Scholarlink

<p>Infectious diseases pose ongoing threats to global health, especially in regions with limited surveillance infrastructure. Traditional monitoring systems often lack predictive capabilities and transparency, hindering timely intervention. To address these challenges, this study proposes a predictive and explainable AI framework that leverages advanced deep learning techniques for accurate disease forecasting. The framework incorporates high-dimensional feature extraction, Least Absolute Shrinkage and Selection Operator (LASSO)-based feature selection, and the state-of-the-art PatchTST model for long-sequence time-series forecasting. Ranger21, a hybrid optimizer combining RAdam, Look Ahead, and gradient centralization, is employed to ensure robust and stable training. An alternative benchmark using the Temporal Fusion Transformer (TFT) validates model reliability. The system's explainability is enhanced through SHAP analysis and attention visualisation, ensuring trust and interpretability. Evaluated on the COVID-19 Global Forecasting (Week 5) dataset and WHO FluNet influenza data, the model achieves a forecasting accuracy of 98.7%, outperforming traditional baselines. To ensure scalability, compression techniques such as knowledge distillation and quantization-aware training are applied, reducing model size by up to 62% with minimal performance loss. This makes the model deployable on edge devices and mobile platforms for real-time, localised health monitoring. Overall, the proposed solution significantly improves the accuracy, interpretability, and deployability of infectious disease forecasting systems, empowering public health authorities with actionable insights for early intervention and resource allocation.</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

Explainable long-sequence time-series forecasting for infectious diseases using optimised transformer models

  • 期刊论文

  • 发表日期：

    2025年11月24日

查看全文
  • Abdulrahman Mathkar Alotaibi

摘要

Infectious diseases pose ongoing threats to global health, especially in regions with limited surveillance infrastructure. Traditional monitoring systems often lack predictive capabilities and transparency, hindering timely intervention. To address these challenges, this study proposes a predictive and explainable AI framework that leverages advanced deep learning techniques for accurate disease forecasting. The framework incorporates high-dimensional feature extraction, Least Absolute Shrinkage and Selection Operator (LASSO)-based feature selection, and the state-of-the-art PatchTST model for long-sequence time-series forecasting. Ranger21, a hybrid optimizer combining RAdam, Look Ahead, and gradient centralization, is employed to ensure robust and stable training. An alternative benchmark using the Temporal Fusion Transformer (TFT) validates model reliability. The system's explainability is enhanced through SHAP analysis and attention visualisation, ensuring trust and interpretability. Evaluated on the COVID-19 Global Forecasting (Week 5) dataset and WHO FluNet influenza data, the model achieves a forecasting accuracy of 98.7%, outperforming traditional baselines. To ensure scalability, compression techniques such as knowledge distillation and quantization-aware training are applied, reducing model size by up to 62% with minimal performance loss. This makes the model deployable on edge devices and mobile platforms for real-time, localised health monitoring. Overall, the proposed solution significantly improves the accuracy, interpretability, and deployability of infectious disease forecasting systems, empowering public health authorities with actionable insights for early intervention and resource allocation.