Comparing machine learning and deep learning models to predict cognition progression in Parkinson's disease

Area of Science:

• Neuroscience
• Artificial Intelligence in Medicine
• Computational Biology

Background:

• Cognitive decline in Parkinson's disease (PD) is highly variable, necessitating accurate predictive models.
• Existing probabilistic models for cognitive progression in PD have limitations, and deep learning approaches remain understudied.
• Early and accurate prediction of cognitive status changes is crucial for managing PD progression.

Purpose of the Study:

• To compare the efficacy of traditional sequential models against deep learning techniques for predicting cognitive progression in individuals with and without PD.
• To evaluate the performance of shallow Markov, deep recurrent (LSTM), and nonrecurrent (TFT) models in forecasting cognitive status transitions.

Main Methods:

• Utilized data from the Parkinson's Progression Marker Initiative (PPMI) database, including clinical, demographic, and cognitive assessment data.
• Compared shallow Markov, Long Short-Term Memory (LSTM), and Temporal Fusion Transformer (TFT) models for annual prediction of cognitive status (Normal Cognition, Mild Cognitive Impairment, Dementia) up to three years.
• Employed an ensemble method combining model outputs and evaluated performance using inverse probability weighted (IPW-) F1 scores.

Main Results:

• The Temporal Fusion Transformer (TFT) model demonstrated superior predictive performance (IPW-F1 = 0.468) compared to Markov (0.349) and LSTM (0.414) models.
• An ensemble approach integrating Markov, LSTM, and TFT models further enhanced prediction accuracy (IPW-F1 = 0.502).
• TFT models showed particular strength in predicting rarer cognitive states like Mild Cognitive Impairment (IPW-F1 = 0.496) and Dementia (IPW-F1 = 0.533).

Conclusions:

• Sequential deep learning models, particularly the TFT, excel at predicting clinically significant cognitive transitions in Parkinson's disease.
• The ability of TFT to handle long-term dependencies and complex data makes it highly effective for degenerative condition prediction.
• Further research into deep learning sequential models is warranted for predicting cognitive changes in neurodegenerative diseases.