Machine Learning Prediction of Hoehn and Yahr Scores at 5‐Years Post‐ ¹²³I‐Ioflupane SPECT Imaging in a Real‐World Parkinson's Disease Dataset

Abstract

Background

Parkinson's disease (PD) progression is highly heterogeneous, complicating clinical management and prognostication. While machine learning models have been developed using research datasets such as Parkinson's Precision Medicine Initiative (PPMI) and Parkinson's Disease Biomarkers Program (PDBP), their clinical translatability is limited due to differences in routinely collected data. The Hoehn and Yahr (H&Y) scale is commonly used in clinical practice to stage PD, yet most predictive models focus on less practical measures.

Objectives

This study developed and validated machine learning models to predict H&Y scores at 5 years post‐123I‐ioflupane single‐photon emission computed tomography (SPECT) imaging, leveraging both a real‐world dataset and a subset of the PPMI cohort. The goal was to assess the utility of routinely collected clinical and imaging data for prognostic modeling.

Methods

Data from medical records and imaging were harmonized from 343 real‐world patients and 134 PPMI patients, resulting in a merged dataset with 83 overlapping features. Random Forest and Gradient Boosting models were trained to predict 5‐year H&Y scores using varying amounts of longitudinal data and imaging features.

Results

Models using 2 years of clinical follow‐up data achieved the highest predictive accuracy. The most important predictors were early H&Y scores, gait symptom severity, and select imaging features.

Conclusions

Machine learning models can predict 5‐year H&Y scores in PD using real‐world clinical data, but imaging features add limited prognostic value. This study demonstrated that implementing machine learning models, when using real‐world data, did not significantly improve the already known gap between prognostic modeling and real‐world implementation. Improvement of models is, however, a promising prospect and further studies are encouraged.

Parkinson's disease (PD) is a progressive neurodegenerative disorder characterized by motor and non‐motor symptoms. Understanding the progression of PD is crucial for clinical management and improving patient outcomes. Predicting the progression of PD can help tailor treatment plans and provide better prognostic information to patients and clinicians. ¹ The heterogeneity of PD progression which includes variations in motor and cognitive symptoms—presents significant challenges in clinical practice.

There have been numerous studies that have used either the Parkinson's Progression Markers Initiative (PPMI) dataset ² and/or the Parkinson's Disease Biomarkers Program (PDBP) dataset ³ to develop machine learning or deep learning models to predict the progression of PD. The PPMI dataset includes comprehensive clinical variables, and ¹²³I‐ioflupane SPECT imaging data collected from over 1500 participants, including healthy controls and PD patients. The PDBP dataset includes clinical, genetic, T1‐weighted MRI, and biomarker data from PD, Lewy Body Dementia, and related disorders from over 4000 individuals. Despite the richness of these datasets, they do not reflect the information that is routinely collected in a clinical environment. These research datasets comprise various metrics including the Unified Parkinson's Disease Rating Scale (UPDRS) and the Montreal Cognitive Assessment (MoCA) scores, which are used to measure disease severity and progression but are rarely used clinic. ⁴ , ⁵ Clinically, the Hoehn and Yahr (H&Y) scale is more commonly used to stage PD and gives a gross representation of the severity of the patient's physical symptoms. ⁶

Studies aimed at predicting progression of PD can generally be grouped into three categories: those that predict cognitive progression, those that predict motor progression and those that predict a combination of both. Predicting cognitive progression has been done in numerous ways, such as predicting the UPDRS I ⁷ or MoCA score at follow‐up, ⁸ , ⁹ , ¹⁰ or classifying patients at a future time point as having or not having dementia. ¹¹ , ¹² Similarly, papers predicting motor progression have been largely focused on predicting the UPDRS III score at a follow‐up time point ⁷ , ¹³ , ¹⁴ , ¹⁵ , ¹⁶ , ¹⁷ or change in UPDRS III over time. ¹⁸ The few studies which predicted a combination of cognitive and motor progression predicted either the combined UPDRS I, II and III score ¹⁹ or classified patients as either having generally “favorable” or “unfavorable” outcomes after a set period of time. ²⁰ Recently, a model was developed to predict the H&Y score 4 years after baseline imaging using features from ¹²³I‐ioflupane SPECT imaging alone. ²¹ Other studies have aimed to predict specific clinical events such as the onset of freezing of gait ²² or levodopa‐induced dyskinesia. ²³

As another method of predicting progression, machine learning and clustering methods have been applied to larger datasets in order to define subtypes based on the pace of disease progression. ²⁴ , ²⁵ , ²⁶ , ²⁷ , ²⁸ , ²⁹ Then, a model can be used to predict the subtype for an individual patient. However, while this may seem an intuitive conceptualization, there are not clinically accepted definitions for fast, medium and slow progression in the context of PD. Therefore, their translation into the clinical environment may be limited.

SPECT imaging using ¹²³I‐ioflupane has been utilized in several studies to assess its potential in predicting PD disease progression. ¹⁵ , ¹⁶ , ²¹ However, predictive models have primarily used the PPMI dataset, in which images were collected at the time of diagnosis in de novo patients, which does not necessarily represent a real‐world scenario. In clinical practice, patients may receive ¹²³I‐ioflupane SPECT imaging years after symptom onset to confirm a diagnosis. Therefore, it is still unclear if ¹²³I‐ioflupane SPECT imaging features can be used along with routinely collected clinical data to predict progression in a real‐world setting.

The aim of this study was to train a model to predict the H&Y score of PD patients at 5 years post‐¹²³I‐ioflupane SPECT imaging using a combined dataset, including both a subset of the PPMI dataset and a real‐world dataset which was collected in a true clinical context. By leveraging both datasets together, this study seeks to test the ability of standard machine learning methods to predict progression of patients from non‐systematic information that is obtained in a real‐world setting. Additionally, the H&Y score was chosen as the target to predict since it is routinely used in a clinical setting.

Materials and Methods

Real‐World Dataset

The real‐world dataset used in this work consists of clinical data and ¹²³I‐Ioflupane SPECT images for PD patients collected at Bispebjerg Hospital. Approval from the Danish Patient Safety Authority (ref 31‐1521‐278) and the Danish Data Protection Agency (ref P‐2020‐372) was obtained to access and collect data between 2000 and 2019. Patients who had an initial diagnosis of PD and at least 5 years follow‐up at the Department of Neurology were initially included in the study (n = 373). In the Danish electronic health record system, all medical records from the 5‐year period after the ¹²³I‐Ioflupane SPECT imaging were manually reviewed and curated. Three persons (MNJ, MPH, MZS) with a medical degree were curating the data retrospectively from patient files using the same approach and algorithm. Ambiguities were decided by a movement disorder specialist, AL. All relevant data were compiled into a newly constructed and anonymized database in the electronic case report software solution REDCap. ³⁰ , ³¹ Data quality was monitored by AL and MNL, both continuously and after data retrieval. The first data entry point (referred to as “Baseline”) was set at the date of the clinical entry point closest to the date of the ¹²³I‐Ioflupane SPECT imaging. In all patients, the first data entry point (“Baseline”) and the date of the ¹²³I‐Ioflupane SPECT imaging were within 12 months of one another. Patients with a normal ¹²³I‐Ioflupane SPECT image (n = 24), less than five visits over the 5 years (n = 1) or more than 12 months between the first data entry point and the time of imaging (n = 5) were excluded. Therefore, the final dataset included 343 patients.

Data from medical records related to PD, including subjective symptoms, objective assessments (motor evaluations and clinical characteristics), neuropsychological tests, laboratory results, imaging reports, and medication status/changes, were recorded at the corresponding time of documentation. If a patient was admitted to the neurology department, all relevant information from the admission was consolidated into a single registration timepoint. The same approach was applied to multiple neurology department contacts within the same month. Hospitalizations related to medication adjustments, deep brain stimulation (DBS) surgery, percutaneous endoscopic gastrostomy (PEG) tube placement, joint replacements, myocardial infarctions, cerebral infarctions, seizures, and head trauma were recorded. If, at any given timepoint, data from medical records raised suspicion of different parkinsonian syndromes, it was reviewed by a movement disorder specialist (AL) to determine the diagnosis.

Clinical data were organized into three sections: Baseline, Follow‐up and Endpoint. Baseline variables included information such as patient, age, demographics and initial diagnosis. Follow‐up variables included information about symptoms such as gait, rigidity, tremors among many others. Endpoint was the final diagnosis after 5 years. As clinical information for each follow‐up timepoint was entered into the REDCap database, H&Y scores were also assigned based on symptom information. In total, 955 variables were evaluated in the REDCap database for each patient.

Briefly, SPECT images were acquired 3 h after injection of 185 MBq ¹²³I‐Ioflupane. Image reconstruction was performed using iterative reconstruction with scatter and non‐uniform attenuation correction. A designation of normal or abnormal was recorded in the REDCap database based on the conclusion that was registered in the patient files, and which was given to the clinician at the time of the imaging.

PPMI Dataset

The PPMI dataset is a comprehensive, open‐access study aimed at identifying biomarkers of PD progression. Patients included recently diagnosed, unmedicated PD patients. The dataset covers demographics, medical and PD history, motor and non‐motor assessments (such as MDS‐UPDRS, Hoehn & Yahr staging, cognitive and mood evaluations), medication use, and additional features like sleep and olfactory tests. It also includes imaging (¹²³I‐Ioflupane SPECT, MRI) and biospecimen data (CSF, blood, urine) gathered at baseline and regular follow‐up visits over several years, providing a deeply phenotyped resource for studying disease progression. Data were collected every 3 months for the first 2 years and then subsequently every 6 months. ² Images, captured on a clinical SPECT/CT scanner 3–4 h after ¹²³I‐Ioflupane injections at doses between 111 and 185 MBq, were collected at both baseline and after 1 year. A subset of patients with formal PD diagnoses and abnormal ¹²³I‐Ioflupane SPECT imaging were included in this study (n = 134) as well as their images at baseline. These patients were part of the earliest cohort included in PPMI.

Image Features

The specific binding ratio (SBR) values, which assess the density of dopamine transporters in the brain compared to a reference region, were calculated from each ¹²³I‐Ioflupane SPECT image in both datasets. Additionally, asymmetry values, indicating differences between right and left sides, standard deviations from the age‐matched mean SBR of the normal database (z‐score), ³² as well as ratios of uptake in caudate versus putamen were computed. Images were processed using DaTQUANT 2.0 (GE Healthcare, Chicago, IL, USA), an FDA‐approved and EU CE‐marked software to compute these values in a standardized way. In the application, the images were first registered to a fixed template. Then, the metrics were calculated based on the contours from the template regions of interest. A full list of imaging features can be found in the Table S1.

Merging Datasets

For this study, the real‐world dataset was combined with the previously described subset of cases from the PPMI dataset. Although there is significant overlap in the information collected in the real‐world and PPMI datasets, the variable names, definitions, and conventions differed. Therefore, manual conversion tables were created to find the overlapping features in both datasets. Additionally, if a variable was missing at all follow‐up timepoints for a patient, it was labeled as “all_missing.” Variables which had a label of “all_missing” for 60% or more of patients were excluded from analysis to avoid over‐imputation of missing data. The resulting overlapping 83 features, 11 baseline variables, 42 follow‐up variables and 30 imaging variables, were used as input to the models. Additionally, the follow‐up data from the real‐world dataset were not collected at regularly scheduled intervals since they represent true clinical follow‐up, which depended on patients, their symptoms and their progression. Therefore, to align the study visits for both datasets, the real‐world data was interpolated onto the PPMI time points. An overview of the process for merging the two datasets is shown in Figure 1. More details on the individual steps can be found in the Supplementary Material.

Figure 1.

Flowchart describing how the PPMI and real‐world datasets were merged.

After the remaining variables were cleaned and the two datasets were merged, there remained variables with missing values. Therefore, imputation methods were applied. For example, the missing values for height and weight were imputed by using the mean value grouped by the sex. The missing values for other numerical baseline variables were imputed by the mean value. For categorical baseline variables, the missing values were imputed by the mode value. For the follow‐up variables, the missing values were imputed by forward‐filling values based on the previous value for that patient. This approach was chosen in consultation with clinicians at Bispebjerg Hospital, as in the real‐world dataset clinical information was often only documented when a change in symptoms was observed; therefore, an unrecorded variable at a given visit was considered most likely to reflect no clinically discernible change since the preceding visit.

Predicting H&Y Scores

Longitudinal data from both the real‐world and PPMI datasets were used to predict H&Y scores at 5 years post ¹²³I‐Ioflupane SPECT imaging using machine learning methods. Various permutations of input data were tested including the dataset (real‐world, PPMI or combined) and the amount of longitudinal data (6 months from baseline, 1 year from baseline or 2 years from baseline). Additionally, both classification and regression models were tested for the task. In the regression task, a regression model was fitted to the input data to directly predict the H&Y score at the fifth year. In the classification task, the H&Y scores at the fifth year are categorized to six classes, one for each possible H&Y score (0–5). The general method is shown in Figure 2. The combined dataset, consisting of the PPMI and real‐world data, was split into 80% training and 20% test data.

Figure 2.

Flowchart demonstrating how information is used in the models to predict H&Y scores at 5 years post imaging.

Random Forest (RF) and Gradient Boosting (GB) models were tested. Hyperparameter selection was performed using a randomized search five‐fold cross‐validation (CV) strategy to find the best model parameters while reducing the risk of overfitting. ³³ The $R^2$ and accuracy were used to evaluate the performance for the regression and classification tasks, respectively.

Results

Data

The final dataset used to develop and test the models consisted of 134 patients from the PPMI cohort and 343 patients in the real‐world dataset. Patient characteristics are listed in Table 1. The distribution of selected variables is shown in Figure 3. In total, 11 baseline variables describing the patient at or near the time of imaging and 42 follow‐up variables were used, see Table S1. In the follow‐up variables, there are 13 timepoints, which are named BL, V01, V02, V03, etc. up to V12, where V represents the “visit.” The full schedule of follow‐up visits is outlined in the Table S2 and the method of interpolating the real‐world dataset onto the PPMI timepoints is shown in the Figure S1.

TABLE 1.

Patient characteristics in both the PPMI and real‐world datasets. Entries are listed as mean ± SD, number (percentage of total for each dataset) or median [lower quartile, upper quartile], where appropriate

Variable	PPMI	Real‐world
Number of patients	134	343
Age at baseline	61.6 ± 9.8	63.7 ± 11.0
Male	88 (65.7%)	210 (61.2%)
Female	46 (34.3%)	133 (38.8%)
Height (cm)	173.0 [165, 180]	177 [164.8, 179]
Weight (kg)	81.5 [71.6, 93.3]	77.1 [65.9, 82.4]
Baseline H&Y 0/1/2/3/4/5*	0/60/74/0/0/0	0/61/167/94/17/4
5 Year H&Y 0/1s/2/3/4/5*	0/12/109/10/2/1	0/12/119/111/65/36

Figure 3.

Violin plots showing distribution of variables for Real‐world and PPMI datasets.

The distribution of H&Y scores at baseline and 5 years post imaging are shown for both the PPMI and real‐world datasets in Figure 4. Overall, the real‐world population shows a greater shift in H&Y scores over 5 years, indicated by a higher number of patients having a H&Y score of 3, 4 or 5 than the PPMI patients at 5 years. A greater number of patients in the PPMI cohort had H&Y scores which remained stable over 5 years. For the combined dataset, the majority of patients either do not have an increase in H&Y score over 5 years (ΔHY ≤0, n = 242) or only increase the score by one (ΔHY = 1, n = 158) (see Supplementary Material, Table 3). A detailed breakdown of scores can be seen in the Tables S4, S5 and S6.

Figure 4.

H&Y distribution at baseline and 5‐years after baseline imaging.

TABLE 3.

Results (accuracy: mean (SD)) of the classification model tested using 6 months, 1 year and 2 years of clinical follow‐up information as input. Clinical = clinical follow‐up information (including H&Y scores). IF = imaging features from baseline ¹²³I‐Ioflupane SPECT imaging. H&Y only = H&Y scores at regular time points. RF = Random Forest. GB = Gradient Boost. Grayed cells represent the most accurate combination of model and input variables for a given amount of follow‐up timepoints included (6 months, 1 year or 2 years)

Method	6 Months			1 Year			2 Years
	Clinical	Clinical + IF	H&Y only	Clinical	Clinical + IF	H&Y only	Clinical	Clinical + IF	H&Y only
RF	0.624 (0.028)	0.621 (0.019)	0.626 (0.023)	0.651 (0.036)	0.648 (0.018)	0.665 (0.038)	0.705 (0.025)	0.700 (0.022)	0.728 (0.036)
GB	0.613 (0.029)	0.600 (0.025)	0.632 (0.028)	0.655 (0.031)	0.648 (0.016)	0.653 (0.035)	0.726 (0.025)	0.709 (0.018)	0.726 (0.029)

Predicting H&Y Scores

In this paper, a regressor and classifier were both tested for predicting the H&Y score at 5 years post baseline imaging. For the regression task, the results show that when only using 6 months or 1 year of clinical follow‐up information, the best results were obtained using the clinical variables alone ($R^2$= 0.548 and $R^2$= 0.610, for 6 months and 1 year, respectively). When using 2 years of follow‐up clinical information with imaging features, the most accurate result achieved was $R^2=$ 684, only slightly improved from using clinical variables alone. The results shown in Table 2 reflect those using the merged data set, comprised of patients from the PPMI and real‐world datasets. The model was also tested on the PPMI and real‐world datasets independently and those results are shown in the Supplementary Material. For the classification task, each patient was classified according to its predicted H&Y score at 5 years after baseline imaging. The accuracy was calculated to evaluate the classification model. Unlike the regression task, overall best results were achieved using H&Y follow‐up variables alone (Table 3).

TABLE 2.

Results ($R^2$: mean (SD)) of the regression model tested using 6 months, 1 year and 2 years of clinical follow‐up information as input. Clinical = clinical follow‐up information (including H&Y scores). IF = imaging features from baseline ¹²³I‐Ioflupane SPECT imaging. H&Y only = H&Y scores at regular time points. RF = Random Forest. GB = Gradient Boost. Grayed cells represent the most accurate combination of model and input variables for a given amount of follow‐up timepoints included (6 months, 1 year or 2 years).

Method	6 Months			1 Year			2 Years
	Clinical	Clinical + IF	H&Y only	Clinical	Clinical + IF	H&Y only	Clinical	Clinical + IF	H&Y only
RF	0.548 (0.038)	0.510 (0.048)	0.464 (0.053)	0.610 (0.068)	0.596 (0.055)	0.542 (0.067)	0.680 (0.057)	0.678 (0.074)	0.616 (0.033)
GB	0.538 (0.041)	0.503 (0.060)	0.464 (0.055)	0.607 (0.068)	0.593 (0.057)	0.541 (0.060)	0.683 (0.057)	0.684 (0.066)	0.611 (0.032)

The top 15 most important features from the best model (regression GB model with 2 years of clinical follow‐up information and imaging features) are shown in Figure 5, in order of relative importance. The list demonstrates that the H&Y score contributes the most to the prediction, although clinical variables describing gait symptom severity and rigidity as well as imaging features such as the left caudate SBR Z‐score and left anterior putamen were also important. For further definition of these variables, see Table S1.

Figure 5.

Feature importance for the most accurate model (GB regression with 2 years of clinical follow‐up information and imaging features) used to predict H&Y scores at 5 years post‐123I‐Ioflupane SPECT imaging.

Discussion

This study leverages a unique real‐world dataset of ¹²³I‐Ioflupane SPECT images and 5 years of clinical follow‐up information from PD patients to predict progression. Most previous studies attempting to predict progression have utilized the PPMI or PDBP datasets, which are highly curated with vast amounts of information that is not routinely collected in a clinical setting. Although they have provided insights into important clinical features, which can be used to distinguish patients which progress more quickly, there remains an open question as to what insights can be gained based on data that is currently collected in a clinical setting. Additionally, many of these studies use scores such as UPDRS and MoCA to measure progression, even though these tests are time‐consuming to administer and thus are not commonly used in a clinical setting. Therefore, the primary focus of this study was to predict the progression of PD patients using the H&Y score, which is widely used in clinical settings, using a real‐world retrospective dataset.

Recently, research frameworks have been proposed for staging individuals with neuronal synuclein disorders, including PD, based on a combination of genetic traits, biomarkers, and clinical factors. ³⁴ Key biological anchors for these schemes are biomarkers for alpha‐synuclein (S+/−), currently determined using a seed amplification assay on cerebrospinal fluid samples, and striatal neurodegeneration (N+/−), currently determined using ¹²³I‐ioflupane SPECT imaging. Alpha‐synuclein status was unavailable for the PPMI data set at the time of downloading and is currently unavailable for the real‐world data set. However, all participants were N+ in both data sets, and a very recent report from the PPMI study ³⁵ indicates that 99% of participants in the PD cohort used in this present study were S+ at baseline. Thus, assuming no causal genetic mutation, the PPMI participants included in this study were potentially Stage 2B or higher in the NSD‐ISS framework ³⁴ and G‐S + N+ in the SynNeurGe framework. ³⁶

The results indicate that, while ¹²³I‐Ioflupane SPECT imaging is known to be useful for diagnosing PD and differentiating it from other neurological phenocopies, quantitative outcomes of scans that were visually abnormal at baseline do not provide significant added prognostic value when predicting the H&Y score of patients 5 years after imaging. According to the best model in this study, the most important variables for prediction include H&Y scores at follow‐up visits in the first 2 years, gait symptom severity, rigidity and imaging features, such as Putamen/Caudate SBR values. Symptoms such as gait and Putamen/caudate SBR values were expected to play a role since the H&Y score is determined based on the patient's motor symptoms. ⁶

These findings are consistent with recent longitudinal research, which demonstrated that while ¹²³I‐Ioflupane SPECT imaging is valuable for initial diagnosis, its utility in tracking disease progression is limited when used alone. ³⁷ This limitation may be further compounded by the heterogeneity of PD itself. For instance, the emerging “gut‐first” or RBD‐dementia subtype suggests that, in some patients, pathology originates in the enteric nervous system and progresses centrally, with REM sleep behavior disorder and cognitive symptoms preceding motor decline. In such cases, the substantia nigra pars compacta may be affected later, potentially reducing the early prognostic value of dopaminergic imaging. Stratifying patients by subtype in future studies may help improve the specificity and clinical utility of prognostic models. ³⁸

There exists a clinical trade‐off between the inclusion of longitudinal information in the dataset and the accuracy of the prediction models. While an increased amount of longitudinal data as input enhances the model's prognostic capability, it also necessitates a longer waiting period for patients to receive their prognosis. Additionally, as more follow‐up information becomes available, clinicians can make more accurate predictions independently, thereby diminishing the added value of such models. The study also demonstrates that prognostic models may not perform as well with H&Y scores compared to models trained to predict changes in UPDRS scores, as was done in many previous studies. This may be due, in part, to the fact that the H&Y score does not capture subtle changes in patients due to its gross scale (integer values 1–5). One recent study by Jiang et al. did, however, demonstrate high accuracy (0.835) when classifying the H&Y score (0, 1, ≥2) 4 years after diagnosis where features from ¹²³I‐Ioflupane SPECT imaging had significant prognostic power. ²¹ This study included both healthy control (n = 38) and PD patients (n = 123). However, authors restricted the number of predicted classes to three: H&Y = 0 (normal), H&Y = 1 (mild PD) and H&Y ≥ 2 (moderate/late‐stage PD). The work done in the current study predicted a greater number of classes encompassing the full range of H&Y scores: 0–5. Future studies could be aimed at predicting fewer H&Y classes using this real‐world dataset. In preliminary analyses, image‐only models were also explored, excluding H&Y scores and other clinical variables, but these did not achieve adequate predictive accuracy. This likely reflects the difficulty of predicting the full 0–5 H&Y scale from baseline imaging alone.

The PPMI and real‐world datasets represent two different clinical cohorts. Patients recruited into the PPMI trial received ¹²³I‐Ioflupane SPECT at or near the time of diagnosis (and follow‐up time points) whereas patients in this real‐world dataset received imaging at various points in their clinical progression, not necessarily at the time of diagnosis. In this analysis, “baseline” was defined as the time of imaging. Therefore, because images were acquired at different clinical time points in each dataset, there was a clear difference in the baseline and five‐year H&Y scores between the patient groups (see Fig. 4 and Table S4, S5 and S6). A greater number of patients showed progression of PD – a higher H&Y score at 5 years compared to baseline—in the real‐world dataset than the PPMI dataset. Rather than viewing this difference solely as a confounder, we consider it reflective of the contrast between a highly selected research cohort and routine clinical practice, and therefore important for assessing model performance across a broader spectrum of disease severity. The dataset‐specific analyses presented in the Supplementary Material (Table S7 and S8) further help to contextualize the effect of cohort composition by showing model performance separately and in combination.

The information collected in the two studies also differed significantly, which posed a challenge in combining the datasets. For example, diagnosis categories differed between the two datasets. Therefore, having an abnormal ¹²³I‐Ioflupane SPECT image was used as inclusion criteria rather than diagnosis. As a result, the Real‐World dataset included cases with atypical PD (see Table S3), which could impact results since those patients may progress differently. A manual process of selecting concordant features from both datasets was used, which resulted in a relatively small number of variables fed into the prediction model. The small number of features may also have limited the robustness and accuracy of the models tested. As a future step, work can be done to quantify qualitative variables to increase the number of features, and amount of clinically relevant information, used in the model.

A potential limitation of the study was the fact that multiple readers worked to review and enter data into the REDCap database, which could produce variability due to differences of interpretation of doctor's unstructured notes. However, the use of the REDCap database itself reduced the potential for bias since it requires methodical entry of data and standardized definitions. Additionally, entries were reviewed by a single expert to correct any errors. Therefore, we assume that any bias resulting from multiple observers is negligible. One additional challenge of this work is the relative class imbalance. The majority of patients either did not have an increase in H&Y score over 5 years or only increased by one (see Table S4, S5 and S6). Therefore, testing anomaly detection methods could help identify patients who will progress more rapidly since there is a large class imbalance between those (in these datasets) who progress “quickly” (ΔHY >1, n = 77) and those who remain stable or progress slowly.

This initial study demonstrated that there are limitations using machine learning models to predict the H&Y scores for PD patients 5 years post ¹²³I‐Ioflupane SPECT imaging. For this task, a real‐world dataset, comprising imaging and clinical follow‐up variables over 5 years, was combined with the PPMI dataset by manually selecting overlapping features. Imaging and clinical follow‐up features did not provide significant improvement over using the H&Y scores from early timepoints alone. The results demonstrated that the principal contribution of this study is not the identification of novel predictors or immediate clinical utility, but rather the evaluation of prognostic modeling in a real‐world clinical dataset that is less curated and more heterogeneous than commonly used research cohorts. In this setting, the modest predictive performance highlights an important translational gap between model development in controlled datasets and implementation in routine clinical practice, underscoring the need for more robust and clinically adaptable approaches. Additionally, this real‐world PD dataset represents a unique and rich set of information which, in future studies, could help facilitate the translation of previous PD prediction models to a real‐world setting.

Author Roles

(1) Research Project: A. Conception, B. Organization, C. Execution; (2) Statistical Analysis: A. Design, B. Execution, C. Review and Critique; (3) Manuscript Preparation: A. Writing of the First Draft, B. Review and Critique.

X.Y.: 1C, 2A, 2B, 2C, 3B.

M.N.J.: 1C, 2C, 3B.

B.L.: 1C, 2A, 2B, 2C, 3B.

M.P.H.: 1C, 3B.

M.Z.S.: 1C, 3B.

R.V.M.: 1C, 3B.

J.W.: 1A, 1B, 2A, 2C, 3B.

M.L.: 1B, 2A, 2C, 3B.

A.L.: 1A, 1B, 1C, 2A, 2C, 3B.

R.M.: 1B, 2A, 2B, 2C, 3A, 3B.

Ethics Statement

The retrospective collection of patient data was approved by the Danish Patient Security Authority (31‐1521‐278) and the Capital Region Data Protection (P‐2020‐372). Informed patient consent was not necessary for this work since it was a retrospective registry study. We confirm that we have read the Journal's position on issues involved in ethical publication and affirm that this work is consistent with those guidelines.

Financial Disclosures and Conflicts of Interest

Author disclosures are available in the Supporting Information.

Supporting information

Figure S1. Figure illustrating methods for interpolating the real‐world dataset onto the PPMI schedule of follow‐up visits.

TABLE S1. Description and range (or categorical values) of all clinical features used in prediction models. Ranges are shown for the combined dataset, including patients from the PPMI and real‐world datasets.

TABLE S2. Scheduled PPMI visit numbers and time points. BL = baseline, V = visit, m = month.

TABLE S3. Table of clinical subtypes included in the real world and PPMI datasets.

TABLE S4. Merged dataset (PPMI + Real‐world) H&Y scores at baseline and year five.

TABLE S5. PPMI dataset H&Y scores at baseline and year five.

TABLE S6. Real‐world dataset H&Y scores at baseline and year five.

TABLE S7. Results (R ²: mean (stdev)) of the regression model with the (a) real‐world dataset alone, (b) the PPMI dataset alone and (c) the merged dataset tested using 6 months, 1 year and 2 years of clinical follow‐up information. Clinical = clinical follow‐up information (including H&Y scores). IF = imaging features from baseline ¹²³I‐Ioflupane SPECT imaging. H&Y only = H&Y scores at regular time points.

TABLE S8. Results (accuracy: mean (stdev)) of the classification model with the (a) real‐world dataset alone, (b) the PPMI dataset alone and (c) the merged dataset tested using 6 months, 1 year and 2 years of clinical follow‐up information. Clinical = clinical follow‐up information (including H&Y scores). IFs = imaging features from baseline ¹²³I‐Ioflupane SPECT imaging. H&Y only = H&Y scores at regular time points.

Supporting Information: COI_disclosure.

Data Availability Statement

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.