Received: 26 October 2016 Revised: 9 April 2017 Accepted: 10 April 2017

DOI: 10.1111/jep.12767

ORIGINAL ARTICLE

Usefulness of a decision tree model for the analysis of adverse

drug reactions: Evaluation of a risk prediction model of
vancomycin‐associated nephrotoxicity constructed using a data
mining procedure
Shungo Imai BS1 | Takehiro Yamada PhD2 | Kumiko Kasashi PhD3 |

Masaki Kobayashi PhD2 | Ken Iseki PhD4,5

1
Pharmacist, Department of Pharmacy,
Hokkaido University Hospital, Sapporo, Japan Abstract
2
Associate Professor, Department of
Objectives: Several publications concerning decision tree (DT) analysis in medical fields have
Pharmacy, Hokkaido University Hospital,
Sapporo, Japan
recently demonstrated its usefulness for defining prognostic factors in various diseases. How-
3
Vice-Director, Department of Pharmacy, ever, there are minimal reports on the predictors of adverse drug reactions. We attempted to
Hokkaido University Hospital, Sapporo, Japan use DT analysis to discover combinations of multiple risk factors that would increase the risk of
4
Director, Department of Pharmacy, Hokkaido
nephrotoxicity associated with vancomycin (VCM). To demonstrate the usefulness of DT analysis,
University Hospital, Sapporo, Japan
5
Professor, Division of Pharmasciences, we compared its predictive performance with that of multiple logistic regression analysis.
Faculty of Pharmaceutical Sciences, Hokkaido
Method: A single‐centre, retrospective study was conducted at Hokkaido University Hospital.
University, Sapporo, Japan
A total of 592 patients, who received intravenous administrations of VCM between November
2011 and April 2016, were enrolled. Nephrotoxicity was defined as an increase in serum creatinine
Correspondence
Ken Iseki, Division of Pharmasciences, Faculty of ≥0.5 mg/dL or a ≥50% increase in serum creatinine from the baseline. Risk factors for VCM neph-
of Pharmaceutical Sciences, Hokkaido rotoxicity were extracted from previous reports. In the DT analysis, a chi‐squared automatic inter-
University, Kita 12‐jo Nishi 6‐chome, Kita‐ku,
action detection algorithm was constructed. For evaluating the established algorithms, a 10‐fold
Sapporo 060‐0812, Japan.
Email: ken‐i@pharm.hokudai.ac.jp cross validation method was adopted to calculate the misclassification risk of the model. Moreover,
to compare the accuracy of the DT analysis, multiple logistic regression analysis was conducted.

Results: Eighty‐seven (14.7%) patients developed nephrotoxicity. A VCM trough concentra-

tion of ≥15.0 mg/L, concomitant medication (vasopressor drugs and furosemide), and a duration
of therapy ≥14 days were extracted to build the DT model, in which the patients were divided
into 6 subgroups based on variable rates of nephrotoxicity, ranging from 4.6 to 69.6%. The pre-
dictive accuracies of the DT and logistic regression models were similar (87.3%, respectively),
indicating that they were accurate.

Conclusion: This study suggests the usefulness of DT models for the evaluation of adverse
drug reactions.

KEY W ORDS

data mining, decision tree analysis, nephrotoxicity, vancomycin

1 | I N T RO D U CT I O N processes, data mining algorithms can find the best fitting model that
searches through the data set to detect patterns. These patterns may
Data mining is defined as extracting information from large sets of include interactions between variables, as well as interactions within
data. In addition, it is a method of predictive analysis that explores data subsets of variables. Decision tree (DT) analysis, which is a flowchart‐
to discover underlying patterns and relationships in complex data sets like tree framework, is a typical technique of data mining used to
and enables one to construct predictive models.1 Through automated construct predictive models.2 The DT model is commonly used to

J Eval Clin Pract. 2017;1–7. wileyonlinelibrary.com/journal/jep © 2017 John Wiley & Sons, Ltd. 1
2 IMAI ET AL.

establish classification systems based on multiple covariates or to who received intravenous administrations of VCM between November
develop prediction algorithms for a target variable. It classifies a popu- 2011 and April 2016. The inclusion criteria were as follows: (1) aged
lation into segments, ie, “branches,” which form an inverted tree. The 18 years and older, (2) dosing period was at least 3 days, and (3) the
algorithm is nonparametric and can efficiently deal with large and com- VCM trough concentration was measured after the third day
plicated data sets without imposing a complicated parametric struc- of administration. We then excluded all patients who underwent
ture.3 In medical fields, recent research concerning DT analysis haemodialysis, continuous haemodialysis flow, or had nephrotoxicity
demonstrated its usefulness for defining prognostic factors in various before the VCM trough concentration was measured.
diseases and operations such as haematopoietic stem cell transplanta- Nephrotoxicity was defined as an increase in serum creatinine
tion and breast cancer, which is easily applicable to clinical practice.4,5 (Scr) of ≥0.5 mg/dL or a ≥50% increase in Scr from the baseline.18 To
However, there are minimal reports on DT analysis that detail the evaluate VCM‐associated nephrotoxicity, we used the maximum Scr
predictors of adverse drug reactions.1,6 Combinations of risk factors during the administration period.
can increase the risk of adverse drug reactions. Thereafter, for clini- Risk factors for VCM‐associated nephrotoxicity that have been
cians and pharmacists, the flowchart‐like algorithm (called reported in previous studies were categorized as follows: patient
“trees”) created by DT models can be interpreted and applied to their age ≥ 52 years,19 gender (male/female), duration of VCM treatment
clinical practice, which are favourable over conventional statistics ≥14 days,12 body mass index (BMI) ≥ 25, Scr ≥ 1.0 mg/dL,13 baseline
models such as multiple logistic regression analysis. 7
creatinine clearance (CCr) ≤ 86.6 mL/min,13 concomitant medica-
Vancomycin (VCM) is a first line of therapy for serious Staphylo- tions14-16 (NSAIDs, furosemide, amphotericin B, aminoglycoside anti-
coccus aureus infections. It is highly active against most of the gram‐ biotics, and vasopressor drugs), residence in intensive care unit13
positive bacterial strains and penicillin‐resistant strains (ICU), and average VCM trough concentration ≥ 15 mg/L.10,11,14
of staphylococci. Nephrotoxicity is a typical side effect of VCM. There 8
Although a body weight of ≥100 kg19 was a reported risk factor of
are many reports of risk factors for VCM‐associated nephrotoxicity, nephrotoxicity, it was very uncommon in Japanese patients. Therefore,
such as larger VCM exposures9-12 (ie, higher troughs [>15 mg/L]; larger instead of body weight, information on patients with obesity (BMI ≥ 25)
AUC (area under the blood concentration‐time curve), and prolonged was extracted. On the basis of the the Japanese criteria for obesity
durations); certain hosts13-15 (ie, history of acute kidney injury, disease,20 we set the cut‐off value of obesity to BMI ≥ 25. Data on
preexisting renal insufficiency, critically ill, or septic); and concurrent the vasopressor drugs adrenaline, etilefrine, noradrenaline, olprinone,
16
medications (ie, nonsteroidal anti‐inflammatory drugs (NSAIDs), milrinone, dopamine, and dobutamine were extracted.
furosemide, amphotericin B, and aminoglycoside antibiotics). It is asso- Creatinine clearancer was calculated using the Cockcroft‐Gault
ciated with prolonged hospitalizations, increased mortality, and the equation. When calculating the CCr, adjusted body weight (ABW)
need for renal replacement therapy.17 was used if a patient's BMI was >25, by using the following formula:
Previous studies have performed multiple logistic regression ABW = IBW + 0.4 (actual body weight—IBW), with IBW (men) = 50 + 2.3
analysis for the detection of factors related to VCM‐associated (inches above 60) and IBW (women) = 45 + 2.3 (inches above 60).12
9-17
nephrotoxicity. However, these reports failed to indicate combina-
tions of risk factors that may increase the risk for nephrotoxicity. Using
the DT model, clinicians and pharmacists can allocate patients to spe-
2.2 | Ethics
cific subgroups based on the rate of nephrotoxicity by simply following This study was conducted in accordance with the guidelines for the
the flowchart‐like tree framework. Once created, this algorithm may care for human study, and the study protocol was approved by
help to predict adverse drug reactions. In the Hokkaido University the ethics committee of the Hokkaido University Hospital.
Hospital, the number of cases of VCM intravenous administration
is approximately 200 cases per year, and this amount has been
suggested to be sufficient for DT analysis.1,4,5 Hence, we selected
2.3 | Statistical analysis
VCM‐associated nephrotoxicity for validation of the capability of the Comparisons of the demographic and clinical characteristics of
DT model for the evaluation of adverse drug reactions. patients' were unpaired, and all the tests of significance were two‐
In the present study, we used DT analysis to explore whether tailed. Continuous variables were compared using the Mann‐Whitney
combinations of risk factors would increase the incidence of U test. Pearson chi‐square or Fisher exact test were used to compare
nephrotoxicity in patients who received VCM. Furthermore, this study the categorical variables. P ≤ .05 was considered statistically significant.
was performed to demonstrate the usefulness of DT analysis for the All the potential risk factors whose P value was less than .1 in uni-
evaluation of adverse drug reactions by comparing its predictive variate analysis were applied to the multiple logistic regression analy-
performance with that of multiple logistic regression analysis. sis. A stepwise approach was used to enter new terms into the
logistic regression analysis, in which nephrotoxicity was the dependent
outcome variable and .05 was set as the limit for the acceptance or
2 | METHODS
removal of new terms. For evaluating the fitness of the logistic regres-
sion model, the Hosmer‐Lemeshow test was used. Statistical analyses
2.1 | Patients were performed using SPSS statistics Version 23 (IBM, Tokyo, Japan).
A single‐centre, retrospective study was conducted at Hokkaido The DT analysis was performed using SPSS Decision Trees Version
University Hospital. The subjects of this study were 592 patients 23. According to the chi‐squared automatic interaction detection
IMAI ET AL. 3

(CHAID) algorithm, classification tree models predicting nephrotoxicity min, odds ratio [OR] 0.64, 95% confidence interval [CI], 0.40‐1.01).
were constructed. CHAID are nonparametric procedures that make no Therefore, we excluded CCr from the multiple logistic regression
assumptions of the underlying data.21 The CHAID algorithm analysis.
determines how continuous and/or categorical independent factors In the multiple logistic regression analysis (Table 4). duration of
best combine to predict a binary outcome based on “if‐then” logic by therapy ≥14 days, concomitant medications (furosemide
portioning each independent factor into mutually exclusive subsets and vasopressor drugs), and VCM trough concentration ≥15.0 mg/L
based on the homogeneity of the data. It was run in duplicate with were extracted as the independent predictors of VCM‐associated
parent nodes defined at 20 subjects, child nodes defined at 10 sub- nephrotoxicity. The Hosmer‐Lemeshow statistic had a P value of .27.
jects, and significance22,23 set at (αmerge, αsplit, and P value) ≤.05.

3.3 | DT analysis
2.4 | Model development and evaluation
As shown in Figure 1, the DT analysis was constructed with 12 risk fac-
The present study adopted a 10‐fold cross validation method to tors of VCM‐associated nephrotoxicity that have been published in
evaluate the misclassification risk of the DT model. The 10‐fold cross previous reports. Four predictive variables were automatically
validation method involves randomly separating the acquired data sets extracted using this analysis to produce a total of 6 patient subgroups
into 10 data sets that are equal in sample size. The DT model is to build the DT. Then the order of splitting variables was determined in
constructed on the basis of a training data set. The rest of the 9 data order of the strength of relation to nephrotoxicity, based on the
sets were used as testing data for verifying model effectiveness. A CHAID algorithm.3
total of 10 repeated empirical tests were conducted, where each Vancomycin trough concentration ≥15.0 mg/L was selected as the
subset was used as the test data. Multiple logistic regression analysis first splitting variable. The possibility of nephrotoxicity was 35.0% for
was used to compare the accuracy of the DT analysis with the conven- patients with a VCM trough concentration of ≥15.0 mg/mL and 8.4%
tional statistical method.7,24 with a VCM trough concentration of <15.0 mg/mL. Among the

3 | RESULTS
TABLE 1 Patient characteristics
Characteristic Data (n = 592)
3.1 | Patient characteristics
Age (years), mean ± SD 60.8 ± 15.9
The clinical characteristics of the 592 patients who received VCM are Range 19‐96
shown in Table 1. Their mean CCr was 91.2 ± 42.9 mL/min and mean Gender (male/female) 382/210
VCM trough concentration was 12.2 ± 4.6 mg/L. Eighty‐seven Body weight (kg), mean ± SD 58.2 ± 12.0
(14.7%) patients developed nephrotoxicity. Range 31.6‐127.0
Comparison of the patient characteristics between those with Body mass index, mean ± SD 22.1 ± 4.2
nephrotoxicity and those without nephrotoxicity is shown in Table 2. Baseline serum creatinine (mg/dL), 0.74 ± 0.35
Patients who developed nephrotoxicity did not differ significantly mean ± SD
from those without nephrotoxicity in terms of age, gender, body Number of days to reach the maximum 11.9 ± 7.4
Scr (days), mean ± SD
weight, baseline Scr, baseline CCr, NSAIDs and aminoglycoside
Baseline creatinine clearance (mL/min), 91.2 ± 42.9
antibiotics use, and days to initial therapeutic drug monitoring (TDM).
mean ± SD
However, they showed significantly different BMI, duration of therapy,
Duration of therapy (days), mean ± SD 11.6 ± 8.9
furosemide use, amphotericin B and vasopressor drugs use, residence
Concomitant medications, n (%)
in ICU, and VCM trough concentrations.
NSAIDs 264 (44.6)
The concomitant medications aminoglycoside antibiotics and
Furosemide 181 (30.6)
amphotericin B were excluded from the analysis of risk factors,
Amphotericin B 5 (0.84)
because the number of cases was too low to perform logistic analysis
Aminoglycoside antibiotics 16 (2.7)
and DT analysis.
Vasopressor drugs 66 (11.1)
Residence in intensive care unit, n (%) 74 (12.5)
3.2 | Multiple logistic regression analysis Number of times of TDM during the 2 (1‐7)
administration period, median (range)
In the univariate analysis (Table 3). baseline CCr ≤86.6 mL/min, dura- Days to initial TDM (days), mean ± SD 3.6 ± 1.0
tion of therapy ≥14 days, concomitant medications (furosemide and Days to second TDM (days), mean ± SD 8.1 ± 3.2
vasopressor drugs), residence in ICU, and VCM trough concentration Days to third TDM (days), mean ± SD 13.3 ± 5.2
≥15.0 mg/L were significant factors (whose P value less than .1). These VCM trough concentration (mg/L), 12.2 ± 4.6
factors were applied to the multiple logistic regression analysis. mean ± SD
In a previous report, pre‐existing renal insufficiency had been With nephrotoxicity, n (%) 87 (14.7)
13
defined as a risk factor of VCM‐induced nephrotoxicity. However, Abbreviations: NSAIDs, nonsteroidal anti‐inflammatory drugs; SCr, serum
our results were inconsistent with this data (baseline CCr ≤ 86.6 mL/ creatinine; SD, standard deviation; VCM, vancomycin.
4 IMAI ET AL.

TABLE 2 Comparison of patient characteristics between those with nephrotoxicity and those without nephrotoxicity
With Nephrotoxicity Without Nephrotoxicity
Description (n = 87) (n = 505) P Value

Age (years), median (range) 62 (19‐96) 64 (19‐89) .164

Gender (male), n (%) 54 (62.1) 328 (65.0) .604
Body weight (kg), median (range) 58.4 (35.9‐89.4) 57.7 (31.6‐127.0 .790
Body mass index, median (range) 22.9 (13.8‐33.9) 21.5 (12.7‐49.1) .037*
Baseline serum creatinine (mg/dL), median (range) 0.64 (0.28‐2.6) 0.68 (0.16‐4.0) .110
Baseline creatinine clearance (mL/min), median (range) 95.9 (29.3‐223.2) 83.7 (16.4‐512.6) .104
Duration of therapy (days), median (range) 13 (4‐71) 9 (3‐83) <.001*
Concomitant medications, n (%)
NSAIDs 45 (51.7) 219 (43.4) .147
Furosemide 51 (58.6) 130 (25.7) <.001*
Amphotericin B 3 (3.5) 2 (0.40) .025*
Aminoglycoside antibiotics 4 (4.6) 12 (2.4) .273
Vasopressor drugs 25 (28.7) 41 (8.1) <.001*
Residence in intensive care unit, n (%) 20 (23.0) 54 (10.7) .001*
Days to initial TDM (days), median (range) 3 (3‐6) 3 (3‐10) .705
VCM trough concentration (mg/L), median (range) 15.5 (7.4‐42.1) 11.2 (2.2‐28.7) <.001*

Abbreviations: NSAIDs, nonsteroidal anti‐inflammatory drugs; VCM, vancomycin.

*P values ≤.05 were considered statistically significant.

TABLE 3 Univariate analysis of risk factors for nephrotoxicity patients with a VCM trough concentration of ≥15.0 mg/mL, concomi-
tant medication of vasopressor drugs was selected as the second
Characteristic OR 95% CI P Value
splitting variable. Patients who used vasopressor drugs had a higher
Age ≥52 years 0.95 0.57‐1.58 .838
probability of developing nephrotoxicity (with vasopressor drugs,
Gender (male) 0.88 0.55‐1.41 .604
69.6% vs without vasopressor drugs, 28.2%). Among patients who
Body mass index ≥25 1.42 0.84‐2.42 .193
did not use vasopressor drugs, concomitant medication of furosemide
Baseline serum creatinine ≥1.0 mg/dL 0.84 0.44‐1.61 .595
was selected as the third splitting variable. Patients who
Baseline creatinine clearance ≤86.6 mL/min 0.64 0.40‐1.01 .057*
used furosemide had a higher probability of developing nephrotoxicity
Duration of therapy ≥14 days 2.68 1.68‐4.26 <.001*
(with furosemide, 43.2% vs without furosemide, 21.2%). Among
Concomitant medications
patients with a VCM trough concentration of <15.0 mg/mL,
NSAIDs 1.40 0.89‐2.21 .149
concomitant medication of furosemide was selected as the second
Furosemide 4.09 2.55‐6.55 <.001*
splitting variable. Patients who used furosemide had a higher
Vasopressor drugs 4.56 2.60‐8.02 <.001*
probability of developing nephrotoxicity (with furosemide, 18.0% vs
Residence in intensive care unit 2.49 1.41‐4.42 <.002*
without furosemide, 4.6%). Among patients who used furosemide,
VCM trough concentration ≥15 mg/L 5.87 3.63‐9.49 <.001*
duration of therapy ≥14 days was selected as the third splitting
Abbreviations: 95% CI, 95% confidence interval; NSAIDs, nonsteroidal variable. Patients with duration of therapy ≥14 days had a higher
anti‐inflammatory drugs; OR; odds ratio, VCM, vancomycin.
probability of developing nephrotoxicity (with duration of therapy
*P values ≤0.1 were included in the multiple logistic regression analysis.
≥14 days, 32.1% vs with duration of therapy <14 days, 8.0%).

TABLE 4 Multivariable analysis of independent risk factors for 3.4 | Validation of the DT model
nephrotoxicity
The predictive accuracies of the DT and logistic regression models
Characteristic OR 95% CI P value were similar (87.3%, respectively).
Concomitant medications The misclassification risk of the DT model, estimated by 10‐fold
Furosemide 2.79 1.63‐4.80 <.001* cross validation, was 13.2 ± 1.4%.
Vasopressor drugs 2.85 1.48‐5.48 <.002*
Duration of therapy ≥14 days 2.50 1.49‐4.22 <.001*
VCM trough concentration ≥15 mg/L 5.92 3.53‐9.93 <.001* 4 | DISCUSSION
Abbreviations: 95% confidence interval; OR, odds ratio; 95% CI; VCM,
vancomycin, In the present study, we used a DT analysis to explore the
*P values ≤.05 were considered statistically significant. combinations of several risk factors and their influence on
IMAI ET AL. 5

FIGURE 1 The decision tree model for the prediction of vancomycin‐associated nephrotoxicity. YES: with nephrotoxicity, NO: without
nephrotoxicity

nephrotoxicity associated with VCM. Moreover, we evaluated the use- “nephrotoxicity‐induced high‐trough concentration.” Thus, it is unclear
fulness of DT analysis for the evaluation of adverse drug reactions. if the high trough levels are the cause or the result of nephrotoxicity.
In this study, 87 (14.7%) patients developed nephrotoxicity. This This is an important limitation of our study.
incidence of nephrotoxicity was also similar to the previous reports. Generally, pre‐existing renal insufficiency has been defined as a risk
However, the rate of nephrotoxicity was highly variable, ranging from factor of VCM‐induced nephrotoxicity.13 However, in the univariate
5% to 43%.25 analysis (Table 3). our results were inconsistent with this (baseline
In the multiple logistic regression analysis (Table 4). four factors CCr ≤ 86.6 mL/min, OR 0.64, 95% CI, 0.40‐1.01). This is because when
(duration of therapy ≥14 days, concomitant medications [furosemide the CCr was calculated, the ABW was used for overweight patients
and vasopressor drugs], and VCM trough concentration ≥15.0 mg/L) (BMI > 25), but Scr was not adjusted. In several reports, with elderly
were extracted as independent predictors of nephrotoxicity. Without patients with low Scr, rounding the Scr to 1.0 mg/dL or 0.6 mg/dL
the VCM trough concentration ≥15.0 mg/L, their OR were similar to was recommended.28,29 If we used an adjusted Scr, baseline CCr would
12-15
previous reports. not have been extracted as a risk factor of nephrotoxicity. In this study,
In a systematic review and meta‐analysis,13 it was reported that a we used actual Scr, because the adjustment of Scr was not well
VCM trough concentration of ≥15.0 mg/L was associated with established. We would like to investigate the sensitivity analysis for
increased probability of nephrotoxicity (OR 2.67, 95% CI, 1.95‐3.65). CCr calculations in future research.
Our result are consistent with these; however, the OR (OR 5.92, The Hosmer‐Lemeshow statistic had a P value of 0.27, indicating a
95% CI, 3.53‐9.93) obtained in our study was higher than in previous good fit.
reports. This was probably caused by differences in the target trough In the DT model, 4 factors (duration of therapy ≥14 days,
range between the study groups. In previous reports, the therapeutic concomitant medications (furosemide and vasopressor drugs], and
range of the VCM was set to 15 to 20 mg/L13,16,18; however, the VCM trough concentration ≥15.0 mg/L) were extracted as splitting
target trough level of VCM in our hospital was set to 10 to 20 mg/L variables. Higher VCM trough levels were the most common risk factor
(including skin and soft tissue infections, bacteraemia, endocarditis, of nephrotoxicity16; therefore, the VCM trough concentration of
pneumonia, bone and joint infections, and central nervous system ≥15.0 mg/L was extracted as the first splitting variable. However, as
infections were set to 15‐20 mg/L) based on the TDM practice described above, some limitations are apparent. Among patients with
guidelines in Japan.26 Therefore, patients in this study with VCM a VCM trough concentration of ≥15.0 mg/mL, concomitant medication
trough level 15 to 20 mg/L were critically ill, and nephrotoxicity was of vasopressor drugs was selected as the second splitting variable.
therefore a common occurrence.13,15,27 Another possibility is that When concomitant vasopressor drugs were in use, 69.6% of patients
our target trough range (10‐20 mg/L) was lower than previous reports, developed nephrotoxicity. Based on the Japanese TDM practice guide-
so the VCM trough concentration of ≥15.0 mg/L might be a lines,26 the target trough range was set to 15.0 to 20.0 mg/L, which
6 IMAI ET AL.

also represented critically ill patients. Therefore, patients using vaso- transplantation using a machine learning algorithm: a European Group
pressor drugs had unstable haemodynamics. This population includes for Blood and Marrow Transplantation Acute Leukemia Working Party
Retrospective Data Mining Study. J Clin Oncol. 2015;33(28):3144‐3151.
the patients who suffered from the most serious illness such as severe
5. LeBlanc M, Crowley J. A review of tree‐based prognostic models. Cancer
septic shock and therefore had a high probability of developing
Treat Res. 1995;75:113‐124.
nephrotoxicity.
6. Hammann F, Gutmann H, Vogt N, Helma C, Drewe J. Prediction of
Concomitant medication of furosemide was selected as the adverse drug reactions using decision tree modeling. Clin Pharmacol
second and third splitting variable. Furosemide is a widely used Ther. 2010;88(1):52‐59.
diuretic. Indeed, 181 patients (30.6%) in this study were prescribed 7. Tseng WT, Chiang WF, Liu SY, Roan J, Lin CN. The application of data
with furosemide. Our results suggested that the VCM trough concen- mining techniques to oral cancer prognosis. J Med Syst. 2015;39(5):59.
tration of ≥15.0 mg/mL and patients not using vasopressor drugs, 8. Liu C, Bayer A, Cosgrove SE, et al. Clinical practice guidelines by the
concomitant medication of furosemide can lead to nephrotoxicity. infectious diseases society of america for the treatment of methicillin‐
resistant Staphylococcus aureus infections in adults and children. Clin
Duration of therapy ≥14 days was selected as the third splitting
Infect Dis. 2011;52(3):e18‐e55. https://doi.org/10.1093/cid/ciq146.
variable. Duration of therapy is one of the factors that contribute to
9. Bosso JA, Nappi J, Rudisill C, et al. Relationship between vancomycin
VCM exposure. Several previous reports suggested that patients with trough concentrations and nephrotoxicity: a prospective multicenter
nephrotoxicity tended to have longer VCM exposures and that the risk trial. Antimicrob Agents Chemother. 2011;55(12):5475‐5479.
of nephrotoxicity increased with time.12,13,25 Our results suggested 10. Hermsen ED, Hanson M, Sankaranarayanan J, Stoner JA, Florescu MC,
that long‐term administration of VCM is a risk factor of nephrotoxicity Rupp ME. Clinical outcomes and nephrotoxicity associated with vanco-
mycin trough concentrations during treatment of deep‐seated
for patients using furosemide, even with a VCM trough concentration
infections. Expert Opin Drug Saf. 2010;9(1):9‐14.
of <15.0 mg/L.
11. Hidayat LK, Hsu DI, Quist R, Shriner KA, Wong‐Beringer A. High‐dose
The predictive accuracies of the DT and logistic regression
vancomycin therapy for methicillin‐resistant Staphylococcus
models were similar (87.3%, respectively). It was shown the predictive aureus infections: efficacy and toxicity. Arch Intern Med. 2006;
performance of each model was equivalent. These accuracies were 166(19):2138‐2144.
considered to be favourable, compared to DT analyses in other 12. Jeffres MN, Isakow W, Doherty JA, Micek ST, Kollef MH. A retrospec-
fields.7,30,31 As this is the first report of a DT analysis for nephrotoxicity tive analysis of possible renal toxicity associated with vancomycin in
patients with health care‐associated methicillin‐resistant Staphylococ-
associated with VCM, predictive accuracies were not comparable to
cus aureus pneumonia. Clin Ther. 2007;29(6):1107‐1115.
previous studies.
13. Lodise TP, Lomaestro B, Graves J, Drusano GL. Larger vancomycin
Another limitation of our study is that we could not evaluate doses (at least four grams per day) are associated with an increased
several factors that have reported previous studies (ie, septic shock, incidence of nephrotoxicity. Antimicrob Agents Chemother.
history of acute kidney injury, APACHE II scores, and total daily dose). 2008;52(4):1330‐1336.

In the future research, it will be important to build more practically 14. Minejima E, Choi J, Beringer P, Lou M, Tse E, Wong‐Beringer A.
Applying new diagnostic criteria for acute kidney injury to facilitate
applicable models that include these factors.
early identification of nephrotoxicity in vancomycin‐treated patients.
Antimicrob Agents Chemother. 2011;55(7):3278‐3283.

5 | C O N CL U S I O N S 15. Cappelletty D, Jablonski A, Jung R. Risk factors for acute kidney injury
in adult patients receiving vancomycin. Clin Drug Investig.
2014;34(3):189‐193.
The predictive performance of the DT model and logistic regression
16. van Hal SJ, Paterson DL, Lodise TP. Systematic review and meta‐
model was equivalent. Application of the DT model to clinical practice analysis of vancomycin‐induced nephrotoxicity associated with dos-
can improve the prediction of adverse drug reactions. This study ing schedules that maintain troughs between 15 and 20 milligrams
suggests the usefulness of DT models for the evaluation of adverse per liter. Antimicrob Agents Chemother. 2013;57(2):734‐744.

drug reactions. 17. Ricci Z, Cruz D, Ronco C. The RIFLE criteria and mortality in acute
kidney injury: a systematic review. Kidney Int. 2008;73(5):538‐546.
18. Rybak MJ, Lomaestro BM, Rotschafer JC, et al. Vancomycin therapeu-
C O N F LI C T O F I NT E R ES TS tic guidelines: a summary of consensus recommendations from the
infectious diseases Society of America, the American Society of
The authors declare no conflict of interests. Health‐System Pharmacists, and the Society of Infectious Diseases
Pharmacists. Clin Infect Dis. 2009;49(3):325‐327.
RE FE R ENC E S 19. Hall RG 2nd, Hazlewood KA, Brouse SD, et al. Empiric guideline‐
1. Hiramatsu N, Kurosaki M, Sakamoto N, et al. Pretreatment prediction recommended weight‐based vancomycin dosing and nephrotoxicity
of anemia progression by pegylated interferon alpha‐2b plus ribavirin rates in patients with methicillin‐resistant Staphylococcus aureus bacter-
combination therapy in chronic hepatitis C infection: decision‐tree emia: a retrospective cohort study. BMC Pharmacol Toxicol. 2013;14:12.
analysis. J Gastroenterol. 2011;46(9):1111‐1119. 20. Takahashi H, Mori M. Characteristics and significance of criteria for
2. Linden A, Yarnold PR. Using data mining techniques to characterize obesity disease in Japan 2011. Nihon Rinsho. 2013;71(2):257‐261.
participation in observational studies. J Eval Clin Pract. 2016; https:// 21. Asakawa Y, Ohtaka M, Maekawa S, et al. Stratifying the risk of lymph
doi.org/10.1111/jep.12515. node metastasis in undifferentiated‐type early gastric cancer. World J
3. Song YY, Lu Y. Decision tree methods: applications for classification Gastroenterol. 2015;21(9):2683‐2692.
and prediction. Shanghai Arch Psychiatry. 2015;27(2):130‐135. 22. Rodríguez AH, Avilés‐Jurado FX, Díaz E, et al. Procalcitonin (PCT) levels
4. Shouval R, Labopin M, Bondi O, et al. Prediction of allogeneic hemato- for ruling‐out bacterial coinfection in ICU patients with influenza: a
poietic stem‐cell transplantation mortality 100 days after CHAID decision‐tree analysis. J Infect. 2016;72(2):143‐151.
IMAI ET AL. 7

23. Lo‐Ciganic W, Zgibor JC, Ruppert K, Arena VC, Stone RA. Identifying 29. Winter ME. Creatinine Clearance. The 3rd Edition of. Basic Clinical Phar-
type 1 and type 2 diabetic cases using administrative data: a tree‐ macokinetics. Vancouver, WA: Applied Therapeutics, Inc Press; 1994.
structured model. J Diabetes Sci Technol. 2011;5(3):486‐493. 30. Nakayama N, Oketani M, Kawamura Y, et al. Algorithm to determine
24. McCabe RM, Grutsch JF, Braun DP, Nutakki SB. Fatigue as a driver of the outcome of patients with acute liver failure: a data‐mining analysis
overall quality of life in cancer patients. PLoS One. 2015;10(6): using decision trees. J Gastroenterol. 2012;47(6):664‐677.
e0130023. https://doi.org/10.1371/journal.pone.0130023. 31. Melillo P, Izzo R, Orrico A, et al. Automatic prediction of cardiovascular
25. Carreno JJ, Kenney RM, Lomaestro B. Vancomycin‐associated and cerebrovascular events using heart rate variability analysis. PLoS
renal dysfunction: where are we now? Pharmacotherapy. 2014; One. 2015;10(3):e0118504. https://doi.org/10.1371/journal.pone.
34(12):1259‐1268. 0118504.
26. Matsumoto K, Takesue Y, Ohmagari N, et al. Practice guidelines for
therapeutic drug monitoring of vancomycin: a consensus review of
the Japanese Society of Chemotherapy and the Japanese Society of How to cite this article: Imai S, Yamada T, Kasashi K,
Therapeutic Drug Monitoring. J Infect Chemother. 2013;19(3):365‐380. Kobayashi M, Iseki K. Usefulness of a decision tree model for
27. Lodise TP, Patel N, Lomaestro BM, Rodvold KA, Drusano GL. Relation- the analysis of adverse drug reactions: Evaluation of a risk
ship between initial vancomycin concentration‐time profile and
prediction model of vancomycin‐associated nephrotoxicity
nephrotoxicity among hospitalized patients. Clin Infect Dis.
2009;49(4):507‐514. constructed using a data mining procedure. J Eval Clin Pract.
28. Smythe M, Hoffman J, Kizy K, Dmuchowski C. Estimating creatinine 2017;0:1–7. https://doi.org/10.1111/jep.12767
clearance in elderly patients with low serum creatinine concentrations.
Am J Hosp Pharm. 1994;51(2):198‐204.