Random forest prediction of Alzheimer's disease using pairwise selection from time series data

Area of Science:

• Neuroscience
• Biostatistics
• Machine Learning

Background:

• Alzheimer's disease research generates time-dependent data with missing and irregular samples.
• Standard time series methods require regularly sampled data, necessitating pre-processing.
• Accurate prediction of Alzheimer's disease progression is crucial for patient management.

Purpose of the Study:

• To develop and evaluate a novel random forest method for analyzing irregularly sampled time-dependent data in Alzheimer's disease studies.
• To predict key Alzheimer's disease progression metrics including diagnosis, ADAS-13 scores, and normalized ventricles volume.
• To compare the performance of the proposed method against a benchmark Support Vector Machine (SVM) predictor.

Main Methods:

• Utilized a random forest model to learn relationships between data points at varying time intervals.
• Input vectors incorporated time series history summaries, demographic, and genetic data.
• Employed data from the TADPOLE grand challenge and the Alzheimer's Disease Neuroimaging Initiative (ADNI) for validation.

Main Results:

• Achieved a mean Area Under the Curve (mAUC) of 0.82 for diagnosis prediction.
• Obtained a Balanced Classification Accuracy (BCA) of 0.73 for diagnosis.
• Outperformed a benchmark SVM predictor which yielded an mAUC of 0.62 and BCA of 0.52.

Conclusions:

• The developed random forest method is effective for analyzing irregular time-series data in Alzheimer's disease research.
• The approach demonstrates superior performance compared to traditional methods like SVM for disease progression forecasting.
• This method offers a promising tool for predicting Alzheimer's disease evolution using complex, real-world datasets.