An innovative supervised longitudinal learning procedure of recurrent neural networks with temporal data...

Longitudinal data in health informatics studies often present challenges due to sparse observations from each subject, limiting the application of contemporary deep learning for prediction. This issue is particularly relevant in predicting birthweight, a crucial factor in identifying conditions such as macrosomia and large-for-gestational age (LGA). Previous approaches have relied on empirical formulas for estimated fetal weights (EFWs) from ultrasound measurements and mixed-effects models for interim predictions.

The proposed novel supervised longitudinal learning procedure features a three-step approach. First, EFWs are generated using empirical formulas from ultrasound measurements. Second, nonlinear mixed-effects models are applied to create augmented sequences of EFWs, spanning daily gestational timepoints. This augmentation transforms sparse longitudinal data into a dense parallel sequence suitable for training recurrent neural networks (RNNs). A tailored RNN architecture is then devised to incorporate the augmented sequential EFWs along with non-sequential maternal characteristics.

The RNNs are trained on augmented data to predict birthweights, which are further classified for macrosomia and LGA. Application of this supervised longitudinal learning procedure to the Successive Small-for-Gestational-Age Births study yields improved performance in classification metrics. Specifically, sensitivity, area under the receiver operation characteristic curve, and Youden's Index demonstrate enhanced results, indicating the effectiveness of the proposed approach in overcoming sparsity challenges in longitudinal health informatics data.

The integration of mixed-effects models for temporal data augmentation and RNNs on augmented sequences shows effective in accurately predicting birthweights, particularly in the context of identifying excessive fetal growth conditions.