# A Cost Analysis of Hematopoietic Colony-Stimulating Factors

# A Cost Analysis of Hematopoietic Colony-Stimulating Factors

### Introduction

Myelosuppression remains the major dose-limiting toxicity of systemic

cancer chemotherapy. The risk of infection and infection-related

mortality increases in direct proportion to the degree and duration

of neutropenia observed [1]. The onset of fever in the setting

of neutropenia generally requires immediate hospitalization and

administration of broad-spectrum antibiotic therapy [2]. Previous

efforts to enhance hematologic recovery following systemic chemotherapy

have had only limited success [3].

Several hematopoietic colony-stimulating factors (CSFs) have been

characterized over the past decade [4], and two are now readily

available: human recombinant granulocyte-colony stimulating factor

(G-CSF, filgrastim, Neupogen) and yeast-derived granulocyte-macrophage

colony-stimulating factor [GM-CSF, sargramostim, Leukine] [5,6].

Several studies have shown that these agents can reduce the severity

and duration of neutropenia associated with cancer chemotherapy

[7-11]. A prospective, double-blind, placebo-controlled trial

of G-CSF utilized prophylactically has demonstrated a significant

reduction in the risk of febrile neutropenia and need for hospitalization

in patients with solid malignancies receiving combination chemotherapy

[12]. Randomized clinical trials have demonstrated that CSFs administered

after the onset of febrile neutropenia accelerate myeloid recovery,

reducing the duration of neutropenia [13]. It remains uncertain,

however, to what degree CSFs so administered reduce duration of

fever or hospitalization.

The approval of the hematopoietic CSFs for use in patients receiving

cancer chemotherapy has resulted in widescale use of these agents

in a variety of clinical settings, contributing to increasing

health care costs. As further studies define additional indications

for CSFs, even greater utilization of these agents can be anticipated.

Nevertheless, clinical and economic uncertainty exists as to the

optimal use of CSFs in patients receiving different cancer chemotherapy

regimens in each tumor type, considering the wide variation in

patients' risk for neutropenia. Recent cost studies have been

conducted in an effort to define, measure, and compare the relevant

positive and negative economic consequences of the use of CSFs

to prevent infections in patients receiving chemotherapy [14-16].

The American Society of Clinical Oncology (ASCO) has recently

adopted general guidelines for the use of CSFs based on these

and other studies in the literature [17]. This paper reviews and

extends the cost model previously presented for the use of hematopoietic

CSFs in patients receiving cancer chemotherapy and presents specific

guidelines based on this model. The administration of hematopoietic

CSFs may be associated with a reduction in health care costs for

hospitalization due to febrile neutropenia if utilized within

the specific guidelines presented.

### Methods

A standard model based on decision theory was developed for this

cost analysis of the use of hematopoietic CSFs in patients receiving

systemic cancer chemotherapy (Table 1). The decision choices consisted

of no CSF; CSF

administered prophylactically after the completion of chemotherapy

with continuation through the period of neutropenia; and CSF administered

therapeutically only if febrile neutropenia occurs. The model

assumes that all patients experiencing febrile neutropenia will

be hospitalized and treated empirically with parenteral antibiotics.

The role for outpatient management of selected patients with febrile

neutropenia has not yet been fully defined.

Baseline probabilities for hospitalization risk and survival with

and without CSF, along with the durations of hospitalization and

CSF administration, were based on the prospective randomized trial

of G-CSF in patients with small-cell lung cancer receiving systemic

chemotherapy [12]. Baseline probabilities of hospitalization and

survival in patients with febrile neutropenia receiving therapeutic

CSF are assumed to be the same as for patients receiving no CSF

(Table 2). Since there was no reduction in the median duration

of hospitalization in patients treated with G-CSF in either the

prophylactic or therapeutic randomized trials, no reduction in

hospitalization duration was assumed at baseline with either choice

[12,13]. As with all variables studied in this model, this baseline

assumption was then varied over the range of possible values in

a sensitivity analysis.

Costs considered in this model include the cost of hospitalization

for febrile neutropenia and the cost of CSF per treatment cycle.

Baseline daily costs of hospitalization were estimated from data

available at our own institution, based on fixed daily costs for

room, antibiotics, fluids and tubing, and professional fees. These

represent minimal cost estimates, since they do not reflect the

cost of diagnostic and monitoring tests (x-rays, scans, blood

tests), cultures, drug levels, and consultations, which can increase

the hospitalization costs substantially. Baseline CSF costs include

agent and administration costs derived from our own institution.

The expected excess cost for each specific choice was calculated

from the sum of the products of the costs and probabilities of

each outcome. The expected cost per treatment cycle represents

the excess cost associated with hospitalization for febrile neutropenia

and treatment with CSFs. Chemotherapy costs and other items likely

to be identical in the three treatment choices were not considered.

Sensitivity analysis provides an estimate of the expected cost

for a range of values of one or more variables for each decision

choice. Sensitivity analysis permits the calculation of thresholds

when the expected cost for two treatment options are the same.

In multiway sensitivity analysis, each function or curve represents

a series of thresholds for a combination of two variables indicated

on the axes. A family of threshold curves may be generated as

the value of a third factor is varied.

Using Monte Carlo analysis, the model was analyzed repeatedly,

each time sampling from the assumed distributions of the main

variables. Probability distributions for the main variables were

derived from the randomized controlled trial and local institutional

data. Monte Carlo simulations consisted of 1,000 sequential samples.

In this study, the distributions of outcomes or thresholds then

serve as a measure of variability upon which to base a level of

confidence in the decision outcome or threshold estimate. The

distribution function of the differences in outcome between two

choices is distributed as sample mean differences, allowing for

statistical inference. Tests of significance were based on a *t*

statistic with n-1 degrees of freedom, with n representing the

number of samples in the simulation.

### Results

Utilizing the baseline probability and cost assumptions in Table

2, the model generated an expected excess cost per treatment cycle

of $5,500 for no CSF, $4,750 for prophylactic CSF, and $6,875

for therapeutic CSF. Sensitivity analyses for the three choices

were performed for each of the study variables. Figure 1 displays

one-way sensitivity analyses for the control probability of hospitalization

(Figure 1A) and for the cost of hospitalization per day (Figure

1B). The excess cost associated with prophylactic CSF increases

at a slower rate than with the other two strategies for both variables.

The thresholds for each variable are the values at the point where

the cost line for prophylactic CSF crosses each of the other lines.

At these points, the total excess cost is the same for each group

being compared.

Model thresholds for the decision between no CSF and prophylactic

CSF are presented in Table 3, and those for the decision between

prophylactic CSF and therapeutic CSF are shown in Table 4. The

strategy favored on a cost basis at values above the threshold

is indicated in the tables to the right of the thresholds, while

the strategy favored at values below the threshold is indicated

to the left of the thresholds.

Figure 2 displays three-way sensitivity analyses based on variations

in daily hospital cost and duration of hospitalization. The region

above each threshold curve represents values of the variables

favoring the use of the hematopoietic CSFs on a cost basis. Figure

2A varies the risk of hospitalization in the prophylactic group

as a proportion of the control risk. Conditions favoring the use

of CSF represented by the area above each curve increase as the

risk of hospitalization in patients receiving CSF decreases. Alternatively,

as the proportional risk of hospitalization with prophylactic

CSF increases, the conditions associated with a net cost advantage

for CSF lessen. The incremental change in the area above the threshold

curve is greater with higher proportional risk of hospitalization.

Figure 2B varies the cost of CSFs per day. The conditions favoring

the use of CSF on a cost basis increase as the daily cost of CSF

lessens. The area above the threshold curves favoring the use

of CSF in this setting increases at approximately equal increments

as the daily cost of CSF decreases.

Estimates of excess cost per treatment cycle based on Monte Carlo

analysis were $7,923 ± $484 (SEM), $6,612 ± $289, and

$9,812 ± $533 for the control, prophylactic, and therapeutic

CSF arms, respectively. The distribution of cost differences based

on Monte Carlo analysis favors prophylactic CSF over no CSF, with

a median difference of $1,070 (Figure 3A). The distribution of

cost differences also favors prophylactic CSF over therapeutic

CSF, with a median difference of $2,671 (Figure 3B). Neither of

these differences was statistically significant, however, based

on the assumed variability of each parameter.

Threshold analysis based on Monte Carlo simulation also demonstrated

considerable variability in threshold measures of the main variables

(Table 5). The distribution of threshold estimates for each of

the variables in the model was studied. Each of the distributions

was skewed, suggesting that the median probably represents a better

measure of central tendency than the mean.