Since Tibshirani[5] first introduced the Least Absolute Shrinkage and Selection Operator (Lasso), this regularization technique has been extensively applied across diverse scientific domains. In finance and economics, Lasso has demonstrated superior predictive accuracy over traditional OLS and probit models in forecasting credit ratings and crude oil yields, particularly when handling large-scale datasets with complex technical indicators [7][8]. Beyond economics, Lasso has become a critical tool in biomedical research for high-dimensional feature selection. For instance, researchers have utilized Lasso to identify influential genetic markers and clinical variables linked to Alzheimer’s disease and schizophrenia, effectively resolving multicollinearity and overfitting issues in clinical datasets[9]. These applications underscore Lasso’s capacity to extract a parsimonious set of significant predictors from redundant information.

The Mixed Data Sampling (MIDAS) framework, introduced by Ghysels[6], provides a robust methodology for integrating data sampled at different frequencies, such as daily and monthly observations. In energy and macroeconomics, MIDAS has been shown to significantly improve forecasting performance for crude oil prices and GDP growth by internalizing high-frequency market signals[10] . To address the “curse of dimensionality” in multi-source datasets, recent studies have integrated regularization techniques into the MIDAS structure. Notably, Marsilli[4] developed the Lasso-MIDAS model, demonstrating its efficacy in identifying key economic drivers from numerous quarterly indicators. Subsequent extensions, such as MIDASSO (integrating Elastic Net) and Group-Lasso U-MIDAS, have further enhanced the precision of risk assessments and default predictions, proving that penalized MIDAS models can effectively manage both frequency mismatch and high-dimensional noise[11][12] .

In summary, while Lasso regression and MIDAS models have been extensively validated in financial and macroeconomic forecasting, their integrated application within the public health domain—specifically for acute cerebrovascular events—remains limited. Existing epidemiological studies often rely on single-frequency data, which may lead to the omission of high-frequency environmental triggers. By bridging this methodological gap, the present study utilizes a Lasso-MIDAS framework to simultaneously manage frequency discrepancies and high-dimensional indicator selection. This approach not only enhances the precision of stroke hospitalization identification but also provides a more robust tool for analyzing the complex interactions between multi-source environmental and socioeconomic drivers.