Suites

Suites are dataset-specific adapters. They turn raw clinical rows into benchmark-ready PatientRecord objects with labels, metadata, prompts, and suite descriptions.

Suites are the clinical boundary

If a dataset uses different units, label definitions, or clinical meaning, it should get its own suite rather than being forced into an existing one.

Available Suites

Suite	Status	Dataset	Tasks
CKD Suite	Available in v0.1	UCI CKD dataset	detection, staging, synthetic progression
Diabetes Suite	Coming soon	Not available yet	Not available yet
Hypertension Suite	Coming soon	Not available yet	Not available yet

What A Suite Owns

Each suite is responsible for the complete path from source data to model-facing clinical cases:

Responsibility	Purpose
Dataset loading	Reads the local source file or supported dataset object
Schema validation	Rejects unsupported columns, labels, and value formats early
Preprocessing	Handles missing values, categorical normalization, scaling, and feature cleanup
Feature engineering	Derives clinical fields such as eGFR, CKD stage, or ambiguity labels
Task generation	Produces labels for detection, staging, progression, or future tasks
Synthetic generation	Adds stress cases where the suite explicitly supports them
Metadata reporting	Documents source, feature set, task, record counts, and limitations

Suite Configuration

Suites use SuiteConfig to keep runs reproducible:

from krisis.data.base import FeatureSet, SuiteConfig, Task
from krisis.data.ckd.suite import CKDSuite

suite = CKDSuite(
    data_path="datasets/ckd/ckd_full.csv",
    config=SuiteConfig(
        task=Task.STAGING,
        features=FeatureSet.FULL,
        n_synthetic=80,
        seed=42,
        test_size=0.2,
    ),
)

Key fields:

Field	Meaning
task	Which benchmark task to generate
features	Whether the model sees the full or reduced feature set
n_synthetic	Number of synthetic stress cases to add
seed	Random seed for reproducible splitting and generation
test_size	Fraction of real source data used as held-out evaluation rows

The 80/20 Split

For CKD, Krisis uses the configured test_size to split the source dataset. With the default test_size=0.2:

• 80% of real rows are used to fit preprocessing and synthetic generation
• 20% of real rows are held out for evaluation
• synthetic rows are generated from the training distribution, not the test rows
• evaluation uses the held-out real rows plus the requested synthetic rows

This keeps the synthetic generator from learning directly from the evaluation split.

Suite Output

Calling suite.load() returns PatientRecord objects. Each record contains:

• model-facing features
• ground-truth label for the selected task
• clinical metadata used for analysis
• should_abstain metadata where the task supports deferral scoring

Calling suite.describe() returns report metadata, including:

• source dataset
• task
• feature set
• real and synthetic record counts
• label distribution
• stage distribution
• clinical source notes
• suite-specific assumptions

Adding New Suites

A new suite should be added when a domain has a distinct dataset, schema, task definition, or clinical interpretation. A good suite should include:

1. a validator for supported source data
2. deterministic preprocessing
3. documented feature engineering
4. task-specific label generation
5. suite metadata for reports
6. focused tests
7. a dataset page and suite page in the docs