Clustering of patient-derived laboratory data: A privacy-preserving distributed approach to enhance clinical laboratory quality control and regional external quality assessment

BACKGROUND: Traditional quality control (QC) has inherent limitations, and there are still various challenges existing in patient-based real-time QC (PBRTQC) and patient-based quality assurance (PBQA). We developed an integrated PBRTQC-PBQA model using density-based clustering and evaluated its effectiveness as a regional clinical laboratory quality assurance tool. METHODS: A 20-month multicentre study across seven hospitals analysed five haematological analytes(white blood cell count [WBC], red blood cell count [RBC], haemoglobin [HGB], platelet count [PLT], and mean corpuscular volume [MCV]). Custom Clustering-PBRTQC-PBQA(CPP) models were developed using log/Box-Cox transformation, 3D spatial construction with companion analytes determined by distribution morphology, Mahalanobis distance filtering (chi-square = 0.3, ~ 47% exclusion rate), and iterative OPTICS clustering to identify hospital-specific core reference populations. A privacy-preserving distributed architecture transmitted only aggregated statistical summaries. Median/EWMA harts (PBRTQC) and Z-score/relative bias charts (PBQA) were created alongside a closed-loop quality improvement system and calibration event verification. RESULTS: The CPP model reduced EWMA alert episodes by 25.0% (92-69) while increasing 1-3S alert episodes by 31.2% (64-84) compared with conventional moving-average PBRTQC, demonstrating selective elimination of systematic bias signals while preserving sensitivity to random errors (net change: 180-173 episodes, -3.9%). Overall calibration consistency was 82.3% (95% CI: 74.1%-88.7%). Among 52 confirmed true-positive alerts, 37 (71.2%) were detected without concurrent internal QC out-of-control signals, demonstrating the model's capacity to capture low-magnitude biases that conventional IQC misses. CONCLUSION: Density-based clustering underpins the PBRTQC-PBQA model's effectiveness in deriving hospital-specific core reference values, filtering unconfirmed QC alerts, validating calibration, and enhancing regional test consistency. The model's ability to detect systematic biases not captured by IQC underscores its value as a complementary quality assurance layer. Its lightweight design supports regional deployment.