Using NLP and Neural Networks to Autonomously Identify Severe Asthma and Determine Study Eligibility in a Large Healthcare System

Asthma is a heterogeneous disease. The heterogeneity of asthma is supported by clinical observations and genome wide association studies (GWASs) that have identified over 200 asthma susceptibility loci in the DNA. These genetic 'hot spots' are near inflammatory cytokines, growth factors, and other inflammatory proteins knowingly linked to airway inflammation, including cytokines IL-4, -5, -13, -25, -33, and TSLP.

Novel monoclonal antibody therapies have drastically changed the treatment of moderate-to-severe asthma. Novel monoclonal antibody therapies introduced in the last 7 years have greatly advanced treatment options for moderate-to-severe asthma patients. These therapies effectively reduce or eliminate severe exacerbations, prevent hospitalizations, and improve patients' quality of life. However, many severe asthma patients, particularly those living in underserved areas, are still being overtreated with steroids and undertreated with monoclonal antibodies.

The 21st Century Cures Act will Change the Landscape of Research. The 21st Century Cures Act reinforced the use of real-world data (RWD) and real-world evidence (RWE) to support clinical trials, aid in drug coverage decisions, develop national treatment guidelines as well as standardized decision support tools. An underutilized source of RWE/D are electronic health records (EHR). Machine Learning (ML), AI, and natural language processing (NLP) are developing technologies that will greatly advance our ability to leverage data in EHR systems.

The study aims to use new technologies (ML, AI, NLP), to autonomously identify moderate to severe asthma populations within an EHR system, describe differences in treatment patterns across different populations, and determine trial eligibility.

Primary Objectives Please ensure you detail primary objectives Aim 1. Determine and validate a diagnosis of severe asthma (SA) using predictive features obtained from the Scripps Health EHR.

• Aim 1a: Use ML applied to structured EHR data to predict SA. Use the opinion of 2 specialty-trained physicians and ATS guidelines to determine model accuracy.
• Aim 1b: Use NLP applied to unstructured text to predict SA. Determine model accuracy as above in Aim 1a.
• Aim 1c: Use a combination of ML applied to structured data to predict SA. Determine model accuracy as above in Aim 1a.

Inclusion criteria

• Demographics: Males ~ 40%, Blacks ~ 5-10%, Hispanic ~15-30%, Urban ~80-90%

Exclusion criteria

• None