Structural ICD-10 coding gaps and economic burden of underrecognized chronic conditions: a machine learning analysis of MEPS 2019–2022 and CDC PLACES

Objective To quantify the economic burden of chronic and invisible illness using federal open data, document structural ICD-10 coding gaps, and evaluate machine learning (ML) performance for individual-level healthcare expenditure prediction. Materials and methods We analyzed four consecutive MEPS waves (2019–2022; n = 17,909–22,552 adults/year) with the 2024 CDC PLACES dataset (45 states). Outcomes included total annual expenditure, out-of-pocket costs, and work-loss days. ICD-10-CM codes in MEPS Medical Conditions files were searched for invisible illness coverage. XGBoost and Random Forest predicted log-transformed expenditure from 14 features; ordinary least squares (OLS) served as baseline. Geographic burden was scored using nine composite CDC PLACES indicators. Results Adults with ≥1 chronic condition incurred 3.3 times higher mean annual expenditures and 80% more work-loss days. Autoimmune patients faced the highest costs ($24,148–$26,630/year): 3.0–3.7× higher than adults without that specific condition, and 5.8–7.6× higher than adults with no chronic condition at all. Because the MEPS public-use file reports ICD-10-CM diagnoses only as 3-character categories, we re-ran the search at that granularity: Lyme disease remained at zero records across all four years, and long COVID in every year its code existed, but the categories containing fibromyalgia and ME-CFS were not uniformly zero and cannot be attributed to these two conditions specifically. XGBoost achieved R 2 = 0.30–0.32 under cross-validation, a 13–17% improvement over OLS; activity limitation was the dominant predictor. Southern states ranked highest on composite burden; disability and uninsured rates were positively correlated across the 43 states with complete indicator data (r = 0.37, p = 0.014). Discussion Lyme disease and long COVID show a genuine structural coding gap in MEPS; fibromyalgia and ME-CFS could not be confirmed present or absent, because the file merges their codes with unrelated diagnoses. ML improves expenditure prediction but cannot compensate for upstream data gaps. High disease burden co-occurring with low insurance coverage in Southern states demands targeted policy responses. Conclusion Federal open data fail to capture Lyme disease and long COVID; their coding granularity cannot resolve fibromyalgia or ME-CFS from broader categories. Reforms in ICD coding, survey design, and data linkage are needed to make these populations visible to health policy and service planning.