Transient Intervals of Significantly Different Whole Brain Connectivity Predict Recovery vs. Progression from Mild...

The high dimensionality and complexity of time-varying measures of functional brain connectivity have created an environment in which a very rich transformation of the data remains difficult to map into disease states without some form of reduction (averaging, clustering, statistical blindness to the multivariate interactions between features that modulate their contributions). In this work, employing a recently developed architecture for long short-term memory classifiers that supports use of gradient-based model interpretability techniques, we predict progression or recovery from mild cognitive impairment (MCI) from an instantaneous (windowless) wavelet-based measure of dynamic functional network connectivity. This time-attention LSTM (TA-LSTM) model achieves 0.79 AUC on the task of predicting which MCI patients who will recover (RMCI) vs. those who will progress (PMCI) to AZD within a three-year timeframe. Using a common gradient-based model interpretation technique, saliency analysis, on this TA-LSTM points to potentially important predictive dynamic biomarkers, including the duration of the highly salient time intervals and the average connectivity patterns within these highly salient intervals.