Multivariate model predicts immune imbalance in recurrent pregnancy loss and recurrent implantation failure.

This work was supported by the Carlos III Institute of Health, Ministry of Economy and Competitiveness (Spain), awarded in the 2023 call under the Health Strategy Action 2021–2024, within the National Research Program oriented to Societal Challenges, part of the Technical, Scientific, and Innovation Research National Plan 2021–2024 (grant reference PI23/00709).

Patients were divided into RPL and RIF according to the inclusion criteria and compared with the control group (HC). No differences were observed in terms of age ( Table 1 ). On the contrary, differences between the number of miscarriages, number of cycles, and number of transfers between patients and controls were observed due to the inclusion criteria themselves. Epidemiological characteristics of the study groups (RPL and RIF) and the control group (HC). The statistical test used was an ANOVA of one factor for three or more study groups or Fisher’s exact test for proportions. P -value <0.05 indicates differences between at least two groups. Data are expressed as mean ± SD or as n (%). NS, not significant; NA, not applicable. A higher proportion of patients with endometriosis in the RIF group compared to RPL was observed. The total percentage of NK cells and subpopulations of NK cells according to CD56 and CD16 expression were evaluated. Those CD56 ++ are termed regulatory NK cells and CD56 + CD16 + cytotoxic NK cells. In turn, we also identified as natural killer T (NKT) cells those that expressed CD3 and CD56. We did not observe statistically significant differences in any of these subpopulations. However, upon analysing the surface receptors expressed on NK cells, disparities were indeed identified. Both the frequency of cells expressing the marker and the mean fluorescence intensity (MFI) were considered. A significant increase in the percentage of NKp30 + cytotoxic NK cells in the RIF group (80.95 (69.90–87.73)) compared to HC (70.80 (56.20–77.28) ( P = 0.003), as well as a significant increase in the MFI of this marker in RIF (1,174 (858.50–1,446)) compared to HC (852.50 (593.8–1,078)) ( P = 0.002), was observed ( Fig. 1 ). A decrease in the percentage of TIGIT + cytotoxic NK cells in the RIF group (65.70 (54.75–75.63)) compared to HC (71.30 (66.25–86.33)) ( P = 0.017) was noted. This decrease was also observed in TIGIT MFI in the RIF group (747.5 (478.5–983)) compared to HC (923.50 (743.3–1,524)) ( P = 0.008). Relevant NK cell surface markers in HC, RPL, and RIF. The box plots show median and interquartile range. Group comparisons were performed using the Kruskal–Wallis test with Dunn’s post hoc test. * P < 0.05; ** P < 0.01; NS, not significant. On the other hand, a significant decrease in the frequency of CD69 + cytotoxic NK cells in the RPL group (1.75 (0.82–3.33)) compared to HC (2.87 (1.94–6.18)) ( P = 0.018) was noted. Regarding NKT subpopulation, there was a significant decrease in CD69 + in both RPL (5.37 (3.70–10.53)) and RIF (5.86 (3.93–11)) groups compared to HC (9.57 (6.29–13.43)) ( P = 0.010; P = 0.016, respectively). Regarding monocyte subpopulations, a significantly higher frequency of classical monocytes in RPL (77.90 (70.48–84.25)) and RIF (77.75 (72.03–84.33)) groups was observed compared to HC (73.10 (69.08–77.18)) ( P = 0.020; P = 0.017, respectively) ( Fig. 2 ). Monocyte subpopulations and surface marker expression in HC, RPL, and RIF. The box plots represent median and interquartile range. The Kruskal–Wallis test with Dunn’s post hoc test was used. Exact P -values are shown for statistical trends. * P < 0.05; ** P < 0.01; NS, not significant. Regarding markers, a significantly increased frequency of classical CX3CR1 + monocytes in the RPL group (10.45 (5.59–16.33)) compared to HC (5.15 (3.12–12.78)) ( P = 0.053) was found. There was a statistical trend indicating a higher percentage of classical CD69 + monocytes in the RIF group (16.40 (9.67–25.50)) compared to the HC group (11.15 (7.98–18.65)) ( P = 0.061). Regarding the MFI, we found a decrease in CCR5 MFI in intermediate monocytes between the RPL group (591.5 (33.18–1,367)) and the HC group (1,206 (563.5–1,616)) ( P = 0.034). Finally, differences in the activated CD11b marker were found in both percentage and MFI ( Fig. 2 ). First, a significant increase in CD11bACT + intermediate monocytes in the RPL (65.60 (36.93–84.60)) and RIF (73.10 (36.45–90.55)) groups was observed compared to HC (47.40 (32.45–61.65)) ( P = 0.039; P = 0.044, respectively). Similarly, an increase in CD11bACT + non-classical monocytes in RPL (13.15 (5–24.75)) and RIF (12.50 (6.24–25.60)) groups was found compared to HC (5.12 (2.24–12.43)). An increase of this marker in classical monocytes between RIF (39.70 (11.15–86.45)) and HC (18.80 (7.39–50.68)) groups was also noted, although with a statistical trend ( P = 0.054). In the same way, the analysis at the single cell level showed a significant increase in the MFI of CD11b in all three monocyte subpopulations when comparing between patient and control groups ( Fig. 2 ). In particular, in classical monocytes, the MFI of CD11bACT was 791 (479.8–1,349) in the RPL group and 872 (529–2,007) in the RIF group, while in HC, it was 572.5 (396.5–1,040). These differences were statistically significant between RIF and HC groups ( P = 0.026), and a statistical trend between RPL and HC was found ( P = 0.064). In intermediate monocytes, the MFI was 1,360 (798.8–2,165) in RPL and 1,365 (746–2,466) in RIF, while in HC, it was 940 (615.5–1,221). These differences were statistically significant in the RPL ( P = 0.041) and RIF ( P = 0.039) groups compared to HC. In non-classical monocytes, the MFI was 362 (275.5–570.3) in RPL and 376 (241.5–548) in RIF, while in HC, it was 225 (80.13–317.8). These differences were statistically significant in RPL ( P = 0.001) and RIF ( P = 0.003) groups compared to HC. MDSCs are a heterogeneous population of leucocytes with significance in cancer and inflammation-related contexts. The percentage of MDSCs in peripheral blood was evaluated, and a significant increase of this population in RPL (7.90 (6.30–10.60)) ( P = 0.001) and RIF (7.7 (6.1–10.1)) ( P = 0.012) compared to HC (5.55 (4.22–7.75)) was observed ( Fig. 2 ). Evaluation of plasma cytokines was performed in a subgroup of both RPL ( n = 22) and RIF ( n = 29) patients and HC ( n = 17). Elevated levels of IL-9 in RIF patients (782.40 (744.30–888.60) compared to HC (732.80 (665.10–798.20) ( P = 0.044) were observed. Similarly, increased levels of MIP-1β in RIF (256.40 (227.60–281.80)) compared to HC (225.30 (197.40–245.60)) ( P = 0.029) were found. Conversely, we noted a tendency towards reduced TGF-β3 levels in RPL (195.30 (98.17–312.5)) compared to HC (302 (184.30–409.60)) ( P = 0.094) ( Fig. 3 ). Circulating relevant cytokine levels in HC, RPL, and RIF. Data are shown as box plots (median and interquartile range). Kruskal–Wallis test with Dunn’s post hoc test. Exact P -values indicate trends. * P < 0.05; ** P < 0.01; NS, not significant. Due to the limitations of classical flow cytometry analysis, more advanced unsupervised analysis techniques were used to identify novel subpopulations with a different marker expression. UMAP was first used for dimensionality reduction and visualization of NK cells and monocytes based on the full marker panels. In NK cells, UMAP visualization revealed the expected segregation according to CD3 expression, with further separation of CD3 − cells into CD56 ++ and CD56 + CD16 + subsets (Supplementary Fig. 5A). Similarly, UMAP projection of monocytes showed separation into CD14 ++ CD16 − and another CD14 − CD16 + , with variable expression of the rest of the markers (Supplementary Fig. 6A). Therefore, FlowSOM clustering was applied with the number of clusters set to 15 (Supplementary Figs 5B and 6B). In the NK cell setting, five clusters (clusters 1, 3, 4, 10, and 12) were underrepresented in both RIF and RPL groups compared to the HC group (Supplementary Fig. 5C). These clusters corresponded to phenotypically distinct NK subsets differing in CD56, CD16, activation (CD69), inhibitory (TIGIT and TIM3), cytotoxic (perforin), and activating receptor (NKp30) expression (Supplementary Table 2). In the monocyte panel, two clusters showed marked between-group differences (Supplementary Fig. 6C). Cluster 5 was enriched in the RIF group and was characterized by CD14 + CD16 − HLA-DR + CCR2 + , whereas cluster 11, reduced in the RPL group, was characterized by CD14 ++ CD16 − monocytes with high CCR2 and CCR5 expression (Supplementary Table 3). After identification, we performed a whole-group analysis by manual gating for validation ( Fig. 4 ). In the NK clusters, a reduction in cluster 4 in the RPL (8.03 (4.49–14.85)) and RIF (6.34 (3.84–13.3)) groups compared to HC (13 (10.15–32.53)) ( P = 0.001; P < 0.0001, respectively) and cluster 3 in the RPL (1.07 (0.57–1.72)) and RIF (0.97 (0.53–2.39)) groups compared to HC (1.89 (0.92–3.75)) ( P = 0.017; P = 0.047, respectively) was found. A reduction in cluster 1 between the RPL (5.78 ((4.33–8.97)) and HC (9.38 (5.53–12.13)) groups ( P = 0.011) was also detected. Frequency of FlowSOM-defined clusters in HC, RPL, and RIF. The box plots show median and interquartile range. Kruskal–Wallis test with Dunn’s post hoc test. Exact P -values indicate trends. * P < 0.05; ** P < 0.01; NS, not significant. In the case of monocytes, an increase in cluster 5 in the RPL group (98.55 (96.77–99.4)) compared to HC (95.7 (93.33–97.60)) ( P < 0.0001) was found. On the other hand, a reduction in cluster 11 in the RPL group (0.98 (0.14–2.13)) compared to HC (2.47 (1.16–4.85)) ( P = 0.001) was found, indicating that these phenotypes are present across groups but vary in abundance rather than being exclusive. Initial analysis encompassed 111 variables were analysed across all women in the HC, RPL, and RIF groups, using four tubes. Comparative analysis between groups revealed 17 statistically significant variables differentiating the HC and RPL cohorts ( n = 86) and 18 differentiating the HC and RIF cohorts ( n = 78), as determined by the Mann–Whitney U test. The overall analytical workflow is summarized in Fig. 5 . Overview of the analytical workflow. Given the high dimensionality of the dataset relative to the sample size, we employed a feature selection strategy to mitigate the risk of model overfitting and to enhance the clinical translatability of our findings. Therefore, the Boruta algorithm (detailed in the Statistical Analysis section) was applied to select the most relevant variables. This reduction simplifies the final test profile for clinical application and enhances both predictive power and interpretability. In the comparison of the RPL group versus the controls, Boruta classified ten variables as important and four as tentative, discarding only three variables from the original 17 (Supplementary Fig. 7). Based on the prior correlation analysis and a review of the data for statistical significance, correlated variables were subsequently removed to simplify the model while preserving its biological relevance. The parameters included in the multivariate analysis were activated CD11b in non-classical monocytes, MDSCs, cluster 5, cluster 3, and cluster 4. Multivariable logistic regression using these parameters achieved an AUC of 0.95 (95% CI: 0.884–0.998) with an overall accuracy of 90.7% ( Table 2 ). Summary of multivariate logistic regression analysis between RPL and HC groups. AUC, area under curve; CI, confidence interval; SP, specificity; SN, sensitivity; AC, accuracy; NPV, negative predictive value; PPV, positive predictive value; MDSCs, myeloid-derived suppressor cells; Act, activated. In the comparison between the RIF group and the control group, Boruta excluded six of the 18 initially significant variables, classifying seven as important and five as tentative (Supplementary Fig. 8). Following the same correlation-based refinement, the final model retained NKp30 + cytotoxic NK cells, the MFI of activated CD11b in non-classical monocytes, MDSCs, and cluster 4. This model achieved an AUC of 0.85 (95% CI: 0.757–0.934) with an accuracy of 79.5% ( Table 3 ). Summary of multivariate logistic regression analysis between RIF and HC groups. AUC, area under curve; CI, confidence interval; SP, specificity; SN, sensitivity; AC, accuracy; NPV, negative predictive value; PPV, positive predictive value; MDSCs, myeloid-derived suppressor cells; Act, activated; MFI, mean fluorescence intensity; cyt, cytotoxic NK cells. To evaluate model stability and correct for potential optimism, bootstrap resampling was performed on the training datasets. The optimism-corrected AUC for the RPL model was 0.92 (95% CI: 0.90–0.94), indicating high stability and limited overfitting. In contrast, the RIF model showed an optimism-corrected AUC of 0.66 (95% CI: 0.64–0.72), suggesting moderate discrimination and a higher sensitivity to sample variability. Accordingly, the results from the RIF model should be interpreted cautiously pending validation in larger cohorts. An independent validation cohort was assembled, comprising 37 RPL and 17 RIF patients, to assess the original diagnostic models without any recalibration. For each subject, the predefined predictor variables were entered into the corresponding multivariate logistic regression equations and classification probabilities were computed using a cut-off of 0.5. In the RPL cohort, 27 of 37 patients (73.0%) were correctly classified, while in the RIF cohort, 16 of 17 patients (94.1%) were accurately identified. Although specificity, AUC, and calibration could not be assessed due to the absence of healthy controls in the validation cohort, the observed sensitivities support a promising early-stage signal in an independent patient population.

Between 2021 and 2024, women experiencing RPL or RIF were referred to our Reproductive Immunology Unit for immunological assessment. Eligible patients met the ESHRE definitions for RPL (spontaneous loss of ≥2 pregnancies, excluding ectopic and molar gestations) or RIF (failure to achieve clinical pregnancy after transfer of ≥4 high-quality embryos across ≥3 fresh or frozen cycles in women <40 years). A healthy control group comprised non-pregnant women with confirmed fertility (≥2 live births) and no history of miscarriage. All participants underwent standardized clinical, hormonal, cytogenetic, and imaging evaluations to exclude non-immunological causes of RPL/RIF. Immunological screening included indirect immunofluorescence for antinuclear antibodies and assays for anticardiolipin, anti-β2-glycoprotein, lupus anticoagulant, anti-thyroid, and anti-transglutaminase-2/anti-deamidated gliadin antibodies. Baseline blood samples were collected for immunophenotyping and cytokine profiling. The protocol was approved by the institutional ethics committee (FIS PI19/01450 and FIS PI23/00709), and written informed consent was obtained from all participants. Lymphocyte subsets and myeloid subsets underwent examination through multiparametric flow cytometry utilizing a single-platform analytical approach (FACSCanto II BD Biosciences, USA). In brief, blood samples were collected using suitable EDTA/heparinized vacuum tubes and promptly processed within a 2 h window following collection. Peripheral blood samples were stained with the following monoclonal antibodies: HLA-DR-Brilliant Violet-510 (BV510; BD Biosciences); CD14-APC (BioLegend, USA); CD45-APC-Cy7 (BioLegend); CD11b-Phycoerythrin (PE; BioLegend), CD33-Brilliant violet 421 (BV421; BD Biosciences), CD3-PerCP-Cy5.5 (BD Biosciences), CD4-Pe-Cy7 (BD Biosciences), CD8-APC-Cy7 (BD Biosciences), CD127-Alexa Fluor 647 (AF647; BD Biosciences), CD25-PE (BD Biosciences), CD14-APC Cy7 (BD Biosciences), CD56-BV510 (BD Biosciences), CD16-FITC (BD Biosciences), TIGIT-PE (BioLegend), TIM3-APC Cy7 (BioLegend), CD69-PE Cy7 (BD Biosciences), NKp30-Alexa Fluor 647 (AF647; BD Biosciences), Perforin-BV421 (BD Biosciences), CD11b ACT-PE (BioLegend), CCR2-Alexa Fluor 647 (BD Biosciences), CCR5-BV421 (BD Biosciences), CX3CR1-PerCP Cy5.5 (BioLegend), and CD45-APC Cy7 (BD Biosciences). In brief, blood samples were incubated 20 min with monoclonal antibodies. After that, lysing solution (BD Biosciences) was added for 15 min. After two washes, samples were measured using BD FACSDiva software (BD Biosciences) and analysed in FlowJo, LLC (version 10.6.1; Becton Dickinson). At least 200,000 events were measured for each sample. Various gating strategies were devised based on the specific population under examination, including NK cells (Supplementary Fig. 1 (see section on Supplementary materials given at the end of the article)), monocytes (Supplementary Fig. 2), MDSCs (Supplementary Fig. 3), and T reg cells (Supplementary Fig. 4). The flow cytometry panels were optimized and locked before the inclusion of the validation cohort. Cytokine concentrations in plasma were measured using a magnetic bead-based multiplex assay (Bio-Rad, USA), strictly following the manufacturer’s instructions. Plasma was first centrifuged at 1,000 g for 15 min at 4°C, and supernatants were diluted 1:4. For each well of a 96-well plate, 50 μL of diluted plasma was combined with 50 μL of bead suspension and incubated at room temperature on a shaker (850 rpm) for 30 min. Wells were then washed three times with 100 μL of wash buffer before the addition of 25 μL biotinylated detection antibody, followed by a further 30 min incubation under the same agitation conditions. After three additional washes, 50 μL streptavidin-phycoerythrin was added and incubated for 10 min at 850 rpm. A final triple wash was performed, and beads were resuspended in 125 μL assay buffer for acquisition on a Luminex MAGPIX system (Luminex, USA). The cytokine panel analysed comprised eotaxin, FGF-basic, G-CSF, IFN-γ, IL-1β, IL-1ra, IL-4, IL-7, IL-8 (CXCL8), IL-9, IL-10, IL-13, IL-17, IP-10 (CXCL10), MCP-1 (CCL2), MIP-1α, MIP-1β (CCL4), PDGF-BB, RANTES (CCL5), TNF-α, and TGF-β isoforms 1–3. Abbreviations and limits of detection for each analyte are provided in Supplementary Table 1. Prior to unsupervised clustering, we processed flow cytometry data from 66 samples (22 RPL, 22 RIF, and 22 healthy controls) by excluding debris, doublets, and dead cells via FSC/SSC gating and selecting CD56 + NK cells or CD14 + monocytes and then concatenating 3,000 events per sample. Dimensionality reduction was performed using uniform manifold approximation and projection (UMAP) (neighbours = 15, min_dist = 0.1, and metric = ‘euclidean’) in FlowJo to project high-dimensional marker expression into two dimensions, preserving global data structure and revealing cellular heterogeneity. Finally, flow cytometry-based self-organizing map (FlowSOM) clustering ( k = 15) was applied to the same dataset, leveraging its self-organizing map algorithm to robustly integrate multiple markers and delineate distinct cell populations in an unsupervised manner. Statistical analyses and graphical representations were performed using GraphPad Prism version 5.0 and RStudio (R version 4.2.3). An initial exploratory assessment of the data was conducted, and quantitative variables were evaluated for normality by means of the Shapiro–Wilk test (for n < 30) or the Kolmogorov–Smirnov test (for n ≥ 30). Variables demonstrating a normal distribution were compared by Student’s t -test (for two groups) or one-way ANOVA (for three or more groups), with results reported as mean ± standard deviation. When normality assumptions were not met, the Mann–Whitney U test or Kruskal–Wallis test followed by Dunn’s post hoc test was applied, and data were presented as median (interquartile range). Categorical variables were analysed using the chi-square test or Fisher’s exact test. Statistical significance was defined as P < 0.05 (two-sided). Variable selection was performed by applying the Boruta algorithm, which utilizes a random forest classifier to discern those features most discriminatory between pathological and control cohorts. Thereafter, variables deemed significant by Boruta were entered into a multivariate logistic regression model, and the model’s overall classification performance was assessed. Finally, a combined-variable receiver operating characteristic (ROC) curve was generated to calculate the area under the curve (AUC), thereby quantifying the predictive capacity of the model in terms of sensitivity and specificity. To assess model stability and correct for potential optimism due to sample size, we performed bootstrap resampling. For each model, we resampled the original training cohort with replacement 1,000 times. In each bootstrap sample, we refit the final multivariable logistic regression model using the predefined predictors and calculate the AUC both in the bootstrap sample (‘apparent performance’) and in the original cohort (‘test performance’). The optimism was defined as the mean difference between apparent and test AUC across all resamples and was subtracted from the original AUC to obtain an optimism-corrected AUC. Percentile 95% confidence intervals were derived from the distribution of test AUC values across resamples.

Our integrative immunophenotypic analysis reveals that RPL and RIF share an overall pro-inflammatory myeloid profile yet diverge in NK cell signalling patterns, and these readouts can be used to construct highly discriminative diagnostic models. Despite comparable absolute and subset-specific NK cell counts across groups, pronounced qualitative differences in receptor expression distinguished RPL from RIF. In RIF, cytotoxic CD56 dim CD16 + NK cells showed a marked upregulation of the activating receptor NKp30 (both in frequency and in MFI) relative to healthy controls, mirroring previous reports that link heightened peripheral NKp30 expression with implantation failure and inflammatory RIF/RPL phenotypes ( Comins-Boo et al. 2021 , Zhang et al. 2021 ). Because NKp30 engagement promotes perforin- and granzyme-mediated cytotoxicity, as well as pro-inflammatory IFN-γ release ( Moretta et al. 2001 ), its overexpression may amplify a pro-inflammatory environment and disrupt endometrial receptivity during the implantation window. Concomitantly, the checkpoint receptor TIGIT, whose expression normally rises throughout uncomplicated gestation to restrain NK effector function ( Wang et al. 2021 ), was significantly downregulated in the same RIF cohort. The resulting NKp30 high /TIGIT low signature suggests an imbalance between activating and inhibitory signalling that could tip peripheral NK cells towards excessive activation. Such a disequilibrium agrees with functional studies in which TIGIT + NK cells display a reduced degranulation and cytokine output, supporting a tolerogenic role during early pregnancy ( Bi & Tian 2019 ). By contrast, RPL was characterized by diminished expression of the early activation marker CD69 on both cytotoxic NK and NKT cells. Reduced CD69 has been associated with lower responsiveness to implantation cues and can be experimentally induced by pre-implantation factor, a peptide whose downregulation has been implicated in miscarriage ( Roussev et al. 2013 ). Collectively, our findings support a model in which RIF is driven by hyper-responsive, poorly restrained NK cells, whereas RPL may reflect insufficient NK activation or defective trafficking at the critical peri-implantation period. Our data did not reveal circulating NK cell expansions (despite reports of elevated peripheral NK percentages in some RPL and RIF populations), which may reflect differences in gating strategies, luteal-phase sampling, or the modest size of the present study. Taken together, the findings highlight that receptor-level alterations can occur independently of cell number changes and may offer greater mechanistic insight than enumeration alone, yet they also emphasize the need for multicentre studies employing harmonized protocols to reconcile quantitative discrepancies in the literature. Nevertheless, our conclusions are limited by the exclusive use of peripheral blood; decidual NK cells, which exhibit a distinct transcriptomic and receptor repertoire, were not examined. Functional assays (e.g. cytotoxicity and cytokine profiling) and longitudinal sampling across the menstrual cycle will be required to determine whether the NKp30 + /TIGIT low and CD69 low signatures causally contribute to implantation failure or are epiphenomena of systemic inflammation. For example, upregulation of the NKp30–BAG6 axis in uNK has been reported in endometriosis, reflecting tissue-specific activation at the implantation site that may not be detectable in blood ( Shi et al. 2025 ). In addition, higher expression of inhibitory receptors, such as KIR2DL1, has been suggested, proposing that endometriotic lesions may skew NK cells towards a more inhibitory phenotype. However, in our cohort, we did not observe significant differences in peripheral blood in either activating receptors (NKp30: endometriosis yes vs no, median 80.1 vs 77.9; P = 0.985) or inhibitory receptors (TIGIT: 69.4 vs 65.15; P = 0.687) within the RIF group. These findings indicate that, in this context, endometriosis is not a major confounder of our peripheral immune panel; rather, the differences between RPL and RIF may reflect a baseline pro-inflammatory state and/or system-level alterations in myeloid and NK regulatory circuits. We acknowledge that the endometrial/decidual compartment may display patterns distinct from peripheral blood; thus, we propose paired tissue–blood studies to disentangle local versus systemic contributions to implantation. A consistent pro-inflammatory shift was detected across the circulating myeloid compartment of both patient groups. Quantitatively, classical CD14 ++ CD16 − monocytes were expanded in RPL and RIF relative to fertile controls, compatible with previous reports that describe an enrichment of this subset in recurrent reproductive failure and link it to a baseline inflammatory milieu ( Comins-Boo et al. 2022 ). Comins-Boo et al. proposed that monocyte-derived markers could help identify patients with an immune-driven phenotype. In the present work, we broaden this approach in two key ways. First, we include not only monocyte activation states but also NK cell receptor status, MDSC abundance, and regulatory T cell levels, thereby capturing multiple axes of innate and regulatory immunity relevant to implantation biology. Second, we evaluate these models in a larger cohort with an independent validation set. Qualitatively, monocytes displayed broad upregulation of the activated integrin CD11b, particularly in intermediate and non-classical subsets, and CD11b MFI rose across all three subsets. These observations may reflect a systemic inflammatory status that can adversely affect the maternal–fetal interface and pregnancy outcome. Inflammation is a double-edged sword in pregnancy: while it is necessary for implantation and delivery, excessive or inappropriate inflammation can be detrimental ( Kwak-Kim et al. 2009 ). In contrast, a selective reduction in the MFI of CCR5 on intermediate monocytes distinguished the RPL cohort. Because CCR5 is normally upregulated on CD16 + monocytes during healthy gestation ( Rees et al. 2024 ), its attenuation here may reflect impaired chemotactic responsiveness to CCL4/CCL5 gradients within the endometrium. The apparent discrepancy between unsupervised analysis and manual gating in cluster 5 is explained by the unsupervised analysis design. Unsupervised tools (FlowSOM/UMAP) group individual cells from many patients that are pooled and downsampled to reveal patterns. This is great for discovery, but it does not directly report how many of that cell type each patient has. Manual gating does the opposite: it applies the same thresholds to each sample and gives frequencies per patient. Because of this, the unsupervised step can make a subset look as if it appears ‘only in’ one group, simply because the clustering places its centre where separation is strongest in the pooled data. When we check with manual gating across the whole cohort, we see that these subsets are present in multiple groups but are more or less frequent in one group. Technical choices also matter: cluster labels are not ontologies, so cluster 5 and cluster 11 should be viewed as phenotypic neighbourhoods, not exclusive entities. Parallel to these monocyte alterations, both pathological groups exhibited elevated frequencies of circulating MDSCs. A study has documented MDSC reduction in RPL, even though the comparison was with healthy pregnant women ( Nair et al. 2015 ). Whether this rise reflects a compensatory tolerogenic response that fails to counterbalance inflammation, or instead reflects dysfunctional MDSCs incapable of exerting suppression, remains to be elucidated. Functional assays will be required to clarify their net contribution. The diagnostic algorithm was built in three sequential steps (feature selection, multivariate modelling, and internal validation). Beyond predictive performance, a primary goal of the modelling strategy was clinical interpretability, favouring transparent coefficients that can be readily translated into a diagnostic score usable at the bedside. Starting from 111 cytometric and clinical variables, univariate screening identified 17 markers that discriminated HC from RPL and 18 that distinguished HC from RIF. Because many markers were biologically correlated, the Boruta feature selection algorithm, a random forest-based wrapper, was applied to each comparison ( Degenhardt et al. 2019 ). This approach yielded a manageable set of candidate predictors (RPL: 10 confirmed and 4 tentative; RIF: 7 confirmed and 5 tentative), allowing subsequent refinement while preserving clinically meaningful signals. For the RPL model, five non-redundant variables were entered: activated CD11b on non-classical monocytes, circulating MDSCs, and three FlowSOM-defined clusters (5, 3, and 4). This model demonstrated strong discrimination (AUC of 0.95; 95% CI: 0.884–0.998), with 91.3% sensitivity, 87.5% specificity, and 90.7% overall accuracy. In RIF, a four-variable model (the frequency of NKp30 + cytotoxic NK cells, the MFI of activated CD11b on non-classical monocytes, MDSCs, and cluster 4) achieved an AUC of 0.85 (95% CI: 0.757–0.934) with 74.1% sensitivity, 87.5% specificity, and 79.5% accuracy. These effect sizes equal or exceed those of previously published multi-marker panels for reproductive failure, which typically report AUCs between 0.75 and 0.85 while using larger, less interpretable variable sets. Our correlation analyses are directionally consistent with this framework (data not shown). In RPL, results suggested pro-inflammatory correlations involving NK-related cluster 3 with IL-8 and TNF-α. In RIF, correlations involving NKp30 tended to increase with pro-inflammatory cues (for example, TNF-α/IL-8), whereas MDSCs showed weaker or inverse couplings with TGF-β isoforms, consistent with insufficient counter-regulation. Although exploratory, these patterns align with distinct immune imbalances underlying RPL and RIF. Endometriosis was more prevalent in the RIF cohort than in RPL (9/54 vs 1/62), raising the possibility that disease-related alterations in NK and myeloid compartments could confound our immune signatures ( Reis et al. 2023 ). To address this, we added endometriosis as a covariate to the multivariable logistic models and repeated performance analyses. In RPL, model discrimination was unchanged after adjustment (AUC 0.95) and when excluding endometriosis (AUC 0.95; ΔAUC −0.001). In RIF, AUC moved minimally with adjustment (AUC 0.881; ΔAUC +0.03) and upon exclusion (0.85; ΔAUC +0.001). Thus, within the limits of our sample, endometriosis did not materially account for the multivariate signal captured by the models. We acknowledge that the number of endometriosis cases was modest and that peripheral phenotypes may not fully reflect endometrial/decidual immunity. When applied unchanged to an independent validation cohort (37 RPL and 17 RIF patients), the models correctly classified 73% of RPL cases and 94% of RIF cases. As healthy controls were unavailable in this cohort, discrimination metrics, such as AUC, specificity, and calibration, could not be evaluated, and these validation results should therefore be considered preliminary. Logistic regression was chosen over more complex machine learning algorithms because its odds ratios are readily interpretable by clinicians and lend themselves to a point-based scoring system for eventual bedside use. Moreover, each retained marker has a plausible mechanistic link to implantation biology (CD11b-mediated adhesion and NKp30-driven cytotoxicity), facilitating biological validation and targeted intervention studies. Activated CD11b (Mac-1) on non-classical monocytes marks an integrin-licensed state that promotes adhesion to endothelium and vascular patrolling; this activation is typically reinforced by inflammatory cues, such as TNF-α/IL-1β/IL-6, and operates within chemokine pathways that recruit patrolling monocytes to endometrium (for example, CCL2–CCR2) ( Medrano-Bosch et al. 2023 ). These mechanisms fit with the endothelial interactions required at the decidual interface ( Lin et al. 2023 ). MDSCs reflect an immunosuppressive myeloid programme. Their expansion is driven by GM-CSF or G-CSF and restrains excessive effector responses, features compatible with the tolerogenic environment needed for implantation ( Groth et al. 2019 ). NKp30 is an activating receptor on natural killer cells that couples target recognition to IFN-γ/TNF-α release. Changes in NK cell receptor signalling and function have been linked to abnormal arterial remodelling and implantation failure, positioning NK features as biologically plausible markers in reproductive disorders ( Xie et al. 2022 ). Our aim is to turn these models into a practical diagnostic score that clinicians can use alongside routine tests. The idea is straightforward: each marker in the model contributes a few points; the sum of points corresponds to a patient’s probability of immune dysregulation. A single threshold (pre-specified) would classify results as ‘likely immune imbalance’ vs ‘unlikely’. The score and its cut-off would then be tested prospectively in independent cohorts and, if necessary, recalibrated. Finally, a pragmatic study should ask whether score-guided immunomodulation improves meaningful outcomes (for example, live-birth rate). The size of the HC group was relatively small compared with the RPL and RIF groups. A limited control arm can, in principle, affect model stability by i) inflating apparent specificity if controls happen to be immunologically homogeneous and ii) reducing our ability to fully characterize calibration against true-negative cases. We evaluated model robustness through bootstrapping-based internal validation and by applying the final models to an independent validation cohort. The absence of healthy controls in the validation phase forbids full calibration and specificity assessment; external and internal validation with balanced case–control sampling is therefore essential. Finally, longitudinal sampling across the menstrual cycle and early pregnancy, incorporation of functional assays (cytokine release and degranulation), and comparison with alternative classifiers (e.g. gradient boosting and Bayesian networks) will determine whether predictive performance can be further improved without sacrificing interpretability. In summary, the Boruta-plus-logistic-regression workflow yielded compact, mechanism-anchored models that classified RPL and RIF with high accuracy in both derivation and validation cohorts, laying the groundwork for prospective trials that test whether modulatory treatments guided by these signatures can enhance live-birth outcomes.

Recurrent pregnancy loss (RPL) and recurrent implantation failure (RIF) represent two of the most distressing complications in reproductive medicine. According to the latest European Society of Human Reproduction and Embryology (ESHRE) guidelines, RPL is defined as the spontaneous loss of two or more pregnancies, with an estimated prevalence of 1–2% of couples, rising to about 5% when three consecutive miscarriages are considered ( Quenby et al. 2021 , ESHRE Guideline Group on RPL et al. 2023 ). Recent literature stresses that the definition of RIF remains blurred and is steadily moving away from the long-standing, cycle-based criterion of ‘no clinical pregnancy after ≥4 good-quality embryo transfers’ ( Ata et al. 2021 , Somigliana et al. 2022 , ESHRE Working Group on Recurrent Implantation Failure et al. 2023 , Gill et al. 2024 ). The 2023 ESHRE good-practice recommendations acknowledge the heterogeneity of earlier definitions and propose a pragmatic working threshold of at least three embryo transfer attempts or a cumulative transfer of four morphologically high-grade embryos, while insisting that embryo quality, maternal age, and paternal factors be considered before assigning the label of RIF ( Cimadomo & de Los Santos 2023, Pirtea et al. 2023 ). The clinical and psychosocial burden of these conditions is substantial. Couples facing RPL or RIF endure repeated treatment cycles, invasive testing, and prolonged uncertainty, with well-documented sequelae that include heightened anxiety, depressive symptoms, and complicated grief responses ( Li et al. 2012 , Cuenca 2023 ). Establishing clearer mechanistic biomarkers and robust diagnostic criteria is therefore essential to improve counselling, personalize therapy, and ultimately reduce the physical and psychological impact of repeated reproductive failure. Successful implantation and early pregnancy rely on finely tuned communication between the embryo and a maternal immune system that is both active and precisely regulated ( Mor & Cardenas 2010 ). Mounting evidence indicates that disruption of this equilibrium (particularly within the innate immune compartment) contributes to both RPL and RIF ( Kwak-Kim et al. 2014 , Mukherjee et al. 2023 ). Peripheral and uterine NK cells, which normally adopt a low-cytotoxic, cytokine-modulating phenotype, display functional drift in complicated pregnancies. Recent reviews and meta-analyses describe either heightened cytotoxicity or defective checkpoint signalling in RPL and RIF ( Guan et al. 2025 ). Parallel disturbances have been documented in myeloid lineages. Classical and intermediate monocytes from affected women exhibit activation markers and altered chemokine receptor profiles, while single-cell transcriptomics reveal enrichment of inflammatory monocyte–macrophage signatures in the endometrium of RIF patients ( Vishnyakova et al. 2019 , True et al. 2022 ). Myeloid-derived suppressor cells (MDSCs) (normally expanded in healthy pregnancy to dampen excessive inflammation) are variably reported as decreased in miscarriage cohorts ( Nair et al. 2015 ), underscoring their context-dependent role. Finally, regulatory T cells (T Reg ), key arbiters of fetal tolerance, often decline numerically or exhibit diminished suppressive capacity in both conditions, further skewing the immune milieu towards a Th1-dominant, pro-inflammatory state ( Keller et al. 2020 ). Collectively, these findings support a pathogenic model in which qualitative remodelling of NK-cell receptors, monocyte activation, and impaired immunoregulatory circuits may converge to impair decidualization, trophoblast invasion, and vascular remodelling, thereby predisposing implantation failure or pregnancy loss. Given the cumulative evidence that RPL and RIF arise from distinct yet partially overlapping defects in innate-immune regulation, we postulated that a peripheral blood-based signature incorporating NK cell receptor status, monocyte activation cues, MDSC abundance, and T Reg levels would discriminate each disorder from non-pregnant healthy controls more reliably than any single marker. By integrating deep immunophenotyping with machine learning-guided model construction, we aim to deliver a compact diagnostic panel that can be readily adopted in reproductive clinics, provide mechanistic insight into immune dysregulation underlying implantation failure, and lay the groundwork for future trials in which therapy is stratified according to an individual’s immune signature.

The authors declare that there is no conflict of interest that could be perceived as prejudicing the impartiality of the work reported.

NSIM and SSR were involved in the study design, experimental procedures, analysis, interpretation of data, and manuscript preparation. EFM, AV, IC, and LPS performed patient recruitment and maintained clinical history. MFA, MGF, and RGL interpreted the data and prepared the manuscript. All authors read and approved the final version of the article.