This prospective method-comparison study was conducted in the biology laboratory of the Abidjan Heart Institute, a tertiary cardiovascular referral centre in Côte d’Ivoire, from June to August 2026. Reporting followed the Guidelines for Reporting Reliability and Agreement Studies (GRRAS) [18].

Patients of all ages receiving long-term VKA therapy and attending the laboratory for scheduled INR monitoring were eligible. Patients were enrolled consecutively as they came for routine INR monitoring, and each eligible monitoring episode was included until the available capillary test strips ran out. No sample-size calculation was performed in advance, since recruitment was mainly limited by the cost and availability of the strips. Each monitoring episode was analysed as a separate observation.

The dataset first included 132 entries. One was a summary line rather than a patient visit, so it was excluded, leaving 131 paired monitoring episodes for analysis.

Patient-level linkage identified 101 unique patients: 78 contributed one visit, 16 contributed two visits, and seven contributed three visits. Figure 1 summarises the study and analysis populations.

For each monitoring episode, venous blood was collected first, followed immediately by capillary testing; the two measurements were obtained less than two minutes apart so that both reflected essentially the same clinical state. Citrated venous blood was analysed on a Sysmex CS-2500 automated coagulation analyser using a chronometric method with Dade® Innovin® thromboplastin reagent (Siemens Healthineers). Calibration was performed according to the laboratory’s routine procedure using a Siemens calibrator. Internal quality control was performed daily at two levels using Dade® Ci-Trol® Level 1 and Dade® Ci-Trol® Level 2 control materials (Siemens Healthineers). The lot-specific ISI used during the study could not be retrieved from the archived laboratory records. Under the laboratory’s routine reporting procedure, INR values above 6 were systematically reported as “>6”; this was treated as a reporting threshold and not assumed to represent the intrinsic analytical limit of the analyser/reagent combination. Capillary INR was measured with the CoaguChek® XS Pro device (Roche Diagnostics) using a 10 - 15 µL fingerstick whole-blood sample. The manufacturer-stated INR measuring range for the CoaguChek XS Pro system is 0.8 - 8.0. A single trained operator performed all capillary measurements. The device’s integrated internal quality-control function was used; no external liquid control solution was used during the study. Laboratory personnel performing the venous INR assay were not informed of the capillary result.

For each episode, the patient’s therapeutic INR interval had already been

Figure 1. Study flow and analysis populations. The primary continuous agreement analysis used 129 exact numerical pairs because two venous INR results remained right-censored as >6. All 131 episodes contributed to treatment-decision concordance. The value 6.00 was assigned to the two censored results only in the post-hoc sensitivity analysis.

documented in the medical record before the paired venous and capillary measurements were obtained. This pre-existing interval was applied independently to both measurements using a prespecified directional rule. A value below the lower boundary indicated a dose increase; a value within the interval, including either boundary, indicated dose maintenance; and a value above the upper boundary indicated a dose reduction. For the final analysis, treatment categories were recalculated directly from the recorded therapeutic interval. These were algorithm-derived analytical categories rather than observed clinician decisions, so no discretionary treatment-category assessor was used. Across the 131 monitoring episodes, the documented therapeutic intervals were 2.0 - 3.0 in 63 episodes, 2.5 - 3.5 in 55, 3.0 - 4.5 in 10, and 2.0 - 4.0, 3.0 - 4.0, and 3.5 - 4.5 in one episode each.

Before analysis, the merged source file was reconciled with the available study records. One confirmed capillary INR transcription error was corrected from the source record, and exact high-range capillary values were restored where they could be recovered. Each correction was retained in an audit trail. Hospital record numbers were normalised to link repeated visits deterministically; direct identifiers were then removed and replaced by sequential study identifiers.

The primary continuous agreement analysis used paired numerical INR values. Five venous values that had previously been treated as censored were recovered as exact measurements from the merged source file. Two venous observations remained right-censored as >6. Because an exact pairwise difference cannot be calculated from a right-censored value, these two observations were excluded from the primary continuous analysis, leaving 129 exact pairs. They remained in the treatment-decision analysis because both were clearly above their documented therapeutic ranges. In a post-hoc sensitivity analysis only, the censoring threshold of 6.00 was assigned to these two observations so that all 131 episodes could be analysed numerically.

Among the 101 patients, 23 had more than one INR monitoring visit. The primary agreement analysis therefore used a linear mixed-effects Bland-Altman model for the paired difference d_ij = capillary INR − venous INR, with a fixed intercept representing the overall mean bias and a patient-specific random intercept representing between-patient variation. The model estimated between-patient and within-patient residual variance components, and the repeated-measures 95% limits of agreement were calculated as the estimated mean bias ± 1.96 times the square root of the sum of these two variance components [19]-[21]. Patient-cluster bootstrap resampling was performed by sampling patients with replacement and retaining all monitoring episodes contributed by each selected patient. Twenty thousand bootstrap replicates were used to obtain cluster-aware confidence intervals for the bias, the limits of agreement, and Cohen’s kappa. Analyses were performed in Python 3.13 using statsmodels 0.14.6 for the mixed-effects model and SciPy 1.17.0 for ancillary statistical calculations; the patient-cluster bootstrap and concordance calculations were implemented in custom Python code.

Pearson’s correlation coefficient was reported only as a descriptive measure of linear association and not as evidence of interchangeability.

Possible proportional bias was examined by regressing the paired difference on the pairwise mean, while heteroscedasticity was assessed from the association between the absolute difference and mean INR. Because the differences between the two methods became more variable as INR increased, we added a post-hoc Bland-Altman analysis based on log ratios as a sensitivity check. No therapeutic-boundary proximity margin had been prespecified, so we did not introduce an arbitrary threshold around the therapeutic limits. Instead, heteroscedasticity was explored descriptively in post-hoc strata defined by the mean of the paired INR measurements: <2.0, 2.0 - 3.0, 3.01 - 4.0, and >4.0. Within each stratum, the mean difference and conventional 95% limits of agreement (mean difference ± 1.96 SD) were calculated as descriptive subgroup summaries; the repeated-measures model remained the primary agreement analysis. Treatment decisions were compared across all 131 monitoring episodes using the proportion of matching decisions and Cohen’s kappa [22][23]. A disagreement was considered major only when the two methods suggested opposite dose changes—an increase with one and a reduction with the other. All remaining disagreements were classified as minor. Linear- and quadratic-weighted kappa values were also calculated as sensitivity analyses. Because no clinically acceptable difference or agreement margin had been defined in advance, the observed limits were interpreted descriptively rather than as evidence of complete interchangeability.

The protocol was submitted to the competent national ethics committee before recruitment began. Before the first participant was enrolled, the investigators received oral notification that the protocol had received a favourable opinion. Recruitment and paired study sampling began only after that notification, with the informed-consent procedures specified in the protocol already in place. Formal written authorization was subsequently issued on 8 July 2026 (reference 00137/26/ MSHPCMU/CNESVS-km). No participant was enrolled or underwent paired study sampling before the oral favourable opinion had been communicated and consent procedures were in place. Informed consent was obtained from adult participants. For participants younger than 18 years, consent was obtained from the parent accompanying the minor.

During manuscript preparation, OpenAI ChatGPT (accessed August 2026) was used to assist with English-language editing, organization of the text, and reference checking. It was not used for participant recruitment, data collection, or statistical analysis. The authors reviewed and revised all AI-assisted content and remain responsible for the final manuscript.

The analysis included 131 monitoring episodes from 101 patients. Seventy-eight patients contributed one episode, 16 contributed two, and seven contributed three. At the episode level, mean age was 46.8 years (SD 17.2; range 8 - 84); 13 episodes (9.9%) involved participants younger than 18 years, and 85 (64.9%) involved women. Acenocoumarol was used in 100 episodes (76.3%), and a mechanical prosthetic valve was the indication for VKA therapy in 72 (55.0%) (Table 1).

Table 1.Characteristics of the analysed monitoring episodes.

*Age and sex are summarised at the monitoring-episode level because some patients contributed more than one visit. Percentages for episode-level characteristics use 131 monitoring episodes as the denominator; the final row uses the 101 unique patients. Patient-level clustering was incorporated into the repeated-measures analyses.

The primary continuous agreement analysis included 129 exact paired measurements from all 101 patients; two venous observations remained right-censored as >6. Capillary and venous INR values were closely associated (Pearson r = 0.934, p < 0.001; Figure 2), although correlation was not used to judge interchangeability.

The repeated-measures Bland-Altman model estimated a mean capillary-minus-venous bias of +0.087 INR units, with 95% limits of agreement from −0.887 to +1.061 (Figure 3). The estimated between-patient variance of the paired differences was 0.00895 and the within-patient residual variance was 0.23796. Patient-cluster bootstrap 95% confidence intervals were −0.004 to +0.170 for the mean bias, −1.288 to −0.559 for the lower limit, and +0.760 to +1.358 for the upper limit. Five of the 129 exact differences (3.9%) fell outside the model-based limits.

The difference between methods increased modestly with the pairwise mean (slope 0.066 INR units per one-unit increase in mean INR; p = 0.046), while absolute dispersion rose more clearly as INR increased (Spearman rho = 0.512; p < 0.001). On the log-ratio scale, the geometric mean capillary/venous ratio was 1.030, with 95% ratio limits of agreement from 0.773 to 1.373. Residual heteroscedasticity was reduced (Spearman p = 0.052).

Figure 2. Capillary versus venous INR measurements for the 129 episodes with exact paired numerical results. The dashed line is the line of identity and the solid line is the descriptive linear regression. Pearson r = 0.934, p < 0.001. Two venous results retained as >6 were excluded from the primary continuous analysis.

Figure 3. Repeated-measures Bland-Altman agreement for 129 exact pairs from 101 patients. The solid line shows the estimated mean bias (+0.09 INR units); dashed lines show the repeated-measures 95% limits of agreement (−0.89 to +1.06). Two venous observations retained as >6 were excluded from the primary continuous analysis.

Post-hoc descriptive strata showed widening dispersion: mean INR < 2.0 (n = 40), bias +0.05 with 95% limits of agreement −0.41 to +0.50; INR 2.0 - 3.0 (n = 51), +0.10 with limits −0.46 to +0.66; INR 3.01 - 4.0 (n = 20), +0.04 with limits −0.89 to +0.97; and INR > 4.0 (n = 18), +0.17 with limits −2.00 to +2.34. These subgroup analyses were post hoc and descriptive.

The post-hoc sensitivity analysis assigned a value of 6.00 to the two right-censored venous observations and therefore included all 131 episodes. The mean bias was +0.115 INR units, with repeated-measures 95% limits of agreement from −0.957 to +1.186. Because the original laboratory results were reported as >6 rather than as exact values, these estimates were considered secondary.

The two methods produced the same algorithm-derived treatment category in 108 of 131 episodes (82.4%). Cohen’s kappa was 0.731, with a patient-cluster bootstrap 95% confidence interval of 0.627 to 0.826 based on 20,000 resamples. The remaining 23 episodes (17.6%) were discordant, but all were adjacent-category disagreements. No episode had one method indicating a dose increase while the other indicated a dose reduction. Linear- and quadratic-weighted kappa values were 0.792 and 0.856, respectively (Figure 4).

Figure 4. Treatment-decision concordance between capillary and laboratory INR for all 131 monitoring episodes. Rows show the laboratory decision and columns the capillary POC decision. Diagonal cells represent concordant decisions (108 episodes). Overall agreement was 82.4%; Cohen’s kappa was 0.731 (patient-cluster bootstrap 95% CI 0.627 - 0.826). All 23 discordances were minor, with no opposite dose directions.