Scholarlink

<p>The goal of the proposed work is to develop a medical predictive engine based on an enhanced LLaMA framework for accurate disease prediction and drug recommendation. However, the existing systems are primarily focused on retrieving and summarising answers from the users’ queries, but they fail to diagnose diseases and provide appropriate recommendations. Therefore, a need to develop a highly efficient model that gives contextual answers and accurate recommendations, minimising the computational overhead and memory consumption. The proposed model relies on User Queries, Health records, and Medical Datasets (MedQuAD dataset, iCliniq Dataset, and UCI ML Drug dataset) that are pre-processed by removing irrelevant data, repeated and null data. Then the tokenisation gives a meaningful token to each word, and the Bidirectional Encoder Representations from Transformers (BERT) and XLNet-based transformer models are used in the word embedding layer, which finds the contextual relationship of each token. The Position encoding of the Large Language Model Meta AI (LLaMA) model has the null-initialised attention with the self-attention layer and a low-dimensional adaptation self-attention framework in the self-attention layer. This feature helps to address the aforementioned limitations. In addition to that, the performance will be affected due to the random initialisation of prompts; therefore, the zero initialisation adopts a zero-gating factor within the attention for stable training in the early training stages. And the low-dimensional adaptation self-attention stimulates the performance of the LLaMA model. The Bert and XlNet-based embedded layer enhance the learning performance of the LLaMA for learning contextual information. The performance of the proposed method is fine-tuned and improved by using an adaptive Sculptor Optimisation Algorithm. In the proposed work, the users’/doctors’ questions are evaluated with the context in the Feature Vector to give an appropriate answer. The implementation is performed using the Python platform. Experimental results show that the proposed work yields the highest accuracy of 0.978 for Dataset 1, an accuracy of 0.987 for Dataset 2, and an accuracy of 0.98 for Dataset 3, respectively.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Dual transformer-based LLaMA framework for efficient disease prediction and drug recommendation in healthcare

  • 期刊论文

  • 发表日期：

    2026年2月17日

查看全文
  • Amit Shrirang Phadtare,
  • Narayan Kulkarni,
  • Manoj Devare

摘要

The goal of the proposed work is to develop a medical predictive engine based on an enhanced LLaMA framework for accurate disease prediction and drug recommendation. However, the existing systems are primarily focused on retrieving and summarising answers from the users’ queries, but they fail to diagnose diseases and provide appropriate recommendations. Therefore, a need to develop a highly efficient model that gives contextual answers and accurate recommendations, minimising the computational overhead and memory consumption. The proposed model relies on User Queries, Health records, and Medical Datasets (MedQuAD dataset, iCliniq Dataset, and UCI ML Drug dataset) that are pre-processed by removing irrelevant data, repeated and null data. Then the tokenisation gives a meaningful token to each word, and the Bidirectional Encoder Representations from Transformers (BERT) and XLNet-based transformer models are used in the word embedding layer, which finds the contextual relationship of each token. The Position encoding of the Large Language Model Meta AI (LLaMA) model has the null-initialised attention with the self-attention layer and a low-dimensional adaptation self-attention framework in the self-attention layer. This feature helps to address the aforementioned limitations. In addition to that, the performance will be affected due to the random initialisation of prompts; therefore, the zero initialisation adopts a zero-gating factor within the attention for stable training in the early training stages. And the low-dimensional adaptation self-attention stimulates the performance of the LLaMA model. The Bert and XlNet-based embedded layer enhance the learning performance of the LLaMA for learning contextual information. The performance of the proposed method is fine-tuned and improved by using an adaptive Sculptor Optimisation Algorithm. In the proposed work, the users’/doctors’ questions are evaluated with the context in the Feature Vector to give an appropriate answer. The implementation is performed using the Python platform. Experimental results show that the proposed work yields the highest accuracy of 0.978 for Dataset 1, an accuracy of 0.987 for Dataset 2, and an accuracy of 0.98 for Dataset 3, respectively.