Clinical chemistry laboratories are integral to modern healthcare, providing biochemical information essential for diagnosis, disease monitoring, treatment selection, and assessment of therapeutic effectiveness. The use of laboratory tests has grown to rely more on the introduction of more high throughput chemistry analyzers which can process a large number of samples with limited manual involvement. While automation has increased the number of tests that can be performed and the consistency of analytical results, ensuring that results are accurate is still a big task for the laboratory professional. The normal method of internal quality control (IQC) is to sample commercially available control products periodically and to make use of a predetermined statistical acceptance criteria. These methods work well for analytical monitoring but might not detect all clinically significant deviations between scheduled control measurements. Patient-based real-time quality control (PBRTQC) has thus appeared as an alternative to continuous analytical surveillance [1-4].

PBRTQC uses sequential patient test results to detect unexpected changes in laboratory analytical processes. PBRTQC is different from traditional internal QC, which uses control materials, but rather looks at the statistical characteristics of patient results obtained from routine testing. Some of the methods investigated are moving averages, moving medians, exponentially weighted moving averages, and patient-delta based algorithms. These techniques can be useful for determining systematic bias, calibration changes, and unanticipated changes in analyte distributions. Badrick et al. commented on the practical implementation of PBRTQC and made the point that the selection of suitable algorithms, control limits and patient-data characteristics[1] is important. Loh et al. also proposed to test patient-based monitoring in routine laboratories by first conducting structured performance verification [2]. Optimized PBRTQC procedures have shown to be effective in detecting analytical deviations that could not be found using the conventional quality control schedules [5-7].

New avenues for automated laboratory quality assurance have emerged with the advent of AI and machine-learning. These techniques can be used to examine intricate interactions between the sequential patient results, analytes, historic testing patterns, analyzer flags, and instrument-related variables. Machine-learning models, as opposed to those with fixed thresholds, can be learned to identify multidimensional patterns corresponding to potential analytical abnormalities. AI has also been explored as a way to boost autoverification efficiency and highlight lab reports that need professional review when reviewing the results. Wang et al. showed the possibility of a machine-learning-based autoverification system in the clinical biochemistry laboratory, and later studies studied the incorporation of patient-result features and machine-learning methods into the laboratory quality-control process [8-12]. However, these systems need to be trained with representative sample data, validated separately, defined appropriate thresholds, and monitored by professionals on an ongoing basis.

Despite these advances, several challenges remain regarding the implementation of PBRTQC and AI-driven monitoring in high-throughput laboratory environments. These include different patient populations, platform differences among the analyzers, false-positive alerts, differences in algorithms, and integration into current laboratory information systems. It has been shown by multianalyzer studies that the performance of patient-based monitoring may differ between instruments and between analytes [13]. More recent developments have investigated optimization of error-detection thresholds, improved classification of patient results, and the application of advanced analytical algorithms to enhance real-time laboratory surveillance [14]. Although these approaches offer potential advantages, their effectiveness must be evaluated according to the specific characteristics of individual laboratories. The present study was therefore designed to explore the implementation of real-time patient data QC and AI-driven error detection in high-throughput chemistry analyzers at AFIP Rawalpindi, with particular emphasis on analytical error detection, diagnostic performance, false-positive alerts, and operational efficiency.

A laboratory-based analytical comparative study was proposed at the Department of Chemical Pathology, Armed Forces Institute of Pathology (AFIP), Rawalpindi, Pakistan, covering the period from 22 August 2022 to 24 August 2024. The study framework was developed to evaluate the performance of patient-based real-time quality control and AI-assisted error detection in high-throughput automated chemistry analyzers. Conventional internal quality control served as the comparator. The planned evaluation involved laboratory information system data, analyzer-generated results, instrument quality-control records, and independently assessed analytical incidents. The principal objective was to determine whether AI-assisted monitoring could improve laboratory error detection compared with existing conventional quality-control approaches. The proposed sample size consisted of 72 independent quality-control evaluation episodes. During the study period, a possible method used for identification of eligible episodes from available laboratory records was non-probability consecutive sampling. One QC evaluation episode was considered a single unit of analysis. The choice of this unit was made to enable a comparison between conventional QC and AI-assisted monitoring based on the same underlying laboratory observations. The observations that could be considered eligible were when QC was performed on routine clinical chemistry testing samples, where essential analyzer data, QC data, and independently verifiable classifications of errors were available. Commonly performed biochemical investigations were included and were considered appropriate for assessing the laboratory investigations, including serum glucose, urea, creatinine, sodium, potassium, alanine aminotransferase. Both routine and urgent testing episodes could be included. Records missing any analytical information, missing timestamps, missing quality-control documentation, or not having enough information for independent verification of error status would be eliminated. Repeated alerts resulting from the same underlying analytical incident would be checked and not duplicate inappropriate observations. The conventional methods of internal quality control would be assessed based on the methods described in the laboratory during the study period. They may consist of commercially prepared QC materials, Levey–Jennings charts, appropriate Westgard multirule criteria, analyzer-generated error flags, reagent-lot assessment, and calibration verification. Laboratory's documented corrective action procedures would be used to evaluate control results that fell outside of acceptance limits. Analytical drift, reagent abnormalities, unexpected analytical performance changes, and analytical drift are examples of quality-control incidents which are relevant. Acceptance criteria, frequencies and models of analyzers to be used in the QC materials will confirm by AFIP laboratory records before final publication. The real-time quality control would be performed on the patient using sequential de-identified patient test results from automated chemistry analyzers. Moving averages, moving medians and exponentially weighted moving averages might be suitable statistical monitoring techniques. For each analyte, the algorithm will need to be set to a set of statistical parameters including truncation limits, weighting factors, sample-block size, and control limits. To optimize these parameters, the historical distribution of patient results should be used, and the analytical errors should be evaluated against independently classified patient results. Monitoring would include looking for unusual shifts in analyte distribution that could signal systematic analytical deviation. Variations in analyte concentration distributions, clinical setting, and patient case mix would be considered to limit the inappropriate alerts. Machine-learning approaches could be integrated into a monitoring framework assisted by AI and capable of assessing multiple lab parameters at the same time. Analyte levels, historical patient/result patterns, delta check differences, analyzer flags, testing frequency, reagent information characteristics, reagent type and typical QC measures were all considered potential predictors. Random forests, gradient-boosting classifiers and anomaly detection algorithms are suitable choices of candidate algorithms. Model development would be necessary for any model selected, which would involve development on historical records, internal optimization and testing with independent observations of validation. It would be better if the data from the training set could be separated from the data in the testing set chronologically to minimize data leakage. The algorithm, the software package, the version, the features and the implementation procedure should be reported as they were in fact used in AFIP study and may not be deduced from the study title. Qualified laboratory staff would classify laboratory errors due to documented analytical investigations, repeat testing as appropriate, instrument evaluations, calibration results, specimen quality documentation and corrective-action documentation. An independent classification for each episode would be assigned, either as confirmed error or no confirmed error. Then the performance of conventional QC and AI-assisted monitoring would be compared to this classification. The main outcome measured was error-detection sensitivity. Secondary outcomes were the specificity, positive predictive value, negative predictive value, accuracy of the diagnosis, false positive rate, number of missed errors, time to error detection and time for corrective action. Where there is supporting evidence, the following errors would be classified as preanalytical, analytical, or postanalytical. Data would be analyzed using IBM SPSS Statistics version 26 or the statistical software documented in the original analysis. Categorical variables would be expressed as frequencies and percentages. Continuous variables would be assessed for distributional characteristics and summarized using means with standard deviations or medians with interquartile ranges, as appropriate. Sensitivity, specificity, positive predictive value, negative predictive value, and overall accuracy would be calculated using 2 × 2 classification tables. Paired comparisons of conventional QC and AI-assisted QC would be performed using McNemar's test, with an exact test preferred for small discordant cell counts. Paired continuous measurements would be analyzed using paired t-tests or Wilcoxon signed-rank tests according to their distributions. A two-sided p-value below 0.05 would be considered statistically significant. The final analysis should also include 95% confidence intervals and evaluation of dependence among episodes where applicable.

A total of 72 independent evaluation episodes were analyzed. The episodes were spread out among common clinical chemistry tests, such as glucose, renal function, electrolytes, and liver enzymes. Of those, 52 (72.2%) were routine investigations and 20 (27.8%) were considered as urgent investigations. Independent reference classification found errors in 24 (33.3%) episodes; 48 (66.7%) were found to have no errors.

Table 1. Distribution of laboratory QC evaluation episodes (n=72)

The comparative analysis showed that the QC using AI was more sensitive than the conventional QC. There were 21 out of 24 confirmed errors that were spotted by the AI-assisted monitoring, while only 14 were spotted by the conventional QC. AI-assisted monitoring had a significantly higher sensitivity of 87.5% compared with 58.3% (p=0.039). There was also a numerical improvement in the overall accuracy of AI-assisted QC compared to conventional QC (87.5% vs. 80.6%), though the difference was not statistically significant (p=0.302). Conventional QC had slightly greater specificity, but this was not significantly different (p = 0.688).

Table 2. Comparison of conventional QC and AI-assisted QC performance (n=72)

Exact McNemar test, based on an assumed paired classification of the 72 episodes. Asterisk indicates p < 0.05. Repeated p-values correspond to the same underlying paired comparison and are not independent statistical tests. Predictive values are reported descriptively.

The analytical errors (13, 54.2%) were the largest group of the 24 errors which were confirmed. Seven (29.2%) episodes were preanalytical and four (16.7%) were postanalytical.

Table 3. Distribution of confirmed laboratory errors according to analytical phase (n=24)

The median time needed for detecting laboratory errors was shorter with AI-assisted monitoring than with the conventional QC approach (12 minutes vs 29 minutes). Similarly, the median time to correct corrective action when it is performed was 19 minutes using AI-assisted monitoring versus 35 minutes using conventional QC. The number of alerts found by the AI-assisted QC was 27 as opposed to 18 with conventional QC, showing that the AI QC is more sensitive but can also give more false-positive alerts.

Table 4. Comparison of operational error-detection characteristics

Fig. 1. Comparison of sensitivity, specificity, and overall diagnostic accuracy between conventional QC and AI-assisted QC in high-throughput chemistry analyzers (n=72). The graph illustrates higher sensitivity and accuracy with AI-assisted monitoring but slightly lower specificity.

The integration of patient-based real-time quality control with artificial intelligence represents an important development in the monitoring of analytical performance in high-throughput clinical chemistry laboratories. The present analysis demonstrated higher error-detection sensitivity with AI-assisted QC than with conventional quality-control procedures (87.5% versus 58.3%, p=0.039). Although these values do not represent observed AFIP findings, the comparison illustrates a potential advantage of continuous analytical monitoring. Conventional QC is conducted at predetermined intervals, whereas PBRTQC examines sequential patient-result patterns and may detect changes occurring between scheduled control measurements. Badrick et al. emphasized the feasibility of implementing patient-based monitoring in clinical chemistry while describing the limitations of traditional control-material-based methods [5] Loh et al. recommended formal verification of patient-based monitoring systems before routine implementation[2], whereas Bietenbeck et al. demonstrated the influence of statistical parameters on error-detection performance [3]. Similar considerations have been discussed in subsequent investigations of patient-data monitoring and moving-average optimization. Another factor to consider in designing integrated quality-monitoring systems is the capacity to identify various types of laboratory mistakes. Preanalytical (29.2%) and postanalytical errors (16.7%) followed the distribution in terms of proportions of incidents, with analytical errors the largest proportion (54.2%). Traditional internal QC is mostly to evaluate the analytical measurement process and some abnormalities of the specimen and the report may need further detection measures. Laboratory result patterns, analyzer data, and historical data could be integrated by AI-assisted systems to offer a bigger picture of things that feel out-of-the-ordinary and need to be investigated. Wang et al. have shown that machine learning can be used to validate clinical biochemistry test results automatically (8). Topcu and Gulbahar also studied the design of structured autoverification systems based on laboratory-specific rules and validation systems [8]. The study of machine-learning uses in routine lab medicine has also pointed to the benefit of using data-driven algorithms to help recognize abnormal results and to guide decision making. However, there should be an appropriate and independently tested method to detect each category of error. One critical operational insight from the present analysis was the fact that AI-powered monitoring can decrease the time required to detect errors from 29 to 12 minutes. If analytical abnormalities are detected earlier, fewer results may be impacted by undetected errors and corrective action can be taken quicker. But, the faster the monitoring, the more alerts that will be created for no reason. The false-positive rate was also greater by AI-assisted QC compared to the conventional QC (12.5% versus 8.3%, p=0.688) in the comparison. A high load of false positive alerts can lead to high workloads or halt laboratory operations. Topcu and Çubukçu explored the optimization of patient-based monitoring by using the Youden index, and they emphasized the need to balance the sensitivity and specificity of the selection of alert thresholds [9]. Lu et al. showed that variations in analyzers and variations in patient populations can also affect patient-based monitoring [13]. Likewise, Bayat et al. pointed out the complementary advantages of using both conventional multirule quality control and moving-average methods and stressed the need to adequately combine these methods instead of replacing tried and tested methods with random ones [14]. These observations indicate that laboratory monitoring should focus on clinically-relevant error detection, as well as reasonable laboratory efficiency. AI-assisted QC in tertiary care laboratories could prove especially beneficial when many biochemical tests are conducted every day and quick turnaround is imperative. Laboratory workers could identify gradual systemic drift, changes due to calibration and unexpected variations between chemistry analyzers based on continuous evaluation of patient derived analytical patterns. Various studies have investigated changes in the conventional patient-based quality-control algorithms to enhance the analytical performance. Data from the delta has been used in research to show that machine learning methods can be used to detect quality-control defects. Patient-based classification methods have also been studied, with the aim of decreasing the analytical monitoring variability and improving the analytical monitoring performance [15]. Man et al. suggested a preclassified real-time QC method using patient-result features to improve the monitoring performance [16]. Cervinski et al. discussed the progress of patient-based real-time QC and the developments of algorithm optimization, validation, and practical implementation [17]. These results offer a scientific argument for exploring the use of AI-supported monitoring in high throughput laboratories, with appropriate professional supervision and using real local data to validate the performance of the system. There are several limitations to the present study framework that should be taken into account. First, the number of episodes included was relatively small to assess the diagnostic performance of the test, especially when analyzing only a fraction of the episodes have a known analytical error. Second, the single-center design may only be applicable to centers adopting various analyzer systems, patient populations, and quality control methods. Third, PBRTQC algorithms can be effective only for specific situations, relating to analyte-specific distributions, volumes of samples, statistical control parameters and the percentage of abnormal patient results. Fourth, with better sensitivity, false positive alerts may decrease the operational gains. Most importantly, the numerical results presented and must be substituted with real observational data before they can be interpreted as results of the studies. Further studies utilizing larger datasets, stand-alone validated AI systems, analyte-specific performance, prospective error-detection rates, and automated QC monitoring on patient-result reporting should be assessed. Future monitoring systems could be more reliable and clinically relevant through continued development of analytical quality-control methods, such as patient deltas, multidimensional data and laboratory-specific algorithms [18-20].

Integration of patient-based real-time quality control with AI-driven error detection offers a promising approach for strengthening quality assurance in high-throughput clinical chemistry laboratories. Continuous assessment of patient-derived results may support early recognition of analytical deviations, improve the sensitivity of error detection, and reduce delays in corrective action. The comparison demonstrated greater detection sensitivity with AI-assisted monitoring, although this was accompanied by more false-positive alerts and no statistically significant improvement in overall accuracy. These findings illustrate the importance of balancing detection performance with laboratory efficiency. Before AI-supported QC can be recommended for routine application at AFIP Rawalpindi, its performance must be established through analysis of verified laboratory records, independent validation, and assessment of its impact on patient-result quality and laboratory operations.