A comparative study of machine learning methods for predicting the evolution of brain connectivity from a baseline...

Area of Science:

• Neuroscience
• Computational Biology
• Machine Learning

Background:

• Predicting brain network (connectome) evolution by forecasting connectivity weight changes aids in early detection of neurological disorders.
• Forecasting connectomic anomalies and neurological disorders is an under-explored area in predictive connectomics.
• While machine learning (ML) excels in computer vision, ML methods for predicting brain connectivity evolution from a single timepoint are scarce.

Purpose of the Study:

• To address the gap in predictive connectomics, this study organized a Kaggle competition.
• The competition aimed to develop advanced machine learning pipelines for predicting brain connectivity evolution.
• The goal was to forecast brain connectivity maps at a future timepoint (t1) using baseline data (t0).

Main Methods:

• Twenty teams developed ML pipelines combining data pre-processing, dimensionality reduction, and learning algorithms.
• The longitudinal OASIS-2 dataset was utilized for training and evaluating the ML models.
• Model generalizability and scalability were assessed using random data splits and 5-fold cross-validation.

Main Results:

• Methods were ranked using mean absolute error (MAE) and Pearson Correlation Coefficient (PCC).
• Performance was evaluated across different data perturbation strategies, including single random splits and cross-validation.
• A rank product approach determined the final ranking, with statistical significance values provided for each pipeline.

Conclusions:

• Twenty ML pipelines and the connectomic dataset are publicly available on GitHub to support open science.
• The competition's outcomes are expected to drive further development of predictive models for brain connectivity.
• These advancements may also inform the prediction of other network types, such as genetic networks.