{
  "abstract": "Background Standard fibrosis scoring tools including FIB-4, perform poorly in lean individuals with MASLD, yet the extent of this failure across racial/ethnic groups and whether machine learning can overcome it remains unknown. We systematically evaluated fibrosis prediction performance in lean versus non-lean adults stratified by race/ethnicity in a nationally representative elastography-validated cohort.Methods We analysed 15,587 adults from NHANES 2017-2023 with valid transient elastography. Significant fibrosis was defined as LSM ≥7.0 kPa and lean as BMI <25 kg/m 2. XGBoost classifiers incorporating 17 clinical and metabolic variables were trained with stratified cross-validation within lean and non-lean subgroups per racial/ethnic group. SHAP values identified feature contributions within each stratum, compared against FIB-4 by AUC and ΔAUC.Results Lean fibrosis prevalence varied substantially by race: 8.5% in Non-Hispanic (NH) White, 12.3% in NH Black, 5.1% in Hispanic, and 3.7% in NH Asian adults.In non-lean individuals, machine learning substantially outperformed FIB-4 across all groups (ΔAUC +0.161 to +0.220), with BMI, waist circumference, and AST as dominant features. In lean individuals, this advantage disappeared entirely: ML failed to outperform FIB-4 in NH Black (ΔAUC -0.010), Hispanic (ΔAUC -0.105), and NH Asian adults (ΔAUC -0.127), with FIB-4 marginally superior in the latter two groups. (IDDF2026-ABS-0276 Figure 1)SHAP analysis revealed near-zero feature contributions across all clinical variables in lean subgroups indicating that standard metabolic markers carry minimal fibrosis signal in lean adults regardless of scoring method. NH Black lean adults carried the highest absolute lean fibrosis burden (12.3%) yet the lowest predictive signal, suggesting fibrosis in this subgroup is driven by unmeasured determinants, likely, structural inflammation and social determinants of health, invisible to current clinical panels. (IDDF2026-ABS-0276 Figure 2)Conclusions Lean MASLD-associated fibrosis represents a genuine clinical blind spot that machine learning cannot resolve with standard clinical inputs. Both FIB-4 and ML fail in lean adults, with the greatest unmet need in NH Black adults, carrying a 12.3% lean fibrosis prevalence undetectable by any current scoring approach. These findings support the use of direct elastography screening in lean adults, particularly in Black populations, rather than risk-stratification triage.Abstract IDDF2026-ABS-0276 Figure 1Abstract IDDF2026-ABS-0276 Figure 2",
  "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-0276 Do fibrosis risk models work equally in lean adults? A race-stratified machine learning analysis in MASLD",
  "uid": "1c066b90-a4ca-5401-8ed2-a55d36ae93e8"
}