{
  "abstract": "Background FIB-4 and the NAFLD Fibrosis Score were derived predominantly from European-ancestry cohorts and may perform inequitably across racial and ethnic groups. We applied machine learning with SHAP explainability to characterise race-specific fibrosis risk architectures and quantify FIB-4 performance disparities in a nationally representative elastography-validated population.Methods We analysed 14,566 adults from NHANES 2017–2023 with valid transient elastography (LSM ≥7.0 kPa = significant fibrosis). XGBoost classifiers incorporating 17 clinical and metabolic variables were trained with 5-fold cross-validation within each racial/ethnic group. SHAP values quantified feature importance per group. FIB-4 and NFS performance were compared against ML across Non-Hispanic (NH) White, NH Black, Hispanic, and NH Asian adults by AUC, calibration, and sensitivity at established thresholds (>1.3, >2.67).Results Fibrosis prevalence varied by race: 16.8% NH White, 17.7% NH Black, 16.2% Hispanic, and 9.6% NH Asian. FIB-4 demonstrated poor discrimination across all groups (AUC 0.584–0.635). Machine learning substantially outperformed FIB-4 in every group (AUC 0.733-0.826; ΔAUC +0.140 to +0.207), with the largest gain in NH Asian adults, ML AUC 0.823 versus FIB-4 AUC 0.605 ( IDDF2026-ABS-0275 Figure 1-4). At the >1.3 threshold, FIB-4 identified only 32.9% of confirmed fibrosis in NH Asian adults, and just 5.6% at the >2.67 threshold, missing 19 of every 20 cases (IDDF2026-ABS-0275 Figure 6). Post-hoc calibration demonstrated near-perfect alignment between predicted probabilities and observed fibrosis rates across all groups, while FIB-4 showed no meaningful risk stratification.SHAP analysis revealed fundamentally distinct risk architectures across groups (IDDF2026-ABS-0275 Figure 5): waist circumference and BMI dominated in NH White and Hispanic adults; age, BMI, and poverty-income ratio were differentiating drivers in NH Black adults; and insulin, HDL, and transaminase ratios, rather than BMI, drove fibrosis risk in NH Asian adults, consistent with metabolically obese normal-weight physiology.Conclusions FIB-4 performs poorly across all racial/ethnic groups, with critical underdetection in NH Asian adults harbouring insulin- and inflammation-mediated fibrosis pathways invisible to BMI-centric scoring ( IDDF2026-ABS-0275 Figure 4-6). Machine learning with race-aware feature weighting offers substantial and clinically meaningful improvements across all groups (IDDF2026-ABS-0275 Figure 1-4). These findings challenge universal FIB-4 application and support the development of ethnicity-specific fibrosis risk models.Abstract IDDF2026-ABS-0275 Figure 1Abstract IDDF2026-ABS-0275 Figure 2Abstract IDDF2026-ABS-0275 Figure 3Abstract IDDF2026-ABS-0275 Figure 4Abstract IDDF2026-ABS-0275 Figure 5Abstract IDDF2026-ABS-0275 Figure 6",
  "authors": [
    {
      "affiliations": [
        "New York Medical College, United States"
      ],
      "name": "Jeril Lasington"
    },
    {
      "affiliations": [
        "Rutgers University, United States"
      ],
      "name": "Lawin Steve Mathew Lasington"
    },
    {
      "affiliations": [
        "Boston University, United States"
      ],
      "name": "Swamynathan Umamaheshwaran"
    }
  ],
  "title": "IDDF2026-ABS-0275 One score, four populations: racial fairness of FIB-4 in MASLD fibrosis detection",
  "uid": "f7681394-0ba3-5f8f-a93f-ad77ac308a21"
}