Scholarlink

This study presents a unique system designed to predict diseases through a system capable of functioning offline, making it accessible in areas without reliable Internet access due to terrain, physical, or economic factors, such as rural, remote, or low-income regions. The system is intended to serve as a preliminary diagnostic tool before professional medical assistance can be accessed. The dataset, developed with the assistance of a medical professional, focuses on symptoms that users can observe and describe themselves without requiring clinical tests or complex medical terminology. Various deep learning models, including bidirectional encoder representations from transformers (BERTs), feedforward neural networks (FNNs), long short-term memory networks (LSTMNs), recurrent neural networks (RNNs), and convolutional neural networks (CNNs), were compared to determine which model is best suited for this application. Notably, FNNs and BERTs produced nearly identical results, with BERTs reaching a final accuracy of 99.00% and a validation accuracy of 98.21%, while FNNs achieved a slightly higher final accuracy of 99.02% and a validation accuracy of 98.14%. FNNs allow the application to be approximately 500 megabytes in size, compared to the 3 gigabytes required for a BERT-based solution, making FNNs a more practical and accessible option for the target audience. This research highlights the potential of using simpler models for effective, low-resource artificial intelligence-driven healthcare solutions, particularly in underserved areas.

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

Comparative Analysis of Deep Learning Techniques for Disease Prediction Using a Symptom-Based Dataset

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Ankush Thalla,
  • S. Sathish Kumar,
  • G. Arun Sampaul Thomas,
  • Beulah J. Karthikeyan

摘要

This study presents a unique system designed to predict diseases through a system capable of functioning offline, making it accessible in areas without reliable Internet access due to terrain, physical, or economic factors, such as rural, remote, or low-income regions. The system is intended to serve as a preliminary diagnostic tool before professional medical assistance can be accessed. The dataset, developed with the assistance of a medical professional, focuses on symptoms that users can observe and describe themselves without requiring clinical tests or complex medical terminology. Various deep learning models, including bidirectional encoder representations from transformers (BERTs), feedforward neural networks (FNNs), long short-term memory networks (LSTMNs), recurrent neural networks (RNNs), and convolutional neural networks (CNNs), were compared to determine which model is best suited for this application. Notably, FNNs and BERTs produced nearly identical results, with BERTs reaching a final accuracy of 99.00% and a validation accuracy of 98.21%, while FNNs achieved a slightly higher final accuracy of 99.02% and a validation accuracy of 98.14%. FNNs allow the application to be approximately 500 megabytes in size, compared to the 3 gigabytes required for a BERT-based solution, making FNNs a more practical and accessible option for the target audience. This research highlights the potential of using simpler models for effective, low-resource artificial intelligence-driven healthcare solutions, particularly in underserved areas.