Neighborhood-Level Non-Medical Factors Alone Poorly Predict Proxy PASC Phenotype : Limited Performance of Logistic...

Early in the SARS-CoV-2 pandemic, acute mortality and hospitalization showed marked disparities across socially disadvantaged groups. In some cases, individuals who initially tested positive for Post-Acute Sequelae of SARS-CoV-2 infection (PASC) also began to experience lingering symptoms after initially recovering from SARS-CoV-2. At the time of the 2023 BRFSS survey the CDC defined PASC as self-report of any symptoms lasting ≥3 months at the time of interview that were not present before having COVID-19. An estimated 6.7% of Delaware adults self-reported symptoms consistent with PASC and 24.5% of those Delawareans reported significant activity limitations. This condition is a chronic disease and the Delaware State Health Improvement Plan (SHIP) has established a priority outcome to reduce the risk of chronic diseases. One of the goals of SHIP is to promote a combination of methods to support and identify chronic diseases and support healthy outcomes, these goals tie directly to non-medical health factors. Although there are notable poverty, food insecurity and health care access disparities at the county level in Delaware, it remains unclear whether neighborhood-level non-medical health factors alone can identify areas with high PASC prevalence so those the most in need can receive targeted resources to address this chronic condition.

To evaluate whether combined ZCTA-level non-medical health factors, without patient-level clinical predictors, can discriminate against a machine-learning-derived PASC phenotype.

We defined a proxy PASC outcome via a validated ML-derived probability score. established by Pfaff et al. using N3C data. This model used a variety of data points from individuals who sought care at PASC clinics and assigned each individual a score that estimated the probability of a PASC diagnosis. Our cohort was run through this same model to identify individuals with PASC. We analyzed 4,521,889 adults with SARS-CoV-2 in N3C. PASC status was assigned using the probability model developed by Pfaff et al., which incorporates clinical attendance, symptoms, comorbidities, and healthcare utilization, with a cutoff ≥ 0.90. The predictors included neighborhood-level race-ethnicity, income, socioeconomic status, education, and insurance variables; composite risk scenarios were evaluated using logistic regression and XGBoost.

Logistic regression performed at chance levels (AUC ~ 0.50). XGBoost showed limited test-set discrimination (AUC 0.685 - 0.689), but cross-validated performance was weak (CV AUC 0.516-0.523) and severity for PASC was poor (recall 0.05-0.06).

Neighborhood-level non-medical health factors alone were insufficient to discriminate the proxy PASC phenotype at the individual level. Considering the state of Delaware's documented county level disparities in poverty and healthcare access, future predictive models should integrate clinical characteristics, comorbidity burden, and biomarkers at the individual level. This method would more accurately identify those at risk of developing PASC and support equitable resource allocation in the communities that are most impacted by this chronic disease.