# A data-driven SSM/PCA analysis approach for differential diagnosis of parkinsonism using 11C-PE2I PET

## TL;DR

This paper introduces a new method using PET data and machine learning to improve the accuracy of diagnosing Parkinson's-related disorders.

## Contribution

The novel approach combines ensemble-based SSM/PCA with dynamic 11C-PE2I PET data for robust differential diagnosis of parkinsonism.

## Key findings

- Combining SBR and R1 metrics achieved 90% balanced accuracy in classifying parkinsonian disorders.
- Ensemble-SSM/PCA improved stability and accuracy compared to single-reference methods.
- SBR primarily differentiates patients from controls, while R1 helps distinguish between PD, DLB, and PSP.

## Abstract

•Dynamic 11C-PE2I SSM/PCA using R1 and SBR for differential diagnosis of parkinsonism.•Single-reference SSM/PCA can be unstable in clinical datasets with uncertain labels.•Ensemble-SSM/PCA improves robustness through repeated reference sampling.•High balanced accuracy achieved on an independent hold-out test set.

Dynamic 11C-PE2I SSM/PCA using R1 and SBR for differential diagnosis of parkinsonism.

Single-reference SSM/PCA can be unstable in clinical datasets with uncertain labels.

Ensemble-SSM/PCA improves robustness through repeated reference sampling.

High balanced accuracy achieved on an independent hold-out test set.

Scaled Subprofile Modelling using principal component analysis (SSM/PCA) is a multivariate analysis technique primarily used in 18F-FDG PET brain studies to produce disease-specific patterns (DPs) and scalar scores aiding neurological diagnosis. SSM/PCA relies on well-characterized reference groups, posing challenges in real-world clinical datasets where diagnoses may be uncertain. A data-driven ensemble approach may offer a more robust alternative to random sampling when reference groups are unavailable.

To apply SSM/PCA to dynamic 11C-PE2I-PET data for differential diagnosis of parkinsonism using a Monte Carlo cross-validation-inspired framework with ensemble prediction.

Dopamine transporter availability, expressed as the specific binding ratio (SBR) relative to cerebellar gray matter and relative cerebral blood flow (R1) images from 47 healthy controls and 316 patients who underwent dynamic11C-PE2I-PET on a Discovery MI PET/CT scanner were included. Patients had a single most probable diagnosis of Parkinson’s disease (PD), dementia with Lewy bodies (DLB), or progressive supranuclear palsy (PSP) based on clinical information and the PET reading. A stratified 80/20 training/testing split was applied, repeated across 100 seeds, to generate DPs used for training ensemble classification models. Classification accuracy was assessed on the test-set.

Combining SBR and R1 improved accuracy yielding a balanced accuracy of 90%, with SBR primarily differentiating patients from healthy controls and R1 for differentiating between PD, DLB and PSP.

Our results highlight the potential of an ensemble-based SSM/PCA method to assist differential diagnosis of parkinsonism. Future work will focus on including additional atypical parkinsonian disorders.