Skip to content

CKD Dataset

The Krisis CKD suite uses the UCI Machine Learning Repository Chronic Kidney Disease dataset as its benchmark source.

Dataset Role

The dataset provides the source clinical rows for:

  • CKD detection
  • eGFR-derived staging
  • synthetic progression stress testing

The dataset is not bundled with the Python package. Users must provide their own local copy of the UCI CKD CSV via CKDSuite(data_path="...").

Download

Download the dataset from the UCI Machine Learning Repository:

Krisis expects a local CSV path:

suite = CKDSuite(data_path="datasets/ckd/ckd_full.csv")

Why The Dataset Is Useful

The UCI CKD dataset contains a mix of:

  • renal markers such as serum creatinine and blood urea
  • urine markers such as albumin and specific gravity
  • hematologic markers such as hemoglobin and packed cell volume
  • comorbidity indicators such as hypertension and diabetes
  • missing values and categorical fields that require preprocessing

This makes it useful for testing whether LLMs can reason over messy clinical feature sets rather than clean textbook cases.

Derived Fields

Krisis derives:

  • sex: synthetic when the source CSV does not provide it
  • egfr: CKD-EPI 2021-style estimate
  • ckd_stage: stage derived from eGFR
  • should_abstain: deferral-alignment label for ambiguous/conflicting cases

Feature Engineering Criteria

Krisis engineers CKD features using explicit, documented clinical criteria:

Engineered field Criteria used Reference
sex Generated when absent using a reproducible serum-creatinine-conditioned heuristic: sc <= 0.7 biases female, 0.7 < sc <= 0.9 is treated as ambiguous, and sc > 0.9 biases male. This exists only because the UCI CKD dataset does not include sex. CKD-EPI creatinine constants and documented Krisis assumption
egfr Computed from serum creatinine, age, and sex using the race-free CKD-EPI 2021 creatinine equation. National Kidney Foundation CKD-EPI 2021 equation
ckd_stage Derived from eGFR using KDIGO GFR categories: G1 >=90, G2 60-89, G3 30-59, G4 15-29, G5 <15 mL/min/1.73m². Krisis combines G3a/G3b by default, with optional split-stage support in code. KDIGO 2024 CKD Guideline
should_abstain Derived for evaluation only. Marks cases where eGFR is close to a staging threshold or where the binary CKD label conflicts with the eGFR-derived stage. This label is hidden from the model and used for deferral-alignment scoring. Krisis benchmark design

Synthetic sex generation

The UCI CKD dataset does not include sex. Krisis generates sex using a reproducible creatinine-conditioned process so eGFR can be computed. This is documented in suite output.

Included Tasks

  • detection: CKD vs not CKD
  • staging: CKD stage classification derived from eGFR
  • progression: synthetic progression stress test

Important Limitations

  • The dataset is small.
  • The dataset is cross-sectional, not longitudinal.
  • The progression task is synthetic.
  • The dataset is not representative of all CKD populations.
  • The benchmark is not clinical validation for patient care.

Do not overclaim

Krisis v0.1 evaluates LLM behavior on a CKD benchmark derived from UCI data. It does not prove clinical safety, diagnostic performance, or real-world deployment readiness.

For suite-level usage and schema requirements, see CKD Suite.

References