Uncertainty aware multimodal model

Abstract

This study explores how deep learning can help match patients with Irritable Bowel Syndrome (IBS) based on their profile and metabolite data. A new model that combines a Variational Autoencoder (VAE) with a Long Short-Term Memory (LSTM) network is proposed, designed to learn underlying patterns in the data while also capturing uncertainty in its predictions. To see how well it performs, it is compared it to two adapted baseline models — UISNet and TeMiNet — using a range of evaluation metrics. Overall, UISNet came out on top, showing the fastest convergence, lowest prediction errors, and the best ability to explain the data. TeMiNet trained quickly but didn’t generalize well to new data. Our VAE-LSTM model showed stable and consistent learning throughout, but its accuracy and ability to capture variance were limited — likely due to underfitting and the use of fixed hyperparameters. Although the model is designed to estimate uncertainty, this study didn’t evaluate how accurate or meaningful those estimates were. That said, the VAE-LSTM still shows promise. With some fine-tuning, better uncertainty evaluation, and more thorough testing, it has the potential to support more personalized approaches to IBS treatment and patient matching in the future.