Globally local: Hyper-local modeling for accurate forecast of COVID-19.

IMPORTANCE: Assumption of a well-mixed population during modeling is often erroneously made without due analysis of its validity. Ignoring the importance of the geo-spatial granularity at which the data is collected could have significant implications on the quality of forecasts and the actionable clinical recommendations that are based on it. OBJECTIVE: This paper's primary objective is to test the hypothesis that the characteristic dynamics defining the trajectory of the pandemic in a region is lost when the data is aggregated and modeled at higher geo-spatial levels. DESIGN: We use publicly available confirmed SARS-CoV-2 cases and deaths from January 1st, 2020 to August 3rd, 2020 in the United States at different geo-spatial granularities to conduct our experiments. To understand the impact of this hypothesis, the output of this study was implemented in Tampa General Hospital (TGH) to provide resource demand forecast. RESULTS: The Mean Absolute Percentage Error (MAPE) in the forecast confirmed cases can be 30% higher for modeling at the state-level than aggregating model results at the scale of counties or clusters of counties. Similarly, modeling at a state-level and crafting policy decisions based on them may not be effective - county-level forecasts made by partitioning state-level forecasts are 3x worse for confirmed cases and 20x worse for deaths relative to the same model at the county level. By leveraging these results, TGH was able to accurately allocate clinical resources to tackle COVID-19 cases, continue elective surgical procedures largely uninterrupted and avoid costly construction of overflow capacity in the first two epidemic waves. CONCLUSIONS AND RELEVANCE: Accurate forecasting at the county level requires hyper-local modeling with county resolution. State-level modeling does not accurately predict community spread in smaller sub-regions because state populations are not well mixed, resulting in large prediction errors. Actionable decisions such as deciding whether to cancel planned surgeries or construct overflow capacity require models with local specificity.

Use of Stratified Cascade Learning to predict hospitalization risk with only socioeconomic factors.

BACKGROUND AND OBJECTIVE: Published models predicting health related outcomes rely on clinical, claims and social determinants of health (SDH) data. Addressing the challenge of predicting with only SDH we developed a novel framework termed Stratified Cascade Learning (SCL) and used it for predicting the risk of hospitalization (ROH). MATERIALS AND METHODS: The variable set includes 27 SDH and "age" and "sex" for a cohort of diabetic patients. The SCL model uses three sub-models: SM1 (whole training set) stratifies training set into "predictable" and "unpredictable" subsets, SM2 (built on whole training set) classifies test set patients into "predictable" and "unpredictable", and SM3 (built on only the "predictable" subset) predicts the ROH for the patients classified as "predictable" by SM2. RESULTS: The SCL model does not improve either the AUC or the NPV of the basic classifier, but materially improves accuracy and specificity measures at the expense of lowering sensitivity for the "predictable" subset. Optimization of the risk thresholds of the sub-models does not noticeably change the AUC and NPV but further improves the accuracy and specificity at the expense of further lowering sensitivity. CONCLUSION: Since the SLC model yields low sensitivity it fails to predict high risk patients. But it yields high specificity that can be useful when the objective is to eliminate low-risk patients as candidates for further testing or treatment. The use of the SCL is not limited to healthcare, it can be applied to any predictive modeling problem when reliable predictions can only be made for a fraction of incoming data.