{
  "generated_by": "research/rebuild_nhanes_model.py",
  "data_release": "NHANES 2017-2018 with 2019 Linked Mortality File",
  "status": "Exploratory observational analysis; not clinically validated",
  "sources": {
    "demographics": {
      "url": "https://wwwn.cdc.gov/Nchs/Data/Nhanes/Public/2017/DataFiles/DEMO_J.xpt",
      "sha256": "c0b46e0345ea19404928656277c8b0d10b0cca348a9b2fe4fc3c67e8b7ee73ec"
    },
    "vitamin_d": {
      "url": "https://wwwn.cdc.gov/Nchs/Data/Nhanes/Public/2017/DataFiles/VID_J.xpt",
      "sha256": "5a688e2107758b9344a222823fbdce4838928bbec70e1fd2f34e3ef5c45a95ac"
    },
    "medical_conditions": {
      "url": "https://wwwn.cdc.gov/Nchs/Data/Nhanes/Public/2017/DataFiles/MCQ_J.xpt",
      "sha256": "79c50c805a377fcd61d44fae8b86a16af0a5741f8b1aeebda918d7d743c68f98"
    },
    "mortality": {
      "url": "https://ftp.cdc.gov/pub/health_statistics/NCHS/datalinkage/linked_mortality/NHANES_2017_2018_MORT_2019_PUBLIC.dat",
      "sha256": "42989d4fe35754d696b770f26553b0a309c110a9e26ab0ee504be3eae5987523"
    }
  },
  "cohort": {
    "participants": 515,
    "events": 35,
    "low_group": 65,
    "low_group_events": 7,
    "high_group": 450,
    "high_group_events": 28,
    "max_follow_up_months": 37,
    "definition": "Adults 20+ reporting a prior cancer diagnosis (MCQ220=Yes), linkage eligible, with measured total 25(OH)D and follow-up time."
  },
  "kaplan_meier": {
    "months": [
      0,
      3,
      6,
      9,
      12,
      15,
      18,
      21,
      24,
      27,
      30,
      33,
      36
    ],
    "low_under_20_ng_ml": [
      1.0,
      0.98462,
      0.95385,
      0.93846,
      0.90769,
      0.90769,
      0.90769,
      0.90769,
      0.90769,
      0.87138,
      0.87138,
      0.87138,
      0.87138
    ],
    "high_at_least_20_ng_ml": [
      1.0,
      0.99556,
      0.98444,
      0.97333,
      0.96889,
      0.95189,
      0.94626,
      0.94626,
      0.94264,
      0.9322,
      0.92486,
      0.9133,
      0.9133
    ]
  },
  "cox": {
    "features": [
      "age",
      "female",
      "vitamin_d_per_10"
    ],
    "coefficients": {
      "age": 0.0931210136551759,
      "female": -0.14260970880899196,
      "vitamin_d_per_10": -0.3666646483140719
    },
    "means": {
      "age": 66.46796116504855,
      "female": 0.5339805825242718,
      "vitamin_d_per_10": 3.4396815533980587
    },
    "baseline_months": [
      0,
      3,
      6,
      9,
      12,
      15,
      18,
      21,
      24,
      27,
      30,
      33,
      36
    ],
    "baseline_survival": [
      1.0,
      0.996596,
      0.988549,
      0.981565,
      0.976866,
      0.967676,
      0.964554,
      0.964554,
      0.962509,
      0.953836,
      0.949776,
      0.94355,
      0.94355
    ],
    "concordance_index": 0.723,
    "vitamin_d_hr_per_10_ng_ml": 0.693,
    "vitamin_d_hr_ci95": [
      0.5087,
      0.9442
    ],
    "vitamin_d_p": 0.020115
  },
  "machine_learning": {
    "task": "Cross-sectional prediction of measured total 25(OH)D <20 ng/mL among the analytic cancer-history cohort.",
    "features": [
      "age",
      "sex",
      "race_ethnicity",
      "exam_season",
      "family_income_to_poverty_ratio"
    ],
    "validation": "Five-fold stratified cross-validation; all preprocessing refit inside each fold.",
    "positive_cases": 65,
    "prevalence": 0.1262,
    "models": {
      "logistic_regression": {
        "roc_auc": 0.652,
        "average_precision": 0.2263,
        "brier_score": 0.1067
      },
      "hist_gradient_boosting": {
        "roc_auc": 0.6657,
        "average_precision": 0.2291,
        "brier_score": 0.1118
      }
    },
    "intended_use": "Research benchmark only; not a clinical testing-priority model."
  },
  "limitations": [
    "Observational association cannot establish causality or treatment benefit.",
    "Only 35 deaths and at most 37 months of follow-up; estimates are imprecise.",
    "The Cox model is unweighted and does not implement NHANES complex-survey variance estimation.",
    "Cancer types and treatment histories are heterogeneous.",
    "The public-use mortality file protects confidentiality and has analytic limitations documented by NCHS."
  ]
}