SEARCH

SEARCH BY CITATION

Keywords:

  • thombocytopenia;
  • bleeding;
  • platelet transfusion;
  • clinical prediction rule

Abstract

  1. Top of page
  2. Abstract
  3. MATERIALS AND METHODS
  4. RESULTS
  5. DISCUSSION
  6. REFERENCES

BACKGROUND

The correlation between platelet count and bleeding has been well described, although no formal methods for applying this information to clinical decisions are available. The authors developed a clinical prediction rule to guide the prophylactic use of platelet transfusions among patients with lymphoma or solid tumors.

METHODS

The Bleeding Risk Index (BRI) was developed from logistic regression analysis of a randomly selected 750-chemotherapy cycle derivation set using data from Day 1 of cycles. The sensitivity and specificity of a BRI-based prophylaxis strategy were compared in a 512-cycle validation set with two strategies based on initiation of prophylaxis when platelet counts fell below thresholds of 20,000 per μL or 10,000 per μL.

RESULTS

Factors that were predictive of bleeding included any prior episode of bleeding (odds ratio [OR], 5.6; 95% confidence interval [95% CI], 2.2–14.0), treatment with a drug affecting platelet function (OR, 5.1; 95% CI, 2.0–12.6), bone marrow metastases (OR, 4.3; 95% CI, 1.7–10.8), a baseline platelet count < 75,000 per μL (OR, 3.5; 95% CI, 1.4–8.9), genitourinary or gynecologic malignancy (OR, 3.3; 95% CI, 1.3–8.2), a Zubrod performance status score > 2 (OR, 3.4; 95% CI, 1.4–8.5), and treatment with agents that were highly toxic to the bone marrow (OR, 2.2; 95% CI, 1.0–5.4). Compared with 20,000 and 10,000 platelet threshold strategies, the BRI-based strategy provided the best trade-off between sensitivity for major bleeding episodes (80%) and specificity for any bleeding (84%).

CONCLUSIONS

Patients with lymphoma or solid tumors who are at high risk of bleeding can be identified reliably on Day 1 of a chemotherapy cycle. An individualized, BRI-based approach to bleeding prophylaxis provides a highly sensitive and specific alternative to traditional, nonindividualized platelet threshold strategies. Cancer 2002;94:3252–62. © 2002 American Cancer Society.

DOI 10.1002/cncr.10603

The association between platelet count and bleeding was described first in 1962; however, no formal methods of applying this information to clinical decision-making have been developed.1 Although no threshold was identified that separated patients who definitely required prophylactic platelet transfusion from those who did not, a standard of care evolved: Prophylactic platelet transfusions were given below a threshold of 20,000 platelets per μL. Despite widespread use of this strategy, its applicability to patients with solid tumors or lymphoma has been questioned widely based on the observation that many patients tolerate platelet counts < 20,000 per μL, whereas others develop bleeding at higher platelet counts.2–7 A number of investigators have suggested that a threshold of 10,000 platelets per μL should be used.6–11 Several recent clinical trials have demonstrated the potential for a 10,000 platelet per μL threshold in patients with acute leukemia.12–14 Platelet growth factors further complicate this issue by providing new options for bleeding prophylaxis.15–17 In an effort to guide prescription of prophylactic platelet transfusions, we developed a clinical prediction rule using data from 608 patients with lymphoma or solid tumors who received 1262 cycles of chemotherapy.

MATERIALS AND METHODS

  1. Top of page
  2. Abstract
  3. MATERIALS AND METHODS
  4. RESULTS
  5. DISCUSSION
  6. REFERENCES

Patient Population

A random sample, which was stratified by underlying neoplasm, was selected from patients with lymphoma or solid tumors who developed chemotherapy-induced thrombocytopenia (platelet count < 50,000 per μL) at The University of Texas M. D. Anderson Cancer Center between January 1, 1994, and December 31, 1995. To avoid ascertainment bias resulting from care delivered by other providers, the sample was limited to patients who received all of their care at our center. Patients with acute or chronic leukemia, bone marrow transplantation recipients, and patients age < 16 years were excluded. To permit longitudinal examination of the risk of bleeding, for each patient, all chemotherapy cycles that resulted in chemotherapy-induced thrombocytopenia during the 2-year study period were evaluated. Cycles during the study period that did not result in chemotherapy-induced thrombocytopenia and cycles during which the patient was managed at another treatment center were excluded.

Data Collection

Pertinent history and clinical information were obtained by review of the medical record. Laboratory, blood bank, and pharmacy data were transferred electronically from institutional data bases to avoid transcription errors. Abstractors were blinded to the predictive variables that were studied. Zubrod performance status18 and Charlson comorbidity scores19 were computed for Day 1 of each chemotherapy cycle to account for the influence of severity of illness.

Variables and Definitions

The unit of analysis was a chemotherapy cycle, defined as the interval between the first day of chemotherapy and either 1) the start of the next cycle; 2) the day of platelet recovery (> 100,000 per μL) if no further therapy was received; or 3) 60 days or death, whichever came first, if no further therapy was received and if no platelet recovery occurred. Thrombocytopenia was defined as a platelet count < 50,000 per μL. Bone marrow metastases were identified by bone marrow biopsy: It was assumed that patients who did not undergo a bone marrow biopsy were free of bone marrow metastases.

The dependent (outcome) variable for all analyses was an episode of bleeding. For descriptive analyses, episodes of bleeding were categorized as either minor (bleeding from the skin or mucosa and bleeding from other sites not requiring red blood cell transfusion; World Health Organization [WHO] Grades 1 and 2) or major (frank hemorrhage or major organ bleeding requiring red blood cell transfusion; WHO Grades 3 and 4). Examples of minor bleeding include petechiae; microscopic hematuria; oozing from wounds, nose, gums, or vagina; and blood-streaked sputum or vomitus. Examples of major bleeding include frank bleeding from wounds, epistaxis or vaginal bleeding, macroscopic hematuria, and hemorrhage from a major organ. Because of the difficulties of distinguishing between major bleeding episodes and minor bleeding episodes from retrospective review of medical records, major episodes of epistaxis, vaginal, and oral bleeding were distinguished from minor bleeding episodes at the same sites on the basis of the requirement for red blood cell transfusion. However, it should be noted that patients with minor bleeding from other sites occasionally received red blood cell transfusions.

For all regression analyses, the dependent variable, bleeding, was dichotomized. We chose to combine major and minor bleeding because of our concern that the difference between some major and minor episodes of gastrointestinal or genitourinary bleeding among patients with solid tumors in the outpatient setting may have been the speed with which bleeding was identified and therapy begun. If that is true, then some minor bleeding episodes may be considered potential major bleeding episodes, and the most conservative approach is to consider their prevention equally important.

We hypothesized that a variety of clinical factors could influence the risk of bleeding.20 Therefore, the following independent (predictor) variables were tested for their association with bleeding. Host-related variables were tested, including age, gender, Zubrod performance status, Charlson comorbidity score, history of hypertension or peptic ulcer disease, any prior history of bleeding, or trauma. Laboratory parameters also were tested, including platelet count, total bilirubin, and serum creatinine on Day 1 of the chemotherapy cycle.

The disease-related variables examined included diagnosis, extent of disease, presence of bone marrow metastases, and presence of central nervous system (CNS) metastases or pathology. Treatment-related variables included chemotherapy agent received, number of prior cycles of chemotherapy, administration of an agent that potentially may affect platelet function, prior radiation therapy, and the presence of a catheter. Chemotherapy agents were examined individually and in combination for their association with the risk of bleeding. After this analysis, agents that were considered highly toxic to the bone marrow, including cisplatin, carboplatin, lomustine, carmustine, dacarbazine, and mitomycin C, were examined for their association with the risk of bleeding, although some of these agents were not used frequently. Agents that were considered likely to affect bleeding or platelet function, including anticoagulants, steroids, nonsteroidal anti-inflammatory agents, penicillins or cephalosporins, psychotropic drugs (tricyclic antidepressants and phenothiazines), antihistamines, cardiovascular drugs (i.e., nitroglycerin, propranolol, nifedipine, quinidine), and histamine Type 2 antagonists, also were examined individually and in combination for their association with the risk of bleeding. After that analysis, only heparin, penicillins or cephalosporins, psychotropic drugs, and antihistamines were considered associated with an increased risk of bleeding.

Confounding factors that occurred during therapy and may have decreased the risk of bleeding also were tested. These included the administration of prophylactic platelet transfusions and therapy with vitamin K.

Statistical Analysis

The sample was divided randomly into a 750-cycle data set (among 375 patients) from which the clinical prediction rule was derived (derivation set) and a 512-cycle data set (among 233 patients) that was used to validate the model (validation set). No patient contributed cycles to both the derivation set and the validation set. The sample size of 750 cycles was chosen to allow for 10 bleeding episodes per variable (on the basis of predictions that the bleeding rate would be 10% and that the multiple-variable model would have 7 or 8 variables).21

The prediction rule was developed in five steps using data from the derivation set. 1) First, we conducted a descriptive analysis of the association between bleeding and each of the independent or confounding factors listed above. Individual factors associated with bleeding episodes were then identified using logistic regression, adjusting for within-patient correlation. 2) Factors identified in the initial analysis were included in a logistic regression analysis to develop a multiple-variable model. A multilevel model was fit with cycles nested within patients. Variables were added to the model (in ascending order by P value, with the most significant first) during sequential regression analyses until a comparison of log-likelihood ratios failed to show significant improvement in the model. This analysis yielded six statistically significant factors. 3) A weight was assigned to each of the six factors by dividing its regression coefficient by 0.7662 (the smallest coefficient for a significant variable) and rounding to the nearest integer. This maneuver resulted in integer weights (rather than decimal regression weights) for all factors and a weight of 1 for the least significant factor. 4) For each chemotherapy cycle in the derivation set, a risk score was computed by summing the weights. 5) Finally, a three-level risk index was developed in the derivation set by comparing the rate of bleeding for each summed risk score to the following clinically significant risk thresholds. The threshold for low risk was 0% risk of major bleeding and < 10% combined risk of major plus minor bleeding (above 10%, the risk of major and minor bleeding exceeds most estimates of the combined risk of major plus minor reactions to platelet transfusions). Moderate risk was > 0% risk of major bleeding and < 10% combined risk of major plus minor bleeding. High risk was > 10% combined risk of major plus minor bleeding.

The risk of acute reactions to platelet transfusions was chosen as the benchmark for the rule, because platelet transfusion prophylaxis involves a trade-off between the risk of bleeding and the risk of reaction. These risks are similar to bleeding in several ways. Like bleeding, a few reactions are very severe, whereas most are easily controllable. Also like bleeding, transfusion reactions that are not controlled promptly may escalate to major episodes with serious clinical outcomes.

The value of 10% was obtained by observing all platelet transfusion recipients for potential transfusion reactions, prospectively, over a 1-month period in our outpatient transfusion clinic. All reactions, whether major or very minor, were included. For example, transient chills or very minor rashes were included. Like many instances of petechiae, these reactions rarely require medical treatment.

Descriptive analyses were computed using BMDP Dynamic software (version 7; BMDP Statistical Software, Inc., Los Angeles, CA). All regression analyses were computed with SAS software (version 6.12; SAS Institute, Cary, NC) using the GENMOD General Estimating Equations procedure with an unstructured covariance matrix. This procedure adjusts for the within-patient correlation caused by multiple cycles of chemotherapy per patient. The model's performance in the derivation and validation sets also was compared using receiver operating characteristic (ROC) curves. The ROC curve illustrates the trade-off between sensitivity and specificity across all threshold values for the positivity of the prediction rule. The area under the ROC curve is a measure of the discriminative ability of the model. The area under the ROC curve (AUC) value can be interpreted in this study as the probability that a patient with bleeding will have a higher risk score from the model than a person without bleeding. Similarity between AUC values for the derivation and validation sets suggests reliability (reproducibility) of the model.

A strategy for prophylactic use of platelet transfusions and platelet growth factors was developed by comparing the rates of bleeding at various platelet nadirs for patients with low, moderate, and high risk of bleeding with the risk of adverse reactions to platelet transfusions (10%). The potential value of the risk index-based strategy was then compared with the potential value of two alternative, threshold-based prophylaxis strategies (the traditional threshold of 20,000 platelets per μL and the recently proposed threshold of 10,000 platelets per μL). For this analysis, the potential impact on quality of care was emphasized over the potential impact on cost of care. The measure of impact on quality of care was whether use of the strategy would have resulted in prophylactic transfusion before a major bleeding episode occurred (the sensitivity). For example, use of a strategy that called for transfusion when the platelet count fell below 20,000 per μL would have resulted in poor quality care in patients who developed bleeding before platelet counts fell below 20,000 per μL but would have resulted in good quality care in patients who developed bleeding after the platelet count fell below 20,000 per μL. The measure of impact on cost was whether use of the strategy would have identified patients properly who did not experience bleeding (the specificity), thus minimizing the frequency of unnecessary prophylaxis.

RESULTS

  1. Top of page
  2. Abstract
  3. MATERIALS AND METHODS
  4. RESULTS
  5. DISCUSSION
  6. REFERENCES

The study included 608 patients who received 1262 chemotherapy cycles that resulted in chemotherapy-induced thrombocytopenia. Their average age was 48 years. Lymphomas, sarcomas, and neoplasms of the breast or genitourinary tract were most common (Table 1). Most patients had received previous chemotherapy; 24% had received > 10 chemotherapy cycles. Half of the patients had widely metastatic disease, whereas 4% of patients were receiving adjuvant chemotherapy in the absence of clinical evidence of disease.

Table 1. Characteristics of the Patients and the Chemotherapy Cycles
CharacteristicPatients (n = 608)Cycles (n = 1262)
No.%No.%
  1. NED, no evidence of disease.

Type of neoplasm    
 Lymphoma1722834227
 Breast821417814
 Sarcoma751218815
 Genitourinary591017414
 Melanoma579917
 Lung498847
 Gastrointestinal153242
 Gynecologic305534
 Unknown primary264544
 Other437746
Male gender2924862950
No. of prior chemotherapy cycles    
 None9115917
 1–52634355344
 6–101121828623
 > 101422433226
Extent of disease    
 NED234907
 Local disease1532528823
 Single metastatic site1232025620
 Disseminated metastases3095162850

Episodes of Bleeding

Bleeding developed during 111 cycles (9%). No patient had more than one episode during a single cycle. Bleeding was minor in 68 episodes, including 19 episodes in which bleeding was confined to the skin. The most common sites of major bleeding were the nose (13 cycles), the gastrointestinal tract (18 cycles), and the genitourinary tract (9 cycles). A mean of 2 units of red blood cells was required during 21 minor bleeding episodes (31%) compared with a mean of 4 units of red blood cells during 21 major bleeding episodes (49%). Hospitalization was required in 33 cycles with bleeding (30%) for an average of 4 days (range, 1–days). Four patient deaths were attributed in part to major bleeding episodes; one patient hemorrhaged at multiple sites, and one patient each had gastrointestinal, bladder, and adrenal hemorrhage. CNS hemorrhage occurred during only 1 cycle, although CNS tumors or metastases were present during 129 cycles.

Risk Factors for Bleeding

The risk of bleeding was related to both the depth and the duration of thrombocytopenia (Table 2). Bleeding was significantly more common during cycles in which thrombocytopenia persisted for more than 14 days than during cycles in which the platelet count recovered in fewer than 7 days (21% vs. 6%, respectively; P < 0.0001). Similarly, bleeding episodes were significantly more common during cycles in which the platelet nadir fell below 20,000 per μL than during cycles in which the nadir exceeded 20,000 per μL (13% vs. 5%, respectively; P < 0.0001). This pattern was independent of the duration of thrombocytopenia. Bleeding was more common during cycles complicated by fever (11%) than during cycles without fever (7%), although this difference was not statistically significant (P = 0.14). Bleeding was significantly more common during cycles in which invasive procedures were performed during thrombocytopenia (19% vs. 8%; P = 0.003).

Table 2. Risk of Bleeding According to Extent and Duration of Thrombocytopenia
Platelet nadir (per μL) and bleeding episodesThrombocytopenia for < 7 days (%)Thrombocytopenia for 7–14 days (%)Thrombocytopenia for > 14 days (%)
> 20,000   
 No. of cycles5698645
 Minor bleeding-no. (%)8 (1)3 (3)6 (13)
 Major bleeding-no. (%)11 (2)3 (3)2 (4)
 Any bleeding-no. (%)19 (3)6 (7)8 (17)
> 10,000–20,000   
 No. of cycles17611772
 Minor bleeding12 (7)5 (4)7 (10)
 Major bleeding4 (2)3 (3)6 (8)
 Any bleeding16 (9)8 (7)13 (18)
5000–10,000   
 No. of cycles354857
 Minor bleeding1 (3)6 (13)13 (23)
 Major bleeding1 (3)4 (8)4 (7)
 Any bleeding2 (6)10 (21)17 (30)
< 5000   
 No. of cycles152022
 Minor bleeding1 (7)3 (15)3 (14)
 Major bleeding2 (13)2 (10)1 (5)
 Any bleeding3 (20)5 (25)4 (18)

Unfortunately, the ultimate depth and duration of thrombocytopenia are not known when platelet transfusions and platelet growth factors are prescribed, and invasive procedures during thrombocytopenia usually are unplanned. Thus, these factors would not be useful in bleeding prophylaxis decision-making. Therefore, related factors that commonly are available at the time of decision-making were examined in the derivation set for their predictive value.

Previous insults to the bone marrow, such as radiation therapy or numerous chemotherapy cycles, were associated with higher risks of bleeding (P = 0.18 and P = 0.18, respectively) (Table 3). Likewise, factors suggesting poor bone marrow function at baseline, such as the presence of bone marrow metastases or a platelet count < 75,000 per μL, were predictive of bleeding (P < 0.0001 for both).

Table 3. Correlates of Risk of Bleeding in the 750-Cycle Derivation Set
CharacteristicNo. of cyclesAny bleeding No. (%)Any bleeding Adjusted % (95% CI)aP valueb
  • 95% CI, 95% confidence interval.

  • a

    Rates and confidence intervals were adjusted to account for within-patient correlation among patients with multiple cycles.

  • b

    P value compares adjusted rates.

No. of prior chemotherapy cycles    
 0544 (7)9 (2–28)
 1–525515 (6)6 (2–17)
 6–1018417 (9)11 (4–27)0.18
 > 1025729 (11)10 (7–15)
Less toxic chemotherapy regimen40632 (8)9 (6–12)
Toxic chemotherapy regimen34433 (15)9 (4–21)0.75
No bone marrow metastases57733 (6)6 (4–8)
Bone marrow metastases17332 (19)18 (8–36)< 0.0001
Baseline platelet count (per μL)    
 > 125,00052034 (7)7 (5–9)
 100,000–125,000903 (3)5 (1–17)
 75,000–99,000534 (8)8 (2–26)
 50,000–74,000368 (22)25 (9–53)< 0.0001
 < 50,0005116 (31)32 (13–59)
No prior radiation therapy53941 (8)8 (6–11)
Prior radiation therapy21124 (11)12 (5–26)0.18
Extent of disease    
 No evidence of disease612 (3)4 (1–16)
 Local disease1496 (4)4 (1–13)
 Single metastatic site1417 (5)6 (2–16)
 Disseminated metastases39950 (13)13 (10–17)0.001
Type of neoplasm:    
 Lymphoma20922 (11)10 (1–88)
 Breast1309 (7)7 (< 1–84)
 Sarcoma933 (3)3 (< 1–70)
 Genitourinary9013 (14)17 (< 1–93)
 Melanoma584 (7)8 (< 1–87)
 Gynecologic406 (15)20 (< 1–95)
 Unknown primary272 (7)7 (1–38)
 Lung291 (3)4 (< 1–85)
 Other745 (7)8 (< 1–86)0.12
No prophylactic platelet transfusion45331 (7)8 (5–10)
Prophylactic platelet transfusion29734 (11)11 (5–22)0.09
Baseline Zubrod score ≤ 267348 (7)7 (5–10)
Baseline Zubrod score ≥ 37717 (22)23 (10–45)0.0001
No prior bleeding episode71652 (7)8 (6–10)
Any prior bleeding episode3413 (38)27 (1–54)0.0006
Drug affecting platelet function    
 No70850 (7)7 (5–10)
 Yes4215 (36)34 (14–62)< 0.0001
No necrotic tumor site53146 (9)9 (7–12)
Necrotic tumor site21919 (9)9 (4–20)0.99

Several host factors were predictive of bleeding (Table 3). A previous history of bleeding was a highly significant predictor of bleeding (P = 0.0006), and the presence of disseminated metastases and a baseline Zubrod performance status score > 2 also were highly significant predictors of bleeding (IP = 0.001 and IP = 0.0001, respectively). The administration of agents highly toxic to the bone marrow was associated with a higher risk of bleeding, particularly during cycles in which two or more of these agents were received (IP = 0.001). Concurrent therapy with agents that affect platelet function and clotting, such as penicillins or cephalosporins (IP = 0.0004), heparin (IP = 0.005), tricyclic antidepressants or phenothiazines (IP < 0.0001), and antihistamines (IP = 0.0001), when considered together, resulted in a significantly higher rate of bleeding (34% vs. 7%; IP < 0.0001). Steroids and nonsteroidal anti-inflammatory agents were not associated with significantly higher rates of bleeding: No patient received aspirin.

Multiple-Variable Model

Logistic regression analysis of the 750-cycle derivation set identified 7 factors that were predictive of bleeding: 1) any prior bleeding episode (odds ratio [OR], 5.6; 95% confidence interval [95% CI], 2.2–14.0), 2) treatment with a drug affecting platelet function (OR, 5.1; 95% CI, 2.0–12.6), 3) bone marrow metastases (OR, 4.3; 95% CI, 1.7–10.8), 4) a baseline platelet count < 75,000 per μL (OR, 3.5; 95% CI, 1.4–8.9), 5) genitourinary or gynecologic malignancy (OR, 3.3; 95% CI, 1.3–8.2), 6) Zubrod performance status score > 2 (OR, 3.4; 95% CI, 1.4–8.5), and 7) treatment with agents highly toxic to the bone marrow (OR, 2.2; 95% CI, 1.0–5.4). The model based on these factors provided a good fit for the derivation set data (Pearson chi-square test, 739.7120; P = 0.9956).

The ROC curve analysis was used to illustrate the accuracy of the model and to compare the model's performance in the derivation and validation sets (Fig. 1). The AUC was 0.83 for the derivation set and 0.73 for the validation set, suggesting that, in 73–83% of cases, a patient with bleeding would have a higher risk score from the model than a patient without bleeding. These values also demonstrate only minor deterioration in the model's performance during prospective testing, suggesting the reliability of the model.

thumbnail image

Figure 1. Comparison of receiver operating characteristic (ROC) curves for derivation and validation sets. The heavy dashed line describes a model with no discriminating ability; the higher the curve above the dashed line, the more accurate the rule. AUC: area under the ROC curve. The AUC ranges from 0.5 to 1.0: An AUC of 0.5 represents a useless test (i.e., a test as good as a coin-flip), and an AUC of 1.0 represents a perfect test. The AUC value can be interpreted as the probability that a patient with bleeding will have a higher risk score than a person without bleeding. Similarity between curves and AUC values suggests reliability of the model.

Download figure to PowerPoint

Clinical Prediction Rule Performance

A prediction rule, the Bleeding Risk Index (BRI), was developed from the multiple-variable model as described above (see Materials and n Methods) and was applied prospectively to the validation set (Table 4). The BRI successfully divided patients into three clinically distinct groups with increasing risks of bleeding—from 1–3% in the low-risk group to 23–31% in the high-risk group (Table 5). Differences between the derivation set and the validation set were minimal.

Table 4. Scoring the Bleeding Risk Index
FactorPointsa
  • a

    Points are summed on Day 1 of each chemotherapy cycle. Patients with 0 points are in the low-risk group, patients with 1–3 points are in the moderate-risk group, and patients with ≥ 4 points are in the high-risk group.

Any prior bleeding episode2
Receiving penicillin, cephalosporin, antihistamine, heparin, tricyclic, antidepressant, or phenothiazines2
Bone marrow metastasis2
Baseline platelet count < 75,000/mL2
Genitourinary or gynecologic neoplasm2
Baseline Zubrod score ≥ 32
Chemotherapy regimen, including any of the following: cisplatin, carboplatin, carmustine, lomustine, dacarbazine, mitomycin C1
Table 5. Performance of the Bleeding Risk Index
Predicted riskNo. of cycles (%)Any bleeding % (95% CI)Major bleeding % (95% CI)
Derivation setValidation setDerivation setValidation setDerivation setValidation set
  1. 95% CI, 95% confidence interval.

Low198 (26)132 (26)1 (0–2)3 (< 1–6)0 (0–2)0 (0–3)
Moderate436 (58)305 (60)6 (4–8)8 (5–11)3 (1–4)2 (< 1–3)
High116 (16)75 (14)31 (22–40)23 (13–32)15 (8–21)13 (5–21)

The ability of the BRI to discriminate between clinically important risk groups was illustrated particularly well when the observed rate of bleeding was correlated with the observed platelet count nadirs for each of the three risk groups and the clinically significant risk threshold of 10%, the point above which the risk of bleeding exceeds the risk of platelet transfusion reaction (Fig. 2). High-risk patients with platelet counts that fall below 20,000 per μL are at such high risk of bleeding (27–60%, depending on the nadir) that every available intervention should be used to avoid platelet counts this low (Fig. 2). These patients (14% of patients with thrombocytopenic solid tumors) should receive platelet transfusions if platelet counts fall below 20,000 per μL. It is possible that platelet growth factors may benefit these patients; however, further prospective clinical trials are needed to demonstrate the effectiveness of platelet growth factors in bleeding prophylaxis.

thumbnail image

Figure 2. Correlation between the risk of bleeding, platelet count nadir, and the predicted risk group. The dashed lines represent the clinically significant threshold and the 10% risk of major and minor reactions after platelet transfusions.

Download figure to PowerPoint

In contrast, low-risk patients (26% of patients with thrombocytopenic solid tumors) were at very low risk of bleeding unless the platelet count fell to 5000 per μL. Platelet growth factors are very unlikely to benefit these patients, and prophylactic platelet transfusions probably are not necessary unless the platelet count falls below 5000 per μL. Moderate-risk patients' risk of bleeding was similar to that of low-risk patients, except that prophylactic platelet transfusion appeared to be indicated at a platelet count of 10,000 per μL rather than 5000 per μL.

Clinical Application of the BRI

These findings suggested that an individualized approach to bleeding prophylaxis may provide higher quality and lower cost care compared with the application of a single threshold value to all patients with solid tumors. To test this hypothesis, we compared the potential clinical value of a BRI-based strategy (initiation of platelet transfusion based on individual risk) with that of two different threshold-based strategies (initiation of prophylaxis when the platelet count fell below 20,000 per μL or 10,000 per μL) in the validation set. The measure of quality of care, as stated above, was whether the strategy would have triggered prophylaxis before a major bleeding episode actually occurred (the sensitivity for major bleeding). Based on this metric, the strategy based on a threshold of 10,000 platelets per μL was inferior to both the BRI-based strategy and the strategy based on a threshold of 20,000 platelets per μL (sensitivity, 53% vs. 80% or 53% vs. 87%, respectively) (Table 6). Thus, the strategy based on a threshold of 10,000 platelets per μL was eliminated from further consideration. Use of the strategy based on a threshold of 20,000 platelets per μL would have resulted in misclassification of two patients, one with major vaginal hemorrhage and one with bladder hemorrhage, each of whom was treated successfully in the outpatient setting with 2 units of red blood cells. The use of the BRI-based strategy would have resulted in misclassification of those two patients and another patient who developed major epistaxis, which was treated in the outpatient setting with 2 units of red blood cells. In contrast, a strategy based on a threshold of 10,000 platelets per μL would have resulted in the misclassification of all three of those patients as well as one patient with major bilateral epistaxis, who required 8 days of hospitalization and transfusion of 16 units of red blood cells; two patients with gastrointestinal hemorrhage, one who was hospitalized for 15 days and the other who was hospitalized for 4 days; and a patient with pulmonary hemorrhage who was hospitalized for 8 days. All patients who would have been misclassified recovered.

Table 6. Comparison of Three Strategies to Prevent Bleeding. Validation Set
Criterion for evaluationBleeding risk index based strategy (%)aPlatelet threshold of 20,000/μL (%)Platelet threshold of 10,000/μL (%)
  • a

    Transfuse high-risk patients at 20,000 platelets per μL, moderate-risk patients at 10,000 platelets per μL, and low-risk patients at 5000 platelets per μL.

Sensitivity for major bleeding808753
Specificity for any bleeding845486

The monetary cost of prophylaxis was not studied; however, the specificity of each strategy was used to estimate the potential impact on cost. High specificity implies that prophylaxis resources (whatever their cost) will not be expended on patients who do not develop bleeding. The BRI-based strategy was more specific (84%) than a strategy based on a threshold of 20,000 platelets per μL (54%). Misclassifications by the BRI would have resulted in prophylactic platelet transfusions during 73 cycles in which there was no bleeding. The use of a threshold of 20,000 platelets per μL would have resulted in prophylaxis with platelet transfusions during 216 cycles in which there was no bleeding.

DISCUSSION

  1. Top of page
  2. Abstract
  3. MATERIALS AND METHODS
  4. RESULTS
  5. DISCUSSION
  6. REFERENCES

In this analysis of more than 600 patients with lymphomas or solid tumors, we identified factors that predict bleeding episodes during chemotherapy-induced thrombocytopenia. Most prominent among these were correlates of the depth and duration of thrombocytopenia, which have been known as predictive of bleeding in patients with leukemia since 1962.1 The predictive value of these factors was confirmed subsequently in patients with lymphomas and solid tumors.2, 3 To support decisions to prescribe platelet transfusions, the BRI uses correlates of prolonged and profound thrombocytopenia that are available on Day 1 of a chemotherapy cycle, including platelet count, bone marrow metastases, and therapy with agents highly toxic to the bone marrow. However, unlike their leukemic counterparts, who often bleed spontaneously, patients with lymphomas or solid tumors often bleed from tumor sites.2–5 The impact of genitourinary or gynecologic primary tumors on the risk of bleeding in the BRI underscores this well-known clinical fact. The absence of an association between gastrointestinal malignancies and bleeding episodes would seem to contradict this finding. However, it should be noted that there were only 24 cycles among 15 patients in our entire data set. It is unlikely that this small sample of patients would provide a precise estimate of bleeding risk. It is also interesting that any prior bleeding episode, regardless of cause, predicted bleeding during thrombocytopenia. However, it is important to note that there were no patients in this sample who had experienced postsurgical bleeding. It is possible that such episodes would not lead to subsequent bleeding and should not be considered in computing the risk score. A validation study of the BRI is planned for the future that should provide an opportunity to examine this issue.

It has been suggested that patients who experience profound thrombocytopenia during chemotherapy should receive platelet growth factors prophylactically during subsequent cycles to reduce their substantial risk of transfusions during subsequent cycles.15, 16 Although our analysis focused on the risk of bleeding, it is possible that patients at particularly high risk of bleeding may benefit from platelet growth factor prophylaxis. However, the effectiveness of platelet growth factors in this setting has not been demonstrated and should not be assumed. Given the paucity of bleeding prophylaxis efficacy data and the frequency of adverse events (approximately 30%) in recipients of the commercially available agent, the use of platelet growth factor prophylaxis was not included in the BRI-based strategy for high-risk patients.

Clinical Value of the BRI

A number of criteria have been suggested for judging the value of clinical prediction rules. Among these are 1) the predicted outcome should be clinically relevant, 2) the clinical determination of the predictors should be feasible, 3) the rule should be accurate, and 4) the rule should have the potential to affect clinical practice.22 The BRI fulfills each of these criteria, as described below.

Clinical relevance of the outcome

The predicted outcome, a bleeding episode, is potentially fatal. Previous authors have developed a risk model for thrombocytopenia using the requirement for platelet transfusion as the predicted outcome.23 Although our models share some factors (such as poor performance status and low baseline platelet count), the BRI predicts bleeding, the relevant outcome for prophylaxis decision-making.

An episode of serious or prolonged bleeding also is costly, resulting in numerous therapeutic platelet transfusions and, frequently, hospitalization. Furthermore, bleeding during thrombocytopenia often is justification for delay or dose reduction of subsequent cycles of chemotherapy, which may decrease the antineoplastic effect of chemotherapy in some populations.24

Feasibility of clinical determination of the predictors

The predictive factors used in the BRI are available routinely on Day 1 of every cycle of chemotherapy among patients with solid tumors or lymphomas. Factors that are not available, such as the depth and duration of thrombocytopenia or the presence of fever during thrombocytopenia, were not tested.

Accuracy of the rule

The ability of a clinical prediction rule to classify patients accurately is among its most important features. The sensitivity and specificity of the BRI were tested prospectively in a large sample (512 cycles) using the split-sample technique. In this analysis, the BRI provided the best trade-off between sensitivity for major bleeding and specificity compared with threshold-based strategies. Misclassification was uncommon, and, more importantly, all instances of misclassification except three would have resulted in prophylaxis of patients who did not experience bleeding rather than lack of prophylaxis for patients who did experience major bleeding. Thus, the rule would be considered clinically accurate, because patients who experienced major bleeding were identified despite somewhat reduced potential cost savings (because of unnecessary prophylaxis in some patients).

Potential impact on clinical practice

Although the BRI's risk factors are familiar to experienced oncologists, its approach to bleeding prophylaxis is unique and innovative. Traditional, threshold-based strategies result from a population perspective. Current debate focuses on which threshold should be used for all patients. In contrast, we have a taken an individual patient perspective and developed a tool for individualized decision-making, as clinical medicine is practiced. If our experience is validated in other settings, then this innovative perspective has the potential to have a major impact on clinical practice.

Limitations and Future Directions

Among the limitations of this study was its development using a retrospective review of the medical records of patients at a single comprehensive cancer center. It is possible that petechiae and other minor bleeding episodes were under-reported in the medical records and that the risk of bleeding reported herein, correspondingly, was under-estimated. Further validation of the BRI, ideally in a community setting, will be crucial to demonstrate its generality and utility. A prospective clinical trial is planned to compare the outcomes and cost of BRI-based care with standard practice.

A second limitation of the study was the use of prophylactic platelet transfusions in some patients but not others. This factor was accounted for in the logistic regression analysis, and its presence most likely would bias results toward no difference. However, this limitation underscores the importance of a controlled, prospective trial to test the safety and utility of a BRI-based strategy.

Although they were not a specific limitation of this study, in general, clinical prediction rules are limited in their ability to incorporate very rare factors and in their applicability to other, related clinical problems. This fact of life has several important implications for patients with thrombocytopenia. First, rare disorders that predispose patients to bleeding (i.e., von Willebrand disease and hemophilia) were not present among our patients and were not incorporated in the model, but they likely would increase dramatically the risk of bleeding regardless of the BRI score for affected patients. These conditions should be considered carefully in transfusion decisions. More importantly, although BRI scores may accurately predict bleeding among patients with autoimmune thrombocytopenia, the BRI-based transfusion strategy probably should be modified for use in this population. Our strategy was designed to be conservative clinically with the goal of avoiding both major and minor bleeding episodes. A separate strategy probably is indicated for patients with autoimmune thrombocytopenia, in whom the avoidance of platelet transfusions is an equally important goal. Both of these examples illustrate that a clinical prediction rule cannot replace clinical judgment.

Finally, it is important to note that there are clinically important differences between bleeding episodes classified as WHO Grade 3 and Grade 4. Differences between episodes classified as WHO Grade 1 and Grade 2 are equally important. However, because of the retrospective nature of this study, it was not possible to discriminate between WHO Grades 2–4 for many episodes of bleeding on any basis other than the use of red blood cell transfusion. Prospective evaluation of the BRI will permit thorough documentation of bleeding episodes and, perhaps, an opportunity to improve its ability to discriminate among these events.

REFERENCES

  1. Top of page
  2. Abstract
  3. MATERIALS AND METHODS
  4. RESULTS
  5. DISCUSSION
  6. REFERENCES