Using predictive modeling approaches to examine the relative impact of clinical and public health factors in predicting disease

Background: With the growth of open data, data sharing, and data integration occurring with the concomitant advances in data mining, the opportunity to merge and integrate disparate data sources becomes more feasible and beneficial. Clinical data, aligned with community-based metrics and environmental factors, can generate a large, multidimensional data source to test hypotheses about the relative importance and impact of these different health factors.

Objective/purpose: This presentation will share the systematic approach used by the authors to prepare and align disparate data sources, generate predictive analytics models, down-select and identify the most relevant models for comparing the impact of clinical and public health factors, and determine the relative importance of various factors identified by different predictive modeling tools.

 Methods: The authors used various database, data manipulation, and statistical analysis tools to gather, process, and analyze data from open source and clinical depositories. Cross-validation and training/testing protocols helped substantiate the reliability of the outcomes for prediction.

Results: The use of various types of regression methods as well as other machine learning toolsets provided insight on which clinical measures are helpful in predicting specific disease conditions. In addition, using a model comparison approach, community health conditions and reported public health metrics were shown to provide a significant impact beyond the effects of clinical indicators.

Discussion/Conclusions: Using data mining approaches on merged datasets have helped document the value of examining more than clinical data: community and public health metrics can provide an enhanced understanding of complex, multifactorial disease processes and help guide public health policy and intervention planning.