Prodromal clinical, demographic, and socio-ecological correlates of asthma in adults: a 10-year statewide big data multi-domain analysis

Abstract

Objectives: To identify prodromal correlates of asthma as compared to chronic obstructive pulmonary disease and allied-conditions (COPDAC) using a multi domain analysis of socio-ecological, clinical, and demographic domains.

Methods: This is a retrospective case-risk-control study using data from Florida’s statewide Healthcare Cost and Utilization Project (HCUP). Patients were grouped into three groups: asthma, COPDAC (without asthma), and neither asthma nor COPDAC. To identify socio-ecological, clinical, demographic, and clinical predictors of asthma and COPDAC, we used univariate analysis, feature ranking by bootstrapped information gain ratio, multivariable logistic regression with LogitBoost selection, decision trees, and random forests.

Results: A total of 141,729 patients met inclusion criteria, of whom 56,052 were diagnosed with asthma, 85,677 with COPDAC, and 84,737 with neither asthma nor COPDAC. The multi-domain approach proved superior in distinguishing asthma versus COPDAC and non-asthma/non-COPDAC controls (area under the curve (AUROC) 84%). The best domain to distinguish asthma from COPDAC without controls was prior clinical diagnoses (AUROC 82%). Ranking variables from all the domains found the most important predictors for the asthma versus COPDAC and controls were primarily socio-ecological variables, while for asthma versus COPDAC without controls, demographic and clinical variables such as age, CCI, and prior clinical diagnoses, scored better.

Conclusions: In this large statewide study using a machine learning approach, we found that a multi-domain approach with demographics, clinical, and socio-ecological variables best predicted an asthma diagnosis. Future work should focus on integrating machine learning-generated predictive models into clinical practice to improve early detection of those common respiratory diseases.

Keywords:

• Asthma
• COPD
• prediction
• multi-domain
• machine learning

Additional information

Funding

This work was supported by the grant UL1TR001427 and UL1TR002389 from NIH-NCATS, by the University of Florida (UF) One Health Center, and by UF “Creating the Healthiest Generation” Moonshot initiative, which is supported by the UF Office of the Provost, UF Office of Research, UF Health, UF College of Medicine and UF Clinical and Translational Science Institute. In addition, JF is supported by NIH-NCATS under University of Florida Clinical and Translational Science Awards KL2TR001429 and UL1TR001427.