, stratifying patients with different progressions, and designing personalized clinical trials (Fereshtehnejad & Postuma, 2017 ).Although increasing evidence supports the existence of different subtypes within PD (Fereshtehnejad et al., 2017 ;Horsager et al., 2020 ;Lewis et al., 2005 ;Thenganatt & Jankovic, 2014 ;Zhang et al., 2019 ), current classifications are primarily based on crosssectional data, which may mix the classification of subtypes with varying disease stages (Fereshtehnejad & Postuma, 2017 ;Young et al., 2018 ).To date, no studies have been reported PD subtyping based on longitudinal neur oima ging data to disentangle disease sta ges fr om disease subtypes.
The pathological mechanism of PD is the intracellular accumulation of misfolded alpha-synuclein, a pr esyna ptic pr otein that forms abnormal cytoplasmic a ggr egates (Choi et al., 2022 ;Kalia & Lang, 2015 ;Lau et al., 2020 ).The misfolded alpha-synuclein can spr ead acr oss distinct br ain r egions via the br ain connectome, causing neuronal death and atrophy in related brain regions (Bloem et al., 2021 ;Henderson et al., 2019 ;Lau et al., 2020 ;Rahayel et al., 2022 ).Mouse models have demonstrated that the patterns of alpha-synuclein a ggr egation ar e closel y associated with br ain connectivity (Henderson et al., 2019 ).Studies in humans have also found a relationship between brain connectivity and clinical symptoms as well as atr ophy pr ogr ession in PD (De Micco et al. , 2021 ;Vo et al. , 2023 ).Ho w e v er, humans v ary in their brain Figur e 1: Schematic o verview of the design.( A ) A total of 268 participants with follow up structural MRI and clinical assessments were included in the study.( B ) Hier arc hical clustering was used to classify PD patients into differ ent subtypes.First, VBM w as performed on the T1-MRI scans b y using the C AT 12. T hen, LME was used to estimate the rate of GMV loss.Hier arc hical clustering was employed to identify subtypes of patients, after normalizing the rate of GMV loss.( C ) The differences in the rate of GMV loss and longitudinal change rate in clinical scores were compared between these subtypes to assess their differences .C AT 12 = Computational Anatomy Toolbox 12. connectivity, suggesting that the alpha-synuclein a ggr egation pattern in PD patients also exhibits a high le v el of heter ogeneity (Karahan et al., 2022 ;Zhang et al., 2017 ).Furthermore, other demogr a phic c har acteristics, suc h as genetic factors and lifestyles also relate to the progression of PD (Guo et al., 2019 ;Li et al., 2023 ;Pu et al., 2022 ;Saunders-Pullman et al., 2018 ).These pr e vious findings suggest that there may be different subtypes of PD with distinct patterns of brain atrophy.
In the present study, we aimed to identify PD subtypes with differ ent r ates of br ain atr ophy deriv ed using voxel-based morphometry (VBM) with long itudinal imag ing data.VBM provides an automated, quantitative analysis of gray matter distribution with high regional specificity (Ashburner & Friston, 2000 ).PD involv es axonal degener ation and neur onal cell death, whic h ar e indexed by gray matter atrophy (Banwinkler et al., 2022 ;Yau et al. , 2018 ;Zeighami et al. , 2015 ).We hypothesized that the rate of gray matter volume (GMV) loss can reflect heterogeneity in PD, enabling its use for subtyping and exploration of subtype differences.We first fitted the rate of GMV loss for each patient at regional le v el, then explor ed whether these r egional r ates of GMV loss could classify PD patients into different subtypes.Furthermor e, the observ ed subtypes wer e v alidated thr ough a systematic assessment of their differences in clinical progression, cerebrospinal fluid (CSF) biomarkers, and dopamine transporter (DAT) binding deficit on single-photon emission computed tomogr a phy (SPECT) imaging.

Participants
The P arkinson's Pr ogr ession Mark ers Initiati ve (PPMI) is a landmark observational, longitudinal database consisting of neur oima g ing, biolog ical tests, and clinical and behavioral assessments (Marek et al., 2018 ).PPMI included individuals diagnosed with PD and healthy controls (HC).Recruitment criteria of patients with PD included: age ≥30, PD diagnosis within 2 years, Hoehn and Yahr Stage I-II at baseline, untreated with PD medications (levodopa, dopamine agonists , MAO-B inhibitors , or amantadine), and with one of at least two from resting tremor, bradykinesia, or rigidity (must have either resting tremor or bradykinesia) or a single asymmetric resting tremor or asymmetric bradykinesia (Mar ek et al., 2018 ).Furthermor e, all PD patients were confirmed by the positiv e DAT-SPECT (Mar ek et al., 2018 ).To ensur e r obustness of the estimated rates of GMV loss, PD patients were excluded if few er than tw o structural MRIs w ere collected or if their longest scanning interval was < 36 months .HC ha ve no current neurological disorder, no first-degree relative with PD, and normal DAT-SPECT ima ging (Mar ek et al., 2018 ).Additionall y, we excluded HC participants who were younger than 30 years old, failed image quality control, or lacked baseline data.Ultimately, a total of 161 HC and 107 PD participants were included in the study, as outlined in Supplementary Fig. S1 .The ov er all sc hema of the study is presented in Fig. 1 .
The data used in this study wer e r etrie v ed fr om the PPMI database in April 2021.The PPMI study was a ppr ov ed by institutional r e vie w boards at eac h study site, and all participants provided written informed consent before enrolment.

Clinical assessments
A compr ehensiv e set of clinical assessments were collected in the PPMI study, in which deriv ed v ariable definitions and score calculations ar e av ailable.The baseline and annual follow-up clinical scores were used in our analysis.In line with previous studies (Fereshtehnejad et al., 2017 ;Pagano et al., 2018 ;Wang et al., 2020 ) (Brandt et al., 1998 ;Shapiro et al., 1999 ). 4. A utonomic Testing: A utonomic dysfunction was e v aluated using the Scales for Outcomes in Parkinson's Disease-Autonomic total score (SCOPA-AUT) (Visser et al., 2004 ). 5. Slee p Problems: Slee p problems were evaluated using the REM sleep behavior disorder screening questionnaire (RBDSQ) and Epworth Sleepiness Scale (ESS) (J ohns , 1991 ;Stiasny-Kolster et al., 2007 ). 6. Neurobehavior: Non-motor Experiences of Daily Living were assessed using MDS-UPDRS Part I (Goetz et al., 2007 ).Depr ession was e v aluated using the Geriatric Depression Scale (GDS) (Yesav a ge, 1988 ).Tr ait and state anxiety were assessed using the State-Trait Anxiety Inventory (STAI) (Kendall et al., 1976 ).Impulse control disorders and related disorders wer e e v aluated using the Questionnair e for Impulsiv e-Compulsive Disorders in Parkinson's Disease (QUIP) (Knight et al., 1983 ;Weintraub et al., 2006Weintraub et al., , 2009 ) ). 7. Olfactory Testing: Impaired olfaction was evaluated using the age/sex-adjusted University of Pennsylvania Smell Identification Test (UPSIT) (Doty et al., 1995 ).

CSF and SPECT biomarkers
All participants conducted a lumbar puncture for the collection of CSF.Measurements of β-Amyloid 1-42 , total tau protein, and phosphorylated tau protein at Serine 181 were obtained for CSF samples at the University of Pennsylvania using the multiplex Luminex xMAP platform (Luminex Corp: Austin, TX, USA) with r esearc h-use-onl y Fujir ebio-Innogenetics INNO-BIA AlzBio3 imm unoassay kit-based r ea gents (Innogenetics Inc: Harv ard, MA, USA) (Marek et al., 2018 ).CSF alpha-synuclein was analyzed at a central laboratory (Co vance , MA, US) using a commercially available enzyme-linked immunosorbent assay kit (Locascio et al., 2011 ;Mollenhauer et al., 2013 ).

Imaging acquisition and processing
SPECT with the DAT tracer 123I-ioflupane was obtained for most of patients at baseline, 1, 2, and 4 years of follow-up.The specific binding ratios (SBR) was calculated for all the striatal areas using the occipital lobe as a r efer ence r egion (Mar ek et al., 2018 ).Structur al MRI scans wer e acquir ed in the sa gittal plane on 3T scanners at each study site using a magnetization-prepared r a pid-acquisition gr adient ec ho sequence (Mar ek et al., 2018 ).Acquisition parameters were as follows: repetition time = 2300/1900 ms; echo time = 2.98/2.96/2.27/2.48/2.52 ms; inversion time = 900 ms; flip angle: 9 • ; 256 × 256 matrix; and 1 × 1 × 1 mm 3 isotropic voxel.
For morphometric analysis of the imaging data, structural MRI scans wer e pr ocessed using the Computational Anatomy Toolbox 12 ( http:// www.neuro.uni-jena.de/cat/ ) in MATLAB.The default processing pipeline was applied, which includes bias correction of field inhomogeneities, segmentation into gray and white matter and CSF, and normalization using DARTEL.GMV estimates wer e extr acted fr om the left and right hemispheres for 200 cortical and 54 subcortical regions of interest (ROI) of the Neuromorphometrics Atlas (Schaefer et al., 2018 ;Tian et al., 2020 ).Raw images with low quality (CAT12 image quality rating < 70%) were excluded.

Cluster analysis
The rate of GMV loss for each R OI w as evaluated through a linear mixed-effects model (LME) (Fig. 1 ).Subsequently, this value w as standar dized using the z -score method at the population le v el.To classify PD patients into subgroups, we utilized hierarchical clustering, specifically W ard' s linkage method, based on the standardized values and Euclidean distance.Hier arc hical clustering offers a clear understanding of the hier arc hical r elationships among data points, allowing for the r e v elation of potential group structures without the need for a priori determination of the number of clusters.Euclidean distance was computed using all 254 dimensions, ensuring compr ehensiv e consider ation of the data.The W ard' s linkage method, which aims to minimize variance, was used to successively merge the two closest participants, resulting in a hierarchical tree structure of larger clusters.To determine the optimal number of clusters, w e emplo y ed the Calinski-Harabasz criterion ( 1974 ).This criterion e v aluates the clustering effect by considering the within-cluster compactness and between-cluster separ ation.A lar ger Calinski-Har abasz index indicates a better hier arc hical clustering effect, indicating that the sample is mor e a ppr opriatel y divided into two clusters, as depicted in Supplementary Fig. S2 .
To validate the clustering results, we also used k -means clustering on the first two principal components of atr ophy r ates acr oss differ ent br ain r egions .T he Cohen κ a gr eement r ate between hier arc hical clustering and k -means clustering was 0.63, indicating substantial a gr eement betw een the tw o methods .T his suggests that the patterns identified by the two different clustering techniques were consistent.

Sta tistical anal yses
The two-sample t -test was used to test the statistical difference of continuous demogr a phic and clinical scores with the adjustment of potential confounders, including age, sex, years of education, race, and study site effect.The χ 2 test was used to e v aluate gender distribution.For e v ery demogr a phic and clinical featur e, we calculated the mean and standard de viation (SD).False discov ery r ate (FDR) corr ection was a pplied to set the threshold for statistical significance at P < 0.05.All missing data points were excluded fr om our anal yses .T he co variates used in each statistical analysis ar e pr esented in Supplementary Fig. S3 .
To e v aluate the r ate of GMV loss, we used LME using the 'fitlme' function in MATLAB software (v.R2022a, MathWorks).Age, sex, race , sites , follow-up time , gr oup, and the inter action between time and group were included as fixed effects.Intercept and follo w-up time w er e also included for eac h participant as r andom effects.Whene v er GMV serv ed as the outcome measure, the total intr acr anial volume (TIV) was adjusted.The following formulas were used for the models: Model 1 was designed to fit the rate of GMV loss over time, model 2 was designed to examine the differences in GMV loss r ates among differ ent gr oups, and model 3 was designed to examine the differences in clinical progression among different gr oups.Multiple comparison corr ections wer e made using the FDR method, and P v alues < 0.05 wer e consider ed statisticall y significant.

Demographic and clinical characteristics
This study included a total of 161 HC, among whom 104 (64.60%) were male, and 107 PD patients, including 75 males (70.09%).The mean age of PD patients was 60.06 ± 9.98 years, and the av er a ge disease duration at baseline was 6.97 ± 7.24 months .T he mean follow-up time of structural MRI in PD patients is 4.06 years .T heir mean MDS-UPDRS Part I-III scores of PD patients were 4.66 ± 3.56, 5.14 ± 3.73, and 20.63 ± 9.52, r espectiv el y.A compr ehensiv e ov ervie w of the demogr a phic and clinical c har acteristics of the PD patients and HC are displayed in Supplementary Table S1 .There were significant differences in MDS-UPDRS, RBDSQ, GDS, SCOPA-AUT, STAI, and UPSIT between HC and PD at baseline.

Two PD subtypes identified by the rate of GMV loss
We then investigated whether the rates of GMV loss could cluster patients into distinct groups.As shown in Fig. 2 A, the results of hier arc hical clustering r e v ealed the pr esence of two subtypes of PD patients.To visualize the differences between these subtypes, we applied dimensionality reduction to the atrophy rates across v arious br ain r egions and plotted a scatter dia gr am of the first two principal components .T his dia gr am illustr ates the distinct distribution patterns observed among the two subtypes (Fig. 2 B).
To assess the stability of these clusters, we conducted a leaveone-out jack-knife validation.In this analysis, we found that each v alidation r esulted in a highl y stability cluster ed assignment, with Dice's coefficients ranging from 0.89 ± 0.159 (mean ± standard deviation; Supplementary Fig. S4 A, upper panel).Furthermore, the atr ophy r ate patterns within eac h cluster wer e found to be stable, with high spatial correlation coefficients compared to the main results (Subtype 1: r = 0.95 ± 0.149; Subtype 2: r = 0.99 ± 0.0052; all PFDR < 0.05; Supplementary Fig. S4 B, upper panel).Additionally, we performed a 5-fold cross-validation to further test the robustness of our findings.Even with this greater perturbation to the data, eac h v alidation sho w ed a r elativ el y high cluster assignment, with Dice's coefficients r anging fr om 0.87 ± 0.124 (mean ± standard deviation; Supplementary Fig. S4 A, lower panel).The atrophy rate patterns within each cluster also exhibited high spatial corr elation coefficients compar ed to the main results (Subtype 1: r = 0.99 ± 0.0057; Subtype 2: r = 0.93 ± 0.0024; all PFDR < 0.05; Supplementary Fig. S4 B, lo w er panel).Collectiv el y, these r esults demonstrate the stability of the clusters based on the spatial distribution of atrophy rate patterns within each subtype.
These two subtypes comprised 58.88% (63 patients) and 41.12% (44 patients) of the PD patients.Subtype 1 is younger than Subtype 2 at baseline.Additionally, Subtype 1 exhibits milder se v erity of motor and non-motor symptoms compared to Subtype 2. Furthermore, Subtype 1 demonstrates higher DAT-SBR values in v arious subcortical br ain r egions and higher concentrations of CSF biomark ers relati ve to Subtype 2. Ho w ever, after correction for multiple comparisons, these differences do not r eac h statistical significance.A summary of clinical, biological, and cognitive c har acteristics of the two PD subtypes is provided in Table 1 .

Comparison of r a tes of GMV loss between subtypes of PD and HC
We further compared the rates of GMV loss between PD subtypes and HC.As shown in Fig. 3 A, there were differences in the rates of GMV loss among HC and the two PD subtypes.In general, the rates of GMV loss over time was minimal within the HC group.By contrast, subtype 1 displayed a moderate level of GMV loss over time, primarily located in the pr efr ontal and temporal lobes.Howe v er, subtype 2 exhibited widespread higher rates of GMV loss over time in most regions of the brain.When comparing these two subtypes, subtype 2 had a significantly higher rates of GMV loss in nearly all brain regions, particularly in the lateral temporal lobe, hippocampus, and thalamus (as illustrated in Fig. 3 B; P < 0.05, FDR corr ected).These findings underscor e the pr esence of unique and distinguishable patterns of neur odegener ation r ates that ar e linked with the different subtypes of PD.
In addition, to gain more detailed information about the specific atrophy patterns associated with each subtype, we performed a compar ativ e anal ysis of GMV data between eac h subtype at each follow-up visit and the corresponding values obtained from HC at baseline.Our anal yses r e v ealed that subtype 1 sho w ed sig-nificant atr ophy primaril y in the superior frontal gyrus and caudate nucleus, whereas subtype 2 presented with marked atrophy in various regions including the temporal lobe, frontal lobe, and subcortical regions when compared to HC participants (refer to Supplementary Fig. S5 ).
Although there were no significant differences in DAT-SBR in the striatum, caudate, and putamen betw een the tw o subtypes at baseline and the rate of change, their difference in follow-up visits r eac hed significant le v els ( P < 0.05 FDR; Fig. 4 G-I).Subtype  All values presented are the standardized beta coefficients (accompanied by their standard errors), unless explicitly stated otherwise.P v alues r epr esent the significance of the interaction term between 'follow-up time' and 'group' in model 3, with v alues consider ed significant after a ppl ying the FDR correction ( P < 0.05).
1 had higher CSF le v els of alpha-synuclein, t-tau, and p-tau181 than subtype 2, both at baseline and at follo w-up.Ho w e v er, these differences in CSF biomarkers did not survival FDR correction ( Fig. S7 ).In addition, we employed Cohen's d to e v aluate the effect size of differences in clinical assessments for the two PD subtypes at baseline and follow-up, r espectiv el y (see Supplementary Tables S2  and S3 ).

Discussion
The current study identified two PD subtypes with distinct rates of br ain atr ophy and clinical pr ogr ession.Subtype 1 was c har acterized by moderate levels of atrophy rates in the prefrontal lobe and later al tempor al lobe, along with slo w er clinical pr ogr ession.By contrast, subtype 2 exhibited higher rates of atrophy across most * P < 0.05, * * P < 0.01, * * * P < 0.005, with the FDR correction applied for the follow-up visits.FDR correction performed in all follow-up times.
br ain r egions, as well as mor e se v er e DAT deficits .T his subtype also demonstrated faster clinical progression in terms of mental, slee p, autonomic, and cogniti ve symptoms.The identification of distinct subtypes may aid in the de v elopment of mor e tar geted and personalized treatment strategies for PD patients.
Our study identified two PD subtypes with different atrophy and clinical pr ogr ession by using longitudinal structur al MRI scans .T he observ ed br ain atr ophy r ates in both PD subtypes wer e notabl y heightened compar ed to those of the contr ol gr oup, thereby confirming that pathological protein aggregation indeed acceler ates neur onal degener ation in PD (Choi et al., 2022 ;Lau et al., 2020 ;Shahnawaz et al., 2020 ).To pr oactiv el y curb the progression of PD, future research endeavours should concentrate on de vising str ategies aimed at mitigating neuron loss or exploring the potential of personalized iPSC-derived dopamine progenitor cell tr ansplantation (Sc hweitzer et al., 2020 ).Subtype 1 was associated with a r elativ el y well-pr eserv ed br ain structur e and slo w er clinical pr ogr ession.By contr ast, subtype 2 is associated with more widespread neurodegeneration, a more severe DAT deficit, and a faster clinical deterioration.These findings align with pr e vious r esearc h showing that PD patients with greater br ain r esources exhibit gr eater compensatory ca pacity, whic h can enhance their ability to maintain function and potentially slows the pr ogr ession of the disease (Arkadir et al., 2014 ;Gregory et al., 2018 ;Nandhagopal et al., 2011 ;Wang et al., 2022 ).Mor eov er, our r esults align with pr e vious findings that plasma neurofilament light c hain (NfL) le v els positiv el y corr elate with the motor and cognitiv e se v erity and pr ogr ession in PD (Lin et al., 2019 ;Ye et al., 2021 ).As NfL is a biomarker for neur odegener ation, higher plasma NfL levels suggest higher le v els of neur odegener ation in the brain (Khalil et al., 2018 ;Wang et al., 2023 ).Collectiv el y, these findings support the notion that PD is a heterogeneous disorder with distinct subtypes exhibiting different patterns of neur odegener ation and clinical pr ogr ession.
Subtype 2 is c har acterized by higher rates of atrophy in subcortical regions and the temporal lobe, suggesting higher rate of misfolded alpha-synuclein accumulation in these regions (Abdelgawad et al., 2023 ).This finding offers a potential explanation for the mor e r a pid pr ogr ession of non-motor symptoms observ ed in this subtype, particularly with regards to de pression, slee p disorders , and cognitive impairment.T he affected br ain r egionswhich play critical roles in functions such as emotion, autonomic control, and cognition-may underlie the faster progression of these symptoms (Rolls, 2015(Rolls, , 2021 ; ;Rolls et al., 2023 ).Ther efor e, the incr eased atr ophy observ ed in these areas could contribute to the mor e r a pid pr ogr ession of non-motor symptoms (Sc ha pir a et al. , 2017 ;W ilson et al. , 2019 ).
After correcting for multiple comparisons, no significant differ ences wer e observ ed in CSF biomarkers or DAT-SBRs between the two PD subtypes .T his ma y be attributed to the small sample size, which limits the statistical po w er to detect significant differences.We calculated effect sizes for the differences between these two subtypes in terms of CSF biomarkers and DAT-SBRs, with all values exceeding 0.2.The effect sizes for putamen-SBR and P-tau r eac hed moder ate le v els, at 0.65 and 0.50, r espectiv el y.This r esult aligns with pr e vious findings that more severe dopamine neuron loss at baseline is associated with more pronounced symptoms (Liu et al., 2018 ).It is consistent with a pr e vious finding that lo w er baseline CSF A β 1-42 and alpha-Synuclein were associated with faster increased motor score (Irwin et al., 2020 ).We observed that subtype 1 had higher le v els of all CSF biomarkers at baseline compared to subtype 2. Given that HC generally have higher CSF biomarker le v els than PD patients (Irwin et al., 2020 ;Kang et al., 2013 ) and considering CSF has important role in clearing brain metabolic waste (Wichmann et al., 2022 ), this might imply that subtype 1 patients possess a r elativ el y str onger ca pacity to eliminate pathological proteins (Kang et al., 2013 ).As a result, subtype 1 appears to demonstrate less accumulation of pathological proteins in the brain, leading to milder brain atrophy and slo w er disease pr ogr ession.This finding suggests a potential connection between CSF clearance efficiency and disease pr ogr ession in PD subtypes , meriting further in v estigation into the underl ying mec hanisms.
Although our discovery of two subtypes of PD was based on longitudinal neur oima ging data, baseline differences in clinical symptoms and biomarkers may aid in early identification.Specificall y, while not initiall y statisticall y significant, measur es suc h as MDS-UPDRS total scores, RBDSQ, UPSIT, SFT, HLVT, putamen-SBR, and P-tau le v els exhibited moder ate-to-lar ge effect sizes (Cohen's d > 0.5), indicating their potential for use in baseline differentiation.Futur e r esearc h should inv estigate the utility of these markers in distinguishing between PD subtypes at baseline.Our findings align with pr e vious studies demonstr ating that PD patients with se v er e RBD, cognitiv e impairment, hyposmia, and dopaminergic deficits on DAT imaging exhibit faster clinical pr ogr ession (P a gano et al., 2018 ;Sc hr a g et al., 2017 ).These r esults support the notion that PD patients with these c har acteristics may r epr esent a distinct subtype associated with a more rapid disease progression (Horsager et al., 2020 ).
PD r emains incur able, making the need for ther a pies that can slow its pr ogr ession incr easingl y ur gent (Armstr ong & Okun, 2020 ).Our findings suggest that subtype 1 exhibits slo w er rates of br ain atr ophy and disease pr ogr ession compar ed to subtype 2. Identifying this subgroup is crucial for guiding the development of targeted therapies.Future research exploring genetic, en vironmental, lifestyle , immune , and metabolic differences between these subtypes may provide insights into the underlying mechanisms and modifiable factors responsible for these distinct patterns of atrophy and clinical progressions (Ortega et al., 2021 ;Pu et al., 2022 ;Zhang et al., 2022 ).Such insights could lead to the de v elopment of ther a peutic str ategies that effectiv el y pr eserv e br ain structur e, slow clinical pr ogr ession, and impr ov e the quality of life for PD patients .Moreo ver, considering the large number of modifiable risk factors associated with PD onset and cognitive decline (Belvisi et al., 2020 ;Guo et al., 2019 ;Tan et al., 2016 ), it will be important to investigate the differences in these risk factors be-tween subtypes to identify potential targets for intervention that can slow PD pr ogr ession.
One strength of this study is the use of a multicenter cohort from the PPMI database, which has comprehensive longitudinal MRI and clinical data to investigate their differences between two subtypes.Ho w e v er, some limitations should be considered.First, we only used a subset of participants who met the inclusion criteria, whic h may intr oduce sample bias.Second, the method used to assess the rates of GMV loss emplo y ed a linear a ppr oac h.Howe v er, giv en that the atr ophy tr ajectories of some br ain r egions may follow non-linear patterns, future studies could use more complex, non-linear models to impr ov e the accuracy of atrophy rate assessment over an extended disease duration.This would allow for a more comprehensive estimation of atrophy rates , pro viding a mor e accur ate r epr esentation of the underl ying pr ocesses .T hird, the mean age of onset in subtype 2 is higher than that of subtype 1 (Table 1 ), although the difference was not significant ( P = 0.129).Older age at onset was associated with fast disease pr ogr ession (Ferr ar a et al., 2016 ).Ther efor e, we cannot completel y rule out the role of age in the progression differences between two subtypes.Finally, this study is a single-cohort study, although PPMI recruited participants fr om m ultiple sites, our subtyping r esults should be validated in other cohorts.

Conclusion
In conclusion, our findings have identified two PD subtypes that exhibit distinct rates of brain atrophy and clinical progression.The identification of these subtypes offers valuable insights into the heterogeneity of the disease and may inform the de v elopment of personalized treatment strategies.Future studies should focus on exploring the neurobiological mechanisms underlying the heterogeneity of PD.

Figur e 2 :
Figur e 2: T he hier arc hical clustering r esults and its visualization.( A ) Dendr ogr am of the hier arc hical cluster of patients with PD. ( B ) Visual r epr esentation of the hier arc hical clustering result based on the principal component (PC) coordinates of the first two dimensions.

Figure 3 :
Figure 3: Patterns of GMV loss rates among both subtypes of PD and HC participants over time.( A ) The patterns of GMV loss rate for each respective gr oup.( B )The differ ences in the r ate of GMV loss between each pair of groups.A higher t value indicates a faster rate of GMV loss or a lar ger differ ence in GMV loss rate between pairs of groups.Only ROI with statistically significant corrected P values, as determined through FDR correction with a threshold of P < 0.05, were visualized in the figures.

Figure 4 :
Figure 4: Longitudinal trajectories of clinical scores and SPECT biomarkers for two PD subtypes .T he asterisks represent the statistical significance of the comparison between the two subtypes in the clinical variables at different follow-up visits .T he significance levels are indicated as follows:* P < 0.05, * * P < 0.01, * * * P < 0.005, with the FDR correction applied for the follow-up visits.FDR correction performed in all follow-up times.

Table 1 :
Characteristics of HC and two PD subtypes.

Table 2 :
Longitudinal changes in clinical scores of two PD subtypes ( n = 107).