Patient-based real-time quality control for serum sodium: benchmark datasets, trained models and evaluation results for traditional, machine learning and deep learning methods

Research compendium for a study of patient-based real-time quality control (PBRTQC) applied to serum sodium. One record, one DOI, covering every publication arising from the study. What the study did. Built a standardized benchmark of analytical errors with known ground truth, then evaluated three families of detection methods against it on identical data: traditional PBRTQC algorithms, machine learning classifiers, and deep learning architectures. Because every method saw the same 232 datasets, the same error magnitudes, and the same evaluation metrics, the comparison is like-for-like, which published PBRTQC performance figures generally are not. Benchmark datasets. 232 datasets — 116 systematic-error and 116 random-error — each of 500,000 serum sodium results arranged as 400 nominal operating days of 1,250 results. Errors were injected across a fixed 116-level magnitude ladder spanning 0.5% to 200% in both directions, with a fine zone at 0.5% resolution covering the clinically relevant region and a coarse zone extending to 200% as a positive control. Each dataset contains exactly 340 error-bearing days and 60 clean days, and carries day-level and episode-level ground truth and an MD5 checksum. Trained models. Machine learning and deep learning models fitted for both error types, with the decision thresholds used to generate the reported results. Evaluation results. Detection performance across all methods, both error types, and every ladder level, including the recalibrated threshold outputs. Figures. Publication figures for all papers, main and supplementary, in PNG and TIFF. Configuration and environment. The injection configuration, magnitude ladder, population statistics, corrections record, Python version, and pinned library versions. Every dataset is exactly reconstructible from the deposited baseline series, the master seed, the configuration file, and the ladder. The source data. Serum sodium results as reported by the laboratory information system of AIIMS Bhubaneswar. The values are unmodified. No patient identifiers, no demographic fields, no clinical annotations, and no dates are retained. Operating days are numbered 1 to 400 and carry no correspondence to real acquisition dates. Release for research use was approved by the Institutional Ethics Committee of AIIMS Bhubaneswar. A note on error types. Systematic and random errors are structurally different here and are not two settings of one model. A systematic episode is a contiguous run of affected results with a well-defined onset. A random episode is a scattered set of affected positions with no onset and no duration, so its manifest schema differs deliberately, and no field expresses duration. Detection delay is undefined in the presence of random error. Code reusing this corpus should branch on error type rather than normalize the column names. Note also that what is called random error here is a fixed magnitude and sign applied at scattered positions: it models sporadic contamination of a result stream, not an increase in analytical imprecision. Structure. 01_config — configuration, magnitude ladder, population statistics, corrections record. 02_datasets — the 232-dataset corpus and the source baseline. 03_models — trained models and thresholds. 04_results — evaluation outputs. 05_figures — publication figures. 06_environment — Python version, pinned requirements, library versions. Access. Files are released on request. Record metadata and documentation are public. Code. The analysis code is not deposited here and is available from the corresponding author on request.