Log scores for probabilistic forecasts across all age group, seasons, weeks and targets from Improved forecasts of influenza-associated hospitalization rates with Google Search Trends

Reliable forecasts of influenza-associated hospitalizations during seasonal outbreaks can help health systems better prepare for patient surges. Within the USA, public health surveillance systems collect and distribute near real-time weekly hospitalization rates, a key observational metric that makes real-time forecast of this outcome possible. In this paper, we describe a method to forecast hospitalization rates using a population level transmission model in combination with a data assimilation technique. Using this method, we generated retrospective forecasts of hospitalization rates for 5 age groups and the overall population during 5 seasons in the USA and quantified forecast accuracy for both near-term and seasonal targets. Additionally, we describe methods to correct for under-reporting of hospitalization rates (backcast) and to estimate hospitalization rates from publicly available online search trends data (nowcast). Forecasts based on surveillance rates alone were reasonably accurate in predicting peak hospitalization rates (within ± 25% of the actual peak rate, three weeks before peak). The error in predicting rates one to four weeks ahead, remained constant for the duration of the seasons, even during periods of increased influenza incidence. An improvement in forecast quality across all age groups, seasons and targets was observed when backcasts and nowcasts supplemented surveillance data. These results suggest that the model-inference framework can provide reasonably accurate real-time forecasts of influenza hospitalizations; backcasts and nowcasts offer a way to improve system tolerance to observational errors.

季节性流感暴发期间，流感相关住院病例的可靠预测可助力卫生系统更好地应对患者激增问题。在美国境内，公共卫生监测系统会收集并发布近乎实时的周度住院率——这是一项关键的观测指标，使得针对该结局开展实时预测成为可能。本文阐述了一种结合人群水平传播模型与数据同化（data assimilation）技术的住院率预测方法。借助该方法，我们针对美国5个流感季内5个年龄组及全体人群的住院率生成了回顾性预测，并量化了短期与季节性两类预测目标的预测精度。此外，本文还介绍了两类辅助方法：一是用于校正住院率漏报情况的反推预测（backcast）方法，二是基于公开在线搜索趋势数据估算住院率的即时预测（nowcast）方法。仅依托监测数据开展的预测在预测住院率峰值时表现出良好精度：可在峰值出现前三周将预测误差控制在实际峰值的±25%以内。即便在流感发病率攀升阶段，提前1至4周预测住院率的误差在整个流感季期间始终保持稳定。当反推预测与即时预测数据补充至监测数据集后，所有年龄组、流感季及预测目标的预测质量均得到显著提升。上述结果表明，基于模型的推断框架可对流感相关住院情况生成精度可靠的实时预测；反推预测与即时预测方法则可提升系统对观测误差的耐受能力。