Young Jin Jeong

Parkinson’s disease (PD) shows heterogeneity in motor progression, complicating prognostic assessment. Dopamine transporter (DAT) imaging provides an in vivo measure of nigrostriatal dopaminergic integrity, whereas wearable gait assessment quantifies motor performance. However, the prognostic value of baseline (BL) DAT-derived measures and wearable gait characteristics remains incompletely understood. We investigated whether BL DAT-derived measures and a limb-focused wearable turning movement-dynamics signature (turning-derived signature) were associated with subsequent motor progression in PD. Twenty-three individuals with PD underwent BL PET imaging, wearable gait assessment, and clinical follow-up. Motor progression was defined using a cohort-specific exploratory threshold of ΔUnified Parkinson’s Disease Rating Scale (UPDRS) Part III ≥ 1.5 points, derived within this cohort using Hoehn and Yahr (H&Y)-stage worsening as the clinical anchor. Twelve participants were classified as motor progressors and 11 as non-progressors. The turning signature yielded an AUC of 0.750, while DAT-model AUCs ranged from 0.462 to 0.500. The clinical model yielded an AUC of 0.962; adding the turning signature yielded an AUC of 0.871. These findings suggest that the BL turning-derived signature contains an exploratory prognostic signal but does not establish incremental value beyond clinical severity. The progression outcome was derived within this small cohort, and external validation is required.

Background: Parkinson's disease (PD) and multiple system atrophy (MSA) are α‐synucleinopathies with overlapping autonomic and sleep dysfunction, but distinct trajectories. Objectives: This study compared sleep and autonomic profiles in PD and MSA and elucidated disease‐specific associations using structural equation modeling (SEM). Methods: We retrospectively analyzed drug‐naïve patients with PD and MSA presenting with sleep complaints. All participants underwent polysomnography (PSG), autonomic function tests, and cardiac [I‐123] metaiodobenzylguanidine (MIBG) scintigraphy. Significant parameters with moderate correlations (Spearman's ρ ≥ 0.35) were included in SEM to examine domain relationships. Results: A total of 136 PD and 67 MSA patients were included. Patients with PD were older, with comparable sex, disease duration, and diabetes mellitus frequency. In PSG, patients with MSA showed lower‐stage N3 sleep, higher apnea–hypopnea index (35.4 vs. 23.2 events/h), and lower minimum oxygen saturation. Autonomic testing revealed that the MSA group demonstrated greater orthostatic blood pressure drops and higher total composite autonomic severity scores. Cardiac MIBG uptake was significantly reduced in the PD group. A directional SEM model in MSA revealed a significant association between autonomic dysfunction and increased sleep disruption ( β = −0.68, P = 0.006), with substantially greater explanatory power than observed in PD ( R ² = 0.54 vs. 0.29). Conclusions: Both autonomic and sleep parameters were significantly more impaired in patients with MSA than in those with PD. SEM further indicated that these domains were closely linked in MSA, with greater autonomic dysfunction associated with poorer sleep quality. This highlights the need for integrated management of both domains.

Background: Parkinson disease (PD) presents diagnostic challenges due to its heterogeneous motor and nonmotor manifestations. Traditional machine learning (ML) approaches have been evaluated on structured clinical variables. However, the diagnostic utility of large language models (LLMs) using natural language representations of structured clinical data remains underexplored. Objective: This study aimed to evaluate the diagnostic classification performance of multiple LLMs using natural language prompts derived from structured clinical data and to compare their performance with traditional ML baselines. Methods: We reformatted structured clinical variables from the Parkinson’s Progression Markers Initiative (PPMI) dataset into natural language prompts and used them as inputs for several LLMs. Variables with high multicollinearity were removed, and the top 10 features were selected using Shapley additive explanations (SHAP)–based feature ranking. LLM performance was examined across few-shot prompting, dual-output prompting that additionally generated post hoc explanatory text as an exploratory component, and supervised fine-tuning. Logistic regression (LR) and support vector machine (SVM) classifiers served as ML baselines. Model performance was evaluated using F1-scores on both the test set and a temporally independent validation set (temporal validation set) of limited size, and repeated output generation was carried out to assess stability. Results: On the test set of 122 participants, LR and SVM trained on the 10 SHAP-selected clinical variables each achieved a macro-averaged F1-score of 0.960 (accuracy 0.975). LLMs receiving natural language prompts derived from the same variables reached comparable performance, with the best few-shot configurations achieving macro-averaged F1-scores of 0.987 (accuracy 0.992). In the temporal validation set of 31 participants, LR maintained a macro-averaged F1-score of 0.903, whereas SVM showed substantial performance degradation. In contrast, multiple LLMs sustained high diagnostic performance, reaching macro-averaged F1-scores up to 0.968 and high recall for PD. Repeated output generation across LLM conditions produced generally stable predictions, with rare variability observed across runs. Under dual-output prompting, diagnostic performance showed a reduction relative to few-shot prompting while remaining generally stable. Supervised fine-tuning of lightweight models improved stability and enabled GPT-4o-mini to achieve a macro-averaged F1-score of 0.987 on the test set, with uniformly correct predictions observed in the small temporal validation set, which should be interpreted cautiously given the limited sample size and exploratory nature of the evaluation. Conclusions: This study provides an exploratory benchmark of how modern LLMs process structured clinical variables in natural language form. While several models achieved diagnostic performance comparable to LR across both the test and temporal validation datasets, their outputs were sensitive to prompting formats, model choice, and class distributions. Occasional variability across repeated output generations reflected the stochastic nature of LLMs, and lightweight models required supervised fine-tuning for stable generalization. These findings highlight the capabilities and limitations of current LLMs in handling tabular clinical information and underscore the need for cautious application and further investigation.

BACKGROUND Parkinson disease (PD) presents diagnostic challenges due to its heterogeneous motor and nonmotor manifestations. Traditional machine learning (ML) approaches have been evaluated on structured clinical variables. However, the diagnostic utility of large language models (LLMs) using natural language representations of structured clinical data remains underexplored. OBJECTIVE This study aimed to evaluate the diagnostic classification performance of multiple LLMs using natural language prompts derived from structured clinical data and to compare their performance with traditional ML baselines. METHODS We reformatted structured clinical variables from the Parkinson’s Progression Markers Initiative (PPMI) dataset into natural language prompts and used them as inputs for several LLMs. Variables with high multicollinearity were removed, and the top 10 features were selected using Shapley additive explanations (SHAP)–based feature ranking. LLM performance was examined across few-shot prompting, dual-output prompting that additionally generated post hoc explanatory text as an exploratory component, and supervised fine-tuning. Logistic regression (LR) and support vector machine (SVM) classifiers served as ML baselines. Model performance was evaluated using F1-scores on both the test set and a temporally independent validation set (temporal validation set) of limited size, and repeated output generation was carried out to assess stability. RESULTS On the test set of 122 participants, LR and SVM trained on the 10 SHAP-selected clinical variables each achieved a macro-averaged F1-score of 0.960 (accuracy 0.975). LLMs receiving natural language prompts derived from the same variables reached comparable performance, with the best few-shot configurations achieving macro-averaged F1-scores of 0.987 (accuracy 0.992). In the temporal validation set of 31 participants, LR maintained a macro-averaged F1-score of 0.903, whereas SVM showed substantial performance degradation. In contrast, multiple LLMs sustained high diagnostic performance, reaching macro-averaged F1-scores up to 0.968 and high recall for PD. Repeated output generation across LLM conditions produced generally stable predictions, with rare variability observed across runs. Under dual-output prompting, diagnostic performance showed a reduction relative to few-shot prompting while remaining generally stable. Supervised fine-tuning of lightweight models improved stability and enabled GPT-4o-mini to achieve a macro-averaged F1-score of 0.987 on the test set, with uniformly correct predictions observed in the small temporal validation set, which should be interpreted cautiously given the limited sample size and exploratory nature of the evaluation. CONCLUSIONS This study provides an exploratory benchmark of how modern LLMs process structured clinical variables in natural language form. While several models achieved diagnostic performance comparable to LR across both the test and temporal validation datasets, their outputs were sensitive to prompting formats, model choice, and class distributions. Occasional variability across repeated output generations reflected the stochastic nature of LLMs, and lightweight models required supervised fine-tuning for stable generalization. These findings highlight the capabilities and limitations of current LLMs in handling tabular clinical information and underscore the need for cautious application and further investigation.

In amyloid brain PET, after parcellation using the finite element method (FEM)-based algorithm FreeSurfer and voxel-based algorithm PMOD, SUVr examples can be extracted and compared. This study presents the classification SUVr threshold in PET images of F-18 florbetaben (FBB), F-18 flutemetamol (FMM), and F-18 florapronol (FPN) and compares and analyzes the classification performance according to computational algorithm in each brain region. PET images were co-registered after the generated MRI was registered with standard template information. Using MATLAB script, SUVr was calculated using the built-in parcellation number labeled in the brain region. PMOD and FreeSurfer with different algorithms were used to load the PET image, and after registration in MRI, it was normalized to the MRI template. The volume and SUVr of the individual gray matter space region were calculated using an automated anatomical labeling atlas. The SUVr values of eight regions of the frontal cortex (FC), lateral temporal cortex (LTC), mesial temporal cortex (MTC), parietal cortex (PC), occipital cortex (OC), anterior and posterior cingulate cortex (GCA, GCP), and composite were calculated. After calculating the correlation of SUVr using the FreeSurfer and PMOD algorithms and calculating the AUC for amyloid-positive/negative subjects, the classification ability was calculated, and the SVUr threshold was calculated using the Youden index. The correlation coefficients of FreeSurfer and PMOD SUVr calculations of the eight regions of the brain cortex were FBB (0.95), FMM (0.94), and FPN (0.91). The SUVr threshold was SUVr(LTC,min) = 1.264 and SUVr(THA,max) = 1.725 when calculated using FPN-FreeSurfer, and SUVr(MTC,min) = 1.093 and SUVr(MCT,max) = 1.564 when calculated using FPN-PMOD. The AUC comparison showed that there was no statistically significant difference (p > 0.05) in the SUVr classification results using the three radiopharmaceuticals, specifically for the LTC and OC regions in the PMOD analysis, and the LTC and PC regions in the FreeSurfer analysis. The SUVr calculation using PMOD (voxel-based algorithm) has a strong correlation with the calculation using FreeSurfer (FEM-based algorithm); therefore, they complement each other. Quantitative classification analysis with high accuracy is possible using the suggested SUVr threshold. The SUVr classification performance was good in the order of FMM, FBB, and FPN, and showed a good classification performance in the LTC region regardless of the type of radiotracer and analysis algorithm.

Alzheimer’s disease (AD), the most prevalent degenerative brain disease associated with dementia, requires early diagnosis to alleviate worsening of symptoms through appropriate management and treatment. Recent studies on AD stage classification are increasingly using multimodal data. However, few studies have applied graph neural networks to multimodal data comprising F-18 florbetaben (FBB) amyloid brain positron emission tomography (PET) images and clinical indicators. The objective of this study was to demonstrate the effectiveness of graph convolutional network (GCN) for AD stage classification using multimodal data, specifically FBB PET images and clinical indicators, collected from Dong-A University Hospital (DAUH) and Alzheimer’s Disease Neuroimaging Initiative (ADNI). The effectiveness of GCN was demonstrated through comparisons with the support vector machine, random forest, and multilayer perceptron across four classification tasks (normal control (NC) vs. AD, NC vs. mild cognitive impairment (MCI), MCI vs. AD, and NC vs. MCI vs. AD). As input, all models received the same combined feature vectors, created by concatenating the PET imaging feature vectors extracted by the 3D dense convolutional network and non-imaging feature vectors consisting of clinical indicators using multimodal feature fusion method. An adjacency matrix for the population graph was constructed using cosine similarity or the Euclidean distance between subjects’ PET imaging feature vectors and/or non-imaging feature vectors. The usage ratio of these different modal data and edge assignment threshold were tuned by setting them as hyperparameters. In this study, GCN-CS-com and GCN-ED-com were the GCN models that received the adjacency matrix constructed using cosine similarity (CS) and the Euclidean distance (ED) between the subjects’ PET imaging feature vectors and non-imaging feature vectors, respectively. In modified nested cross validation, GCN-CS-com and GCN-ED-com respectively achieved average test accuracies of 98.40%, 94.58%, 94.01%, 82.63% and 99.68%, 93.82%, 93.88%, 90.43% for the four aforementioned classification tasks using DAUH dataset, outperforming the other models. Furthermore, GCN-CS-com and GCN-ED-com respectively achieved average test accuracies of 76.16% and 90.11% for NC vs. MCI vs. AD classification using ADNI dataset, outperforming the other models. These results demonstrate that GCN could be an effective model for AD stage classification using multimodal data.

In amyloid brain PET, after parcellation using the finite element method (FEM)-based algorithm FreeSurfer and voxel-based algorithm PMOD, SUVr examples can be extracted and compared. This study presents the classification SUVr threshold in PET images of F-18 florbetaben (FBB), F-18 flutemetamol (FMM), and F-18 florapronol (FPN) and compares and analyzes the classifica-tion performance according to computational algorithm in each brain region. PET images were co-registered after the generated MRI was registered with standard template information. Using MATLAB script, SUVr was calculated using the built-in parcellation number labeled in the brain region. PMOD and FreeSurfer with different algorithms were used to load the PET image, and after registration in MRI, it was normalized to the MRI template. The volume and SUVr of the individual gray matter space region were calculated using an automated anatomical labeling at-las. The SUVr values of eight regions of the frontal cortex (FC), lateral temporal cortex (LTC), mesial temporal cortex (MTC), parietal cortex (PC), occipital cortex (OC), anterior and posterior cingulate cortex (GCA, GCP), and composite were calculated. After calculating the correlation of SUVr using the FreeSurfer and PMOD algorithms and calculating the AUC, the classification ability was calculated, and the SVUr threshold was calculated using the Youden index. The cor-relation coefficients of FreeSurfer and PMOD SUVr calculations of the eight regions of the brain cortex were FBB (0.95), FMM (0.94), and FPN (0.91). The SUVr threshold was SUVr(LTC,min) = 1.264 and SUVr(THA,max) = 1.725 when calculated using FPN-FreeSurfer, and SUVr(MTC,min) = 1.093 and SUVr(MCT,max) = 1.564 when calculated using FPN-PMOD. The AUC comparison showed that there was no statistically significant difference (p>0.05) in SUVr classification results using the three radiopharmaceuticals, specifically for LTC and OC regions in PMOD analysis, and LTC and PC regions in FreeSurfer analysis. The SUVr calculation using PMOD (voxel-based algorithm) has a strong correlation with the calculation using FreeSurfer (FEM-based algorithm); therefore, they are complement with each other. Quantitative classification analysis with high accuracy is possible using the suggested SUVr threshold. The SUVr classification performance was good in the order of FMM, FBB, and FPN and showed good classification performance in the LTC region, regardless of the type of radiotracer and analysis algorithm.

This study presents a deep learning-based restoration method for low-quality amyloid positron emission tomography (PET) images acquired in a short period, which can be generalized across multiple domains. Each of these domains consists of low-quality amyloid PET images acquired in the same environment. Owing to variations in image characteristics, such as contrast, across different acquisition environments, the restoration performance of the deep learning methods can significantly degrade when applied to PET images obtained from unseen domains (i.e., not seen in training). To address the difficulty, we introduce a mapping label and condition the network on this label. This enables the network that takes a low-quality amyloid PET image and the corresponding mapping label as inputs to effectively generate the desired high-quality amyloid PET image. We assign the mapping label as a one-hot vector for each domain and use pairs of PET images from short (2 min) and standard (20 min) scanning times for training. The network, trained with the mapping label, can efficiently restore low-quality amyloid PET images in unseen domains by estimating an unknown mapping label for the unseen domain. We demonstrate the effectiveness of the proposed method through quantitative and qualitative analyses on the several datasets.

Patients with mild cognitive impairment (MCI) have a relatively high risk of developing Alzheimer’s dementia (AD), so early identification of the risk for AD conversion can lessen the socioeconomic burden. In this study, ¹⁸F-Florapronol, newly developed in Korea, was used for qualitative and quantitative analyses to assess amyloid positivity. We also investigated the clinical predictors of the progression from MCI to dementia over 2 years. From December 2019 to December 2022, 50 patients with MCI were recruited at a single center, and 34 patients were included finally. Based on visual analysis, 13 (38.2%) of 34 participants were amyloid-positive, and 12 (35.3%) were positive by quantitative analysis. Moreover, 6 of 34 participants (17.6%) converted to dementia after a 2-year follow-up (p = 0.173). Among the 15 participants who were positive for amyloid in the posterior cingulate region, 5 (33.3%) patients developed dementia (p = 0.066). The Clinical Dementia Rating-Sum of Boxes (CDR-SOB) at baseline was significantly associated with AD conversion in multivariate Cox regression analyses (p = 0.043). In conclusion, these results suggest that amyloid positivity in the posterior cingulate region and higher CDR-SOB scores at baseline can be useful predictors of AD conversion in patients with MCI.

Objective: Amyloid positron emission tomography (PET) with F-18 florbetaben (FBB), F-18 flutemetamol (FMM), and F-18 florapronol (FPN) is being used clinically for the evaluation of dementia. These radiopharmaceuticals are commonly used to evaluate the accumulation of beta-amyloid plaques in the brain, but there are structural differences between them. We investigated whether there are any differences in the imaging characteristics. Methods: A total of 605 subjects were enrolled retrospectively in this study, including healthy subjects (HS) and patients with mild cognitive impairment or Alzheimer’s disease. Participants underwent amyloid PET imaging using one of the three radiopharmaceuticals. The PET images were analyzed visually and semi-quantitatively using a standardized uptake value ratio (SUVR). In addition, we calculated and compared the cut-off SUVR of the representative regions for each radiopharmaceutical that can distinguish between positive and negative scans. Results: In the negative images of the HS group, the contrast between the white matter and the gray matter was high in the FMM PET images, while striatal uptake was relatively higher in the FPN PET images. The SUVR showed significant differences across the radiopharmaceuticals in all areas except the temporal lobe, but the range of differences was relatively small. Accuracy levels for the global cut-off SUVR to discriminate between positive and negative images were highest in FMM PET, with a value of 0.989. FBB PET also showed a high value of 0.978, while FPN PET showed a relatively low value of 0.901. Conclusions: Negative amyloid PET images using the three radiopharmaceuticals showed visually and quantitatively similar imaging characteristics except in the striatum. Binary classification using the cut-off of the global cortex showed high accuracy overall, although there were some differences between the three PET images.