Review

Methods

Protocol Registration and Reporting Guidelines

This systematic review and meta-analysis was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines [9]. The review protocol was prospectively registered in the International Prospective Register of Systematic Reviews (PROSPERO; CRD420261359922) before the initiation of the literature screening.

Search Strategy and Information Sources

A systematic search of the literature was performed across major electronic databases, including PubMed/MEDLINE, Embase, Cochrane Central Register of Controlled Trials (CENTRAL), Scopus, Web of Science, and Science Citation Index Expanded (SCIE). The grey literature was searched to mitigate publication bias. The search strategy utilized a combination of controlled vocabulary (e.g. medical subject headings (MeSH) and Emtree terms) and free-text keywords constructed using Boolean operators (AND, OR, NOT) and truncations. The key search terms included "Leukemia", "Minimal Residual Disease", "Measurable Residual Disease", "Flow Cytometry", "Polymerase Chain Reaction", "Diagnostic Accuracy", and "Prognosis". No language or date restrictions were imposed. Furthermore, forward citation searching (snowballing) and backward reference list checking of the included studies and relevant reviews were performed to identify additional eligible literature.

Eligibility Criteria

The study inclusion was determined using a predefined PICOS (Population, Intervention/Index test, comparator/reference standard, Outcomes, Study design) framework. The population was patients of any age (pediatric or adult) with a confirmed diagnosis of leukemia, including acute ALL, AML, chronic lymphocytic leukemia (CLL), and chronic myeloid leukemia (CML). The intervention/index test was MRD assessment performed using MFC, and the comparator/reference standard was MRD assessment performed using PCR-based methodologies (e.g. RQ-PCR, RT-qPCR, allele-specific oligonucleotide PCR).

For diagnostic test accuracy (DTA), outcomes included sensitivity, specificity, positive/negative predictive values (PPV/NPV), diagnostic odds ratio (DOR), and area under the curve (AUC). For prognostic value, outcomes included overall survival (OS), event-free survival (EFS), progression-free survival (PFS), disease-free survival (DFS), and cumulative incidence of relapse. Non-randomized study types, including prospective and retrospective cohorts, longitudinal observational studies, and diagnostic accuracy studies, were included. We excluded case reports, series with <10 participants, animal studies, and reviews.

Study Selection and Data Extraction

Following the removal of duplicates, two investigators independently screened the titles and abstracts for eligibility. Full-text articles of potentially eligible studies were then independently evaluated, and disagreements were resolved by adjudication by a third reviewer. The inter-rater reliability for study inclusion was quantified using Cohen’s Kappa coefficient (κ) [10].

Data were independently extracted using a standardized pilot-tested extraction form. The extracted variables included study and patient characteristics, leukemia subtype, MRD methodologies, specific targets, timing of assessment, MRD thresholds, 2×2 contingency tables (true positive, false positive, false negative, true negative), and survival data (hazard ratios (HRs) and 95% confidence intervals (CIs)). If HRs were not explicitly reported, they were estimated from Kaplan-Meier survival curves using the Tierney method [11].

Quality Assessment and Risk of Bias

The methodological quality and risk of bias (RoB) of the included studies were appraised by two independent reviewers using tools specific to the study design. For diagnostic accuracy outcomes, the Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2) tool was used to evaluate bias across four domains: patient selection, index test, reference standard, and flow and timing [12].

Statistical Analysis and Data Synthesis

All statistical analyses were conducted using R software version 4.6.0 (R Foundation for Statistical Computing, Vienna, Austria), utilizing the meta, metafor, and mada packages [13].

Diagnostic accuracy meta-analysis: To account for the intrinsic correlation between sensitivity and specificity across varying study thresholds, a bivariate random-effects model was employed to calculate the pooled sensitivities, specificities, positive likelihood ratios (PLR), negative likelihood ratios (NLR), and DORs [14]. A hierarchical summary receiver operating characteristic (HSROC) curve was generated to visualize the overall diagnostic performance and calculate the AUC [15].

Prognostic value meta-analysis: For time-to-event outcomes (OS, EFS, DFS, and PFS), HRs and their corresponding standard errors were natural log-transformed and pooled using the restricted maximum likelihood (REML) random-effects model to account for anticipated between-study clinical heterogeneity [16]. Dichotomous outcomes (e.g. relapse rates) were pooled as risk ratios (RRs) or odds ratios (ORs) using the DerSimonian-Laird random-effects model [17].

Heterogeneity and subgroup analysis: Statistical heterogeneity was quantified using Cochran’s Q test (with p<0.10 indicating significance), the I2 statistic (where values >50% represent substantial heterogeneity), and the τ2 statistic [18]. To explore the sources of heterogeneity, predefined subgroup analyses and univariate meta-regression were conducted based on leukemia subtype (ALL vs. AML), patient age (pediatric vs. adult), MRD assessment timing (e.g. post-induction vs. pre-transplant), and MRD positivity thresholds (e.g. 0.1% vs. 0.01%). Leave-one-out sensitivity analyses were performed to evaluate the robustness of pooled estimates.

Publication Bias and Certainty of Evidence

Publication bias and small-study effects were visually evaluated using contour-enhanced funnel plot. Statistical asymmetry was tested using Egger’s linear regression [19] and Begg’s rank correlation tests [20] when 10 or more studies were available for a specific outcome. The trim-and-fill method was applied to adjust the pooled estimates if publication bias was detected. The overall certainty of the evidence for both diagnostic and prognostic outcomes was graded independently by two reviewers using the Grading of Recommendations Assessment, Development, and Evaluation (GRADE) framework, which was adapted for diagnostic test accuracy [21]. Disagreements were resolved by adjudication by a third reviewer.

Results

Literature Search and Study Selection

The initial database and registry search yielded 3,047 records screened. After removing 669 duplicates and 274 records for other automated reasons, 2,104 unique titles and abstracts were screened. Of these, 1,997 records were excluded for not meeting the predefined eligibility criteria. Full-text retrieval was sought for 107 reports; 83 were not retrieved (predominantly because they were conference abstracts lacking full-length peer-reviewed publications with extractable data), leaving 24 full-text articles assessed for eligibility. One report was excluded because of insufficient extractable data. Twenty-three non-randomized studies met all inclusion criteria and were incorporated into the systematic review, with data from 12 studies providing 2×2 contingency data for the DTA meta-analysis and six studies providing effect sizes for the prognostic meta-analysis (Figure 1).

Figure 1: PRISMA 2020 flow diagram.PRISMA: Preferred Reporting Items for Systematic Reviews and Meta-Analyses.

Characteristics of the Included Studies

The 23 included studies encompassed a diverse international cohort of patients with leukemia treated across various clinical settings [6-8,22-41]. The studies spanned pediatric, adult, and infant populations, capturing a broad spectrum of hematologic malignancies, including B-cell precursor acute lymphoblastic leukemia (BCP-ALL), T-cell ALL, and AML with specific genetic aberrations (e.g. NPM1 mutations and core-binding factor (CBF) translocations). Methodologically, the index test MFC utilized four- to eight-color panels with positivity thresholds set between 0.01% and 0.1%. The reference standard (PCR-based methodologies, including RQ-PCR and RT-qPCR) evaluated leukemia-specific fusion gene transcripts (FGT) or clonal Ig/TCR rearrangements, with sensitivity thresholds of 0.001%-0.01%. Although CLL and CML were included in the initial search strategy, all studies that met the inclusion criteria for paired MFC and PCR assessments with extractable quantitative data were restricted to ALL and AML cohorts. The baseline characteristics of the included studies are presented in Table 1.

Study (Year) | Country | Study Design | Population | Leukemia Subtype | Index Test (MFC/FCM) Threshold | Reference Test (PCR) Threshold | N (Patients/Samples)
Hrabovsky et al. (2018) [6] | Czech Republic | Retrospective | Adult | ALL | 8-color (0.1%) | RQ-PCR (Ig/TCR/FGT)(0.1%) | 56/103
Gaipa et al. (2012) [7] | Europe (AIEOP-BFM) | Prospective | Pediatric | ALL | 4-color (0.01%) | RQ-PCR (Ig/TCR) (0.01%) | 1547/3565
Malec et al. (2004) [8] | Sweden/Netherlands | Retrospective | Pediatric | ALL | 3-/4-color (0.01%) | RQ-PCR (Ig/TCR) (0.01%) | 22/93
Denys et al. (2012) [22] | Netherlands/Belgium | Retrospective | Pediatric | BCP-ALL, T-ALL | 4-/6-color (0.01%) | RQ-PCR (Ig/TCR) (0.01%) | 363/898
Popov et al. (2021) [23] | Russia/Belarus | Retrospective | Infant | KMT2A-r ALL | Multicolor (0.01%) | RT-qPCR (FGT) (0.01%) | 123/942
Ramos Elbal et al. (2023) [24] | Spain | Retrospective | Pediatric | AML | MFC (0.1%) | RT-qPCR / FISH (0.1%) | 20/297
Inaba et al. (2012) [25] | USA (Multicenter) | Prospective | Pediatric | AML | 4-color (0.1%) | RQ-PCR (FGT) (0.1%) | 203/1514
Rossi et al. (2012) [26] | Italy | Retrospective | Adult | AML | 6-color (0.1%) | RQ-PCR (WT1) (90 copies) | 23/46
Karlsson et al. (2022) [27] | Nordic Countries | Retrospective | Pediatric | CBF/MLL AML | MFC (0.1%) | RT-qPCR (FGT) (0.1%) | 15/22
Wery et al. (2024) [28] | Belgium | Retrospective | Adult | AML / MDS | MFC (LAIP) | RT-qPCR (WT1/FGT) | 192/192
Rocha et al. (2019) [29] | Brazil | Retrospective | Pediatric | ALL | 4-color (0.01%) | Standard PCR (Qualitative) | 42/68
Jung et al. (2024) [30] | South Korea | Retrospective | Adult | B-ALL | MFC (0.01%) | qPCR (BCR-ABL) (0.01%) | 182/646
Chang et al. (2026) [31] | China | Retrospective | Pediatric | Ph+ B-ALL | MFC (LAIP) | RT-qPCR (0.01%) | 55/100
Hendricks et al. (2019) [32] | South Africa | Retrospective | Pediatric | B-ALL | MFC (EuroFlow) | qPCR (FGT) | 64/64
Shang et al. (2021) [33] | China | Retrospective | Adult/Ped | AML t(8;21) | 8-color (0.01%) | RQ-PCR (RUNX1) (0.001%) | 124/450
Thörn et al. (2011) [34] | Sweden (Multicenter) | Retrospective | Pediatric | ALL | 3-/4-color (0.01%) | RQ-PCR (Ig/TCR) (0.01%) | 228/726
Gao et al. (2020) [35] | China | Retrospective | Adult | AML (NPM1) | 8-color (0.01%) | RQ-PCR (NPM1) (>0) | 93/370
Neale et al. (2004) [36] | USA | Prospective | Pediatric | B-ALL | 3-color (0.01%) | RQ-PCR (Ig/TCR) (0.01%) | 227/1375
Huang et al. (2017) [37] | Taiwan | Retrospective | Pediatric | B-ALL (Fusion+) | 6-color (0.01%) | RQ-PCR (FGT) (0.01%) | 55/108
Kerst et al. (2005) [38] | Germany (Multicenter) | Retrospective | Pediatric | ALL | 4-color (0.01%) | RQ-PCR (Ig/TCR) (0.01%) | 45/105
Ryan et al. (2008) [39] | Ireland | Prospective | Pediatric | ALL | 3-color (0.01%) | RQ-PCR (Ig/TCR) (0.01%) | 53/151
Theunissen et al. (2017) [40] | Europe (EuroFlow) | Prospective | Pediatric | BCP-ALL | 8-color (0.001%) | RQ-PCR (Ig/TCR) (0.001%) | 178/377
Alm et al. (2017) [41] | Nordic Countries | Retrospective | Pediatric | ETV6-RUNX1 ALL | FACS (0.1%) | RT-qPCR (0.1%) | 29/78

Methodological Quality and Risk of Bias

Methodological quality for the 12 studies providing quantitative diagnostic test accuracy data was appraised using the QUADAS-2 framework, revealing a robust study design across the included literature (Figures 2,3). As QUADAS-2 is designed for diagnostic studies, the remaining 11 studies providing only prognostic outcomes were not assessed with this tool. The reference standard domain (PCR) was rated as having a low risk of bias (100%). However, a high risk of bias in patient selection was identified in 22% of the studies, driven by the exclusion of specific patient subsets or targeted evaluations (e.g. evaluation of specific translocation subgroups) [26,27]. Furthermore, 22% of studies exhibited an unclear risk of bias regarding the index test, as the blinding of the MFC interpreter to the PCR results was not documented [26,35].

Figure 2: Methodological quality assessment using the QUADAS-2 framework.Traffic-light plot of domain-level judgments for each individual study. The complete aggregate summary of risk of bias for all 12 studies included in the diagnostic accuracy meta-analysis is presented in Figure 3.Studies referenced in the plot: Denys et al. (2012) [22], Popov et al. (2021) [23], Inaba et al. (2012) [25], Rossi et al. (2012) [26], Karlsson et al. (2022) [27], Gaipa et al. (2012) [7], Neale et al. (2004) [36], Theunissen et al. (2017) [40], and Gao et al. (2020) [35].

Figure 3: Risk of Bias Summary.Bar plot displays the overall proportion of studies at low, high, or unclear risk of bias and applicability concerns.

Diagnostic Test Accuracy of MFC versus PCR-Based MRD

Data from 12 studies [7,8,22,23,25,33,34,36-40] reporting paired 2×2 contingency tables were synthesized using a bivariate random-effects model, with PCR as the reference standard. MFC demonstrated a moderate pooled sensitivity of 76.2% (95% CI, 55.2-89.3) and an excellent pooled specificity of 96.0% (95% CI, 93.2-97.7) (Table 2; Figure 4). The corresponding false-positive rate was remarkably low at 4.0% (95% CI, 2.3-6.8).

Diagnostic Metric | Pooled Estimate | 95% Confidence Interval | Statistical Model
Sensitivity | 0.762 (76.2%) | 0.552-0.893 | Bivariate (Reitsma) [14]
Specificity | 0.960 (96.0%) | 0.932-0.977 | Bivariate (Reitsma) [14]
Positive Likelihood Ratio (PLR) | 15.45 | 8.00-29.85 | Univariate DSL [17]
Negative Likelihood Ratio (NLR) | 0.237 | 0.168-0.334 | Univariate DSL [17]
Diagnostic Odds Ratio (DOR) | 79.54 | 25.18-251.24 | Univariate DSL [17]
HSROC AUC | 0.965 | pAUC: 0.87* | HSROC Regression [15]

Figure 4: Per-study sensitivity and specificity with exact (Clopper-Pearson) 95% confidence intervals.Coupled forest plots of per-study sensitivity (left) and specificity (right) with exact (Clopper-Pearson) 95% confidence intervals. Dashed lines indicate the bivariate pooled estimates.Studies referenced in the plot: Denys 2012 [22], Popov 2021 [23], Inaba 2012 [25], Malec 2004 [8], Gaipa 2012 [7], Shang 2021 [33], Thörn 2011 [34], Neale 2004 [36], Huang 2017 [37], Kerst 2005 [38], Ryan 2008 [39], and Theunissen 2017 [40].

Overall diagnostic discrimination was excellent, characterized by an HSROC curve of 0.965 (Figure 5). To facilitate clinical interpretation, likelihood ratios and DOR were computed using a univariate DerSimonian-Laird approach. MFC yielded a high PLR of 15.45 (95% CI, 8.00-29.85) and a NLR of 0.237 (95% CI, 0.168-0.334). The pooled DOR was 79.54 (95% CI, 25.18-251.24) (Figure 6), indicating that a positive MFC result strongly increases the probability of PCR-detectable residual disease, serving as a highly reliable rule-in test, whereas a negative MFC result provides only moderate rule-out capacity.

Figure 5: Hierarchical summary receiver operating characteristic (HSROC) curve for MFC versus PCR-based MRD.The dark central point represents the bivariate summary operating point (sensitivity 0.76, specificity 0.96) surrounded by its 95% confidence region; AUC=0.965.

Figure 6: Forest plot of study-level diagnostic odds ratios (DOR) plotted on a logarithmic scale, featuring the REML random-effects pooled estimate.Studies referenced in the plot: Denys et al. (2012) [22], Popov et al. (2021) [23], Inaba et al. (2012) [25], Malec et al. (2004) [8], Gaipa et al. (2012) [7], Shang et al. (2021) [33], Thörn et al. (2011) [34], Neale et al. (2004) [36], Huang et al. (2017) [37], Kerst et al. (2005) [38], Ryan et al. (2008) [39], and Theunissen et al. (2017) [40].

Heterogeneity, Subgroup Analyses, and Meta-Regression

Substantial between-study heterogeneity was observed in the log-DOR under the REML random-effects model (Cochran’s Q=276.67, df=11, p<0.0001; I2=94.4%; τ2=2.55). To elucidate the sources of this heterogeneity, pre-specified univariate meta-regression and subgroup analyses were conducted (Table 3).

Moderator / Subgroup | QM​ Test Statistic | P -value | Residual I2 | R2 (Heterogeneity Explained)
Leukemia Subtype (ALL vs. AML) | QM​=7.27 | 0.0070 | 91.8% | 36.7%
Age Group (Pediatric vs. Adult/Mixed) | QM​=4.39 | 0.111 | 92.7% | 13.6%
MRD Positivity Threshold (e.g., 0.1% vs 0.01%) | QM​=1.36 | 0.507 | 95.2% | 0.0%
Study Sample Size (logN) | Slope | 0.674 | 93.5% | —

Leukemia subtype emerged as the sole statistically significant moderator (QM​=7.27, p=0.0070). Specifically, AML studies exhibited a lower log-DOR than ALL studies (coefficient -3.18, 95% CI, -5.48 to -0.87), accounting for 36.7% of the observed heterogeneity. Variables, including age group (p=0.111), MRD threshold (p=0.507), and study sample size (p=0.674), did not significantly influence the diagnostic performance (Figure 7).

Figure 7: Meta-regression bubble plot of log-DOR against log sample size.Bubble area is proportional to study weight.Studies referenced in the plot: Denys et al. 2012 [22], Popov et al. 2021 [23], Inaba et al. 2012 [25], Malec et al. 2004 [8], Gaipa et al. 2012 [7], Shang et al. 2021 [33], Thörn et al. 2011 [34], Neale et al. 2004 [36], Huang et al. 2017 [37], Kerst et al. 2005 [38], Ryan et al. 2008 [39], and Theunissen et al. 2017 [40].

Sensitivity Analysis and Publication Bias

A leave-one-out sensitivity analysis confirmed the robustness of the diagnostic estimates. Sequentially omitting individual studies left the pooled DOR essentially unchanged, fluctuating within a narrow confidence band around the full-model estimate of ~79.5 (Figure 8).

Figure 8: Leave-one-out sensitivity analysis.Plot demonstrates the stability of the pooled diagnostic odds ratio (log scale) upon sequential omission of individual studies.Studies referenced in the plot: Denys et al. 2012 [22], Popov et al. 2021 [23], Inaba et al. 2012 [25], Malec et al. 2004 [8], Gaipa et al. 2012 [7], Shang et al. 2021 [33], Thörn et al. 2011 [34], Neale et al. 2004 [36], Huang et al. 2017 [37], Kerst et al. 2005 [38], Ryan et al. 2008 [39], and Theunissen et al. 2017 [40].

The publication bias for diagnostic accuracy was visually assessed using a Deeks’-style funnel plot (Figure 9). Statistical evaluation revealed a significant Egger’s linear regression test (t=2.43, df=10, p=0.035), although Begg’s rank correlation test remained non-significant (Kendall’s τ=-0.24, p=0.311). This mixed signal provides weak-to-moderate evidence of small study effects or publication bias. However, given the subtype-related heterogeneity identified in the meta-regression, these asymmetry tests should be interpreted with caution.

Figure 9: Deeks’ funnel plot for the assessment of diagnostic test accuracy publication bias (inverse root effective sample size vs ln DOR).

Prognostic Value of MFC/PCR Discordance

Six survival comparisons from five studies provided extractable data to contrast the prognostic impact of discordant MFC/PCR states with that of concordant reference groups. Pooled on the log scale utilizing a REML random-effects model, the overall HR for survival based on MFC/PCR discordance was 1.27 (95% CI, 0.34-4.70; p=0.720) (Figure 10). Similarly, the pooled RR for relapse (incorporating four studies) was 0.88 (95% CI, 0.40-1.95) (Figure 11).

Figure 10: Forest plot of hazard ratios (HR) for survival outcomes based on MFC/PCR discordance.The direction of the effect size is highly dependent on the utilized reference group (double-positive vs. double-negative).Studies referenced: Gaipa et al. 2012 [7], Shang et al. 2021 [33], Gao et al. 2020 [35], Chang et al. 2026 [31], and Ryan et al. 2008 [39].The notations (a) and (b) corresponding to Gao et al. 2020 [35] are utilized to distinguish between different clinical subgroups or specific analytical thresholds evaluated for survival outcomes within the same primary study.

Figure 11: Forest plot of risk ratios (RR) for leukemia relapse contrasting discordant versus concordant-negative MRD states.Studies referenced: Gaipa et al. (2012) [7], Shang et al. (2021) [33], Gao et al. (2020) [35], and Chang et al. (2026) [31].The notations (a) and (b) corresponding to Gao et al. (2020) [35] are utilized to distinguish between different clinical subgroups or specific analytical thresholds evaluated for survival outcomes within the same primary study.

Both prognostic estimates were statistically non-significant and showed extreme heterogeneity (I2=89.2% for survival; I2=83.8% for relapse risk). This non-significance masks a directional divergence that is dependent on the assigned reference group. Studies contrasting isolated PCR positivity (MFC−/PCR+) against double-positive disease (MFC+/PCR+) found the discordant state to be prognostically favourable (e.g., Gaipa et al. [7] HR 0.31; Shang et al. [33] HR 0.12), while studies contrasting the exact same discordant state against double-negative disease (MFC−/PCR−) found it to carry a substantially higher clinical risk (e.g., Chang et al. [31] HR 9.49; Ryan et al. [39] HR 4.51). Therefore, the prognostic implications of MFC/PCR discordance are not uniform; they are highly context-dependent and influenced by the baseline comparator. Small-study effects for prognostic outcomes were visually mapped, although formal statistical testing was precluded by the limited number of effect sizes (k=6) (Figure 12).

Figure 12: Contour-enhanced funnel plot evaluating small-study effects and publication bias for prognostic hazard ratios.

Certainty of Evidence (GRADE)

The certainty of the synthesized evidence was evaluated using the GRADE framework [21], which was adapted for diagnostic tests (Table 4). The certainty for core diagnostic accuracy metrics (Sensitivity, Specificity, AUC) was deemed moderate, downgraded one level due to severe inconsistency (I2>90%). The certainty for the DOR was rated as low, downgraded for both inconsistency and indirectness (arising from the pooling of mixed ALL/AML and pediatric/adult cohorts). The certainty for prognostic outcomes (EFS/RFS and relapse/CIR) was classified as low to very low, penalized for risk of bias (reliance on estimated HRs from survival curves), severe inconsistency, and imprecision driven by sparse event rates across a limited number of studies.

Outcome | No. of Studies | Risk of Bias | Inconsistency | Indirectness | Imprecision | Publication Bias | Certainty of Evidence
Sensitivity (MFC vs. PCR) (Diagnostic accuracy) | 12 [7,8,22,23,25,33,34,36-40] | Not Serious | Serious (-1)a | Not Serious | Not Serious | Not Serious | ⊕⊕⊕⊝ MODERATE
Specificity (MFC vs. PCR) (Diagnostic accuracy) | 12 [7,8,22,23,25,33,34,36-40] | Not Serious | Serious (-1)a | Not Serious | Not Serious | Not Serious | ⊕⊕⊕⊝ MODERATE
Diagnostic Odds Ratio (DOR) (Diagnostic accuracy) | 12 [7,8,22,23,25,33,34,36-40] | Not Serious | Serious (-1)a | Serious (-1)b | Not Serious | Not Serious | ⊕⊕⊝⊝ LOW
Summary AUC (HSROC) (Diagnostic accuracy) | 12 [7,8,22,23,25,33,34,36-40] | Not Serious | Serious (-1)a | Not Serious | Not Serious | Not Serious | ⊕⊕⊕⊝ MODERATE
EFS/RFS (Discordant vs. Concordant MRD) (Prognostic - survival time) | 5 [7,31,33,35,39] | Serious (-1)c | Serious (-1)d | Not Serious | Serious (-1)e | Not Serious | ⊕⊝⊝⊝ LOWf
Relapse / CIR (Discordant vs. Concordant MRD) (Prognostic - dichotomous events) | 4 [7,31,33,35] | Serious (-1)c | Serious (-1)d | Serious (-1)g | Serious (-1)e | Not Serious | ⊝⊝⊝⊝ VERY LOWf

Discussion

This systematic review and meta-analysis synthesized data from 23 studies to elucidate the comparative diagnostic accuracy and prognostic implications of MFC versus PCR-based MRD assessment in leukemia. The findings confirm that while MFC provides excellent overall discrimination (AUC=0.965) and exceptional specificity (96.0%), its pooled sensitivity remains moderate (76.2%) when compared with PCR. The meta-regression identified a subtype-specific divergence, demonstrating that MFC performs significantly less optimally in AML than in ALL. Furthermore, the prognostic synthesis revealed that inter-method discordance, specifically the MFC−/PCR+ state, does not confer a uniform clinical risk but is highly context-dependent, serving as a critical intermediate-risk state between deep molecular remission and overt macroscopic residual disease.

The high specificity of the MFC observed in our analysis underscores its utility as a reliable rule-in test. A positive MFC result generated a high PLR (15.45), indicating that the flow cytometric detection of aberrant blasts is rarely false-positive when referenced against molecular persistence. The instances of false positivity (MFC+/PCR−, 4.0%) observed in the literature are attributable to the detection of regenerating normal hematopoietic progenitors (e.g. hematogones in B-ALL) that mimic LAIPs or, conversely, the actual loss of PCR molecular targets due to clonal evolution [7,35,40]. The advent of highly standardized next-generation flow cytometry protocols, such as those developed by the EuroFlow Consortium, has mitigated these false-positives by utilizing sophisticated eight-color panels and bulk-lysis procedures to evaluate ≥4×106 cells, pushing MFC sensitivity closer to that of PCR (10-5) [40], complementing international molecular standardization efforts such as the EuroMRD guidelines [42].

However, the moderate pooled sensitivity (76.2%) highlights the limitations of the standard MFC. A negative MFC result cannot rule out the presence of submicroscopic disease, as reflected by the modest NLR (0.237). This limitation was illuminated by our meta-regression, which demonstrated that the DOR for MFC was significantly lower in AML cohorts than in ALL cohorts (p=0.0070). This biological discrepancy is multifactorial in nature. In ALL, PCR targets highly specific, clonally rearranged Ig/TCR genes, whereas MFC targets well-defined B- or T-cell developmental aberrancies [8,38]. AML is characterized by intratumoural heterogeneity, asynchronous maturational arrest, and a high frequency of immunophenotypic shifts following targeted or cytotoxic therapies [24,26,32]. AML LAIPs overlap with dysplastic or regenerating myelopoiesis, which diminishes the sensitivity of flow cytometry [2,26]. For AML, particularly in subtypes defined by distinct molecular aberrations (e.g. NPM1 mutations or RUNX1::RUNX1T1 fusions), PCR remains the indispensable gold standard for deep MRD quantification [33,35].

The central objective of this meta-analysis was to decode the prognostic significance of MFC/PCR discordance, a frequent clinical conundrum. Our pooled survival analysis yielded a non-significant hazard ratio (HR=1.27) with extreme statistical heterogeneity (I2=89.2%), which is not statistical noise; rather, it reflects a vital clinical reality: MFC/PCR discordance is not a biological monolith. When studies compared the isolated PCR-positive state (MFC−/PCR+) against double-positive disease (MFC+/PCR+), the discordant patients exhibited superior survival outcomes [7,33], while when this exact same discordant state was compared against double-negative patients (MFC−/PCR−), it portended a higher risk of relapse [31,39], indicating that the MFC−/PCR+ state represents an intermediate tumor burden, with disease levels low enough to evade MFC detection but high enough to be captured by PCR amplification. Thus, while isolated PCR positivity is clinically preferable to double-positive disease, achieving a dual-negative state (deep molecular remission) remains a prerequisite for long-term cure, particularly prior to allogeneic HSCT [4,28]. However, it is crucial to interpret these findings with caution. The extreme statistical heterogeneity (I²>80%) observed in both diagnostic and prognostic models reflects profound underlying clinical heterogeneity. Because MRD cutoffs, baseline comparators, and patient cohorts varied significantly across studies, our pooled estimates should be viewed as directional trends rather than absolute clinical parameters, thereby reducing the overall certainty of the synthesized evidence.

The clinical implications of these findings advocate for a synergistic rather than competitive application of MRD modalities. MFC is well-suited for early therapeutic time points (e.g. post-induction day 15 or 33) due to its rapid turnaround time, lower cost, and ability to assess treatment kinetics without the need for patient-specific primer design [7,34], whereas PCR should be employed at later milestones (e.g. post-consolidation or pre-transplantation), where disease burden is exponentially lower, and the superior analytical sensitivity of molecular testing is required to guide high-stakes decisions, such as transplant conditioning intensity or the administration of targeted immunotherapies [4,28,35].

This study is strengthened by its rigorous PRISMA-compliant methodology, utilization of hierarchical bivariate models to account for threshold effects, and direct meta-analytical exploration of inter-method discordance on survival. However, the findings of this study must be interpreted in light of several limitations. First, the reliance on predominantly non-randomized, retrospective cohorts introduces selection bias. Second, substantial methodological heterogeneity was present across the included literature, driven by wide variations in MFC gating strategies, antibody panels, PCR techniques, and differing MRD positivity thresholds. This heterogeneity is compounded by the pooling of data across diverse biological environments, most notably pediatric and adult populations. Third, the included studies span a prolonged publication period during which significant technological advancements occurred; pooling older diagnostic technologies (e.g., three- or four-color flow cytometry) with highly standardized contemporary methodologies (e.g., eight-color next-generation flow) contributed to the observed variance and may affect the generalizability of our pooled estimates to modern clinical practice. Fourth, our inability to retrieve a large number of conference abstracts lacking full-text publications may have introduced a degree of publication bias, potentially skewing the synthesis toward studies with statistically significant or favorable outcomes. Finally, while next-generation sequencing (NGS) represents an emerging modality that bridges the high sensitivity of PCR with the broad applicability of MFC, it was not the primary focus of this analysis, though recent data suggest it may eventually supersede standard PCR in clarifying discordant MFC results [30,31].