Available via license: CC BY 4.0
Content may be subject to copyright.
Original Paper
Risk Factors and Predictive Models for Peripherally Inserted
Central Catheter Unplanned Extubation in Patients With Cancer:
Prospective, Machine Learning Study
Jinghui Zhang1,2,3, PhD; Guiyuan Ma1,2, MD; Sha Peng1,2, MD; Jianmei Hou1, MD; Ran Xu1,2, MD; Lingxia Luo1,2,
MD; Jiaji Hu1,2, MD; Nian Yao1,2, MD; Jiaan Wang4, BS; Xin Huang5, BS
1Teaching and Research Section of Clinical Nursing, Xiangya Hospital of Central South University, Changsha, Hunan, China
2Xiangya School of Nursing, Central South University, Changsha, Hunan, China
3National Clinical Research Center for Geriatric Diseases, Xiangya Hospital, Central South University, Changsha, Hunan, China
4Vascular Access Department, Hainan Provincial People's Hospital, Hainan, China
5Department of Nursing, Affiliated Hospital of Qinghai University, Qinghai, China
Corresponding Author:
Guiyuan Ma, MD
Teaching and Research Section of Clinical Nursing
Xiangya Hospital of Central South University
Number 87 Xiangya Road, Kaifu District
Changsha, Hunan, 410008
China
Phone: 86 13026179120
Email: mmgy0906@163.com
Abstract
Background: Cancer indeed represents a significant public health challenge, and unplanned extubation of peripherally inserted
central catheter (PICC-UE) is a critical concern in patient safety. Identifying independent risk factors and implementing high-quality
assessment tools for early detection in high-risk populations can play a crucial role in reducing the incidence of PICC-UE among
patients with cancer. Precise prevention and treatment strategies are essential to improve patient outcomes and safety in clinical
settings.
Objective: This study aims to identify the independent risk factors associated with PICC-UE in patients with cancer and to
construct a predictive model tailored to this group, offering a theoretical framework for anticipating and preventing PICC-UE in
these patients.
Methods: Prospective data were gathered from January to December 2022, encompassing patients with cancer with PICC at
Xiangya Hospital, Central South University. Each patient underwent continuous monitoring until the catheter’s removal. The
patients were categorized into 2 groups: the UE group (n=3107) and the non-UE group (n=284). Independent risk factors were
identified through univariate analysis, the least absolute shrinkage and selection operator (LASSO) algorithm, and multivariate
analysis. Subsequently, the 3391 patients were classified into a train set and a test set in a 7:3 ratio. Utilizing the identified
predictors, 3 predictive models were constructed using the logistic regression, support vector machine, and random forest
algorithms. The ultimate model was selected based on the receiver operating characteristic (ROC) curve and TOPSIS (Technique
for Order Preference by Similarity to Ideal Solution) synthesis analysis. To further validate the model, we gathered prospective
data from 600 patients with cancer at the Affiliated Hospital of Qinghai University and Hainan Provincial People’s Hospital from
June to December 2022. We assessed the model’s performance using the area under the curve of the ROC to evaluate differentiation,
the calibration curve for calibration capability, and decision curve analysis (DCA) to gauge the model’s clinical applicability.
Results: Independent risk factors for PICC-UE in patients with cancer were identified, including impaired physical mobility
(odds ratio [OR] 2.775, 95% CI 1.951-3.946), diabetes (OR 1.754, 95% CI 1.134-2.712), surgical history (OR 1.734, 95% CI
1.313-2.290), elevated D-dimer concentration (OR 2.376, 95% CI 1.778-3.176), targeted therapy (OR 1.441, 95% CI 1.104-1.881),
surgical treatment (OR 1.543, 95% CI 1.152-2.066), and more than 1 catheter puncture (OR 1.715, 95% CI 1.121-2.624). Protective
factors were normal BMI (OR 0.449, 95% CI 0.342-0.590), polyurethane catheter material (OR 0.305, 95% CI 0.228-0.408), and
valved catheter (OR 0.639, 95% CI 0.480-0.851). The TOPSIS synthesis analysis results showed that in the train set, the composite
J Med Internet Res 2023 | vol. 25 | e49016 | p. 1https://www.jmir.org/2023/1/e49016 (page number not for citation purposes)
Zhang et alJOURNAL OF MEDICAL INTERNET RESEARCH
XSL
•
FO
RenderX
index (Ci) values were 0.00 for the logistic model, 0.82 for the support vector machine model, and 0.85 for the random forest
model. In the test set, the Ci values were 0.00 for the logistic model, 1.00 for the support vector machine model, and 0.81 for the
random forest model. The optimal model, constructed based on the support vector machine, was obtained and validated externally.
The ROC curve, calibration curve, and DCA curve demonstrated that the model exhibited excellent accuracy, stability,
generalizability, and clinical applicability.
Conclusions: In summary, this study identified 10 independent risk factors for PICC-UE in patients with cancer. The predictive
model developed using the support vector machine algorithm demonstrated excellent clinical applicability and was validated
externally, providing valuable support for the early prediction of PICC-UE in patients with cancer.
(J Med Internet Res 2023;25:e49016) doi: 10.2196/49016
KEYWORDS
cancer; PICC; unplanned extubation; predictive model; logistic; support vector machine; random forest
Introduction
Peripherally inserted central catheters (PICCs) are commonly
used in patients with cancer who need long-term chemotherapy
and supportive care therapy [1]. PICCs can effectively minimize
vascular irritation caused by chemotherapy drugs, thereby
preventing extravasation and the necessity for repeated punctures
[2,3]. However, PICCs also have their share of disadvantages.
One significant issue is the occurrence of unplanned extubation
(UE) during PICC placement, which can be both frequent and
severe [4]. PICC-UE occurs when the catheter needs to be
withdrawn prematurely due to severe complications or accidental
dislodgment resulting from patient or operator factors [4,5].
The incidence rates for PICC-UE range from 2.5% to 40.7%
[6]. The occurrence of PICC-UE poses a significant risk to
patients with cancer. It not only delays chemotherapy, prolongs
hospitalization, and increases the financial burden on their
families but also impacts the patients’ quality of life and, in
some cases, even threatens their lives [7].
Previous studies primarily focused on risk factors for
PICC-related complications. These complications can be
associated with a variety of factors, including (1) patient-related
factors, such as critically ill bedridden patients, age, and
immunity [8,9]; operator-related factors, such as puncture times,
professional skills, and the use of visualization technology
[10-12]; catheter-related factors, such as catheter material,
catheter lumen, and catheter diameter [13-15]; and treatment
process–related factors, such as chemotherapy, radiotherapy,
different drug types, and other aspects [16-18]. However, there
is limited research on the risk factors for PICC-UE. Existing
studies have primarily centered on accidental dislodgment of
ventilator tubes [4,19], with insufficient attention paid to
PICC-UE. Therefore, it is imperative to identify PICC-UE risk
factors and develop predictive models in patients with cancer
to enhance the safety of PICC usage.
To mitigate the adverse effects of PICC-UE, a promising
strategy is to identify high-risk patients and offer appropriate
advice for extended catheter usage. While risk prediction models
for UE have been developed for intensive care unit (ICU)
patients with ventilator tracheal intubation [20,21], there are no
studies or models that can identify high-risk patients for
PICC-UE. Lee et al [20] developed a risk assessment tool for
evaluating UE of the endotracheal tube, while Hur et al [21]
used 8 years of data to build a predictive model for UE using
various machine learning (ML) algorithms. While both models
exhibited high sensitivity and specificity, they were designed
for predicting UE in ventilator tube cases.
ML algorithms are adept at extracting key features from complex
data sets and are increasingly used in diagnosing and
prognosticating various diseases [22]. In the context of
PICC-related complications, previous studies have used ML
techniques to assess risk [23,24]. Badheka et al [23] identified
high-risk predictors of catheter-related thrombosis in infants
under 1 year using conventional and neural network methods.
Conversely, Liu et al [24] developed a predictive model for
PICC-related vein thrombosis in patients with cancer using the
least absolute shrinkage and selection operator (LASSO) and
random forest (RF) algorithms, which exhibited impressive
performance. However, as far as we know, no specific research
on ML for PICC-UE in patients with cancer has been conducted
yet.
This study aimed to identify PICC-UE risk factors in patients
with cancer, develop and validate ML-based predictive models
for PICC-UE, and promote early intervention to reduce its
incidence and enhance patients’ quality of life. This study
represents the first attempt to identify high-risk PICC-UE
patients and serves as a valuable reference for future research
and medical decision-making. We followed the Guidelines for
Developing and Reporting Machine Learning Predictive Models
in Biomedical Research [25] to report our study.
Methods
Study Design and Participants
This study used data from Xiangya Hospital of Central South
University to build a predictive model for PICC-UE. Prospective
data were collected from various hospital systems from January
1, 2022, to December 31, 2022, including the infusion system,
the in-hospital Hitech electronic case system, and the PICC
catheter integrated case management system. We utilized all
available data to identify independent risk factors. The entire
data set was divided into a train set and a test set using a 7:3
ratio through the random number table method. The train set
was used for model construction, while the test set was used
for internal validation. We collected data from the Affiliated
Hospital of Qinghai University and Hainan Provincial People’s
Hospital to perform additional validation of the model between
June 1, 2022, and December 31, 2022. The external validation
J Med Internet Res 2023 | vol. 25 | e49016 | p. 2https://www.jmir.org/2023/1/e49016 (page number not for citation purposes)
Zhang et alJOURNAL OF MEDICAL INTERNET RESEARCH
XSL
•
FO
RenderX
data were sourced from different hospitals and were independent
of the data used for model construction.
Inclusion criteria were as follows: (1) pathological diagnosis
of oncology; (2) availability of PICC catheterization
information; and (3) voluntary participation with informed
consent. Exclusion criteria were as follows: (1) patients or
caregivers unable to cooperate with the investigation; (2)
patients who missed visits before catheter removal; (3)
incomplete data collection; and (4) abnormal values that affect
judgment.
Sample Size and Sampling
We used the sample size formula designed for cohort studies
to calculate the minimum number of PICC-UE cases needed.
Then, we determined the sample size required to prospectively
enroll patients with cancer with PICC insertions for this study
based on the PICC-UE incidence. We set α=.05 and β=.10 and
obtained μα/2=1.96 and μβ=1.28.
Previous studies [4,5] have identified multiple risk factors for
PICC-UE, and among these risk factors, thrombosis had the
largest minimum sample size requirement for the case group.
In the group without PICC-UE, the incidence of thrombosis
was 8.9% (22/247; P0=.09), whereas in the group with
PICC-UE, it was 27% (12/44; P1=.27). Hence, this study’s case
group (UE cases) requires a minimum sample size of 164. The
incidence of PICC-UE is reported as 9% (11/121) [6]. Based
on this value, the initial sample size needed for a prospective
study was 2448. After accounting for the possibility of missed
visits and increasing the sample size by 20%, the required
sample size is at least 2937.
Instruments
The follow-up data collection schedule and clinical data
collection form for this study were established through a
literature review [4-19,23,24], semistructured interviews, and
research group discussions.
The study investigators enrolled eligible participants who
provided informed consent into a cancer whole-course
management system. One-to-one follow-up through WeChat
(Tencent Holdings Ltd.) was established, with follow-ups
scheduled in advance. Patients were reminded to contact the
investigators immediately in case of any catheter-related
abnormalities. Collected data included observations of catheter
patency; signs of redness, swelling, and pain in the extremity
at the insertion site; blood and fluid leakage at the puncture site;
catheter prolapse and its length; and any other abnormalities.
Additionally, PICC-UE occurrences were monitored, and their
time and reasons were recorded. Follow-up visits were
conducted on the day of placement, as well as on days 1, 7, 14,
21, and every 21 days thereafter.
A total of 33 relevant factors were collected for data analysis,
categorized as follows: (1) general information (gender, age,
tumor type, education, BMI [calculated using height and
weight], alcohol history, mental status, cooperation, and physical
mobility); (2) medical history (history of deep vein thrombosis,
history of central venous placement, diabetes, hypertension,
cardiovascular disease, hyperlipidemia, and surgical history);
(3) laboratory indicators (D-dimer concentration and fibrinogen
concentration); (4) therapy schedule (radiotherapy treatment,
targeted therapy, surgical treatment, anticoagulation,
chemotherapy treatment, and hyperosmolar drugs); and (5)
placement information (limb on the side of placement, puncture
method, puncture times, catheter gauge, catheter lumen, catheter
material, presence of a valve, high-pressure–resistant catheter,
and catheter indwelling time). All variables were collected
through observation using patient IDs and case numbers as the
indexes. Data were obtained from the hospital’s Safe Infusion
System (SIS) database and the Hitech electronic case system.
Detailed explanations of the corresponding variables can be
found in Multimedia Appendix 1.
Criteria for PICC-UE, based on previous studies [4,5], were as
follows: (1) a patient who still requires a PICC catheter, but
experiences early extubation due to severe complications; and
(2) a patient who still requires a PICC catheter, but experiences
accidental catheter dislodgment due to patient or operator
factors. PICC-UE serves as the primary outcome of this study.
Risk Factors Identification and Model Development
We reviewed the prospective data collected and categorized
continuous variables, such as age, into 6 groups: “0-11,”
“12-18,” “19-35,” “36-59,” “60-75,” and “≥76.” The variables
height and weight were used to calculate BMI. D-dimer
concentration and fibrinogen concentration values were
converted into high or low categories. Missing values in the
vector data were removed.
We conducted a univariate analysis of the overall data to identify
variables with 2-sided statistical significance (P<.05). Following
a literature review and expert consultations, we used the LASSO
regression algorithm to include clinically significant variables.
The selected variables underwent multifactorial analysis to
identify independent risk factors for PICC-UE in patients with
cancer.
The model was constructed using prospective data from Xiangya
Hospital of Central South University. Data order was
randomized using a shuffling algorithm for even distribution.
The data were then split into a train set and a test set at a ratio
of 7:3 using the random number table method. The overall data
were used for independent risk factor screening, the train set
for model construction, and the test set for internal model
validation. The risk prediction models were constructed using
the train set, incorporating prescreened independent risk factors.
In this study, 3 ML algorithms, namely, logistic regression (LR),
support vector machine (SVM), and RF, were selected to build
risk prediction models for PICC-UE in patients with cancer.
We compared these models using the area under the receiver
operating characteristic (ROC) curve (AUC) and the TOPSIS
(Technique for Order Preference by Similarity to Ideal Solution)
method [26]. AUC assesses the predictive power of the
PICC-UE model, while the model’s superiority was evaluated
based on the Composite Index (Ci) value in the TOPSIS method.
The model with the highest AUC and Ci values was considered
optimal for predicting PICC-UE and selected as the best model.
J Med Internet Res 2023 | vol. 25 | e49016 | p. 3https://www.jmir.org/2023/1/e49016 (page number not for citation purposes)
Zhang et alJOURNAL OF MEDICAL INTERNET RESEARCH
XSL
•
FO
RenderX
Validation and Model Performance Evaluation
Data from June 2022 to December 2022 from Qinghai
University Hospital and Hainan Provincial People’s Hospital
were used for external validation. The collected data were
randomized using a shuffling algorithm for even distribution.
The optimal model was assessed for discrimination, calibration,
and clinical applicability.
Discrimination assesses the model’s ability to distinguish
between high and low PICC-UE risk in the cancer population,
which we evaluated using the AUC. Calibration indicates the
degree of agreement between the predicted and actual results.
The calibration of the model was assessed using the
Hosmer-Lemeshow test with a calibration curve [27]. Clinical
applicability, which gauges the diagnostic accuracy of the model
in clinical use, was evaluated using decision curve analysis
(DCA) [21]. Additionally, model performance was measured
using sensitivity, specificity, positive predictive value, negative
predictive value [24], and AUC.
Ethical Considerations
The study was approved by the Hospital Ethics Review
Committee (approval number 202204210). We adhered to the
principles of informed consent, data confidentiality, anonymity,
and nonharmfulness. Written informed consent was collected,
and any papers or publications based on the study data will not
reveal personal information about the patients. For younger or
unconscious patients who were unable to participate, data
collection was facilitated by their caregivers.
Statistical Analysis
We excluded data with missing or unusual variables from the
prospective data set. Continuous variables were compared using
independent-sample unpaired (2-sided) ttests or one-way
analysis of variance (ANOVA). Categorical variables were
presented as numbers and proportions and compared using the
chi-square test or Fisher exact test. We collected variables with
bilateral P<.05 statistical significance and then included
variables with potential clinical significance for the LASSO
algorithm based on literature analysis and expert consultation.
We identified independent risk factors for PICC-UE in patients
with cancer through multifactorial analysis. After consulting
with experts in ML algorithms and discussions within the
research group, we chose 3 ML methods to construct the study’s
model: RF, SVM, and LR.
All hypothesis tests with 2-sided P<.05 indicated statistical
significance. The “na.omit” function was used to remove
missing values from the vector data. LASSO primarily used the
“glmnet” package with a 10-fold orthogonal method to define
the penalty function. LR, RF, and SVM were mainly
implemented using “caret,” “randomForest,” “pROC,”
“varImpPlot,” and “e1071,” respectively. The ROC curves were
plotted using the “pROC” packet, and the Hosmer-Lemeshow
test using the “hoslem.test,” “data.table,” and “plyr” data
packages was used for the TOPSIS integrated analysis. The
DCA decision curves were constructed using the “rms” and
“rmda” packets. All the analyses were performed using R
Statistical Software, version 4.1.3 (R Foundation).
Results
Participants Characteristics
A total of 3391 patients were included, with a sample loss rate
of 7.34% (269/3660). This included 2374 in the train set and
1017 in the test set, with 284 PICC-UE cases. The study flow
diagram is presented in Figure 1. Baseline participant
characteristics are presented in Table 1. Importantly, there was
no multicollinearity among the variables, as all variance inflation
factor values were less than 5.0.
J Med Internet Res 2023 | vol. 25 | e49016 | p. 4https://www.jmir.org/2023/1/e49016 (page number not for citation purposes)
Zhang et alJOURNAL OF MEDICAL INTERNET RESEARCH
XSL
•
FO
RenderX
Table 1. Comparison of general information between the control and case groupsa.
PvalueChi-square/Ftest (df)PICC-UE (n=284)Non–PICC-UEb(n=3107)Variables
.01d
6.349c(1)
Gender, n (%)
158 (55.63)1486 (47.83)Male
126 (44.37)1621 (52.17)Female
.01d
14.431c(5)
Age (years), n (%)
3 (1.06)57 (1.83)0-11
5 (1.76)73 (2.35)12-18
15 (5.28)208 (6.69)19-35
140 (49.30)1695 (54.55)36-59
100 (35.21)962 (30.96)60-75
21 (7.39)112 (3.60)≥76
.05
15.289c(8)
Tumor type , n (%)
89 (31.34)794 (25.56)Lung cancer
83 (29.23)745 (23.98)Thymic cancer breast cancer
27 (9.51)464 (14.93)Gastro-colorectal cancer
30 (10.56)407 (13.10)Hematologic tumors
16 (5.63)187 (6.02)Cervical cancer
8 (2.82)146 (4.70)Head-neck tumors
9 (3.17)93 (2.99)Hepatobiliary-pancreatic tumors
8 (2.82)95 (3.06)Intracranial tumors
14 (4.93)176 (5.66)Others
.04d
8.088c(3)
Educational level, n (%)
152 (53.52)1399 (45.03)Illiterate primary and junior high schools
69 (24.30)861 (27.71)Secondary and high school
60 (21.13)788 (25.36)College bachelor’s degree
3 (1.06)59 (1.90)Master’s degree doctorate
<.001d
33.741c(2)
BMI (kg/m2), n (%)
18 (6.34)201 (6.47)<18.5
131 (46.13)1945 (62.60)18.5-24.0
135 (47.54)961 (30.93)>24.0
.15
2.123c(1)
Alcohol history, n (%)
242 (85.21)2739 (88.16)None
42 (14.79)368 (11.84)Yes
.02d
5.529c(1)
Mental status, n (%)
258 (90.85)2930 (94.30)Sobriety
26 (9.15)177 (5.70)Blurred consciousness
.004d
8.528c(1)
Cooperation, n (%)
253 (89.08)2909 (93.63)Cooperative
31 (10.92)198 (6.37)Noncooperative
<.001d
43.276c(1)
Physical mobility, n (%)
J Med Internet Res 2023 | vol. 25 | e49016 | p. 6https://www.jmir.org/2023/1/e49016 (page number not for citation purposes)
Zhang et alJOURNAL OF MEDICAL INTERNET RESEARCH
XSL
•
FO
RenderX
PvalueChi-square/Ftest (df)PICC-UE (n=284)Non–PICC-UEb(n=3107)Variables
232 (81.69)2884 (92.82)Normal
52 (18.31)223 (7.18)Abnormal
<.001d
12.799c(1)
History of deep vein thrombosis, n (%)
250 (88.03)2909 (93.63)None
34 (11.97)198 (6.37)Yes
.05
7.724e(3)
History of central venous placement, n (%)
233 (82.04)2720 (87.54)None
29 (10.21)235 (7.56)1
15 (5.28)107 (3.44)2
7 (2.46)45 (1.45)≥3
<.001d
13.381c(1)
Diabetes, n (%)
253 (89.08)2935 (94.46)None
31 (10.92)172 (5.54)Yes
.70
0.149c(1)
Hypertension, n (%)
261 (91.90)2875 (92.53)None
23 (8.10)232 (7.47)Yes
.08
3.023c(1)
Cardiovascular disease, n (%)
257 (90.49)2897 (93.24)None
27 (9.51)210 (6.76)Yes
<.001d
14.841c(1)
Hyperlipidemia, n (%)
243 (85.56)2864 (92.18)None
41 (14.44)243 (7.82)Yes
<.001d
21.580c(1)
Surgical history, n (%)
137 (48.24)1935 (62.28)None
147 (51.76)1172 (37.72)Yes
<.001d
66.054c(1)
D-dimer concentration, n (%) (mg/dl)
187 (65.85)2632 (84.71)≤0.5
97 (34.15)475 (15.29)>0.5
.005d
10.658c(2)
Fibrinogen concentration, n (%) (mg/dl)
7 (2.46)94 (3.03)Lower
189 (66.55)2315 (74.51)Normal
88 (30.99)698 (22.47)Higher
.06
3.464c(1)
Radiotherapy treatment, n (%)
249 (87.68)2828 (91.02)None
35 (12.32)279 (8.98)Yes
.001d
12.042c(1)
Targeted therapy, n (%)
103 (36.27)1460 (46.99)None
181 (63.73)1647 (53.01)Yes
<.001d
26.409c(1)
Surgical treatment, n (%)
J Med Internet Res 2023 | vol. 25 | e49016 | p. 7https://www.jmir.org/2023/1/e49016 (page number not for citation purposes)
Zhang et alJOURNAL OF MEDICAL INTERNET RESEARCH
XSL
•
FO
RenderX
PvalueChi-square/Ftest (df)PICC-UE (n=284)Non–PICC-UEb(n=3107)Variables
190 (66.90)2483 (79.92)None
94 (33.10)624 (20.08)Yes
.09
2.759c(1)
Anticoagulation, n (%)
258 (90.85)2903 (93.43)None
26 (9.15)204 (6.57)Yes
.16
2.016c(1)
Chemotherapy treatment, n (%)
31 (10.92)433 (13.94)None
253 (89.08)2674 (86.06)Yes
.002d
9.783c(1)
Hyperosmolar drugs, n (%)
106 (37.32)1460 (46.99)None
178 (62.68)1647 (53.01)Yes
.13
5.718e(3)
Limb on side of placement, n (%)
131 (46.13)1570 (50.53)Left upper extremity
141 (49.65)1467 (47.22)Right upper extremity
7 (2.46)37 (1.19)Left lower extremity
5 (1.76)33 (1.06)Right lower extremity
.92
0.170c(2)
Puncture method, n (%)
10 (3.52)98 (3.15)Blind
14 (4.93)164 (5.28)
Blind-MSTf
260 (91.55)2845 (91.57)Bright scan ultrasound-MST
.001d
11.900c(1)
Puncture times, n (%)
253 (89.08)2928 (94.24)1
31 (10.92)179 (5.76)Many times
.26
3.678e(3)
Catheter gauge (Fr), n (%)
2 (0.70)6 (0.19)1.9
5 (1.76)82 (2.64)3
260 (91.55)2847 (91.63)4
17 (5.99)172 (5.54)5
.32
0.984c(1)
Catheter lumen, n (%)
258 (90.85)2763 (88.93)Single chamber
26 (9.15)344 (11.07)Double chamber
<.001d
13.010c(1)
Catheter material, n (%)
151 (53.17)1308 (42.10)Silicone
133 (46.83)1799 (57.90)Polyurethane
.003d
8.821c(1)
Presence of valve, n (%)
146 (51.41)1314 (42.29)None
138 (48.59)1793 (57.71)Yes
.37
0.791c(1)
Whether high-pressure–resistant catheter, n (%)
J Med Internet Res 2023 | vol. 25 | e49016 | p. 8https://www.jmir.org/2023/1/e49016 (page number not for citation purposes)
Zhang et alJOURNAL OF MEDICAL INTERNET RESEARCH
XSL
•
FO
RenderX
PvalueChi-square/Ftest (df)PICC-UE (n=284)Non–PICC-UEb(n=3107)Variables
229 (80.63)2435 (78.37)None
55 (19.37)672 (21.63)Yes
aThe mean catheter indwelling time for all participants is 91.22 (SD 78.88) days, for the non–PICC-UE group is 91.26 (SD 80.15) days, and for the
PICC-UE group is 90.79 (SD 78.95) days (unpaired 2-tailed ttest =.009; P=.92).
bPICC-UE: unplanned extubation of the peripherally inserted central catheter.
cChi-square test
d2-tailed P<.05.
eFisher exact test.
fMST: modified Seldinger technique.
Independent Risk Factor Determination
A total of 19 potential risk factors, including gender, age, and
education level, were initially screened using univariate analysis.
Following consultations with specialists in vascular surgery,
pathology, and venous therapy, catheter lumen and central
venous placement history were added. Thus, there were a total
of 21 independent variables for the LASSO analysis.
In Figure 2, each colored line represents a variable trend that
decreases as the penalty factor λ changes, resulting in the model
incorporating fewer variables. In Figure 3, the dashed line on
the left indicates the λ value associated with the maximum AUC
and the number of features included in the model. On the right,
the dashed line represents a reduction in the number of features
in the model as the standard error increases by 1 to achieve the
maximum AUC. The minimum error is reached at 1SE=0.013,
resulting in the screening of 11 predictor variables.
Figure 2. Cross-validation plot of the LASSO penalty term. LASSO: least absolute shrinkage and selection operator.
Figure 3. LASSO regression coefficients on the different penalty parameters. LASSO: least absolute shrinkage and selection operator.
The 11 predictors identified by LASSO were analyzed using
conditional LR with a fixed αin of 0.05 and αout of 0.10, using
the backward LR method. The results revealed the following
independent risk factors for PICC-UE in patients with cancer
(ranked by importance from high to low): impaired physical
mobility (odds ratio [OR] 2.775, 95% CI 1.951-3.946), elevated
D-dimer concentration (OR 2.376, 95% CI 1.778-3.176),
diabetes (OR 1.754, 95% CI 1.134-2.712), surgical history (OR
J Med Internet Res 2023 | vol. 25 | e49016 | p. 9https://www.jmir.org/2023/1/e49016 (page number not for citation purposes)
Zhang et alJOURNAL OF MEDICAL INTERNET RESEARCH
XSL
•
FO
RenderX
1.734, 95% CI 1.313-2.290), more than 1 catheter puncture (OR
1.715, 95% CI 1.121-2.624), surgical treatment (OR 1.543, 95%
CI 1.152-2.066), and targeted therapy (OR 1.441, 95% CI
1.104-1.881). Protective factors, ranked by importance from
high to low, were valved catheter (OR 0.639, 95% CI
0.480-0.851), normal BMI (OR 0.449, 95% CI 0.342-0.590),
and polyurethane catheter material (OR 0.305, 95% CI
0.228-0.408). Details are presented in Table 2.
Table 2. Multivariate analysis to identify independent risk factors.
PvalueOdds ratio (95% CI)βVariables
.220.712 (0.414-1.224)–.340
BMI<18.5 kg/m2
<.0010.449 (0.342-0.590)–.801
BMI=18.5-24.0 kg/m2
<.0012.775 (1.951-3.946)1.021Physical mobility
.011.754 (1.134-2.712).562Diabetes
<.0011.734 (1.313-2.290).551Surgical history
<.0012.376 (1.778-3.176).866D-dimer concentration
.0071.441 (1.104-1.881).365Targeted therapy
.0041.543 (1.152-2.066).434Surgical treatment
.011.715 (1.121-2.624).540Puncture times
<.0010.305 (0.228-0.408)–1.188Catheter material
.0020.639 (0.480-0.851)–.447Presence of valve
<.0010.130 (0.087-0.193)–2.043Constant
Prediction Model Construction
The train set and the test set were well balanced, with no
statistically significant differences in composition (P>.05 in all
cases). Further details can be found in Multimedia Appendix
2.
The logistic predictive model was constructed using the 10
independent risk factors identified in the previous phase. The
final model included 9 variables with a χ82value of 320.374
and P<.001. SVM modeling was performed with 10-fold
cross-validation and grid search methods, autonomously
determining the optimal number of vector machines and related
parameters using the tune.svm function. The polynomial kernel
function demonstrated the highest prediction accuracy among
the 4 kernel functions. The RF predictive model for patients
with cancer was constructed with a final minimum of 196 trees.
Model Comparison and Validation
The SVM predictive model exhibited the best predictive efficacy
for PICC-UE when considering AUC and Ci values together.
A comparison of the ROC curves of the 3 models is presented
in Figures 4 (train set) and 5 (test set). The 3 models were
assessed using the TOPSIS integrated analysis in the train set
and test set, as depicted in Tables 3 and 4. The RF predictive
model performed the best in the train set, and the overall
performance of the models is as follows: RF model>SVM
model>logistic model. However, it is worth noting that the Ci
value for the SVM model was 0.82, while that for the RF model
was 0.85, with only a 0.03 difference. In the test set, the TOPSIS
integrated analysis revealed that the SVM predictive model had
the best fit, and the models ranked in terms of overall
performance as SVM model>RF model>logistic model. For a
visual comparison (AUC, sensitivity, specificity, accuracy,
positive predictive value, and negative predictive value) of the
3 models, please refer to Figures 6 (train set) and 7 (test set).
These figures demonstrate that both the SVM model and the
RF model outperform the logistic model in terms of predictive
effects.
J Med Internet Res 2023 | vol. 25 | e49016 | p. 10https://www.jmir.org/2023/1/e49016 (page number not for citation purposes)
Zhang et alJOURNAL OF MEDICAL INTERNET RESEARCH
XSL
•
FO
RenderX
Figure 4. AUC (95% CI) plots of train set.
Figure 5. AUC (95% CI) plots of test set.
Table 3. Comparison of the 3 predictive models in the train set.
Composite
Index
Negative predictive
value (%)
Positive predictive val-
ue (%)
Accuracy
(%)
Specificity
(%)
Sensitivity
(%)
Area under the curve
(%)
Model
0.0096.0639.5289.1391.9158.3875.8Logistic regression
0.8298.1491.7697.6899.3679.1990.4Support vector ma-
chine
0.8599.0493.6298.6199.4489.3484.7Random forest
Table 4. Comparison of the three predictive models in the test set.
Composite
Index
Negative predictive value
(%)
Positive predictive value
(%)
Accuracy
(%)
Specificity
(%)
Sensitivity
(%)
Area under the
curve (%)
Model
0.0095.2419.6275.9177.5358.6268.1Logistic regression
1.0098.9397.4798.8299.7888.5187.5Support vector ma-
chine
0.8198.585.8897.4498.7183.9179.6Random forest
J Med Internet Res 2023 | vol. 25 | e49016 | p. 11https://www.jmir.org/2023/1/e49016 (page number not for citation purposes)
Zhang et alJOURNAL OF MEDICAL INTERNET RESEARCH
XSL
•
FO
RenderX
Figure 6. Comparison of the 3 predictive models in the train set.
Figure 7. Comparison of the 3 predictive models in the test set.
We assessed the performance of the best model through
discrimination, calibration capability, and clinical applicability
analysis. The AUC values evaluated the discrimination, and the
SVM model demonstrated strong differentiation with an AUC
of 0.718 for external validation (Figure 8). The
Hosmer-Lemeshow test for goodness of fit resulted in χ82=8.205,
P=.06, which is greater than 0.05, indicating a well-fitting model
for external validation. The calibration curve for the optimal
model is presented in Figure 9. The clinical applicability of this
predictive model is demonstrated by the DCA curve in Figure
10.
Figure 8. External validation of the ROC curve.
J Med Internet Res 2023 | vol. 25 | e49016 | p. 12https://www.jmir.org/2023/1/e49016 (page number not for citation purposes)
Zhang et alJOURNAL OF MEDICAL INTERNET RESEARCH
XSL
•
FO
RenderX
Figure 9. External validation of the Calibration curve.
Figure 10. External validation of the DCA.
Discussion
Principal Findings
Our prospective study is a pioneering contribution to the field,
being the first to develop and validate a predictive model for
PICC-UE in patients with cancer that can guide decision-making
without requiring extensive laboratory testing. We adhered to
the Guidelines for Developing and Reporting Machine Learning
Predictive Models in Biomedical Research for model
development. Our model demonstrates outstanding performance
in predicting PICC-UE in patients with cancer, achieving an
AUC of 0.904 in the train set and 0.875 in the test set.
Importantly, we identified 10 highly correlated independent
risk factors using univariate, LASSO, and multivariate analyses
to build the model, with the 3 most significant risk factors being
physical mobility (P<.001), D-dimer concentration (P<.001),
and diabetes (P=.01).
PICC-UE incidence varies, ranging from 7.5% to 22.0% in
China [28,29] and from 2.5% to 40.7% in other countries [4].
Duwadi et al [30] noted a higher PICC-UE incidence in the ICU
compared with other units, attributing it to the ICU environment
and patient severity. Additionally, PICC-UE rates differed in
studies from different regions [28,29]. In our study, the
incidence of PICC-UE was 8.38% (284/3391), which is lower
than in most previous studies [4,29]. This could be attributed
to our hospital’s intravenous infusion therapy committee,
improved standardized nurse train, rigorous quality control
management, and numerous educational sessions on patient
health management. These differences in incidence may also
be related to variations in inclusion criteria, follow-up methods,
duration, and the sample size in our study. Future prospective
studies with larger, multicenter samples and extended follow-up
may be necessary for further validation.
In terms of general information, medical history, and laboratory
indicators, we discovered that BMI, physical mobility, diabetes,
surgical history, and D-dimer concentration were linked to the
occurrence of PICC-UE. In particular, patients who are
overweight (BMI>24.0 kg/m2) [31], those with reduced physical
activity [32], and individuals with diabetes prone to
hypercoagulation [32] were at a higher risk of catheter
thrombosis. A recent surgical trauma can also stimulate the
release of a significant amount of coagulation factors to aid
wound healing [33], while a prolonged period of postoperative
bed rest can slow blood flow, both of which increase the risk
of coagulation [34]. An elevated D-dimer level is indicative of
a hypercoagulation state, with a concentration exceeding 500
μg/L signifying a high risk of thrombosis [35]. Bertoglio et al
[36] demonstrated that PICC catheter thrombosis is a significant
risk factor for UE. Patients with a low BMI (BMI<18.5 kg/m2)
have compromised immunity and are prone to malnutrition,
increasing their risk of catheter-related complications and the
J Med Internet Res 2023 | vol. 25 | e49016 | p. 13https://www.jmir.org/2023/1/e49016 (page number not for citation purposes)
Zhang et alJOURNAL OF MEDICAL INTERNET RESEARCH
XSL
•
FO
RenderX
need for catheter removal [37]. Excessive physical activity
increases catheter-vessel wall friction, raising the risk of
bloodstream infection and early catheter dislodgment [38].
In terms of therapy schedule and placement information, we
observed that targeted therapy, surgical treatment, puncture
times, catheter material, and the presence of a valve were linked
to the occurrence of PICC-UE. The use of targeted drugs [39]
and multiple punctures [40] can lead to vascular endothelial
damage, exposing subendothelial prothrombotic components,
inducing platelet aggregation, contributing to catheter
thrombosis, and elevating the risk of extubation [41]. Surgical
treatment leading to PICC-UE aligns with the explanation of
the recent surgical history mentioned earlier. Additionally,
patients recovering from postoperative anesthesia are often
unconscious and may inadvertently remove the catheter due to
the foreign body sensation at the catheter placement site [42].
We identified a higher risk of PICC-UE associated with silicone
catheters. This is attributed to the use of new
high-pressure–resistant polyurethane catheters in our hospital,
which incorporate a surface-active macromolecule with fluorine
atom doping. This component inhibits platelet adhesion and
protein procoagulation, ultimately lowering the incidence of
PICC-related thrombosis [13]. Catheter valves effectively
prevent blood regurgitation, reducing both catheter-related
blockages and thrombosis [13].
In this study, we developed a comprehensive predictive model
to assess the risk of high-risk PICC-UE in patients with cancer.
The model’s performance was evaluated using the AUC as a
measure of classification efficacy, and all models in our study
achieved AUC values exceeding 0.7, demonstrating their strong
ability to distinguish high-risk patients. After comparing the
AUC and Ci values, the SVM model emerged as the optimal
choice. Calibration and DCA curves confirmed the SVM
model’s accuracy, stability, generalizability, and clinical
applicability.
The PICC-UE predictive model for patients with cancer
developed in this study using the ML algorithm offers insights
for related research. In the predictor screening process, previous
studies often relied on a single statistical method [43], while
our approach combined univariate analysis, 10-fold
cross-validation LASSO, and multivariate screening, enhancing
precision and rigor. This approach resulted in the creation of a
more concise and accurate predictive model through multiple
rounds of variable filtering. The LASSO method effectively
aggregates features, achieves dimensionality reduction, and
serves as a feature screening tool, preventing issues related to
covariance and overfitting [44].
This study used multiple ML algorithms to construct the
predictive model, a more scientifically rigorous approach
compared with using a single method alone [26]. ML algorithms
are well-suited for managing high-dimensional variables and
their intricate interactions, making full use of the available data
[22]. The test set demonstrated superior predictive performance
in forecasting PICC-UE based on the results from the train set,
significantly enhancing prediction accuracy. This study
compared 3 ML models and selected the best-performing one,
significantly improving the model’s accuracy. We used AUC
and TOPSIS methods for a comprehensive and rigorous
screening of the optimal predictive model. The SVM algorithm
in the optimal model robustly encompasses the data and reduces
the model’s complexity through linear regression with
insensitive loss functions in a high-dimensional feature space
[45]. Importantly, external validation of the model using
independent data demonstrated significant predictive superiority.
Our study has successfully developed a highly predictive model
for the risk of PICC-UE in patients with cancer using the SVM
algorithm. This model enables the development of personalized
precautions for patients with cancer at a high risk of PICC-UE,
such as the regular assessment of physical mobility and the
provision of targeted physical activity guidance for patients
with impaired physical mobility [32]. For patients with abnormal
BMI, dynamic monitoring of BMI and weight adjustment
through exercise and diet should be implemented [32]. Patients
with diabetes require special attention [32], with routine blood
tests on admission and regular monitoring of D-dimer
concentrations [35] to take preventive measures against early
catheter removal. For patients with a history of surgery and
those undergoing surgical treatment or targeted therapy
[33,34,39], close monitoring of the catheter exit site is essential.
Patients should receive instructions for regular catheter
maintenance and be advised to seek medical attention if they
experience any discomfort. Our study concluded that patients
with multiple punctures are at a higher risk of PICC-UE. It is
recommended that the medical department standardizes the
qualifications of PICC placement nurses and conducts regular
training and assessments [40]. Furthermore, medical departments
should exercise strict control over the choice of catheter
materials and the presence of valves in catheters to minimize
catheter-related complications and lower the incidence of
PICC-UE [13].
Limitations and Challenges
This study has some limitations. First, it did not include
individual genetic data, which can be a significant factor in
PICC-UE. Future studies may benefit from incorporating genetic
data to improve predictive accuracy. Second, external validation
was limited by a small data set, which included data from only
2 hospitals. More extensive external validation is required to
thoroughly validate the predictive model. Lastly, we did not
consider how the risk factors and predictive model for PICC-UE
may differ among various subpopulations of patients with
cancer, including different age groups, genders, and cancer
stages.
Despite the limitations, our study has identified 10 independent
predictors, including BMI, mobility, diabetes, surgical history,
and other factors, that are significantly associated with an
increased risk of PICC-UE in patients with cancer. Furthermore,
our SVM predictive model has been externally validated and
demonstrates excellent generalization. The optimal SVM model
achieved a high accuracy of 97.68% in the train set and 98.82%
in the test set, indicating excellent model fitting. The LASSO
algorithm used for risk factor screening effectively prevented
overfitting. Our findings can raise awareness among clinicians
and patients for the early prevention and reduction of PICC-UE
in high-risk cancer populations. Further prospective multicenter
J Med Internet Res 2023 | vol. 25 | e49016 | p. 14https://www.jmir.org/2023/1/e49016 (page number not for citation purposes)
Zhang et alJOURNAL OF MEDICAL INTERNET RESEARCH
XSL
•
FO
RenderX
studies are needed to validate risk factors and establish effective
UE prophylaxis interventions. Our group is in discussions with
a computer company to develop a plug-in for our hospital’s
electronic system. This plug-in aims to automatically capture
independent risk factors for PICC-UE from patient
hospitalization information. Using the optimal prediction model
from this study, patients’ risk of PICC-UE is categorized into
3 levels: red (high risk), yellow (medium risk), and green (low
risk). Using the color-coded cues, health care providers can
implement tailored interventions for high-risk patients while
offering self-monitoring guidance and health education to
medium- and low-risk patients.
Conclusions
In summary, the developed predictive model for assessing the
risk of PICC-UE in patients with cancer has shown excellent
discrimination, high predictive accuracy, and broad applicability
across a range of risk factors. This model serves as a valuable
tool for the early identification of high-risk patients and holds
promise for clinical implementation.
Acknowledgments
We appreciate the assistance and support of all those in charge of the selected hospitals in the data collection process, as well as
the nurses who participated in the data collection for their time. We also thank all patients who participated in this study. This
work was supported by the Clinical Research Fund of the National Clinical Research Center for Geriatric Disorders (grant number
2021LNJJ09), the National Natural Science Foundation of China (grant number 72174210), the Hunan Natural Science Foundation
(grant number 2022JJ70168), and the Changsha Natural Science Foundation (grant number kq2208367).
Data Availability
The data sets generated or analyzed during this study are not publicly available due to the terms of consent and permission to
which the participants agreed but are available from the corresponding author upon reasonable request.
Authors' Contributions
JHZ designed the study, extracted and analyzed the data, and wrote the paper as the first author. SP contributed to the analysis
of the results in a statistical aspect. JMH verified the analytical methods. RX and LXL investigated and supervised the findings
of this work and helped in the language edit. JJH and NY assisted in the support of clinical knowledge and reviewed the paper.
JAW and XH contributed to the data collection of the external validation and reviewed the paper. GYM was in charge of the
overall direction of the study as the corresponding author. All authors gave final approval of the paper for submission. We did
not use generative artificial intelligence in any portion of the manuscript writing.
Conflicts of Interest
None declared.
Multimedia Appendix 1
Definitions for the factors examined in the models.
[DOCX File , 15 KB-Multimedia Appendix 1]
Multimedia Appendix 2
Detailed comparison of general information between the train set and the test set.
[DOCX File , 25 KB-Multimedia Appendix 2]
References
1. Johansson E, Hammarskjöld F, Lundberg D, Arnlind MH. Advantages and disadvantages of peripherally inserted central
venous catheters (PICC) compared to other central venous lines: a systematic review of the literature. Acta Oncol 2013
Jun;52(5):886-892 [doi: 10.3109/0284186X.2013.773072] [Medline: 23472835]
2. Glauser F, Kivrak S, Righini M. [Peripherally inserted central catheters : indications, contraindications, complications].
Rev Med Suisse 2018 Dec 05;14(630):2211-2213 [Medline: 30516889]
3. Sakai T, Kohda K, Konuma Y, Hiraoka Y, Ichikawa Y, Ono K, et al. A role for peripherally inserted central venous catheters
in the prevention of catheter-related blood stream infections in patients with hematological malignancies. Int J Hematol
2014 Dec 24;100(6):592-598 [doi: 10.1007/s12185-014-1677-9] [Medline: 25248408]
4. Silva PSLD, Reis ME, Aguiar VE, Fonseca MCM. Unplanned extubation in the neonatal ICU: a systematic review, critical
appraisal, and evidence-based recommendations. Respir Care 2013 Jul 27;58(7):1237-1245 [FREE Full text] [doi:
10.4187/respcare.02164] [Medline: 23271815]
5. Gao W, Luan X, Sun Y, Zhang M, Li K, Li QH, et al. Experiences of patients with abnormal extubation of PICC tubes: a
qualitative study. Int J Clin Exp Med 2015;8(10):19297-19303 [Medline: 26770567]
J Med Internet Res 2023 | vol. 25 | e49016 | p. 15https://www.jmir.org/2023/1/e49016 (page number not for citation purposes)
Zhang et alJOURNAL OF MEDICAL INTERNET RESEARCH
XSL
•
FO
RenderX
6. Chan RJ, Northfield S, Larsen E, Mihala G, Ullman A, Hancock P, et al. Central venous Access device SeCurement And
Dressing Effectiveness for peripherally inserted central catheters in adult acute hospital patients (CASCADE): a pilot
randomised controlled trial. Trials 2017 Oct 04;18(1):458 [FREE Full text] [doi: 10.1186/s13063-017-2207-x] [Medline:
28978332]
7. Kang JR, Long LH, Yan SW, Wei WW, Jun HZ, Chen W. Peripherally Inserted Central Catheter-Related Vein Thrombosis
in Patients With Lung Cancer. Clin Appl Thromb Hemost 2017 Mar 09;23(2):181-186 [FREE Full text] [doi:
10.1177/1076029615595880] [Medline: 26207019]
8. Fallouh N, McGuirk HM, Flanders SA, Chopra V. Peripherally Inserted Central Catheter-associated Deep Vein Thrombosis:
A Narrative Review. Am J Med 2015 Jul;128(7):722-738 [doi: 10.1016/j.amjmed.2015.01.027] [Medline: 25697969]
9. Pan L, Zhao Q, Yang X. Risk factors for venous thrombosis associated with peripherally inserted central venous catheters.
Int J Clin Exp Med 2014;7(12):5814-5819 [FREE Full text] [Medline: 25664112]
10. Clarke DJ, Hawkins R, Sadler E, Harding G, McKevitt C, Godfrey M, et al. Introducing structured caregiver training in
stroke care: findings from the TRACS process evaluation study. BMJ Open 2014 Apr 15;4(4):e004473 [FREE Full text]
[doi: 10.1136/bmjopen-2013-004473] [Medline: 24736035]
11. Brewer C. Reducing upper extremity deep vein thrombosis when inserting PICCs. Br J Nurs 2012 Jul 25;21(Sup14):S12-S17
[doi: 10.12968/bjon.2012.21.sup14.s12]
12. Xiao W, Lin Q, Chen S, Li S, Lin C, Su S, et al. Catheterization of PICC through a superficial femoral vein for patients
with superior vena cava syndrome using ECG positioning and ultrasound-guided technologies. J Vasc Access 2023 May
27;24(3):397-401 [doi: 10.1177/11297298211035331] [Medline: 34313170]
13. Westergaard B, Classen V, Walther-Larsen S. Peripherally inserted central catheters in infants and children - indications,
techniques, complications and clinical recommendations. Acta Anaesthesiol Scand 2013 Mar 17;57(3):278-287 [doi:
10.1111/aas.12024] [Medline: 23252685]
14. Bertoglio S, Faccini B, Lalli L, Cafiero F, Bruzzi P. Peripherally inserted central catheters (PICCs) in cancer patients under
chemotherapy: A prospective study on the incidence of complications and overall failures. J Surg Oncol 2016 May
29;113(6):708-714 [doi: 10.1002/jso.24220] [Medline: 27020965]
15. Liem TK, Yanit KE, Moseley SE, Landry GJ, Deloughery TG, Rumwell CA, et al. Peripherally inserted central catheter
usage patterns and associated symptomatic upper extremity venous thrombosis. J Vasc Surg 2012 Mar;55(3):761-767
[FREE Full text] [doi: 10.1016/j.jvs.2011.10.005] [Medline: 22370026]
16. Xie J, Xu L, Xu X, Huang Y. Complications of peripherally inserted central catheters in advanced cancer patients undergoing
combined radiotherapy and chemotherapy. J Clin Nurs 2017 Dec 20;26(23-24):4726-4733 [doi: 10.1111/jocn.13825]
[Medline: 28334494]
17. Johansson E, Hammarskjöld F, Lundberg D, Arnlind MH. Advantages and disadvantages of peripherally inserted central
venous catheters (PICC) compared to other central venous lines: A systematic review of the literature. Acta Oncologica
2013 Mar 11;52(5):886-892 [doi: 10.3109/0284186x.2013.773072]
18. Lv S, Liu Y, Wei G, Shi X, Chen S, Zhang X. The anticoagulants rivaroxaban and low molecular weight heparin prevent
PICC-related upper extremity venous thrombosis in cancer patients. Medicine (Baltimore) 2019 Nov;98(47):e17894 [FREE
Full text] [doi: 10.1097/MD.0000000000017894] [Medline: 31764785]
19. Shen BH, Mahoney L, Molino J, Mermel LA. Risk factors for early PICC removal: A retrospective study of adult inpatients
at an academic medical center. PLoS One 2022 Jul 8;17(7):e0264245 [FREE Full text] [doi: 10.1371/journal.pone.0264245]
[Medline: 35802673]
20. Lee JY, Park H, Chung E. Use of electronic critical care flow sheet data to predict unplanned extubation in ICUs. Int J Med
Inform 2018 Sep;117:6-12 [doi: 10.1016/j.ijmedinf.2018.05.011] [Medline: 30032966]
21. Hur S, Min JY, Yoo J, Kim K, Chung CR, Dykes PC, et al. Development and Validation of Unplanned Extubation Prediction
Models Using Intensive Care Unit Data: Retrospective, Comparative, Machine Learning Study. J Med Internet Res 2021
Aug 11;23(8):e23508 [FREE Full text] [doi: 10.2196/23508] [Medline: 34382940]
22. Kermany DS, Goldbaum M, Cai W, Valentim CC, Liang H, Baxter SL, et al. Identifying Medical Diagnoses and Treatable
Diseases by Image-Based Deep Learning. Cell 2018 Feb 22;172(5):1122-1131.e9 [FREE Full text] [doi:
10.1016/j.cell.2018.02.010] [Medline: 29474911]
23. Badheka AV, Hodge D, Ramesh S, Bloxham J, Espinoza E, Allareddy V, et al. Catheter related thrombosis in hospitalized
infants: A neural network approach to predict risk factors. Thromb Res 2021 Apr;200:34-40 [doi:
10.1016/j.thromres.2021.01.009] [Medline: 33529871]
24. Liu S, Zhang F, Xie L, Wang Y, Xiang Q, Yue Z, et al. Machine learning approaches for risk assessment of peripherally
inserted Central catheter-related vein thrombosis in hospitalized patients with cancer. Int J Med Inform 2019 Sep;129:175-183
[doi: 10.1016/j.ijmedinf.2019.06.001] [Medline: 31445252]
25. Luo W, Phung D, Tran T, Gupta S, Rana S, Karmakar C, et al. Guidelines for Developing and Reporting Machine Learning
Predictive Models in Biomedical Research: A Multidisciplinary View. J Med Internet Res 2016 Dec 16;18(12):e323 [FREE
Full text] [doi: 10.2196/jmir.5870] [Medline: 27986644]
26. Gupta H. Assessing organizations performance on the basis of GHRM practices using BWM and Fuzzy TOPSIS. J Environ
Manage 2018 Nov 15;226:201-216 [doi: 10.1016/j.jenvman.2018.08.005] [Medline: 30119045]
J Med Internet Res 2023 | vol. 25 | e49016 | p. 16https://www.jmir.org/2023/1/e49016 (page number not for citation purposes)
Zhang et alJOURNAL OF MEDICAL INTERNET RESEARCH
XSL
•
FO
RenderX
27. Peng S, Wei T, Li X, Yuan Z, Lin Q. A model to assess the risk of peripherally inserted central venous catheter-related
thrombosis in patients with breast cancer: a retrospective cohort study. Support Care Cancer 2022 Feb 25;30(2):1127-1137
[doi: 10.1007/s00520-021-06511-3] [Medline: 34432170]
28. Collins GS, Reitsma JB, Altman DG, Moons KGM. Transparent reporting of a multivariable prediction model for individual
prognosis or diagnosis (TRIPOD): the TRIPOD statement. BMJ 2015 Jan 07;350(jan07 4):g7594 [FREE Full text] [doi:
10.1136/bmj.g7594] [Medline: 25569120]
29. Kang J, Chen W, Sun W, Ge R, Li H, Ma E, et al. Peripherally inserted central catheter-related complications in cancer
patients: a prospective study of over 50,000 catheter days. J Vasc Access 2017 Mar 21;18(2):153-157 [doi:
10.5301/jva.5000670] [Medline: 28218366]
30. Duwadi S, Zhao Q, Budal BS. Peripherally inserted central catheters in critically ill patients - complications and its prevention:
a review. Int J Nurs Sci 2019 Jan 10;6(1):99-105 [FREE Full text] [doi: 10.1016/j.ijnss.2018.12.007] [Medline: 31406874]
31. Al-Asadi O, Almusarhed M, Eldeeb H. Predictive risk factors of venous thromboembolism (VTE) associated with peripherally
inserted central catheters (PICC) in ambulant solid cancer patients: retrospective single Centre cohort study. Thromb J 2019
Jan 25;17(1):2 [FREE Full text] [doi: 10.1186/s12959-019-0191-y] [Medline: 30697126]
32. Yi X, Chen J, Li J, Feng L, Wang Y, Zhu J, et al. Risk factors associated with PICC-related upper extremity venous
thrombosis in cancer patients. J Clin Nurs 2014 Mar 28;23(5-6):837-843 [doi: 10.1111/jocn.12227] [Medline: 23710585]
33. Leung A, Heal C, Perera M, Pretorius C. A systematic review of patient-related risk factors for catheter-related thrombosis.
J Thromb Thrombolysis 2015 Oct 14;40(3):363-373 [doi: 10.1007/s11239-015-1175-9] [Medline: 25680892]
34. Shimizu Y, Kamada H, Sakane M, Aikawa S, Mutsuzaki H, Tanaka K, et al. A novel exercise device for venous
thromboembolism prophylaxis improves venous flow in bed versus ankle movement exercises in healthy volunteers. J
Orthop Surg (Hong Kong) 2017 Nov 14;25(3):2309499017739477 [FREE Full text] [doi: 10.1177/2309499017739477]
[Medline: 29137566]
35. Chen P, Wan G, Zhu B. Incidence and risk factors of symptomatic thrombosis related to peripherally inserted central catheter
in patients with lung cancer. J Adv Nurs 2021 Mar 29;77(3):1284-1292 [doi: 10.1111/jan.14666] [Medline: 33249623]
36. Bertoglio S, Cafiero F, Meszaros P, Varaldo E, Blondeaux E, Molinelli C, et al. PICC-PORT totally implantable vascular
access device in breast cancer patients undergoing chemotherapy. J Vasc Access 2020 Jul 01;21(4):460-466 [doi:
10.1177/1129729819884482] [Medline: 31674857]
37. Bullock AF, Greenley SL, McKenzie GAG, Paton LW, Johnson MJ. Relationship between markers of malnutrition and
clinical outcomes in older adults with cancer: systematic review, narrative synthesis and meta-analysis. Eur J Clin Nutr
2020 Nov 04;74(11):1519-1535 [FREE Full text] [doi: 10.1038/s41430-020-0629-0] [Medline: 32366995]
38. D'Arrigo S, Annetta MG, Musarò A, Distefano M, Pittiruti M. Secondary malposition of a PICC-port due to heavy physical
exercise: A case report. J Vasc Access 2023 May 16;24(3):507-510 [doi: 10.1177/11297298211039445] [Medline: 34399639]
39. Mahajan A, Brunson A, White R, Wun T. The Epidemiology of Cancer-Associated Venous Thromboembolism: An Update.
Semin Thromb Hemost 2019 Jun 30;45(4):321-325 [doi: 10.1055/s-0039-1688494] [Medline: 31041801]
40. Kang J, Sun W, Li H, Ma E, Wang K, Chen W. Peripherally inserted central catheter-related vein thrombosis in breast
cancer patients. J Vasc Access 2016 Apr 09;17(1):67-71 [doi: 10.5301/jva.5000457] [Medline: 26349890]
41. Hou J, Zhang J, Ma M, Gong Z, Xu B, Shi Z. Thrombotic risk factors in patients with superior vena cava syndrome
undergoing chemotherapy via femoral inserted central catheter. Thromb Res 2019 Dec;184:38-43 [doi:
10.1016/j.thromres.2019.10.030] [Medline: 31706066]
42. Wei B, Feng Y, Chen W, Ren D, Xiao D, Chen B. Risk factors for emergence agitation in adults after general anesthesia:
A systematic review and meta-analysis. Acta Anaesthesiol Scand 2021 Jul 07;65(6):719-729 [doi: 10.1111/aas.13774]
[Medline: 33370461]
43. Hao N, Xie X, Zhou Z, Li J, Kang L, Wu H, et al. Nomogram predicted risk of peripherally inserted central catheter related
thrombosis. Sci Rep 2017 Jul 24;7(1):6344 [FREE Full text] [doi: 10.1038/s41598-017-06609-x] [Medline: 28740162]
44. Meng Z, Wang M, Zhao Z, Zhou Y, Wu Y, Guo S, et al. Development and Validation of a Predictive Model for Severe
COVID-19: A Case-Control Study in China. Front Med (Lausanne) 2021 May 25;8:663145 [FREE Full text] [doi:
10.3389/fmed.2021.663145] [Medline: 34113636]
45. Feng Y, Yan X. Support Vector Machine Based Lane-Changing Behavior Recognition and Lateral Trajectory Prediction.
Comput Intell Neurosci 2022 May 10;2022:3632333-3632339 [FREE Full text] [doi: 10.1155/2022/3632333] [Medline:
35592714]
Abbreviations
AUC: area under the curve
Ci: Composite Index
DCA: decision curve analysis
LASSO: least absolute shrinkage and selection operator
LR: logistic regression
OR: odds ratio
J Med Internet Res 2023 | vol. 25 | e49016 | p. 17https://www.jmir.org/2023/1/e49016 (page number not for citation purposes)
Zhang et alJOURNAL OF MEDICAL INTERNET RESEARCH
XSL
•
FO
RenderX
PICC: peripherally inserted central catheter
RF: random forest
ROC: receiver operating characteristic
SIS: Safe Infusion System
SVM: support vector machine
TOPSIS: Technique for Order Preference by Similarity to Ideal Solution
UE: unplanned extubation
Edited by T de Azevedo Cardoso; submitted 15.05.23; peer-reviewed by L Guo, K Gupta; comments to author 10.08.23; revised version
received 24.09.23; accepted 30.10.23; published 16.11.23
Please cite as:
Zhang J, Ma G, Peng S, Hou J, Xu R, Luo L, Hu J, Yao N, Wang J, Huang X
Risk Factors and Predictive Models for Peripherally Inserted Central Catheter Unplanned Extubation in Patients With Cancer:
Prospective, Machine Learning Study
J Med Internet Res 2023;25:e49016
URL: https://www.jmir.org/2023/1/e49016
doi: 10.2196/49016
PMID:
©Jinghui Zhang, Guiyuan Ma, Sha Peng, Jianmei Hou, Ran Xu, Lingxia Luo, Jiaji Hu, Nian Yao, Jiaan Wang, Xin Huang.
Originally published in the Journal of Medical Internet Research (https://www.jmir.org), 16.11.2023. This is an open-access
article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/),
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in the
Journal of Medical Internet Research, is properly cited. The complete bibliographic information, a link to the original publication
on https://www.jmir.org/, as well as this copyright and license information must be included.
J Med Internet Res 2023 | vol. 25 | e49016 | p. 18https://www.jmir.org/2023/1/e49016 (page number not for citation purposes)
Zhang et alJOURNAL OF MEDICAL INTERNET RESEARCH
XSL
•
FO
RenderX
