Rapid and accurate Bayesian diagnosis of heparin-induced thrombocytopenia

Heparin-induced thrombocytopenia (HIT) is an immune-mediated serious adverse effect of heparin treatment, characterized by a severe prothrombotic state and representing a challenging clinical emergency.¹  HIT occurs in 0.2% to 3% of patients treated with heparin; the risk varies with the type of drug and clinical setting.^2-4  It is mediated by circulating anti– platelet factor 4 (PF4)/heparin-antibody complexes, which cross-link FcγRIIa receptors on the platelet surface and induce their activation, aggregation, and procoagulant activity.^1,5  This limb- and life-threatening condition must be rapidly and reliably distinguished from other causes of thrombocytopenia⁶  to immediately stop all heparin sources and switch to alternative nonheparin anticoagulants (argatroban,^7,8  danaparoid,⁹  fondaparinux,¹⁰  or a direct oral anticoagulant¹¹ ).¹  Although unrecognized HIT is associated with high morbidity and mortality,¹²  unnecessary switching of anticoagulant therapy for misdiagnosed HIT is associated with bleeding complications and increased health care costs.¹³ 

When HIT is suspected, clinical probability can be established by use of scoring systems (eg, the validated 4T score).^14,15  Diagnosis is biologically confirmed by the detection of circulating heparin-dependent platelet-activating anti-PF4/heparin antibodies.^16,17  The functional serotonin-release assay (SRA) and the heparin-induced platelet activation assay (HIPA) are considered gold standard tests for HIT diagnosis.^18,19  However, these assays are technically demanding, expensive, and not routinely available.^1,18  Immunoassays (IAs) such as enzyme-linked immunosorbent assays (ELISAs) for detection of anti-PF4/heparin antibodies have been developed. Their seroconversion precedes SRA seroconversion by at least 1 day for most HIT cases,²⁰  and their diagnostic performance is characterized by high sensitivity and low specificity.^21,22  A negative IA result can reasonably rule out HIT, although confirming the diagnosis remains difficult. HIT-dedicated IAs have significantly evolved during the last decade.¹⁸  New techniques such as particle gel,²³  chemiluminescence,²⁴  and lateral flow²⁵  facilitate the rapid detection of anti-PF4/heparin antibodies in plasma or serum.^26,27  Two systematic reviews compared technical and analytical characteristics, as well as diagnostic performances, of rapid IAs and defined the best-suited options for incorporation into diagnostic algorithms.^16,17  Current guidelines for HIT management define the integration of emerging rapid IAs into diagnostic algorithms as a key research priority.²⁸ 

We and others^29,30  have shown that the comparison of quantitative IAs vs a functional test for heparin-dependent platelet-activating antibodies allows: (1) calculation of likelihood ratios (LRs) for result intervals; and (2) identification of cutoff IA results associated with 100% negative predictive values (NPVs) and positive predictive values (PPVs) for a positive functional assay. This enables a Bayesian diagnostic approach, transforming the clinical pretest probability in a posttest likelihood for ruling in or out HIT diagnosis based on the quantitative IA result.^1,19,29-31 

The current study aimed to assess the diagnostic performance of 3 quantitative IAs (Zymutest-HIA-IgG, HemosIL-AcuStar-HIT-IgG, and ID-PaGIA-H/PF4) to derive and prospectively validate a combined clinical and laboratory diagnostic algorithm for guiding prompt clinical management decisions.

We studied 3 cohorts of consecutive patients investigated for suspected HIT at our institution (Figure 1). The results from the retrospective and prospective derivation cohorts were used to develop a diagnostic algorithm for HIT (Figure 2), which was prospectively assessed in the validation cohort.

Between May 2014 and August 2015, a total of 221 consecutive patients with suspected HIT were investigated (Figure 1A). During this period, detection of anti-PF4/heparin antibodies was routinely performed using the Zymutest-HIA-IgG (ELISA). All 46 (20.8%) plasma samples with a positive ELISA were routinely tested by HIPA in the reference laboratory (see the “Assays description” section). In August 2015, ELISA was repeated in stored plasma samples. Passing-Bablok regression analysis showed no statistically significant difference between initial and repeat OD values (slope = 0.968 [95% confidence interval (CI), 0.924 to 1.006]; intercept = 0.009 [95% CI, −0.002 to 0.044]; P = .56), confirming that anti-PF4/heparin antibodies were stable during storage at −80°C.²⁶  This allowed a reliable comparison of the performance of the ELISA with 2 rapid IAs for HIT diagnosis (HemosIL-AcuStar-HIT-IgG, chemiluminescent IA [CLIA] and ID-PaGIA-H/PF4, particle gel IA [PaGIA]) employing stored plasma samples from the retrospective derivation cohort.

Fresh plasma samples from all 305 consecutive patients with suspected HIT sent to our laboratory between September 2015 and December 2016 were simultaneously tested at the initial diagnostic evaluation with the 3 IAs (Figure 1B). HIPA was systematically performed in samples from patients with a high pretest probability (4T score ≥6)^14,15  and/or for those with at least 1 IA result in the intermediate (CLIA ≥0.13-0.99 U/mL [discussed later]) to positive range (official cutoffs for ELISA, CLIA, and PaGIA).

Between January 2017 and November 2019, fresh plasma samples of all 687 consecutive patients with suspected HIT sent to our laboratory (Figure 1C) were investigated with the diagnostic algorithm developed on the basis of the derivation cohorts (Figure 2). HIPA was systematically performed in samples from patients with a high pretest probability (4T score ≥6)^14,15  and/or with at least 1 intermediate (CLIA ≥0.33-0.99 U/mL) or positive (CLIA ≥1.00 U/mL; PaGIA titer ≥1) result.^29,32 

This work was conducted as a systematic quality control study of our diagnostic laboratory and clinical practice. HIT suspicion is usually raised by the physicians caring for an individual patient. The hematologists on duty (a fellow and an attending) subsequently estimate the clinical pretest probability for HIT by calculating the 4T score^14,15  and in difficult cases (eg, cardiac surgery, intensive care unit patients) by also evaluating the whole clinico-pathologic picture.^33-35  Because, in rare cases, HIT is diagnosed by a positive functional assay even when the 4T score is low,^29,35-38  we measure anti-PF4/heparin antibodies in patients with a 4T score ≥2 points (lower only if there is incomplete data). The hematologist on duty may be asked to evaluate a patient with “heparin resistance,”³⁹  defined biologically by the failure to reach the therapeutic target despite administration of >1.5-fold the usually required dose of unfractionated heparin⁴⁰  (ie, >27 IU/kg body weight per hour) and clinically by the occurrence of venous or arterial thrombosis or extension of thrombosis in a patient receiving unfractionated heparin within therapeutic target range³⁹ ; when this request is made, we assess inflammation, antithrombin deficiency, and anti-PF4/heparin antibodies. A final clinical diagnosis of HIT is reached by 2 experienced hematologists (the attending in charge of the patient and the senior author of the current article) based on the combination of estimated pretest probability, quantitative IA results, outcome of HIPA, and posttesting clinical and laboratory evolution. Specifically, we calculate the clinical score described in Table 1 of our previous publication³²  (based on Greinacher at al⁴¹  and Fabris et al⁴² ). We diagnose HIT in the presence of a score ≥4 and a positive HIPA or, in case of a negative HIPA, a score ≥6 points, a positive IA, and decreasing D-dimers under alternative nonheparin anticoagulation.^8,43 

Blood was drawn into 3 mL plastic syringes (Monovette; Sarstedt, Nümbrecht, Germany) containing 0.3 mL 0.106 mol/L trisodium citrate. Plasma was prepared by double centrifugation at 1500g during 10 minutes at room temperature. Plasma samples were then snap-frozen for storage in polypropylene tubes at −80°C.

The Zymutest-HIA-IgG (Hyphen BioMed, Neuville-sur-Oise, France) is a commercially available immunoglobulin G (IgG)-specific ELISA coated with heparin-protamine complexes in which PF4 is provided by a platelet lysate added to the reaction mixture.⁴⁴  Analytical turnaround time (TAT) is ∼3 hours (performed Monday to Friday, 8:00 am to 4:00 pm). The cutoff recommended by the manufacturer is set at ∼0.3 OD (depending on the daily determination of the control sample).

The HemosIL-AcuStar-HIT-IgG (Instrumentation Laboratory GmbH, Munich, Germany) is an automated CLIA with PF4 bound to polyvinyl sulfonate particles.^24,45,46  Anti-PF4/heparin antibodies form a complex with PF4/polyvinyl sulfonate, which is adsorbed on magnetic beads. After separation of the microparticles, an isoluminol-labeled anti-human-IgG-antibody is added. After washing, light emission intensity measured in relative light units is directly proportional to the anti-PF4/heparin-IgG-antibody concentration. TAT is ∼30 minutes (performed round-the-clock). The cutoff recommended by the manufacturer is ≥1.0 U/mL. The reproducibility of results obtained with the HemosIL-AcuStar-HIT-IgG was assessed. The interassay coefficient of variation lay between 3.1% (positive control sample, mean 2.31 U/mL, SD 0.071, n = 14), 4.1% (negative control sample, mean 0.49 U/mL, SD 0.02, n = 14), and 6.2% (clinical plasma sample, mean 1.07 U/mL, SD 0.066, n = 4). The intra-assay coefficient of variation was 3.3% (clinical plasma sample, mean 1.13 U/mL, SD 0.037, n = 11).

The ID-PaGIA-H/PF4 (Bio-Rad, DiaMed SA, Basel, Switzerland) is a PaGIA that detects IgG–, IgM–, and IgA–anti-PF4/heparin antibodies with a TAT of ∼30 minutes²³  (performed Monday to Friday 7:00 am to 6:00 pm; on weekends 8:00 am to 12:00 pm). Ten microliters of plasma are added into a reaction chamber of the ID-card test, followed by 50 μL of polymer particles (red high-density polystyrene beads coated with PF4/heparin-complexes). After incubation for 5 minutes at room temperature, the ID-card is centrifuged for 10 minutes with a dedicated ID-centrifuge (DiaMed SA). Anti-PF4/heparin antibodies cross-link the red polymer particles, which remain on the top of the gel chamber. In the absence of a significant level of anti-PF4/heparin antibodies, the particles sink to the bottom of the gel chamber. The result of the card is read by the laboratory operator. In case of indeterminate or positive result with the undiluted sample, the analysis is repeated in serially diluted plasma samples by using specific diluent II (DiaMed SA).^26,29,32  The titer of the positive result with the highest dilution, followed by an indeterminate or negative result in the subsequent dilution, is reported after confirmation by dual control. According to the manufacturer, the official cutoff is a positive result in the undiluted probe. Interassay and intra-assay reproducibility has been previously assessed.^26,32 

The HIPA⁴⁷  is recognized as 1 of the 2 gold standard tests (along with the SRA) for the detection of heparin-dependent, platelet-activating antibodies.^18,19  Patient plasma is added to washed reactive platelets from selected healthy donors. If functional anti-PF4/heparin antibodies are present, platelet aggregation is observed at low heparin concentration and is suppressed at high concentration. This functional assay (and an ELISA) was performed at the “Thrombozytenlabor,” Institute for Immunology and Transfusion Medicine, University Hospital of Greifswald, Greifswald, Germany.

We report median and interquartile range values. MedCalc (version 15.11.0) was used for calculating Passing-Bablok regression, receiver-operating characteristic (ROC) curves, sensitivity and specificity of different cutoff values, 100% NPV and PPV for a positive HIPA, and LR for intermediate results (“gray zone”) of each IA. Concise information on the Passing-Bablok statistical procedure is presented at www.medcalc.org/manual/passing-bablok_regression.php. A detailed description of the Bayesian analytical approach is found in the “Brief tutorial on ROC analysis and clinical application of Bayes’ theorem” published in the Haematologica journal Web site under “Figures & Data” as the Online Supplementary Appendix to our previous publication.²⁹  Briefly, pretest probability of HIT was assessed by using the 4T score.^14,15  This value was transformed into a posttest probability by combining it with the LR (95% CI limits) of the quantitative IA result (online calculator at http://www.sample-size.net/post-probability-calculator-test-new). For constructing Figure 2, when using the second IA, the first posttest probability was combined with the LR (95% CI limits) of the PaGIA titer.

The study was conducted according to the guidelines of the competent ethical board (Commission cantonale d’éthique de la recherche sur l’être humain). Patient informed consent was waived for this systematic quality control study of a diagnostic laboratory practice (Commission cantonale d’éthique de la recherche sur l’être humain, protocol number 497/15).

The study design is depicted in Figure 1. Overall, we included all 1116 consecutive patients with suspected HIT, whose plasma samples had been sent to our laboratory for diagnostic evaluation. Among the 221 patients from the retrospective derivation cohort, 46 individual plasma samples with a positive ELISA for anti-PF4/heparin antibodies (Zymutest-HIA-IgG) were retrospectively investigated with repeated ELISA, CLIA, and PaGIA. In the prospective derivation cohort, plasma samples from all 305 patients with suspected HIT were simultaneously analyzed with the 3 IAs. Finally, during the prospective validation phase, all 687 patients with suspected HIT were evaluated on the basis of the diagnostic algorithm derived from the analysis of both derivation cohorts (Figure 2).

Patient demographic and clinical characteristics with results of the diagnostic evaluation for suspected HIT are summarized in Table 1. No relevant differences were observed in the 3 cohorts regarding patients’ age, sex, clinical settings, and laboratory results. The prevalence of HIPA-positive cases was 10% (22 of 221) in the retrospective derivation cohort, 8.5% (26 of 305) in the prospective derivation cohort, and 7.7% (53 of 687) in the validation cohort. In the 3 cohorts, the median 4T score was 4 in HIPA-negative patients and 5 in HIPA-positive patients. A low 4T score (0-3 points) was associated with a positive HIPA in 7 (1.7%) of 407 cases in both the derivation and validation cohorts (supplemental Table 3, available on the Blood Web site).

Results of the retrospective and prospective derivation cohorts are summarized in supplemental Tables 1 and 2 and supplemental Figure 1. According to areas under the ROC curves, CLIA and PaGIA performed similarly (P = .59 in the retrospective derivation cohort; P = 1.0 in the prospective derivation cohort). We observed a trend toward a better performance of CLIA and PaGIA compared with ELISA (areas under the ROC curves of ELISA vs CLIA in the retrospective and prospective derivation cohort, P = .0184 and P = .061, respectively; ELISA vs PaGIA, P = .0094 and P = .0647).

Considering the cutoffs recommended by the manufacturers, PaGIA and ELISA exhibited excellent sensitivity (100% and 96.2%, respectively) in the prospective derivation cohort with a high number of false-positive results (PaGIA, 69 of 95 [72.6%]; ELISA, 12 of 37 [32.5%]). In contrast, and worryingly from a clinical point of view, CLIA had a specificity close to 100% but a high number of false-negative results: 4 (18.2%) of 22 in the retrospective derivation cohort, 5 (19.2%) of 26 in the prospective derivation cohort, and 4 (8.5%) of 47 in the prospective validation cohort (supplemental Table 4).

According to cutoffs with a 100% PPV, CLIA predicted 38 (79.2%) of 48 HIPA-positive samples, PaGIA 31 (64.6%) of 48, and ELISA 11 (22.9%) of 48 in the pooled derivation cohorts. As for 100% NPV, CLIA predicted 276 (91.5%) of 303 HIT-negative samples, PaGIA 272 (89.8%) of 303, and ELISA 263 (86.8%) of 303. The remaining samples had results falling into the respective intermediate gray zone between cutoff values for 100% NPV and PPV (37 of 351 [10.5%] with CLIA; 48 of 351 [13.7%] with PaGIA; and 77 of 351 [21.9%] with ELISA). For these intermediate results, the LR for a positive HIPA with each IA was calculated (Table 2).

Based on the results of the derivation cohorts, a diagnostic algorithm was established (Figure 2). From January 2017 to November 2019, a total of 687 consecutive patients were investigated for suspected HIT (Figure 1C). According to the algorithm, CLIA was used as a unique laboratory test in 566 (82.4%) patients, PaGIA as second-line assay in 121 (17.6%), and HIPA as third-line assay in 146 (21.3%). HIPA was positive in 53 instances (7.7%).

HIT was ruled out in 523 (76.1%) patients based on the combination of a low to intermediate pretest probability and a CLIA <0.13 U/mL (Figure 1C). Eight additional patients (1.2%) had a CLIA <0.13 U/mL but a high pretest probability. In 6 of these patients, the algorithm excluded HIT by a negative PaGIA (the exclusion was confirmed by a negative HIPA). The remaining 2 patients had an “undetermined” algorithm prediction and were resolved by a negative HIPA.

A total of 111 patients (16.2%) with CLIA values falling within the intermediate gray zone of our algorithm (0.13-3.0 U/mL) were further investigated by using PaGIA (Figure 2). Table 3 summarizes the individual diagnostic evaluations. In 74 (10.8%) patients, HIT was ruled out based on the combination of pretest probability and CLIA followed by PaGIA. Fifty of these 74 instances were verified by use of HIPA (as discussed in the Methods) (Figure 1C). Seventeen (2.5%) cases had an “undetermined” algorithm prediction even after PaGIA; 16 were solved by HIPA (negative in 15, positive in 1), and laboratory evaluation remained inconclusive in 1 case because of an atypical aggregation in the HIPA. In 20 patients (2.9%), HIT was predicted by the algorithm, based on the combination of pretest probability, CLIA, and PaGIA. In 10 cases, the prediction was confirmed by a positive HIPA. In 10 cases, HIPA remained negative (supplemental Table 5). However, in 8 patients, clinical courses were compatible with HIT and raise the possibility of false-negative HIPA finding.^18-20  In 2 patients, clinical course was compatible with HIT, but results of the ELISA performed in the reference laboratory remained negative. Therefore, no definitive conclusion was possible.

Finally, 45 patients (6.6%) had a CLIA >3.0 U/mL (Table 4). In 2 cases, the pretest probability was low. In 1 case, the combination of the low pretest probability and a negative PaGIA ruled out HIT according to the algorithm (confirmed by a negative HIPA). The other had an “undetermined” algorithm prediction even after PaGIA and a positive HIPA solved the case. In 43 (6.3%) instances, an intermediate to high pretest probability for HIT combined with the CLIA predicted HIT. This was confirmed by a positive HIPA in 41 cases. In 2 cases (0.3%), HIPA remained negative, but clinical course was compatible with HIT (supplemental Table 5). For these samples, PaGIA (not required by the algorithm) was performed post hoc.

Assessing clinical pretest probability in cardiac surgery³³  and ICU^34,35  patients is challenging and could affect the application of a diagnostic algorithm based on the 4T score. Following algorithm predictions were observed in our prospective validation cohort. In cardiac surgery (n = 36), there were 24 HIT exclusions, 4 not solved by the algorithm (all HIPA negative), and 8 HIT predictions (6 confirmed by HIPA and 2 not). Second, in the ICU (n = 106), there were 92 HIT exclusions, 3 were not solved (all HIPA negative), and 11 HIT predictions (7 confirmed by HIPA and 4 not). HIT predictions not confirmed by HIPA are reviewed in the Discussion and summarized in supplemental Table 5.

In summary, the diagnostic algorithm (Figure 2) ruled out HIT diagnosis in 604 patients (87.9%) and accurately predicted a positive HIPA in 51 patients (7.4%) with a 30- to 60-minute analytical evaluation. No false-negative HIT prediction was detected. In 12 (1.7%) cases, HIT was predicted by the algorithm but was not corroborated by HIPA. In 10 instances, the further clinical courses strongly suggested HIT, leaving 2 cases (0.3%) of possible false-positive HIT prediction (supplemental Table 5). The algorithm left 20 cases (2.9%) unresolved: in 17, HIPA was negative, in 2 positive, and in 1 inconclusive.

Untreated HIT is associated with high morbidity and mortality, whereas an unnecessary switch to alternative nonheparin anticoagulants increases the risk of bleeding complications and treatment costs.^1,12,13  Clinical suspicion of HIT therefore requires rapid and accurate laboratory assessment to guide management.^21,48  The current study evaluated the diagnostic accuracy of 3 IAs for detection of clinically relevant anti-PF4/heparin antibodies. Subsequently, we derived and prospectively validated a rapid and accurate diagnostic algorithm for patients with suspected HIT.

Evaluation of the 3 IAs revealed that the sensitivities of the recommended cutoffs differ significantly: 100% for ID-PaGIA-H/PF4, 96.2% for Zymutest-HIA-IgG, and 80.8% for the HemosIL-AcuStar-HIT-IgG (supplemental Table 1). With regard to this last IA, we confirm^38,46  that the official cutoff value set at ≥1.0 U/mL is too high, resulting in a not inconsequential number of false-negative results (13 of 101 HIPA-positive samples) (supplemental Table 4). Of note, Althaus et al⁴⁶  previously observed a sensitivity of 96.2% for the official cutoff of the HemosIL-AcuStar-HIT-IgG, identifying an ideal cutoff according to ROC analysis of 0.57 U/mL, which is similar to our data. Warkentin et al³⁸  reported a sensitivity <100% of the official cutoff of the HemosIL-AcuStar-HIT-IgG.

We show that all 3 IAs perform best by using the magnitude of the quantitative values instead of a positive/negative result based on a fixed cutoff value. In particular, we confirm that the PaGIA, which has a high rate of false-positive results when evaluated only in the undiluted probe (supplemental Table 1),^36,49  is most useful when a positive result is assessed semi-quantitatively according to the measured titer (Table 2).^26,29,32  For CLIA, we confirm that the magnitude of its results is associated with the likelihood for HIT (Table 2; Figure 2), as recently observed by Warkentin et al.³⁸ 

In clinical practice, the most relevant diagnostic performance of an IA for anti-PF4/heparin antibodies is its ability to significantly increase or decrease the pretest probability by predicting the outcome of a functional gold standard assay.^{1,19,21,29-31,36,48}  This diagnostic approach is depicted in supplemental Table 2. In addition, we evaluated the performance of cutoff values with 100% NPV or PPV for a positive HIPA (Table 2).²⁹  We found that the 100% predictive values of the CLIA, which has an analytical TAT of 30 minutes, could solve ∼90% (314 of 351) of cases of suspected HIT (supplemental Table 1). We confirmed that PaGIA titers ≤1 and ≥16 were able to rapidly rule out or rule in a positive HIPA, respectively.^29,32  Thus, the manual semi-quantitative assessment of the titer in a positive PaGIA could rapidly solve ∼85% (303 of 351) of cases of suspected HIT.^26,29,32  In addition, LRs of PaGIA titers are in line with our previous report.²⁹ 

Based on these observations, we derived an in-house diagnostic algorithm expected to solve ≥95% of cases of suspected HIT with an analytical TAT of ≤1 hour (Figure 2). The CLIA is used as the first-line test because of its performance in both derivation studies and efficient automated diagnostic laboratory evaluation. To minimize the risk of false-negative and false-positive results, we set conservative cutoffs for the 100% predictive values for a positive HIPA (<0.13 U/L for NPV and >3.0 U/mL for PPV). The probability by which a quantitative result below/above these cutoffs predicts a negative/positive HIPA depends on the individual pretest probability of HIT and the 95% CI of the LR of the corresponding result intervals (Table 2). For results lying in the CLIA gray zone (0.13-3.0 U/mL), owing to its performance,²⁹  we chose to use the PaGIA as the second-line test and sequentially combine the LR of its titer with the previously assessed clinico-pathologic probability of HIT.

The prospective validation of our diagnostic approach shows an excellent clinical performance by accurately solving 96.8% of cases of suspected HIT with an analytical TAT of 30 to 60 minutes (Tables 3 and 4). Of note, no false-negative results were observed (Figure 1C). This is reinforced by the fact that only 2 of the 20 cases unresolved by the diagnostic algorithm had a positive HIPA, underscoring the cautious nature of our criteria for HIT exclusion. In 12 (1.7%) cases, the algorithm-based prediction of HIT was not confirmed by HIPA. In 10 cases, the patients’ clinical courses, including prompt resolution of thrombocytopenia and reduction of fibrin D-dimers after switching to alternative nonheparin anticoagulation, were judged compatible with HIT, suggesting false-negative HIPA results (supplemental Table 5). Of note, in 7 instances, with high titer confirmatory anti-PF4/heparin antibodies by ELISA, the external reference laboratory could not exclude HIT despite the negative HIPA. The laboratory reports advised starting alternative anticoagulation in case of strong suspicion of HIT, thus replicating the clinical management decision issued by our diagnostic algorithm.

Although the HIPA and the SRA are considered gold standard tests for HIT, a “limited” sensitivity of both methods is recognized.^18,19  For instance, cases of “SRA-negative HIT” have been recently documented by reference HIT laboratories of the Service d'Hématologie-Hémostase, CHRU Tours (France),⁵⁰  Bloodcenter of Wisconsin, Milwaukee (United States),⁵¹  and McMaster University, Hamilton (Canada).⁵²  These observations are in line with those of Warkentin et al,^20,53  who recently observed that anti-PF4/heparin-IgG-antibodies may be detectable by IAs before SRA seroconversion. This raises 2 implications for our study. First, it offers a rationale for false-negative HIPA findings, other than those due to technical issues, such as the presence in patient plasma of substances inhibiting formation of PF4/heparin-complexes (eg, danaparoid⁵⁴ ) and/or activation of donor platelets (eg, anti-P2Y12 ADP-receptor inhibitors). Second, it highlights the impact of diagnostic timing, allowing HIT to be suspected in the time window between immunologic and functional seroconversion.^19,20,53,55 

Because our study was conducted in the context of real-world clinical practice, we did not ask for the functional gold standard in all patients (Figure 1; Table 2). This potential partial verification bias is the major limitation of our work. However, we performed the HIPA more frequently than required by current guidelines (as discussed in the Methods section).^{1,4,28,30,31,36}  Of particular note, our approach allowed us to document that the recommended cutoff of the HemosIL-AcuStar-HIT-IgG (≥1.0 U/mL) generated a high number of false-negative results (supplemental Tables 1 and 4). Our in-house algorithm also allowed us to exclude HIT in the presence of a low or intermediate pretest probability only if CLIA was <0.13 U/mL. This very conservative negative cutoff is in line with a recent publication of Warkentin et al³⁸  showing that a CLIA ≥0.4 U/mL had a sensitivity of 100%. Finally, it has been shown that this type of partial verification bias does not affect the PPV.⁵⁶  These considerations support the validity of our diagnostic algorithm.

Another potential limitation of our work is the use of the HIPA as the laboratory gold standard instead of the SRA, which is the classical test for detecting platelet-activating antibodies. However, as stated in a recent review,¹⁸  both tests are currently accepted as gold standard HIT assays, with high sensitivity and specificity.^18,19  Of note, there are few direct comparisons of these assays; although one publication indicated similar diagnostic performance,⁵⁷  in two other instances the HIPA showed a slightly higher sensitivity for anti-PF4/heparin antibodies compared with the SRA.^58,59  This could explain the slight differences between our data and those published by Warkentin et al³⁸  regarding the performance of the CLIA.

Finally, we did not evaluate the impact of our algorithm on clinical outcomes of patients. However, improving diagnostic accuracy is reasonably expected to decrease the rate of complications due to delayed diagnosis or misdiagnosis.^12,13  Based on the assumptions used in Table 4 of the 2018 American Society of Hematology (ASH) guideline (1000 patients evaluated for HIT suspicion; 11% prevalence of HIT),²⁸  our approach is expected to rapidly exclude HIT in 863 of 890 patients (during the prospective validation, the algorithm excluded HIT in 604 of 623 negative patients) and correctly identify at least 106 of 110 patients with HIT (the algorithm predicted 51 of 53 HIPA-positive samples) at the expense of 3 of 1000 possibly false-positive HIT predictions (2 in our cohort of 687 patients). In 29 of 1000 instances with undetermined algorithm prediction, an expected maximum of 4 would eventually turn out having HIT. Overall, with a rapid analytical TAT of up to 1 hour, ∼973 of 1000 patients would be treated appropriately, and ∼27 of 1000 would possibly unnecessarily receive non-heparin alternative anticoagulation (until resolved by mandatory functional testing). Of note, without false-negative HIT predictions, our algorithm would prevent thrombosis (estimated individual risk in unrecognized HIT, 30%),²⁸  amputations, and death (each 6%) at the price of increased bleeding risk in <3% of patients unnecessarily treated with non-heparin anticoagulation (estimated individual risk, ∼8%-35% over treatment duration). This scenario compares well with the strategy recommended by the ASH guideline that, assuming a delay of 1 to several days until IA results are available (ELISA), would lead to 40.8% patients who do not have HIT receiving non-heparin anticoagulation unnecessarily for varying periods of time and which, of particular note, would miss HIT diagnosis in 10 of 110 patients.

The excellent performance of our diagnostic algorithm can be explained by the following: (1) we used 2 of the 5 IAs that showed a combination of high sensitivity (>95%) and specificity (>90%) in a recent meta-analysis¹⁷ ; (2) both tests are among the 3 rapid IAs found to be particularly well suited for incorporation into diagnostic algorithms by another recent meta-analysis¹⁶ ; and (3) by applying “extreme” cutoff values, we maximized sensitivity (low threshold with 100% NPV) and specificity (high threshold with 100% PPV) of both IAs.¹⁷  In addition, the combination of the IgG-specific CLIA as first-line automated IA with the polyspecific PaGIA for initially unresolved cases increases the discriminative power of each IA by itself.

Our results are of immediate practical relevance because they: (1) respond to the most recent ASH guidelines on HIT, indicating that one of the key research priorities is the integration of emerging rapid IAs into diagnostic algorithms²⁸ ; (2) provide a guidance for interpreting quantitative results of 2 very valuable rapid IAs for anti-PF4/heparin antibodies, HemosIL-AcuStar-HIT and ID-H/PF4-PaGIA (supplemental Table 2)^16,17 ; (3) underscore the utility of quantifying PaGIA according to titer^26,29 ; (4) alert personnel who use the HemosIL-AcuStar-HIT-IgG that the recommended cutoff is too high and generates a high number of false-negative results (supplemental Table 4)⁴⁶ ; (5) support recent observations that diagnostic testing for HIT might be useful in some low pretest probability situations (Figure 2; supplemental Table 3)^36,38 ; and (5) raise the hypothesis that a reason for a negative HIPA may be its too-early timing.^19,20 

In conclusion, we showed that establishing cutoffs with 100% PPV and NPV and LRs for result intervals substantially improved IA accuracy for diagnosing HIT. The approach combining the estimated pretest probability and the quantitative results of 2 sequential IAs was able to rule in or out the diagnosis within 1 hour in >95% of patients with suspected HIT (Figures 1C and 2; Tables 3 and 4). In the remaining cases, management decisions have to be based on individualized judgment while awaiting HIPA results.^1,18  This original multistep strategy combining clinical and laboratory assessment results in rapid and accurate diagnostic evaluation and will improve prompt clinical management of patients with suspected HIT. We intend to verify the potential generalization of this approach and these cutoffs in a multicenter study with the ultimate goal of developing a diagnostic application for smartphones and tablets.⁶⁰ 

This study constitutes the medical master thesis of M.M. and has been partially presented at several scientific meetings (noted in the footnotes).

The authors thank the physicians caring for the patients, the laboratory technicians for accurate diagnostic evaluation (Cindy Celeste, Sabrina Jordi, Vicky Menétrey, Raphael Bauer, and Mathieu di Tomaso), Michel Duchosal for support, Thomas Thiele and Andreas Greinacher for supervising the HIPA assay, Pierre-Alexandre Bart for acting as expert of the master thesis of M.M., and Oscar Marchetti for careful reading of the manuscript.

Contribution: M.M. performed research, analyzed data, and wrote the manuscript; S.B. designed research, analyzed data, and revised the manuscript; M.G.Z. analyzed data and wrote the manuscript. F.M.-R., E.M.-G., and N.N. performed research; F.G. and C.G. performed research and managed the database; M.G. helped to collect data and revised the manuscript; and L.A. designed research, analyzed data, and wrote the manuscript. All authors read and approved the final version of the manuscript.

Conflict-of-interest disclosure: The authors declare no competing financial interests.

Correspondence: Lorenzo Alberio, Division of Hematology and Central Haematology Laboratory, Lausanne University Hospital (CHUV), Rue du Bugnon 46, CH–1011 Lausanne, Switzerland; e-mail: lorenzo.alberio@chuv.ch.