Application of Patient‐Based Real‐Time Quality Control Based on Artificial Intelligence Monitoring Platform in Continuously Quality Risk Monitoring of Down Syndrome Serum Screening

ABSTRACT

Background

Patient‐based real‐time quality control (PBRTQC) has gained attention because of its potential to continuously monitor the analytical quality in situations wherein internal quality control (IQC) is less effective. Therefore, we tried to investigate the application of PBRTQC method based on an artificial intelligence monitoring (AI‐MA) platform in quality risk monitoring of Down syndrome (DS) serum screening.

Methods

The DS serum screening item determination data and relative IQC data from January 4 to September 7 in 2021 were collected. Then, PBRTQC exponentially weighted moving average (EWMA) and moving average (MA) procedures were built and optimized in the AI‐MA platform. The efficiency of the EWMA and MA procedures with intelligent and traditional control rules were compared. Next, the optimal EWMA procedures that contributed to the quality assurance of serum screening were run and generated early warning cases were investigated.

Results

Optimal EWMA and MA procedures on the AI‐MA platform were built. Comparison results showed the EWMA procedure with intelligent QC rules but not traditional quality rules contained the best efficiency. Based on the AI‐MA platform, two early warning cases were generated by using the optimal EWMA procedure, which finally found were caused by instrument failure. Moreover, the EWMA procedure could truly reflect the detection accuracy and quality in situations wherein traditional IQC products were unstable or concentrations were inappropriate.

Conclusions

The EWMA procedure built by the AI‐MA platform could be a good complementary control tool for the DS serum screening by truly and timely reflecting the detection quality risks.

During the period of early‐warning occurrence, the traditional internal quality control (IQC) method did not trigger the quality control rules, such as 1‐3s or 2‐2s. However, the patient‐based real‐time quality control (PBRTQC) exponentially weighted moving average (EWMA) procedure with intelligent QC rules in the artificial intelligence monitoring (AI‐MA) platform could accurately provide warning signals before the emergence of realistic warning events caused by machine malfunction, which proved that the PBRTQC procedure built up by the AI‐MA platform could be used as a complementary quality control tool for the traditional IQC method to monitor the overall process of experiment.

1. Introduction

Down syndrome (DS), also known as trisomy 21‐syndrome, is the commonest chromosomal aneuploid disease in live born babies with low intelligence, and the birth incidence rate of DS was approximately 1 in 800 worldwide [1]. Up to now, the pathogenesis was not clear and currently had no efficient treatment means for DS [1]. Therefore, antenatal screening was developed and gradually became the most common used way to detect DS fetus through detecting some maternal serum markers to distinguish between trisomy 21 and euploid fetuses since 1980s [2, 3]. Based on the importance of prenatal serum screening, all those serum markers were required strict exhaustive quality control according to quality indicators for helping to acquire accurate screening results. Usually, the measurement of dedicated internal quality control (IQC) materials is the most common quality control tool to monitor the performance of routine experiment. However, as we know, IQC products contained many limitations: (i) matrix effect and noncommutability [4]; (ii) providing false‐positive or false‐negative error, due to the characteristic of noncommutability [5]; (iii) poor stability; (iv) inappropriate detection range: The analytical concentration of IQC products is too high or too low, for which real patients' data could not reach; (v) discontinuity of quality monitoring: IQC materials is always detected only one time [6]. These limitations would lead to false alarms or failure to detect true alarms quickly, which finally influence the accuracy of detection results.

According to the Clinical and Laboratory Standards Institute (CLSI) document EP23‐A‐Laboratory Quality Control Base on Risk Management: Approved Guidelines, clinical laboratory should establish an effective quality control plan for particular measuring system, laboratory, and clinical environment. And the application of quality control method based on patient data is encouraged to be used as the technique can eliminate matrix effects and detect systematic error [7]. In this light, the use of patient samples for the purpose of continuous quality control, also known as patient‐based real‐time quality control (PBRTQC), was suggested in medical laboratories by the International Federation of Clinical Chemistry (IFCC) and Laboratory Medicine PBRTQC Working Group [8]. Studies have found that PBRTQC contained many advantages compared with the traditional IQC method, including (i) continuous quality control of the whole detection system; (ii) no matrix effect; (iii) no additional cost; (iv) more sensitive to the identification of pre‐analysis errors; and (v) can be used for monitoring those items without quality control materials [9, 10, 11, 12, 13]. It has been shown for large commercial laboratories that incorporating validated the PBRTQC procedures into the QC plan could result in a 75%–85% reduction in IQC [4, 11]. However, there is no relative research about the PBRTQC application in prenatal serum screening.

Up to now, there are many well‐established PBRTQC techniques like Bull's algorithm, average of normals (AON), moving median, moving average (MA), and exponentially weighted moving average (EWMA) [14, 15, 16, 17, 18, 19, 20]. In those procedures, the EWMA procedure was found and can sensitively identify small variations in the inaccuracy or imprecision of the inspection analysis process due to the introduction of weighting coefficient, which outperforms AON and Bull's algorithm [19, 21]. Unfortunately, although Hoffman and Waid had suggested using the PBRTQC AON method as a complementary quality control method for clinical chemistry testing as early as in 1965, many clinical chemistry laboratories do not routinely use this methodology, mainly because setting up these rather laboratory‐specific PBRTQC procedures is perceived as complex and it cannot be applied in practice without a crucial set of software functionalities [15, 22]. But in consideration of the importance of maternal serum screening, in this research, we tried to build a PBRTQC procedure using artificial intelligence software to continuously monitor the testing quality of maternal serum screening in order to improve the detection accuracy of maternal serum screening by combining with the IQC method.

2. Materials and Methods

2.1. Data Collection

Maternal serum screening contained five serum analytes, including two markers (free β subunit of human chorionic gonadotropin [β‐HCG] and pregnancy‐associated plasma protein A [PAPP‐A]) in the first trimester (9–13⁺⁶ weeks gestation), and three markers (β‐HCG [labeled as m‐β‐HCG, distinguished from the β‐HCG in the first trimester], α‐fetoprotein [AFP], and unconjugated estriol [uE₃]) in the second trimester (15–20⁺⁶ weeks gestation). The determination data of five antenatal screening serum analytes and relative routine IQC data in Xiaolan People's Hospital of Zhongshan from January 4 to September 7 in 2021 were collected, which were detected by the Wallac Auto DELFIA 1235 resolved fluorescence immunoassay analyzer (PerkinElmer) and the corresponding reagent. As target population were pregnant women from outpatients and relatively homogeneous, all serum screening data were collected and not grouped, but except for patients: (i) with twin pregnancies, (ii) cancers, and (iii) liver diseases such as liver cirrhosis. All the units of test results were unified as β‐HCG and m‐β‐HCG for ng/mL, PAPP‐A for mU/L, AFP for U/mL, and uE₃ for nmol/L. This research was approved by the ethical committee of the Xiaolan People's Hospital of Zhongshan. Moreover, all collected data in this research were in strict compliance with relevant laws and regulations as well as IFCC PBRTQC recommendations for desensitization and privacy protection of patient information.

2.2. IQC Management

The IQC materials contained the first (SERO Maternal Health Control‐Early; Wallac Oy) and second (Lyphochek Maternal Serum Control; Bio‐Rad Laboratories Inc.) trimester control materials, and each stage of materials contained three concentration levels (L1, L2, and L3). In this research, the IQC materials of the first trimester contained two different lot numbers (Lot1.2000206645 and Lot2.2000266715) and the second trimester also contained two (Lot1.39190 and Lot2.89400). In the process of experiment, the corresponding stage of IQC materials was performed with daily specimens simultaneously. Only when all IQC results were under control, which followed the Westgard Sigma rules (1‐3s, 2‐2s, and R4s), could the serum screening data be used for the later value risk calculation. Once a level of IQC triggered the above control quality rules, five patient serum samples would be selected and retested together with the next batch of samples. Only when more than 80% of those five serum samples had a coefficient of variation (CV) between two detection times < 10%, could we issue the testing report. Moreover, instrument calibration and maintenance were executed according to the clinical laboratory management specification management requirements.

2.3. PBRTQC Professional Intelligent Software Platform and Quality Control Rules

All the collected data were also monitored by the artificial intelligence monitoring platform (AI‐MA; Shanghai Senxu Medical Technology Co. Ltd.) using the EWMA and MA algorithms. The calculated formulas of EWMA and MA are as follows:

$$\text{EWMA}:{\tilde{Z}}_{t+1}={\tilde{Z}}_t+\lambda {\epsilon }_t=\lambda Z_t+\left(1-\lambda \right){\tilde{Z}}_t$$

${\tilde{Z}}_{t+1}$ was the estimated value of the t +  1 point; ${\tilde{Z}}_t$ was the t point estimated value; $Z_t$ was the actual value of the t point; and $\lambda$ was a weighting coefficient (0 < $\lambda$ ≤ 1).

$$\begin{array}{c}\mathrm{MA}:Z_{\left(t\right)}=\frac{X_{\left(t\right)}}{n}+\frac{X_{\left(t-1\right)}}{n}+\frac{X_{\left(t-2\right)}}{n}+\cdots +\frac{X_{\left(t-n+1\right)}}{n}\hfill \end{array}$$

$Z_{\left(t\right)}$ was the t point calculated average value; $X$ was the raw detecting data; and n was the batch size.

The quality rules used were the intellectual QC rules based on the AI‐MA platform, which were confidential for the corporation, and the traditional Westgard Sigma rules (1‐3s and 2‐2s).

2.4. Quality Control Objectives

The settings of total allowance error (TEA), precision, and accuracy were according to the CLSI 2023 External Quality Assessment Programs in Laboratory Medicine (https://www.nccl.org.cn/planEn). The TEA of all five analytes was required to be at ± 30%. In this research, quality targets of five analytes were set as follows: the TEA of PAPP‐A, β‐HCG and m‐β‐HCG were set as ± 25%, AFP was ± 24%, and uE₃ was ± 20%.

2.5. Establishment of PBRTQC EWMA and MA Procedures

The process of establishing optimal PBRTQC protocols in the AI‐MA platform is illustrated in Figure 1. Briefly, the establishment process contained normal distribution analysis, quality target setting, quality control rule setting, data automation, extraction, parameter setting, intelligent operation, performance verification, optimal PBRTQC method selection, and real‐time operation. Normal distribution analysis of collected data was firstly processed on the AI‐MA platform using the Kolmogorov–Smirnov test method, which helped to confirm the optimal truncated concentration range of test data. Then, other parameters such as the optimal $\lambda$ and n also should be set to make the established PBRTQC program consistently showing the change in the trend of real data. Probability for error detection (Ped), false‐positive rate (FPR), and false‐negative rate (FNR) as the evaluation index for the PBRTQC‐based procedures in the AI‐MA platform were employed. When the Ped > 90%, FPR < 5%, and FNR = 0%, the parameters on that procedure were the optimal parameters that we need.

FIGURE 1.

Flow chart of setting up and verification of the PBRTQC procedure. EWMA, exponentially weighted moving average; IQC, internal quality control; LIS, laboratory information system; MA, moving average; n, batch size; PBRTQC, patient‐based real‐time quality control; λ, weighting coefficient.

2.6. PBRTQC Chart

All the collected patient test result data and IQC data were converted into the z‐score charts. The formula is shown as follows: $z=(x-\overline{x})/\mathrm{SD}$ ($x$ is the daily detected value; $\overline{x}$ is the mean value; and SD is the standard deviation of all collected values).

2.7. Efficiency Evaluation of Established Optimal PBRTQC EWMA Procedures

First, the accumulating CV was calculated through the estimated values that were transformed by the established PBRTQC procedures and then comparing the accumulating CV with TEA we set. Second, comparing the patient data tendency transformed by the PBRTQC EWMA procedure and the traditional IQC data tendency in a z‐score chart. Then, the time node of early warning signal and relative causes were analyzed according to experimental records.

3. Results

3.1. Optimization of the PBRTQC EWMA Procedures in the AI‐MA Platform

From January 4 to September 7 in 2021, 4035 of the first trimester and 3435 of the second trimester serum data of pregnancy women aged from 15 to 45 years were collected. But because the population of maternal serum screening were relatively homogeneous that all were pregnant women from obstetrics clinic, the collected data in our research were not grouped like other studies [23, 24]. To establish the EWMA procedure, the two major parameters—concentration ranges and weighting coefficient—should be firstly confirmed. Taking β‐HCG as an example, the total and truncated concentration range (5–131 ng/mL) and the corresponding different weighting coefficient (0.01, 0.03, 0.05, 0.1, 0.2 and 0.4) are presented in Table 1. According to the results of Table 1, with the truncated concentration range of 5–131 ng/mL and a weighting coefficient of 0.03, the EWMA procedures could show optimal efficiency with 100% Ped, 0% FPR, and 0% FNR. Using the 5–131 ng/mL truncated range, we then compared the efficiency of the EWMA procedures ($\lambda$ was 0.03) between using intelligent quality control rules exploited by the AI‐MA platform and the traditional Westgard Sigma rules (1‐3s and 2‐2s) (Table 2). As we expected, the PBRTQC EWMA procedures with intelligent quality control rules showed optimal efficiency (100% Ped, 0% FPR, and 0% FNR), whereas the efficiency with traditional rules was poor (20.14% Ped, 3.74% FPR, and 79.86% FNR). In addition, the MA procedure (n was set as 20), another common method used for the PBRTQC algorithm, was also built with the above two kinds of quality control rules (Table 2). However, using neither the intelligent nor the Westgard Sigma quality rules could show such excellent efficiency like the EWMA procedures with intelligent control rules. Hence, we employed the PBRTQC EWMA procedures with intelligent quality rules in the AI‐MA platform as the optimal procedures. Accordingly, the optimal truncated concentration range and weighting coefficient for other four analytes (PAPP‐A, m‐β‐HCG, AFP, and uE₃) were also confirmed (Table 3). Except for AFP, for which the optimal $\lambda$ was 0.05, all other (PAPP‐A, m‐β‐HCG and uE₃) $\lambda$'s were set as 0.03. And the optimal truncated concentration range for those analytes are as follows: PAPP‐A, m‐β‐HCG, AFP, and uE₃ accordingly were 111–6628 mU/L, 3.5–32 ng/mL, 18–63 U/mL and 1.7–11 nmol/L, respectively.

TABLE 1.

Ped, FPR, and FNR of the EWMA procedures for β‐HCG with different truncated concentration range and different weighted coefficients.

Truncation range of β‐HCG	λ	No. positive bias of true results	No. positive bias‐based EWMA procedure				No. negative bias of true results	No. negative bias‐based EWMA procedure				No. EWMA procedure		No. true results		Total true results	Ped (%)	FPR (%)	FNR (%)
			EW	TP	FP	FN		EW	TP	FP	FN	FP	FN	TP	TN
Total concentration range	0.01	0	0	0	0	0	144	0	0	0	144	0	144	144	3891	4035	0.00	0.00	100.00
	0.03	0	123	0	123	0	144	190	81	109	63	232	63	144	3891	4035	56.25	5.96	43.75
	0.05	0	143	0	143	0	144	334	133	201	11	344	11	144	3891	4035	92.36	8.84	7.64
	0.1	0	48	0	48	0	144	131	6	125	138	173	138	144	3891	4035	4.17	4.45	95.83
	0.2	0	50	0	50	0	144	81	12	69	132	119	132	144	3891	4035	8.33	3.06	91.67
	0.4	0	60	0	60	0	144	154	11	143	133	203	133	144	3891	4035	7.64	5.22	92.36
5–131 ng/mL	0.01	0	0	0	0	0	144	0	0	0	144	0	144	144	3395	3539	0.00	0.00	100.00
	0.03	0	0	0	0	0	144	144	144	0	0	0	0	144	3395	3539	100.00	0.00	0.00
	0.05	0	146	0	146	0	144	52	49	3	95	149	95	144	3395	3539	34.03	4.39	65.97
	0.1	0	3	0	3	0	144	59	3	56	141	59	141	144	3395	3539	2.08	1.74	97.92
	0.2	0	0	0	0	0	144	51	9	42	135	42	135	144	3395	3539	6.25	1.24	93.75
	0.4	0	23	0	22	0	144	81	8	73	136	95	136	144	3395	3539	5.56	2.80	94.44

Note: Ped (%) = No. TP (positive bias plus negative bias of EWMA procedure)/(No. TP (positive bias plus negative bias of EWMA procedure) + No. FN (EWMA procedure)); FPR (%) = No. FP (EWMA procedure)/No. TN (true results); FNR (%) = No. FN (EWMA procedure)/No. TP (true results).

Abbreviations: EW, early warning; FN, false negative; FP, false positive; No., number; TN, true negative; TP, true positive.

TABLE 2.

Comparison the PBRTQC EWMA and MA procedures built on the AI‐MA platform.

Truncation range	λ/n	Quality rules	No. positive bias of true results	No. positive bias‐based EWMA/MA procedure				No. negative bias of true results	No. negative bias‐based EWMA/MA procedure				No. EWMA/MA procedure		No. true results		Total true results	Ped (%)	FPR (%)	FNR (%)
				EW	TP	FP	FN		EW	TP	FP	FN	FP	FN	TP	TN
β‐HCG (5–131 ng/mL)	0.03	Intelligent	0	0	0	0	0	144	144	144	0	0	0	0	144	3395	3539	100.00	0.00	0.00
		Westgard	0	75	0	75	0	144	81	29	52	115	127	115	144	3395	3539	20.14	3.74	79.86
	20	Intelligent	0	164	0	164	0	144	364	9	355	135	519	135	144	3395	3539	6.25	15.29	93.75
		Westgard	0	423	0	423	0	144	414	79	335	65	758	65	144	3395	3539	54.86	22.33	45.14

Note: Intelligent quality rules were the rules developed and used in the AI‐MA platform; Westgard‐Sigma quality rules were 1‐3s and 2‐2s; Ped (%) = No. TP (positive bias plus negative bias of EWMA/MA procedure)/(No. TP (positive bias plus negative bias of EWMA/MA procedure) + No. FN (EWMA/MA procedure)); FPR (%) = No. FP (EWMA/MA procedure)/No. TN (true results); FNR (%) = No. FN (EWMA/MA procedure)/No. TP (true results).

Abbreviations: EW, early warning; FN, false negative; FP, false positive; No., number; TN, true negative; TP, true positive.

TABLE 3.

Optimal weighting coefficient and truncated concentration range for different analyte.

Analyte	No. test data	λ	Truncation range
β‐HCG	3539	0.03	5–131 ng/mL
PAPP‐A	3501	0.03	111–6628 mU/L
m‐β‐HCG	2971	0.03	3.5–32 ng/mL
AFP	3047	0.05	18–63 U/mL
uE3	3386	0.03	1.7–11 nmol/L

3.2. The Trend of the z‐Score Chart Transformed From Daily Detection Data Based on Optimal PBRTQC EWMA Procedure Was Basically Consistent With That of Traditional IQC Data

Using the optimal parameters we set before for all five analytes, the z‐score charts transformed from all the collected daily detection data–based PBRTQC EWMA procedure for each analyte were drawn and compared with those of traditional IQC data. As shown in Figure 2A,B, the trend of the daily detection data in the z‐score charts (β‐HCG and AFP) transformed by the optimal PBRTQC EWMA procedure was basically consistent with that of the traditional IQC data. Similar consistent trend was also shown in other three items (PAPP‐A, m‐β‐HCG, and uE₃) (Figure S1). Moreover, the accumulation CVs of all five analytes were less than the 1/4TEA (TEA were ± 30%) set by the National Center for Clinical Laboratories (Figure 2C). The CV of AFP and uE₃ was even up to 1/5TEA (Figure 2C). All these indicated the optimal PBRTQC EWMA procedure in the AI‐MA platform possessed well stability, which may be used to monitor the testing quality of maternal serum markers by combing traditional IQC.

FIGURE 2.

The distribution trends of truncated daily data and IQC data using the z‐score charts. Truncated daily detection data of β‐HCG and AFP calculated by the PBRTQC EWMA intelligent procedure are plotted with green solid lines, and different levels of IQC data are shown with different colors of triangles in the z‐score charts. The truncated daily data of analytes (A) β‐HCG and (B) AFP and the corresponding IQC data are shown in the z‐score charts. (C) The percentage of PBRTQC accumulation CV of five analytes; TEA was the allowable total error ± 30%.

3.3. The Optimal PBRTQC EWMA Procedure Accurately Provided Early‐Warning Signal Before the Occurrence of Instrument Failure

By employing the optimal PBRTQC EWMA procedure, the z‐score chart of β‐HCG came out as a negative‐bias early warning signal between June 11 and June 23, 2021 (Figure 3A). Retrospective analysis found that the left reagent mechanical arm was broken in June 18, 2021, which leaded the experiment interrupted. Until June 23, the broken mechanical arm was repaired but poor tightness was found when adding reagent. Hence, the experimental procedure was changed to use the right reagent mechanical arm, and later, the negative bias was finally corrected. At the same time, the PAPP‐A item also was detected. But because its corresponding detection procedure only used the right reagent mechanical arm, no early warning signals were provided during that period. Similarly, in the z‐score chart of uE₃, we also found an early warning signal between June 24 and July 1, 2021 (Figure 3B). Retrospective analysis showed that the sample needles were found position bias and wear blockage, which may lead to the above warning. After new sample needles were installed and the parameters of testing procedure were readjusted on July 8, the bias was also corrected. Combining the realistic warning events, it indicated that the optimal PBRTQC EWMA intelligent procedure in the AI‐MA platform could provide accurate warning signals before the emergence of realistic warning events caused by machine malfunction. Interestingly, during the period of early‐warning occurrence, the IQC did not trigger the quality control rules, such as 1‐3s or 2‐2s, which also proved the PBRTQC EWMA intelligent procedure could be used as a complementary quality control tool for traditional IQC to monitor the overall process of experiment.

FIGURE 3.

Early‐warning signal provided by the optimal PBRTQC EWMA procedure in the AI‐MA platform. Solid line with yellow color in the z‐score charts represent the early‐warning signal. (A) One negative‐bias early‐warning signal of β‐HCG between June 11 and June 23, 2021. (B) One positive‐bias early‐warning signal of uE₃ between June 24 and July 1, 2021.

3.4. Application of EWMA Intelligent Procedure in the AI‐MA Platform Could Help to Identify the Testing Quality When Random Error Occurred in Traditional IQC

When the testing process finished in our routine experiment, sometimes the detection results of one or more than one levels of IQC materials would trigger the 1‐3s or R4s control quality rules. In this case, serum samples of five patients would be selected and retested together with the next batch of samples. Only if more than 80% of those five serum samples had a CV% between two detection times less than 10%, could we issue the testing report. But most of the time, the comparison experiments were successful. As shown in Figure 4, two time points of IQC out of control events were found in PAPP‐A in June 16 and August 20, 2021, but the corresponding daily data in the z‐score chart transformed by the PBRTQC EWMA intelligent procedure did not provide any warning signal and the comparison experiment record at that time was also successful (Table S1). Because the control materials we used were dissolved and packed into Eppendorf tubes, we thought those random errors were probably caused by the quality of control materials as those packed control materials would be stored at ‐30°C for more than one month. Therefore, we believed that the PBRTQC EWMA intelligent procedure could truly reflect the stability of detection system.

FIGURE 4.

Exhibition of the out‐of‐control points of PAPP‐A IQC in the z‐score chart. Triangles marked by red circles represented the out‐of‐control points of IQC. In the z‐score chart, one level and two levels of the PAPP‐A IQC data were out of the 1‐3S quality control line in 2021 June 16 and August 20, respectively.

3.5. The PBRTQC EWMA Intelligent Procedure in the AI‐MA Platform Could Help to Monitor the Detection Accuracy of Experiment When IQC Materials With Inappropriate Concentration Range

According to the requirement of ISO 15189:2012 (https://www.iso.org/standard/56115.html), the use of independent, third party control materials should be considered, either instead of, or in addition to, any materials supplied by the reagent or instrument manufacture. Moreover, the standard also suggested that laboratories should choose concentrations of control materials, wherever possible, especially at or near clinical decision values, which ensure the validity of decisions made. However, the concentration of IQC materials used for the second trimester prenatal screening in our laboratories was not so suitable although it was supplied by the independent third party. In this study, the highest level of clinical concentration of m‐β‐HCG was from 3.5 to 29.5 ng/mL and of uE₃ was from 1.7 to 11 nmol/L, which we chose as the optimal truncated concentration ranges for the PBRTQC EWMA intelligent procedure (Figure 5A,B), but the corresponding lowest concentration level of IQC materials we used for m‐β‐HCG was about 80 ng/mL, and the newest lot number (89400) were almost up to 120 ng/mL, let alone the higher two levels of control materials (Figure 5C). Similarly, the middle and high level concentrations of IQC materials for uE₃ were also too high, which could not be observed routinely at all although the IQC materials showed good stability (the SD values being small) (Figure 5D). However, PBRTQC procedures based on clinical patient data had no such problem and the trend of the z‐score chart transformed from m‐β‐HCG (Figure 5E) and uE₃ (Figure 5F) daily detected data based on PBRTQC procedures showed relatively smooth, which could help to truly monitor and reflect the detection quality. All these indicated that when the concentrations of traditional IQC materials could not meet the clinical needs, the PBRTQC EWMA intelligent procedure we built based on the daily detection data could be a good complementary quality tool to monitor the examining system.

FIGURE 5.

Concentration distribution of collected daily data and the corresponding IQC data. The chart of concentration frequency distribution of (A) m‐β‐HCG and (B) uE₃ from daily data. The IQC concentration of (C) m‐β‐HCG and (D) uE₃; L1, L2, and L3 represented the three concentration levels of IQC; Lot.1 and Lot.2 represent the IQC lot number 39190 and 89400, respectively. The daily detected data of (E) m‐β‐HCG and (F) uE₃ were transformed into the z‐score charts by the optimal PBRTQC EWMA procedure in the AI‐MA platform.

4. Discussion

In this study, we aimed to establish the optimal PBRTQC models for prenatal serum screening and assess their potential role in continuous quality monitoring in daily work. Results demonstrated that PBRTQCs could be applied as an efficient supplementary tool for quality monitoring test of prenatal serum screening by combing with the traditional IQC method.

More recently, with the continuous improvement in laboratory information systems and the development of “big data” technology, professional software for the PBRTQC models was exploited and improved, promoting the application of PBRTQCs in the quality management of clinical biochemistry [25]. However, the PBRTQC method actually still faced many challenges, like the selection and setting up of the PBRTQC procedure, performance verification and optimization, and how to realize dynamic and continuous monitoring. Hence, Badrick et al. [26] suggested that it is necessary to change the statistical process to artificial intelligence method in the future for the application of PBRTQC, like the application of artificial software. In addition, the software should be capable of selecting different algorithms for the best performance since using certain algorithms may positively or negatively affect the PBRTQC performance [25, 27]. In the present study, we firstly build up the PBRTQC models in prenatal serum screening using commercial AI‐MA software, which was developed based on big data processing by artificial intelligence technology. A published study explored the PBRTQC models in clinical laboratory using that commercial AI‐MA software, and results suggested the PBRTQC models built by that platform could be used as a complementary quality control tool for a traditional quality control method [23]. Using the optimal PBRTQC models built by that commercial AI‐MA software, simulation results were concordant with those obtained with serum samples. In general, the application of the intelligent platform could help inspectors more effectively to set up and select the optimal model for the laboratory.

In the previous research [28], specimens from diseases that had significant impacts on the data were excluded. Moreover, data also would be divided into different groups, such as outpatient and inpatient, to build up the corresponding procedures [23, 24]. In our research, we also excluded some data, such as women with twin pregnancies or with liver cancer, since expression levels of makers (β‐HCG and AFP) would be influenced. However, because the population of maternal serum screening was relatively homogeneous in such a way that all were pregnant women from obstetrics clinics, the collected data in our research were not grouped.

A notable finding was that the PBRTQC model using the intellectual QC rules built by that AI‐MA platform is better than the PBRTQC model using the traditional Westgard rules, especially the PBRTQC EWMA models with intelligent QC rules. Using that optimal PBRTQC EWMA model, early‐warning signals induced by the broken of mechanical arm and wear of reagent needles were provided. Similarly, published research also had verified the PBRTQC EWMA intelligent procedure built by the AI‐MA platform, which could sensitively and accurately identify the quality risks arising from reagents, instrument failures and calibration, enabling intelligent monitoring and early warning throughout the time period [29]. van Rossum et al. [22] suggested that PBRTQC can be presented in either Levey–Jennings charts or accuracy/PBRTQC reference plots depending on the optimization method used. The commercial AI‐MA software we used presented patient's data and IQC data with the z‐score charts, which made the chart more intuitive. In addition, the EWMA algorithm was usually suggested to be used for detecting small‐to‐moderate systematic shifts [29, 30]. But in this research, it seems the PBRTQC EWMA intelligent procedure also could help us quickly to identify false random errors (1‐3s, R4s) since relatively smooth data tendency without warning signal were presented through the PBRTQC procedure. Hence, unnecessary duplicate testing and comparison work could be reduced because most of the random errors in this research we found were caused by the instability of IQC material. Moreover, the ability to detect patient's data could not be evaluated if the concentrations of traditional IQC materials are inappropriate, but there is no such problem in the PBRTQC models. In this research, two analytes (m‐β‐HCG and uE₃) of the second trimester confronted such problems as the concentration of IQC materials was too high even though all levels of them contained excellent stability. Hence, Xu and his colleagues suggested that the PBRTQC method could be as a complementary quality control tool for clinical chemical items that IQC materials used were not stable or had no commercial IQC products [12]. van Rossum [31] also recommended using PBRTQC if IQC is insufficient or inefficient. Therefore, although the traditional IQC method may remain the backbone, we still recommended that the PBRTQC models be conducted as a supplementary procedure to guarantee the accuracy of testing results of prenatal serum screening and probably other quantitative laboratory items in the future for containing the following advantages: (i) providing proper real‐time notification of alarms caused by instruments or reagents; (ii) quickly identifying false random error signal caused by IQC materials itself; and (iii) helping to monitor the testing quality when commercial IQC products are inappropriate.

However, there are some limitations in this study. First, the intellectual QC rules to assess the biased data points in the chart were confidential for the corporation and hence could not be presented in detail. In addition, the AI‐MA platform has not yet been operating in our laboratory and further monitoring efficacy could not be provided even though this study should be continued for a longer time. In conclusions, further studies should be conducted to optimize and validate the PBRTQC model to improve the detection accuracy of serum markers and then increasing the stability of the testing report we issued since the accuracy of serum markers would greatly influence the final risk value of DS.

Author Contributions

Xuran Yang: Conceived and designed the experiments, analyzed and interpreted the data, and wrote the paper. Qianlan Chen: Conceived and designed the experiments, performed the experiments, and analyzed and interpreted the data. Jingmao Cheng and Zhifeng Pan: Performed the experiments. Wenting Zheng, Yingliang Liang, and Hui Chen: Conceived and designed the experiments, and contributed reagents, materials, analysis tools, or data. Guanghui Chen: Conceived and designed the experiments, contributed reagents, materials, analysis tools, or data, and wrote the paper. Wandang Wang: Analyzed and interpreted the data, contributed reagents, materials, analysis tools, or data, and wrote the paper.

Ethics Statement

This research was approved by the Ethical Committee of the Xiaolan People's Hospital of Zhongshan. Moreover, all collected data in this research were in strict compliance with relevant laws and regulations as well as IFCC PBRTQC recommendations for desensitization and privacy protection of patient information.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1.

Table S1.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.