Scholarlink

<p>Accurate short-term forecasts of influenza activity are critical for public health preparedness but remain challenging with classical statistical models that depend on long historical data and often fail under data scarcity. Time-series foundation models (TSFMs), pretrained on large heterogeneous datasets, may offer improved accuracy and data efficiency through transferable temporal representations. We analyzed 605 weeks of influenza surveillance data from Hong Kong (2014–2025), including consultation rates, laboratory detections, hospital admissions, severe cases, and school outbreaks. Six TSFMs (Google TimesFM, Amazon Chronos, IBM Granite Tiny Time Mixers [TTM], Timer-XL, Moirai-MoE, and UniTS) were compared with OLS, GLM, SARIMA, and empirical Bayesian models in a rolling-origin design generating one- to four-week forecasts. Forecast accuracy was evaluated using R<sup>2</sup>, RMSE, symmetric MAPE, weighted interval score (WIS) and multivariable residual diagnostics. We further estimated the minimal input window required for R<sup>2</sup> ≥ 0.50 and the maximum credible forecast horizon (R<sup>2</sup> ≥ 0.50 with significant RMSE gain vs. seasonal-naive). Chronos, TimesFM, and SARIMA achieved the best overall one-week performance (R<sup>2</sup> 0.88–0.89; RMSE 30.6–31.3), significantly outperforming OLS (R<sup>2</sup> 0.78; <i>p</i> &lt; 0.0001) and GLM (R<sup>2</sup> 0.79; <i>p</i> &lt; 0.0001). Chronos outperformed other models in two-to-four-week performance. Chronos produced credible forecasts of the multivariate influenza surveillance predictions up to three weeks (R<sup>2</sup> ≥ 0.50; <i>p</i> &lt; 0.05), whereas classical models fell below threshold beyond two weeks. TSFMs required as low as eight weeks of input data to achieve R<sup>2</sup> ≥ 0.50, compared to at least 297 weeks for classical models. TSFMs deliver comparatively more accurate, well-calibrated, and data-efficient epidemic forecasts, capturing temporal and inter-variable dependencies than classical regression models. Their ability to generate more reliable forecasts from minimal historical data compared to classical approaches supports their potential for rapid deployment in seasonal influenza surveillance and emerging pandemics such as COVID-19.</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

Multivariate influenza forecasting using time-series foundation models: a comparative study with classical statistical models using 11 years of surveillance data in Hong Kong

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月31日

查看全文
  • Rachel Yui Ki Chu,
  • Kam Chiu Szeto,
  • Pui Hong Chung

摘要

Accurate short-term forecasts of influenza activity are critical for public health preparedness but remain challenging with classical statistical models that depend on long historical data and often fail under data scarcity. Time-series foundation models (TSFMs), pretrained on large heterogeneous datasets, may offer improved accuracy and data efficiency through transferable temporal representations. We analyzed 605 weeks of influenza surveillance data from Hong Kong (2014–2025), including consultation rates, laboratory detections, hospital admissions, severe cases, and school outbreaks. Six TSFMs (Google TimesFM, Amazon Chronos, IBM Granite Tiny Time Mixers [TTM], Timer-XL, Moirai-MoE, and UniTS) were compared with OLS, GLM, SARIMA, and empirical Bayesian models in a rolling-origin design generating one- to four-week forecasts. Forecast accuracy was evaluated using R2, RMSE, symmetric MAPE, weighted interval score (WIS) and multivariable residual diagnostics. We further estimated the minimal input window required for R2 ≥ 0.50 and the maximum credible forecast horizon (R2 ≥ 0.50 with significant RMSE gain vs. seasonal-naive). Chronos, TimesFM, and SARIMA achieved the best overall one-week performance (R2 0.88–0.89; RMSE 30.6–31.3), significantly outperforming OLS (R2 0.78; p < 0.0001) and GLM (R2 0.79; p < 0.0001). Chronos outperformed other models in two-to-four-week performance. Chronos produced credible forecasts of the multivariate influenza surveillance predictions up to three weeks (R2 ≥ 0.50; p < 0.05), whereas classical models fell below threshold beyond two weeks. TSFMs required as low as eight weeks of input data to achieve R2 ≥ 0.50, compared to at least 297 weeks for classical models. TSFMs deliver comparatively more accurate, well-calibrated, and data-efficient epidemic forecasts, capturing temporal and inter-variable dependencies than classical regression models. Their ability to generate more reliable forecasts from minimal historical data compared to classical approaches supports their potential for rapid deployment in seasonal influenza surveillance and emerging pandemics such as COVID-19.