TOTAL RECORDSSYNTHETIC
—
Rows available for experiments
MONITORING CENTER
Diabetes research overview
One workspace for data preparation, experiment training and exploratory prediction.
✦Research environment
SYNTHETIC COHORT · PIPELINE STATUS
From raw measurements
to reproducible experiments.
Cleaning, leakage-aware preprocessing and model evaluation are handled by a single repeatable training pipeline.
Dataset snapshot
Generated PIMA + laboratory composite cohort
FEATURE COLUMNSCLEANING READY
—
Mixed measurements and proxy fields
MODEL ARTIFACTSNOT CHECKED
—
Saved classification pipelines
INPUT QUALITYREVIEW
—
Highest feature missingness
Target distribution
Class balance in synthetic labels
Loading distribution…
Count per generated status class Source labels excluded from inputs
Quick actions
Continue your workflow
✦
Leakage-aware setup — target columns, source row identifiers and synthetic type-marker fields are excluded from model inputs.
How to interpret this workspace
Training metrics measure whether the pipeline can learn patterns in this generated cohort. They are not evidence of performance on real patients. Validate later using independent, clinically labeled data.
INFERENCE PLAYGROUND
Prediction studio
Explore the two-stage status and conditional type workflow with synthetic training labels.
BULK SCORING
Batch predictions
Upload a CSV to score up to 100,000 rows and export a new file with model outputs.
Upload dataset
Model uses an allowlist of feature columns. Target and provenance columns are ignored.
What happens to your file?
Preprocessing applied before inference
01
Schema normalizationWhitespace headers, invalid numerics and empty fields are normalized.
02
Missingness handlingKnown PIMA zero-as-missing values are handled; model medians are fitted during training.
03
Feature allowlistPatient IDs, source labels, target labels and Sim_* latent markers are never used as inputs.
04
CSV exportPredicted status, maximum class probability and conditional type scenario are appended.
◇
This local demo does not intentionally save uploaded files. Avoid uploading identifiable health records; no production security review has been completed.
TRAINING & EVALUATION
Model laboratory
Four saved pipelines compare clinical-only inputs with Ayurveda-inspired synthetic proxies.
Evaluation snapshot
Grouped synthetic holdout · not clinical validation
Loading model results…
Top feature importance
Choose a trained model
Feature importance appears after training.
Evaluation guardrails
Read before interpreting a score
Labels are synthetic
Model performance measures recovery of generated scenarios, not true diabetes status or type.
Repeated source donors
The cohort reuses PIMA and lab donors. Splits hold out PIMA source rows, but lab donors may recur.
Latent type markers excluded
Simulated autoantibodies and C-peptide are not model inputs, preventing direct access to generated type rules.
Independent validation required
Use separate patient-level clinical records and verified labels before any real-world claim.
Confusion matrix
Rows = actual synthetic labels · columns = predicted synthetic labels
Confusion matrix appears after training.
DATA VALIDATION
Data quality
Review the cohort schema, source provenance and cleaning steps used by the training pipeline.
!
This is a synthetic composite, not a patient-level merge
The two source CSVs did not contain a common patient key. The 100,000 rows were generated through donor resampling and jitter. The data cannot establish clinical type labels or real Ayurveda assessments.
COHORT ROWS
—
Generated composite records
TOTAL FIELDS
—
Features, targets and metadata
PIMA DONORS
—
Repeated source records
LAB DONORS
—
Female donors selected for cohort scope
Feature missingness
Highest missing percentage after basic normalization
Loading quality summary…
Cleaning pipeline
What the system does automatically
✓
Header normalizationTrim whitespace and handle duplicate column names.
✓
Type coercionInvalid numeric values become missing instead of crashing inference.
✓
Range checksImplausible numeric values become missing using broad guard ranges.
✓
Train-only imputationMedian imputation is fit within the training subset, then reused at inference.
✓
Leakage preventionSource labels, provenance IDs, and Sim_* fields are excluded from features.
Field groups
How the main dataset columns are interpreted
Clinical inputs
MODEL INPUTSPIMA measures and lab markers that can be entered as model features. Lab units must be verified.
Ayurveda / symptom proxies
OPTIONAL INPUTSSynthetic Prakriti/Vikriti scores and symptom flags. Not measured by either original source dataset.
Simulated type markers
EXCLUDEDSimulated autoantibodies, C-peptide and clinical clues. Excluded from all prediction feature sets.
Synthetic target labels
LABELS ONLYGenerated status and type scenarios. Used only as training targets; not verified patient labels.