Introduction: Approximately 80% of modifiable factors of population health outcomes are attributed to Social determinants of health (SDOH). Among individuals with diabetes, SDOH factors are crucial in shaping both physical and psychosocial outcomes. This study aims to examine the significance of SDOH factors on diabetes outcomes among adults participating in a quality improvement (QI) program.

Methods: Adults with elevated HbA1c levels (>8.5%) were enrolled in the QI program. Participants were categorized as responders (HbA1c reduction ≥ 0.5%) or non-responders (HbA1c reduction < 0.5%) at follow-up. A total of 41 predictive features were identified from LexisNexis socioeconomic health attributes. ANOVA and Fisher's Exact tests were employed to compare responders versus non-responders. Three machine learning techniques yielded the most efficient features for inclusion in the predictive model.

Results: The study included 475 individuals from the diabetes program; (mean) age 55.5 years, pre-intervention HbA1c (9.0%), post-intervention HbA1c (8.3%), and change in HbA1c (-0.72%). Significant differences between groups were observed in terms of crime and burglary indices and address stability (p < 0.05). Final selected features included household characteristics (age and income) and neighborhood-level attributes (income, home values, and crime). Logistic regression demonstrated improved accuracy (0.55), while Support Vector Regression recorded the highest Mean Squared Error (MSE) of 0.32.

Discussion: A clinically significant decrease in HbA1c is closely associated with patients’ SDOH factors, collectively accounting for over 30% of intervention outcomes. Future interventions must include social support navigators to identify and address patients’ neighborhood stressors and socioeconomic barriers that may impact intervention outcomes.

Missing data are common in large-scale health and biomedical studies. Existing methods for handling missing data for high-dimensional data are either slow or difficult to implement by applied researchers and practitioners. In a national study conducted in 2020-2023, a higher percentage of missing data was encountered duo to the fact that surveys were administered during the COVID-19 era. The first round of the survey responses collected from 761 behavioral health and criminal justice practitioners from 504 counties across the U.S, with over 2000 recorded variables, revealed the presence of different types and degrees of missing data. Using this data, a new methodology is developed and tested by carefully combining statistical methodology and machine learning (ML) computing power, and compared to multiple imputation and other traditional missing data methods.

Supervised ML techniques, including tree-based and trained random forest models, are used for tackling missing data by selecting and using available data features to estimate imputed values. Cross Validation is used to simultaneously select the appropriate prediction model, train the model, and validate the accuracy of the of the predicted imputed values. An additional layer of clustering is added to these ML-based methods to obtain accurate predictions of missing values which will be useful to biomedical/healthcare researchers encountering missing data in their studies. Accounting for different characteristics of the data, this strategy accurately imputes missing data in a hierarchical setting and provides a toolkit for public health researchers who need a robust and easy to use tool for accurately predicting/imputing missing values.

Background

Small Area Estimation (SAE) is a cost-effective methodology that combines multiple data sources to enhance the survey estimator for small geographic areas or subpopulations. As machine learning (ML) continues to gain momentum in data science, more ML applications is seen in the SAE landscape. The objective of this study was to compare classic generalized linear mixed models (GLMMs) with ML algorithms for SAE of county-level obesity prevalence in Mississippi.

Methods

The 2022 Mississippi Behavioral Risk Factor Surveillance System (BRFSS) data were obtained for this study. The 2020 US census data at county level is incorporated as auxiliary data to “borrow strength.” GLMMs were constructed and validated using the survey packages in R to account for weights and YRBS complex sample design; Tree-based ML, and Neural network models are trained and validated using the IBM SPSS Modeler v18.4.0.

Results

The Mississippi county-level obesity prevalence estimated by GLMMs was validated by a three-fold cross-validation approach. An adequate range of variation among counties and satisfactory precision demonstrated by standard error are observed with the GLMM estimates. Gradient-boosted decision trees (GBDTs) generate point estimates similar to those of GLMMs, with slightly wider confidence intervals. Neural network models appear to be less accurate and less robust among the three techniques applied.

Conclusions

The classic GLMMs are among the top choices as a SAE tool. IBM SPSS Modeler makes it intuitive and easy to train and validate GBDTs, which makes it a suitable choice for less experienced ones to start SAE.

Objectives: To develop a novel compartmental mathematical model (SEITIRD) based on the existing SEIR model to predict the rate of emergency department (ED) visits of detected cases of emerging and reemerging infectious diseases (ERID), using COVID-19 and influenza as case studies for model prediction.

Biostatistics, economics Epidemiology Protection of the public in relation to communicable diseases including prevention or control Public health or related research