Key words
buried bumper syndrome - percutaneous endoscopic gastrostomy - ultrasound - artificial
intelligence
Introduction
Percutaneous endoscopic gastrostomy (PEG) is a well-established and safe long-term
feeding method for pediatric patients suffering from malnutrition and inability to
swallow [1]
[2]
[3]. A PEG consists of a polyurethane probe that is inserted through the abdominal wall
into the gastric lumen providing direct access for enteral nutrition. The probe is
held in position by an external fixation plate and an inner holding plate, adjacent
to the inner gastric wall. Besides acute procedural complications, a major long-term
complication is the development of a buried bumper syndrome (BBS) [4]. BBS describes the overgrowth of gastric mucosa and penetration of the inner holding
plate of the PEG into the gastric wall or beyond [1]
[5]. With a prevalence of 2.0–2.9 % in adults [1]
[6]
[7], BBS can cause peritonitis, heavy bleeding, and even gastric perforation [8]
[9]. Therefore, timely diagnosis and rapid endoscopic or surgical PEG replacement are
necessary [8]
[10]. The current recommendations for the diagnosis of BBS include contrast studies via
the PEG or upper endoscopy [11]. Computed tomography (CT) may provide additional detailed imaging information [12]
[13], but the risk of ionizing radiation needs to be considered for pediatric patients.
As an alternative approach, endosonography (EUS) represents a radiation-free, but
invasive [6]
[14] modality, which is not yet accessible for pediatric patients in routine diagnostics
[15]
[16]. Conventional B-mode ultrasound is commonly used in various clinical scenarios as
the standard of care in pediatric medicine. High-resolution US imaging is capable
of differentiating the individual wall layers of the stomach and bowel segments [17]. While the identification of the mucosa (hypo-), submucosa (hyper-), and muscularis
propria (hypoechoic) has already been helpful in the assessment of bowel diseases
[17]
[18]
[19]
[20], it may also allow the exact location of the inner holding plate in BBS [10]. However, imaging approaches are time-consuming [21], while the clinical routine is characterized by time and medical staff shortage.
Therefore, AI-supported methods, which offer the possibility to aid clinical symptom-based
decision making in real time, are becoming increasingly important. In adults, AI implementation
for diagnostic support in gastroenterological diseases is expanding [22]
[23] and the trend to integrate these systems in pediatric routine care is imminent [24]
[25]. The primary objective of this retrospective study was to evaluate the feasibility
of transabdominal ultrasound to visualize the inner holding plate of a PEG and its
relation to the gastric wall for the diagnosis of BBS in children. Furthermore, we
compared the value of transabdominal US for the diagnosis of BBS to a novel symptom-based
AI approach.
Materials and Methods
Study design
This single-center retrospective study was performed in accordance with the declaration
of Helsinki and approved by the local ethics committee (No. 110_20 Bc).
To identify relevant cases, the hospital data warehouse was screened for ultrasound
reports containing the keywords “BBS” or “PEG tube” in the years between 2009–2019.
Regardless of age, gender or primary disease, all US reports of children with PEG
or PEG with jejunal extension (PEG-J) and description of positioning of the inner
holding plate (bumper), were included in the study.
The exclusion criteria were the lack of time-related (within 4 weeks) endoscopic or
CT-guided validation in the case of US reports indicating the presence of a BBS (BBS+),
and patients with a PEG button device or US reports where the same diagnoses was repeatedly
given within a 4-week interval.
Transabdominal US acted as a binary classifier (US positive = diagnosis BBS+ = inner
holding plate not luminal to the gastric mucosa/US negative = diagnosis BBS- = inner
holding plate luminal to the gastric mucosa) dividing the study population into two
groups. The accuracy of the group assignment to BBS+ by ultrasonography was subsequently
verified by the gold standard of endoscopic or CT control. BBS- was proven by negative
endoscopic assessment within 4 weeks or assumed if the inner holding plate was clearly
visible on US without sonographic and clinical signs or evidence of a BBS and by the
absence of any clinical evidence for BBS+ within 4 weeks leading to further diagnostics.
Data acquisition
Relevant patient data were extracted from electronic patient records or written/printed
documents. Data collection included demographic characteristics (e. g., sex, age,
and BMI), patient history (e. g., primary disease and gastroenterological characteristics),
and laboratory parameters (e. g., hemoglobin level and c-reactive protein (CrP)).
The time interval between laboratory results, clinical symptoms, medications, and
the US examination should not exceed seven days in order to be considered.
Ultrasound acquisition
US examinations were performed by n = 11 pediatricians with varying degrees of US
qualification. Images were acquired with seven different US systems (six high-end
US devices and one mid-range device) and probes with frequencies ranging from 5 to
17 Megahertz (MHz). The US probe was positioned at the area of the PEG entry point
(for further details please see Supplementary Table 1).
Statistical Analysis
Categorical variables are given as number and percentage. Metric variables are given
as mean and standard deviation. Non-parametric Kolmogorov-Smirnov and Mann-Whitney-U
test were used for statistical evaluation. Precision rate, recall rate, and f1 score
(harmonic mean between precision and recall) were calculated in order to predict the
accuracy of ultrasound as a diagnostic tool. A correlation matrix was created in order
to describe the strength of the correlation between the collected data. Spearman’s
coefficient (rs) was applied. Missing data points were excluded from the final analysis. P-values
< 0.05 were considered statistically significant. All statistical analyses were performed
by the Python programming language (version 3.7 released in June 2018 by Python Software
Foundation Wilmington, Delaware, United States) and its corresponding libraries (SciPy,
Pandas, Seaborn, Matplotlib and Scikit-learn).
Artificial intelligence (AI)-based model for BBS prediction
Logistic regression, a supervised learning classification AI algorithm, was used.
The aim was to extract patterns out of raw data and compute, based on the clinical
parameter data sets and their occurrences, an automated, accurate AI model capable
of predicting and identifying children with BBS (+). The initial data set was randomly
divided into two sub-data sets SDS1 (clinical data for 2/3 of the patients) and SDS2
(clinical data for 1/3 of the patients). The SDS1 set was used to perform the training
of the AI model and computed the learning process. The SDS2 set was used to verify
the results and test the accuracy of automated BBS (+) prediction. Subsequently, the
whole data set (SDS1 +SDS2) was subjected to a second analysis of the AI model. Precision
rate, recall rate, and f1 score for the developed AI model were calculated and compared
to the findings by ultrasonography.
Results
Patient characteristics
Between 2009 and 2019, n = 82 biologically independent pediatric patients with US
of the PEG/PEG-J were identified according to the study inclusion and exclusion criteria.
The patient cohort included n = 40 female (48.8 %) and n = 42 (51.2 %) male children
with a mean age ± standard deviation (SD) of 5.9 ± 5.6 years at first presentation
([Table 1]). Relevant comorbidities included neurological (n = 42, 51.2 %) and metabolic, cardiological,
nephrological, gastrological and oncological disorders (range 6.1–13.4 %, [Table 1]).
Table 1
Demographic and clinical patient characteristics.
|
n = 82 patients
n (%)
|
mean ± SD
|
|
sex (female, n)
|
40 (48.8)
|
|
|
age (years)
|
82 (100)
|
5.9 ± 5.6
|
|
weight (kg)
|
70 (85.4)
|
17.4 ± 12.0
|
|
height (cm)
|
62 (75.6)
|
97.3 ± 27.8
|
|
BMI (kg/m²)
|
61 (74.4)
|
16.4 ± 3.1
|
|
primary disease
|
|
neurological
|
42 (51.2)
|
|
|
cardiological
|
11 (13.4)
|
|
metabolic
|
9 (11.0)
|
|
gastroenterological
|
7 (8.5)
|
|
oncological
|
6 (7.3)
|
|
nephrological
|
5 (6.1)
|
|
Ø documentation
|
2 (2.4)
|
n = 82 independent pediatric patients. Categorical variables are given as n (%), continuous
variables as mean ± standard deviation (SD).
Distribution of ultrasound examinations among the patient cohort
In total n = 82 pediatric patients underwent n = 124 transabdominal US examinations.
Of those, n = 3 cases with sonographic diagnosis of a BBS+ were excluded due to missing
endoscopic or CT control after US, resulting in n = 121 independent sonographic cases
for further analysis. Of n = 82 patients, n = 55 patients underwent a single sonographic
examination and n = 27 patients had multiple examinations performed within 4 weeks
to 6 years. The US equipment that was used (devices, probes, and image settings) is
summarized in Supplementary Table 1.
Ultrasonographic examinations, imaging precision, and recall rate
The position of the inner holding plate was determined in n = 121 independent sonographic
cases (please see exemplary images in [Fig. 1]). In n = 101 cases the inner holding plate was correctly positioned in the child's
stomach (intragastral, [Fig. 1a]), whereas in n = 2 cases the inner holding plate was not in its original gastric
position but had slipped into the jejunum (intrajejunal). Thus, in n = 103 cases the
inner holding plate had not grown into the mucous membrane and a BBS was sonographically
excluded (BBS-), which was verified in n = 20 cases by endoscopy and in n = 83 cases
based on the further clinical course. In n = 18 cases a BBS was sonographically diagnosed
(n = 16 intragastral and n = 2 extragastral) with n = 17 being endoscopically ([Fig. 1b, c]) confirmed and n = 1 being confirmed by CT ([Fig. 1 d]). Therefore, false positives and false negatives were n = 0, true positives were
n = 18 and true negatives were n = 103. The recall and precision rates were 100 %
and an f1 score of 1 was achieved. In addition, the expertise of the investigator
(pediatricians trained in US and/or DEGUM (Deutsche Gesellschaft für Ultraschall in
der Medizin) level I–III-certified pediatricians) did not appear to affect the accuracy
of BBS diagnosis (Supplementary Table 2).
Fig. 1 Different cases of BBS (+) and BBS (–) are presented as a schematic cartoon, B-mode
ultrasound image, and the respective endoscopy or CT image. green = mucosa = hypoechoic,
yellow = submucosa = hyperechoic, red = muscularis propria = hypoechoic. a normal PEG bumper (intragastral): the inner holding plate of PEG, with the white
double contour in the US image, is located in a correct intragastral position, with
the border of the gastric wall towards the outside. In the reference method (endoscopy)
the PEG bumper is endoluminally visible. b, c buried bumper (intragastral): the inner holding plate of PEG, with the white double
contour, is overgrown by hypoechogenic or inhomogeneous hyper- to hypoechogenic mucosa
and eroded into the stomach wall. In the reference method (endoscopy) the PEG bumper
is endoluminally partially visible. d buried bumper (extragastral): the inner holding plate of PEG, with the white double
contour, is outside the gastric wall, surrounded by hyperechogenic tissue. In the
reference method (CT) the PEG bumper is outside the stomach (marked with a star).
Clinical and gastroenterological findings of n = 121 independent sonographic cases
Considering n = 121 independent US examinations, in n = 92 cases a PEG tube, in n = 20
cases a PEG-J tube and in n = 9 cases no sufficient documentation was found (for details
please see Supplementary Table 3). Documented complications during insertion were post-operative bleeding, abrasions
of mucous membranes, and small rupture of the liver capsule. In n = 69 cases the children
were treated with proton pump inhibitors (PPI) ([Table 2]).
Table 2
Laboratory values and gastroenterological characteristics.
|
laboratory values
|
|
cases
n
|
BBS(–)
n
|
BBS(+)
n
|
BBS (–)
mean±SD
|
BBS (+)
mean±SD
|
p-value
|
|
Hb (mmol/L)
|
103
|
88
|
15
|
7.4 ± 1.6
|
7.7 ± 1.2
|
0.48
|
|
CRP (mg/L)
|
101
|
87
|
14
|
39.4 ± 70.3
|
18.6 ± 34.6
|
0.28
|
|
WBC (10³/µl)
|
103
|
88
|
15
|
11.4 ± 6.9
|
27.5 ± 63.9
|
0.027
|
|
PLT (10³/µl)
|
103
|
88
|
15
|
272.4 ± 162.0
|
269.7 ± 150.7
|
0.95
|
|
gastroenterological characteristics and medication
|
|
time interval between 1st PEG placement and US (years)
|
113
|
96
|
17
|
1.9 ± 3.0
|
5.6 ± 4.1
|
< 0.001
|
|
n = 121 cases
with multiple items possible
|
n = 121
n (%)
|
n = 103
BBS (–)
cases
n (%)
|
n = 18
BBS (+)
cases
n (%)
|
|
|
abdominal pain
|
32 (26.4)
|
28 (27.2)
|
4 (22.2)
|
|
mobilization problems
|
31 (25.6)
|
14 (13.6)
|
17 (94.4)
|
|
inflammation signs
|
35 (28.9)
|
27 (26.2)
|
8 (44.4)
|
|
secretion/exudation
|
22 (18.2)
|
14 (13.6)
|
8 (44.4)
|
|
pus discharge
|
10 (8.3)
|
8 (7.8)
|
2 (11.1)
|
|
vomiting
|
30 (24.8)
|
28 (27.2)
|
2 (11.1)
|
|
leakage of tube feed
|
8 (6.6)
|
7 (6.8)
|
1 (5.6)
|
|
Ø documentation
|
22 (18.2)
|
22 (21.4)
|
0
|
|
proton-pump inhibitors
|
|
Omeprazole
|
60 (49.6)
|
50 (48.5)
|
10 (55.6)
|
|
Esomeprazole
|
9 (7.4)
|
8 (7.8)
|
1 (5.6)
|
|
no medication
|
32 (26.4)
|
28 (27.2)
|
4 (22.2)
|
|
Ø documentation
|
20 (16.5)
|
17 (16.5)
|
3 (16.7)
|
n = 121 independent ultrasound cases. For gastroenterological symptoms, multiple entry
of items was possible. Categorical variables are presented as n (%), continuous variables
are given as mean ± SD. P-values < 0.05 were considered statistically significant.
Ø Documentation: cases in which the requested data was too old (> 7 days older than
the ultrasound examination) or not documented. Hb = hemoglobin, in mmol/L, CRP = C-reactive
protein, in mg/L, WBC = white blood cells in 10³/ul, PLT = platelets/thrombocytes
in 10³/ul.
While in n = 16 BBS+ cases the inner holding plate had grown into the stomach wall,
n = 2 BBS+ cases required surgical intervention due to extragastral positioning of
the plate and/or gastric perforation. General complaints of the patient cohort were
general/localized abdominal pain (n = 32, 26.4 %), mobilization problems of the PEG
tube (n = 31, 25.6 %), inflammation signs (redness, overheating or swelling) (n = 35,
28.9 %), secretion/exudation (n = 22, 18.2 %), pus discharge (n = 10, 8.3 %), vomiting (n = 30,
24.8 %) and/or leakage of tube feed (n = 8, 6.6 %) ([Table 2]).
In the BBS+ group (n = 18) the mobilization problem was the most common symptom (n = 17,
94.4 %). The main symptoms in the BBS- group (n = 103) were vomiting (n = 28, 27.2 %)
and abdominal pain (n = 28, 27.2 %). The BBS+ group showed a significantly longer
time interval between 1st PEG/PEG-J placement and US than the BBS- group (5.6 ± 4.1 years vs. 1.9 ± 3.0 years,
p-value < 0.001). Within laboratory parameters only leucocytes showed a statistically
significant difference between groups (27.5 ± 63.9 × 103/µl vs. 11.4 ± 6.9 × 103/µl, p = 0.027) ([Table 2]).
Correlation between BBS and clinical characteristics
To reveal potential correlations between BBS+, PEG-associated parameters, and clinical
characteristics, a correlation matrix was created. A significantly positive correlation
of the occurrence of BBS+ and mobilization problems of PEG (rs = 0.66, p-value < 0.001), secretion/exudation of the PEG (rs = 0.29, p = 0.002), longer time interval between 1st PEG placement and US examination (rs = 0.38, p < 0.001), and white blood cell count (rs = 0.24, p = 0.016) was observed. No statistically significant negative correlation
between the occurrence of BBS+ and PEG-associated parameters and clinical symptoms
was found ([Fig. 2]).
Fig. 2 Shown is the correlation matrix of the collected study parameters. Numbers are Spearman
correlation coefficients (rs). Fields in orange show a statistically significant positive correlation and fields
in blue show a statistically significant negative correlation between the parameters.
P-values < 0.05 were considered statistically significant. Hb = hemoglobin, in mmol/L,
CRP = C-reactive protein, in mg/L, WBC = white blood cells in 10³/ul, PLT = platelets/thrombocytes
in 10³/ul.
Precision and recall rate of AI-based model in comparison to US
In order to assess precision, recall rate, and f1 score for the AI-based model, the
same clinical characteristics as described above in the correlation matrix were weighted
according to the logistic regression algorithm for the diagnosis of BBS (+). The two
most important identifying characteristics were mobilization problems of the PEG (logistic
regression coefficient value r = 5.36) and time between 1st PEG placement and US examination (logistic regression coefficient value = 3.03) (for
individual rating please see Supplementary Table 4). Applying the AI prediction model to the SDS2 data set (clinical data for 1/3 of
patients) 3 out of 5 BBS (+) children were predicted ([Fig. 3]) and a precision, recall rate, and f1 score of 0.6 were achieved. Applying the same
model to the integral data set (SDS1 +SDS2), 16 of 18 BBS (+) children were predicted
([Fig. 3]). The AI model achieved a precision, recall rate, and f1 score of 0.88, which turned
out to be inferior to the US examination with a precision, recall rate, and f1 score
of 1. The confusion matrices (AI-based model vs. US) and the experimental setup for
creating the AI model are shown in [Fig. 3].
Fig. 3 Shown is the comparison between transabdominal ultrasound (US) and the artificial
intelligence (AI) model to detect BBS (+). Left, the experimental setup and the results
of the AI model (presented by confusion matrices) are shown. First, the clinical data
for 2/3 of the patients (SDS1 set) was used as a training set to create an AI model
capable of detecting BBS. The clinical data for the remaining 1/3 of the patients
(SDS2 set) was randomly extracted and used for the first run of the AI model. 3 of
5 BBS cases were diagnosed correctly by the AI model (recall and precision rate of
0.66). Using the integral data set (SDS1 + SDS2), the AI model correctly identified
16 of 18 BBS cases (precision and recall rate of 0.88). In comparison, transabdominal
ultrasound correctly diagnosed all 18 cases of BBS (precision and recall rate of 1.0).
Discussion
This retrospective single-center study provides systematically collected pediatric
data demonstrating that transabdominal US is suitable to exactly assess the intra-
or extra-gastral position of the inner PEG holding plate for the diagnosis of BBS
in children.
To date, endoscopy is used as the reference standard for the diagnostic and therapeutic
workup in most cases of suspected BBS [10]
[11]
[26]. However, invasive endoscopic procedures are risky in children and multimorbid patients
with PEG tubes [7] – especially when considering the need for sedation and its potential associated
complications. In addition, the time resources required for preparation and performing
of an endoscopy examination are considerably higher than for US. Moreover, for endoscopic
procedures an experienced pediatric gastroenterologist is needed while US is widely
accessible and commonly performed by pediatricians and radiologists. In this study,
all investigators, regardless of their professional training level, were able to correctly
identify the position of the inner PEG holding plate. However, cases with extragastral
positioning of the PEG holding plate were scarce. Furthermore, in the clinical routine
BBS- is relatively easy to assess by verification of proper mobilization of the inner
holding plate in its intragastral position during dynamic US. To reduce the occurrence
of BBS, it is generally recommended for trained nursing staff to regularly mobilize
and maintain the PEG [8]
[10].
From a clinical perspective, subjects with mobilization problems, increased secretion,
longer time interval between 1st PEG placement and US as well as elevated leucocytes showed the highest positive correlation
with the occurrence of pediatric BBS. Similar to our findings, Blumenstein et al.
[4] ranked mobilization problems or blockage of the PEG tube, peritubular leakage, and
abdominal pain among the most frequent symptoms in patients with BBS. Other studies
considered the insertion of PEG-J tubes and the number of gastrostomies as potential
risk factors for the development of BBS [8]. While white blood cell count correlated with the occurrence of BBS, C-reactive
protein (CRP) was insignificant.
To increase cost-efficiency and save time, symptom-based AI models could aid clinical
decision making [27]
[28]. Although the initial implementation of a structured electronic database is also
time-consuming, the basic concept is to integrate automated algorithms without further
expenditure of time afterwards. Furthermore, as previously shown in pediatric patients
with appendicitis, routine parameters could be used in AI models to improve diagnostics
[29]. Oelen et al. described that deep learning-based algorithms could even be superior
to physicians for measuring hip angles [25]. In this study, clinical diagnosis supported by AI was able to identify BBS cases,
however, to a lower degree than conventional US. The inferior results of the AI approach
might be attributed to the small data set and the general nonspecific clinical appearance
of pediatric BBS. However, the AI approach could provide aid during differential diagnostic
considerations and initiation of appropriate US diagnostics with subsequent specific
therapy as far as clinical data sets are digitally explorable. A sequence of AI supported
clinical diagnostics with point-of-care ultrasound (POCUS) [30]
[31] could eventually display the right balance for future settings. While the results
revealed perfect diagnostic accuracy for conventional US and good potential for AI-based
decision support, the results are potentially limited by the lack of endoscopic validation
in BBS- cases. Therefore, especially early-stage BBS, classified as type II by Richter-Schrag
[26], could be missed. To further increase the diagnostic validity of US for the diagnosis
of BBS, a standard procedure and documentation could be implemented as follows: mobilization
of the PEG in its intragastral position during the examination and documentation of
local findings including complications like abscesses of the abdominal wall and/or
retention. If the inner holding plate is not clearly detectable, the examination should
be performed with an empty stomach followed by installation of fluids (water or tea).
In summary, this is, to the best of our knowledge, the first study evaluating the
diagnostic value of transabdominal US and a clinical parameter-based AI approach for
the diagnosis of BBS in children.
Particularly the routine use of US in cases of suspected BBS enables rapid diagnosis
and a personal encounter and could spare pediatric patients more invasive diagnostics.
It should therefore be integrated in the routine clinical workup of suspected BBS.
Author contributions
A.H. and J.J. designed the study. Data were analyzed by C.A., A.H., and J.J. C.A.
completed data entry. Statistical analysis were performed by D.R.. C.A., A.P.R, and
F.K. wrote the first draft of the manuscript. A.L.W. performed language editing. G.S. contributed
ultrasonic data. H.K. contributed endoscopic data. The manuscript was critically reviewed
by all authors.
CA and APR contributed equally as first authors.
AH and JJ contributed equally as last authors.
Funding
Interdisciplinary Center for Clinical Research (IZKF), Erlangen (CSP Program)
