Expediting primary immunodeficiency diagnosis with artificial intelligence

Published: October 12, 2022
 
Many individuals with primary immunodeficiency (PI) suffer prolonged and costly diagnostic odysseys, excessive morbidity and impaired quality of life due to delayed diagnosis. While newborn screening has revolutionized early detection for patients with severe T cell dysfunction, the majority of PI patients will manifest concerns stochastically outside of the newborn period. Some contributing factors for delayed diagnosis include the variable nature of disease presentation, lack of a universally informative biomarker and incomplete provider awareness about PI. Given these challenges, novel methods for disease screening are needed. Artificial/augmented intelligence (AI) approaches offer one such solution to systematically ascertain risk for underlying PI given their proven ability to analyze healthcare data and make predictions at the patient level.

In a recent article in The Journal of Allergy and Clinical Immunology (JACI), Rider et al. describe a validated AI-based method for population-wide PI risk assessment. Their approach leverages a 2-step analytical pipeline featuring the Jeffrey Modell Foundation’s Software for Primary Immunodeficiency Intervention Recognition and Tracking (SPIRIT) Analyzer coupled to a custom machine learning (ML) algorithm. The study was conducted within the Texas Children’s Hospital Health Plan (TCHP) on a cohort of nearly 430,000 individuals who were prospectively assessed at 6-month intervals for a period of 30 months. The investigators characterized every member of the cohort as having “high”, “medium-low” or “no clear risk” at baseline and at each 6-month assessment. All high-risk subjects underwent further health-record scrutiny and analysis with the ML algorithm to determine likelihood of underlying PI and for prioritizing referrals to clinical immunologists.
 
The investigators note that 1036 individuals (0.2% of total cohort) were given a new PI diagnosis during the study period, with a significantly greater proportion of those coming from the initial high-risk group. They also found that subjects identified as high-risk ultimately had more complex PI diagnoses and higher hospitalization rates than those in the medium-low and no-risk groups. In addition, this AI platform was able to predict which high-risk individuals were most likely to benefit from referral. The algorithmic approach, coupled with human oversight, allowed for the distillation of 427,110 individuals down to 36 (0.008%) who ultimately warranted referral with high likelihood for underlying PI.

Importantly, this 2-step AI approach is generalizable and can be embedded into a care coordination workflow to provide a systematic means by which any health system can quantify PI risk at the individual level. The authors note that SPIRIT Analyzer is freely available, and their ML code and models are open-source for use without restriction. The findings and approach outlined by Rider et al. point the way towards an interoperable methodology for calculating ongoing PI risk for each and every member of a population.  Their AI platform is expected to facilitate early PI diagnosis while also driving healthcare value for patients with an existing diagnosis.

The Journal of Allergy and Clinical Immunology (JACI) is an official scientific journal of the AAAAI, and is the most-cited journal in the field of allergy and clinical immunology.

Full Article