Binay Adhikari - Forecasting the number of primary care visits for respiratory symptoms using multiple health administrative data and spatial machine learning approaches

Background: COVID-19 has affected public health, including changes to circulation of other respiratory viruses. Accurate forecasting methods are needed to predict infections at a regional level1. Given the contagious nature of respiratory illness, within region forecasting can be improved using data from surrounding regions. Many current methods ignore spatial correlation and can be improved by using geospatial machine learning. Utilizing modern deep learning methods such as Graph Neural Networks can have the potential to combine different data sources accounting for their spatial and temporal relationships. This approach can forecast surges in primary care visits due to respiratory-related illnesses and enhance the surveillance system. Methods: In this study, we aim to enhance our current machine learning models that use LHA-level MSP data for forecasting respiratory symptom related primary care visits using additional data from the lab test results data hosted in the PANDA-PAWS platform. Using multiple data sources will enhance the predictive power of the models by incorporating different pathways by which individuals with a viral respiratory illness may encounter healthcare. We use the Consolidated Local Health Areas (CLHA) as the spatial unit of analysis to make the forecasting amenable for primary care planning. We will fit a series of machine learning models (Lasso, Random Forest, Light Gradient Boosted Method [LGBM], CatBoost, XGBoost), and compare with modern deep learning graph neural networks for time-series on the combined data and assess their predictive ability. We implement a blocked time series approach for cross-validation and hyper-parameter tuning and withhold 30% of the data for testing. The spatial dependence is accounted for using an inverse-distance-weighted matrix and contiguous adjacency matrix to create spatial lagged features. Model performance is assessed using the Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE) and model predictors are assessed using SHapley Additive exPlanations (SHAP) values.  Results: Our initial results which make use of MSP data for 19+ years age-group show 1% RMSE with XGBoost & CatBoost on out of sample COVID data, 1.2% on Flu, 2.4% on Pneumonia and Flu & 1.5% on Acute Respiratory Infections. Across most of our models and diseases, the population size and the first geospatial lag rank amongst the top features in SHAP plots consistently. Conclusions:   Machine learning can provide improved prediction primary care visits for respiratory symptoms in local health areas by incorporating geospatial epidemiological data. These models show improved performance over more traditional regression and time-series approaches. By incorporating multiple data sources with modern deep learning methods, we can directly incorporate the spatio-temporal properties of the data and provide better forecast of primary care visits for respiratory symptoms for timely intervention.  Data used: British Columbia Ministry of Health [creator]. Medical Services Plan (MSP) Payment Information File. British Columbia Ministry of Health [publisher]. Data Extract. MOH (2020). 2021.https://www2.gov.bc.ca/gov/content/health/health-forms/online-services. British Columbia Ministry of Health [creator]. Client Roster (Client Registry System/Enterprise Master Patient Index). British Columbia Ministry of Health [publisher]. Data Extract. MOH (2020). 2021. https://www2.gov.bc.ca/gov/content/health/health-forms/online-services 

Mediasite Presenter

10/29/2024 9:20:00 PM