Recursive forecasting for vaccination demand with hybrid feature LSTM considering heterogeneous policies and risk perception

Abstract

Vaccination demand changes rapidly during the pandemic, as the public health policy evolves and the pandemic develops among different countries and regions. Preprocessing the original multi-source daily data frames, we characterize the temporal utilities of heterogeneous vaccination policies and the epidemic risk perception, which lead to a boost to our recursive predictive practice. Beyond leveraging these two primary predictors, factors of non-pharmaceutical interventions (NPIs) and fundamental temporal dynamics are fully used to make fine-grained recursive predictions for vaccination demand. We conduct a series of experiments on real-world vaccination-related datasets among different countries and regions and the experimental results show that the proposed hybrid feature recursive model gets almost 94% predictive accuracy in mainland China, which gains an evident advantage over ones fed with features excluding policy-related and perception-related predictors with 7.42% improvement rate (IR) in terms of accuracy. It is far better than the baseline methods with 31.84% absolute IR on average and gets 92.57% average accuracy among different regions. The tailored Policy-specific Risk-aware Long Short-Term Memory (PsRa-LSTM) approach outperforms the state-of-the-art competing benchmark methods in terms of valid evaluation metrics including RMSE, MAE, and MAPE. Utilizing local and global machine learning model explainers, we carry out innovative quantitative contribution analysis to capture reasonable causality between preceding predictors and subsequent prediction targets. Policymakers can optimize the allocation and schedule of supplies and personnel with kinds of vaccines, formulate public health policy including sequential booster vaccination, or provide guidance to public opinion through media campaigns for vaccination.