Objective: The objective of this research is to employ an Artificial Intelligence (AI) approaches to identify patient status and disease states in neurodegenerative diseases (NDDs), such as Parkinson’s Disease (PD), by analyzing cerebrospinal fluid (CSF) proteomics data.

Background: PDs are typically assessed using subjective clinician scores, Unified Parkinson’s Disease Rating Scale (UPDRS), which are prone to biases and variability across different physicians. This is mainly due to variance in how clinicians interpret patient symptoms, which may also be expressed differently by patients, and in how scores are assigned, potentially leading to inaccurate representations of the patient’s true condition [1]. Furthermore, existed efforts in employing machine learning algorithms that analyze these scores often inherit these biases, leading to unreliable patient assessments.

Method: To address this challenge, this study analyzes six years of CSF proteomics data from seven patients diagnosed with PD. Various encoding models were employed, including dense,  Long Short-Term Memory (LSTM), and graph autoencoders autoencoders, after representing patient data in both tabular and graphical forms [figure1]. Then, patient strata and disease states were identified by clustering the created embeddings using agglomerative clustering algorithm. In addition, insights about the key proteomic biomarkers and the alignment of the defined disease states with clinically recognized disease stages were identified by additional statistical and visualization analysis.

Results: The analysis successfully identified two distinct patient strata [figure1, figure2], one of which exhibited significant temporal variation [figure2, right side section). Further examination revealed that both groups exchanged similar disease states over time. Moreover, the most correlated proteins with each new state were identified, and the mapping of the discovered states to existing clinician-based rating scales was also achieved.

Conclusion: This study shows that combining CSF proteomics with artificial intelligence can objectively define Parkinson’s disease progression. It provides a promising alternative to traditional physician-based scores, enhancing understanding of the disease, its progression, and aiding in the development of better clinical trials and disease-modifying therapies.

figure1

figure2

References: [1] Tolosa, E., Garrido, A., Scholz, S. W., & Poewe, W. (2021). Challenges in the diagnosis of Parkinson’s disease. The Lancet. Neurology, 20(5), 385–397. https://doi.org/10.1016/S1474-4422(21)00030-2

« Back to 2025 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/artificial-intelligence-for-understanding-parkinson-disease-state-and-patient-stratification-using-csf-proteomics/