BekzatK/diabetes-qms-benchmark

Diabetes Diagnostic Device Manufacturing QC Benchmark

This dataset contains de-identified manufacturing quality-control (QC), calibration, and analytical process records generated during routine production of diabetes diagnostic medical devices (glucose test strips and HbA1c analyzers). It was compiled to support research on leakage-safe, temporally validated AI-enabled quality-management systems (AI-QMS) for predictive process monitoring.

The data describe manufacturing and analytical process behaviour only. No patient-identifiable information, clinical intervention data, or personal health records are included.

Dataset structure

Config	File	Unit of analysis	Records	Years	Main task
glucose_qc	glucose_qc_ml_dataset.csv	QC series	70	2020–2026	Early QC-failure prediction (59 pass / 11 fail)
hba1c_measurements	hba1c_ml_measurements.csv	Measurement	690	2022–2026	Residual calibration correction / ISO tolerance
hba1c_lots	hba1c_ml_lots.csv	Production lot	40	2022–2026	Lot-level drift and anomaly detection

Excel workbooks (glucose_qc_ml_dataset.xlsx, hba1c_ml_dataset.xlsx) contain the same data plus an embedded data_dictionary sheet.

Field descriptions

glucose_qc — glucose test-strip QC series

Column	Description
serie	Series number / source file name
year	Protocol year (2020–2026)
date	Control date (may be empty in some protocols)
ctrl_mean	Mean of 10 MID-level measurements, mg/dL
ctrl_sd	Standard deviation of the control solution
ctrl_cv	Control-solution coefficient of variation, % (acceptance criterion ≤5)
hct_mean, hct_min, hct_max, hct_range, hct_std	Hematocrit statistics across samples A–D
err_A1…err_D2	Signed percentage error at 8 concentration points
abserr_A1…abserr_D2	Absolute percentage error at the same points
err_mean	Mean signed error (systematic shift)
err_abs_mean	Mean absolute error
err_abs_max	Maximum absolute error (= regression target)
err_std	Dispersion of point-wise errors
err_bias	Systematic shift (= err_mean)
n_points_over5, n_points_over10	Number of points exceeding 5% / 10%
target_reg_maxabs	Regression target: series maximum absolute error, %
target_clf_pass	Classification target: 1 = series acceptable (error ≤10% and CV ≤5%), 0 = fail
fail_reason	Fail cause: none / error / cv / error+cv
instability_score	err_abs_max + ctrl_cv — continuous instability index

hba1c_measurements — HbA1c measurement level

Column	Description
year, lot, date	Protocol / lot identifiers
sample_no, analyzer_rep	Sample number and replicate index
hct	Sample hematocrit, %
ref_hba1c	Regression target: reference HbA1c (G11 / Pentra), %
hb_od, signal_od	Raw instrument optical densities
signal_od2	signal_od² — quadratic feature (as in the factory curve)
conc	HbA1c from the factory instrument calibration, %
err_abs	conc − ref_hba1c (absolute deviation)
err_rel_pct	100·(conc−ref)/ref, %
within_iso	Classification target: 1 = within tolerance (|abs| ≤0.5% or |rel| ≤6%)
calib_a, calib_b, calib_c	Lot factory-curve coefficients

hba1c_lots — HbA1c lot level

Column	Description
year, lot	Lot identifiers
n_meas	Number of venous measurements in the lot
calib_a, calib_b, calib_c	Coefficients of the curve conc = a·OD² + b·OD + c
hct_mean, hct_std, ref_mean, ref_std, sig_mean, sig_std	Lot statistics
cv_mean	Mean concentration CV across triplets, %
da, db, dc	Coefficient drift relative to the 2022 baseline curve (lot 201222)
calib_anom	Anomaly target: 1 = anomalous calibration (|a| > 1500 or a > 0)

Leakage-safe usage (important)

Predictors must be available at or before the QMS decision point. Some columns are derived from the target and must not be used as model inputs:

• Glucose regression: target_reg_maxabs equals err_abs_max, so all err_* / abserr_* / err_abs_max / instability_score columns are excluded. Admissible predictors are the hematocrit profile, control-solution statistics, and year.
• Glucose classification: exclude the final error, fail_reason, and the pass/fail decision from predictors.
• HbA1c regression (ref_hba1c target): train on signal_od, signal_od2, hb_od, and hct only. conc, err_*, and within_iso are derived from the target and are not used as features.
• Lot anomaly detection: calibration coefficients that define the anomaly label are excluded from the anomaly-model predictors.

Recommended validation

Because the data are structured by production year and lot, random splitting can leak process-specific information. Deployment-oriented evaluation should use lot-grouped cross-validation, leave-one-lot-out, leave-one-year-out, and temporal holdout (train ≤2024, test ≥2025), with random splitting kept only as a technical baseline.

Notes and limitations

• Data derive from a single manufacturer with a modest number of series and lots per year; failure events are few (11 for glucose), so metrics should be reported with bootstrap confidence intervals and warrant prospective confirmation.
• The hba1c_measurements table includes all 690 records from 40 lots (2022 onward). A stricter analytical protocol may restrict to measurements with complete optics and reference values (e.g., 2023–2026); apply such filtering downstream.
• This is a process-monitoring resource, not a clinical-diagnosis dataset.

Citation

@article{aiqms_diabetes_manufacturing,
  title   = {Leakage-Safe and Temporally Validated AI-QMS Framework for Predictive
             Quality Management in Diabetes Diagnostic Medical Device Manufacturing},
  author  = {[Bekzat Karimkyzy]},
  year    = {2026}
}

License

Released under CC BY 4.0. You may share and adapt the material with appropriate attribution.