Kernel-Weighted Deep Neural Networks for Sparse and Irregular Longitudinal Prediction

Thomas M. Menino Convention & Exhibition Center 

We propose a kernel-weighted deep neural network (K-DNN) for predicting an individual's future conditional mean response trajectory from a history of predictors in sparse and irregularly measured longitudinal data. To accommodate subject-specific and irregular observation times, we first reconstruct each subject's predictor trajectories nonparametrically, yielding predictor values on a common grid over the historical time window. The proposed architecture has two stages. In Stage 1, early layers form time-specific local subnetworks that learn representations of predictor information at each grid time point. In Stage 2, representations from multiple grid times are integrated in subsequent layers to capture complex temporal dependence and interactions across the history. To further mitigate the impact of data sparsity, we incorporate a kernel-weighting mechanism that prioritizes observations measured in close proximity to the target future time point. This hybrid approach combines the flexibility of deep learning with the local smoothing advantages of kernel methods, providing a robust predictive tool for complex longitudinal profiles.

Deep neural networks

kernel weighting

Longitudinal data analysis 

Section on Nonparametric Statistics