Chronological Diagnostic Algorithm Predicting Neuropathology in Parkinsonism

Area of Science:

• Neurology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Pre-mortem diagnosis of parkinsonism is challenging due to diverse presentations and overlapping conditions.
• Accurate neuropathological diagnosis is crucial for effective treatment and research.

Purpose of the Study:

• To develop and validate a machine learning algorithm for predicting parkinsonism neuropathology.
• To utilize chronological clinical data for improved diagnostic accuracy.

Main Methods:

• Automated abstraction of clinical data from medical records using Generative Pre-trained Transformer 4 (GPT-4) models.
• Training six machine learning models on patient data, including age, sex, family history, and 197 clinical presentations.
• Predicting nine neuropathologic diagnoses, including Lewy body disease (LBD), Alzheimer's disease (AD), progressive supranuclear palsy (PSP), multiple system atrophy (MSA), corticobasal degeneration (CBD), and frontotemporal lobar degeneration (FTLD).

Main Results:

• The CatBoost algorithm achieved an area under the receiver operating characteristic curve (AUC) of 0.83 for predicting neuropathology three years post-onset.
• Key predictors included age at onset, restricted eye movement, and tremor.
• The model demonstrated robustness with incomplete data, achieving an AUC of 0.80 using only 23 of 200 parameters.

Conclusions:

• The developed algorithm serves as a cost-effective and interpretable screening tool for parkinsonism.
• This tool can aid in bridging biomarker testing and the development of molecular-targeted therapies.
• The algorithm provides diagnostic probabilities and visualizations, facilitating clinical decision-making.