Enhancing Medical Question Answering with LSTM-Based Recurrent Neural Networks and Integrated Multi-Task Learning

EasyChair Preprint 15908

Abstract

The purpose of this study is to examine the use of Long-Short Term Memory (LSTM)-based Recurrent Neural Networks (RNNs) for the short text conversation problem—specifically targeted to generate solutions to medical questions. User-generated questions that have been answered by professionals form the training data bases—which are sourced from various online services. WebMD, HealthTap and iCliniq are some of the online services that provide data on which models are trained and evaluated before an optimal dataset is selected. However, neural machine translation models served as the foundation for the models created for this task—with extensions that included transfer learning and multi-task learning. Every model adheres to the encoder-decoder paradigm. In this sense, an encoder creates a latent vector representation of a question, and after being trained end-to-end, it initializes the state of a decoder to generate an answer for that question. A model architecture that accepts a binary input and controls two "modes" for its decoder RNN is one theory. While the latter trains its decoder to generate replies to questions encoded with answers—the former trains it on generic medical/health related text using a "language-model" mode. Therefore, this study proposed model architecture that integrates the work of question category classification with the goal of answer production. In this case, the network uses the encoder's final state to classify the query category and provides the decoder with further input in the form of anticipated class. Finally, a unique RNN-based language system trained on general medical/health-related literature has been taught to support these suggested models during interpretation by aggregating their probabilities at every time-step to enhance the accuracy of the resulting response.

Keyphrases: AI/ML, Artificial Intelligence, CNN, Digital Healthcare, Natural Language Processing, Neural Machine Translation, RNN, character embeddings, deep learning, health related wikipedia articles, healthtech, language model and answer, machine learning, multi-task learning, neural networks

@booklet{EasyChair:15908,
  author    = {Dedeepya Sai Gondi and Veera Venkata Raghunath Indugu and Hemanth Volikatla and Vamsi Krishna Reddy Bandaru},
  title     = {Enhancing Medical Question Answering with LSTM-Based Recurrent Neural Networks and Integrated Multi-Task Learning},
  howpublished = {EasyChair Preprint 15908},
  year      = {EasyChair, 2025}}