Keywords
machine learning - artificial intelligence - pediatric surgery - review
Introduction
Artificial intelligence (AI) and machine learning (ML) are rapidly transforming health
care, with applications ranging from diagnostic image interpretation, predictive modeling,
personalized treatment planning, real-time intraoperative guidance, and outcome prediction.
While numerous studies have explored AI applications in adult medicine and pediatric
radiology, the application of AI in general pediatric surgery remains in its early
stages.[1]
[2]
[3]
[4]
[5] This is likely due to the rarity of many pediatric surgical conditions, which often
require highly individualized approaches. Furthermore, the wide variability in patient
age—from preterm infants to adolescents—as well as differences in size, anatomy, physiology,
and disease presentation pose challenges for standardization. AI/ML models typically
require large and relatively homogeneous datasets, making their application in pediatric
surgery for rare diseases even more complex.[6]
[7]
[8]
In pediatric AI research, datasets are often smaller, and less uniform compared with
those in adult populations, limiting both the development and external validation
of AI/ML models. These limitations underscore the need for AI systems that are interpretable,
rigorously validated, unbiased, and specifically tailored to pediatric populations
or, in some cases, even to individual patients.[6]
[9]
Understanding the current level of AI knowledge and usage among pediatric surgeons
is crucial for the effective implementation of these technologies in clinical practice
in the future. Despite a growing interest, most pediatric surgeons have only basic
knowledge of AI/ML and lack formal training. A recent survey amongst members of the
European Society of Pediatric Endoscopic Surgeons found that 65% of respondents rated
their AI/ML knowledge as basic, 86% had no formal training, yet 95% expressed interest
in further education, particularly through workshops and hands-on learning.[10] Similarly, targeted workshops among pediatric surgeons in training have been shown
to significantly improve participants' AI-related knowledge and confidence.[11] Other studies demonstrate that AI-assisted learning modules can enhance diagnostic
accuracy and accelerate the learning curve in the evaluation of pediatric fractures.[12]
[13]
Despite the before-mentioned challenges, the number of AI-related publications in
pediatric surgery is increasing rapidly ([Fig. 1]). To provide the best care, pediatric surgeons must stay informed about these rapidly
evolving technologies.
Fig. 1 Number of publications found in PubMed with the search terms (“artificial intelligence”
OR “machine learning”) AND “pediatric surgery” published between 2014 and 2025 as
of July 10, 2025. AI, artificial intelligence; ML, machine learning.
This narrative review, therefore, aims to summarize the current literature on the
applications of AI and ML in conditions relevant to pediatric surgery. To achieve
this goal, PubMed was screened, and relevant articles were retrieved and summarized.
Additionally, [Table 1] provides an overview of the most commonly used AI and ML models in these studies,
including their underlying principles, advantages, and limitations.
Table 1
Commonly applied artificial intelligence/machine learning models in pediatric surgery
with their core principles, strengths, and limitations
|
Model
|
Published
|
Core principle
|
Strengths
|
Limitations
|
|
k-Nearest Neighbors
|
1952
|
Classifies items based on the k-closest data points
|
Simplicity, no training phase, high explainability
|
Data-sensitive, computationally expensive for large datasets
|
|
Logistic Regression
|
1958
|
Computes class probability using internal weights and a non-linear sigmoid function
|
Simplicity, explainability, and computationally inexpensive
|
Limited performance with complex, non-linear data
|
|
Ridge Regression
|
1970
|
Linear regression with an additional penalty term (L2 regularization) to reduce the
complexity of internal weights
|
Simplicity, explainability, computationally inexpensive, and reduced risk of overfitting
|
Limited performance with complex, non-linear data, necessitates feature scaling, additional
hyperparameters to tune
|
|
Decision Tree
|
1984/86
|
Build a tree of (binary) decisions to successively split data into distinct subgroups
for classification
|
Simplicity, explainability, and interpretability
|
Limited generalization, error-prone on small datasets
|
|
Ensemble Voting Classifier
|
1990s
|
Aggregate predictions from multiple (potentially different) ML models by voting mechanism
|
Robust, high generalization, scalability
|
Complexity, low interpretability, requires additional design decisions
|
|
Support Vector Machine
|
1995
|
Finds optimal decision boundary (“hyperplane”) to separate data points of different
classes by maximizing the distance between the boundary and the next closest point
of each class
|
Simplicity, effectiveness on linearly separable datasets
|
Requires hyperparameter tuning, low scalability
|
|
Random Forest
|
2001
|
Train multiple decision trees on random subsets of data and features to use in an
Ensemble Voting Classifier
|
Robust, good accuracy on large datasets, and interpretability
|
Computationally expensive on large datasets, requires hyperparameter tuning
|
|
Deep Neural Network
|
2012
|
Artificial neural network with multiple hidden layers between the input and output
that learn progressively more complex, non-linear relations in the data
|
High flexibility, high capability of learning complex, non-linear relations
|
Low explainability and interpretability, reliance on large amounts of labeled data,
computationally expensive to train
|
|
U-Net
|
2015
|
Deep neural network employing a symmetric structure with an encoder (input compression
and feature extraction) and decoder (feature expansion and prediction)
|
High flexibility, high capability in particular for image-based applications
|
Low explainability and interpretability, reliance on large amounts of labeled data,
computationally expensive to train
|
|
eXtreme Gradient Boosting
|
2016
|
Sequentially build decision trees that correct predecessors' error in a weighted ensemble
for prediction.
|
High capabilities on structured (e.g., tabular) data, robust against outliers
|
Low explainability and interpretability, reliance on hyperparameter tuning
|
Abbreviation: ML, machine learning.
Appendicitis
Acute appendicitis (AA) is the most common cause for emergency surgery in children
and adolescents.[14] Early diagnosis is crucial as delayed identification increases the risk of perforation
and its associated morbidities.[15] However, correct diagnosis is still challenging, and even scoring systems, including
the Alvarado score and the Pediatric Appendicitis Score (PAS), are limited in their
clinical impact.[16]
[17] No single history, physical examination, laboratory finding, or score can eliminate
the need for imaging studies such as ultrasound or computed tomography.[17] However, ultrasound is user-dependent, therefore, limiting its diagnostic capacity,
and CT, albeit having a high diagnostic accuracy, requires radiation exposure.[18] Taken together, it is not surprising that there is an increasing number of reports
evaluating AI algorithms in the setting of pediatric AA.
By analyzing ultrasound images, ML models can improve diagnostic accuracy and reduce
subjectivity, even in less-experienced users. Several studies have applied ML to ultrasound
imaging to assist in the diagnosis of pediatric appendicitis. Marcinkevics et al demonstrated
how interpretable ML models, that is, concept bottleneck models, based on ultrasound,
can accurately detect AA in children, while also making the algorithm's reasoning
more transparent for physicians.[19]
[20] Moreover, the same group investigated random forests, logistic regression (LR),
and gradient boosting machines to predict diagnosis, management (surgical vs. conservative),
and severity (complicated vs. uncomplicated) using data from 430 pediatric appendicitis
cases (aged 0–18). The random forest model achieved strong performance with area under
the precision–recall curve (AUPR) values of 0.94 for diagnosis, 0.92 for management,
and 0.70 for severity. These findings were used to develop an online Appendicitis
Prediction Tool for children with suspected appendicitis.[20]
Similarly, in another study using several explainable AI techniques on a dataset consisting
of 465 appendicitis cases and 317 non-appendicitis cases, an interpretable and transparent
ML framework for rapid pediatric appendicitis diagnosis proposes a random forest model
that integrated clinical and laboratory data to quick and safe diagnosis. Among others,
the critical variables were appendix on ultrasonography and appendix diameter, supporting
other reports that have shown that AI/ML models can be successfully used on pediatric
ultrasound images in children suspected of appendicitis.[21] In another study, 50 videos and 6,914 images were used to train an AI neural network,
showing that ML-assisted ultrasound evaluation improved sensitivity in diagnosing
appendicitis in children while being transparent for clinical users, noting, however,
that if failing to detect, examiners might be negatively affected.[22]
To account for the more complex nature of pediatric appendicitis, a range of different
AI models have been developed and tested that combine medical history, laboratory
results, and physical examinations.[23]
[24] One key example is a random forest classifier trained on a wide range of clinical
parameters. This model cannot only predict whether a child has AA, but also assess
the severity and suggest the best treatment approach.[25] Other ML models have been developed to reduce unnecessary surgeries in children
with suspected appendicitis. Trained on a dataset from 551 patients, the random forest
model outperformed the appendicitis inflammatory response (AIR) score and achieved
near-perfect sensitivity while potentially avoiding 17% of negative appendectomies.
It also showed strong accuracy to differentiate complicated (gangrenous appendicitis)
from uncomplicated ones (phlegmonous appendicitis), showing great potential for both
diagnosing and treatment guidance.[16] In another study by Erman et al, an ML model was developed to also predict the severity
of pediatric appendicitis across five grades using clinical and operative data from
1,980 children. The most promising model achieved 70.1% accuracy, offering a novel
preoperative method to personalize treatment.[26] These advances suggest that AI models hold great promise in increasing diagnostic
reliability while reducing the need for ionizing imaging methods like CT in children.
Another practical tool, the AI Pediatric Appendicitis Decision-tree (AiPAD), has been
developed based on a model based on clinical and laboratory parameters without imaging
data. AiPAD was highly accurate in predicting appendicitis and may offer a tool to
be used in diagnosing appendicitis in children without the need for imaging. Nevertheless,
the results of this study are based on a small number of patients.[27]
Another innovative use of AI involves natural language processing (NLP). Instead of
manually reviewing operative and ultrasound reports, NLP can now self-label these
free-text documents automatically—making it easier to assess and track the severity
of appendicitis.[25] Such efforts could make it easier to effectively analyze large datasets. Nevertheless,
such AI solutions require further rigorous validation.
In addition, analysis of data derived from wearable technologies might be of value
to predict postoperative complications. A study by Ghomrawi et al analyzed data from
a consumer-grade wearable device recording multimodal data about daily physical activity,
heart rate, and sleep, analyzed with a balanced random forest classifier in a cohort
of children following appendectomy.[28] The ML model accurately detected 83% of abnormal recovery days in complicated appendicitis
and 70% of abnormal recovery days in simple appendicitis prior to the true report
of a symptom/complication. Such results could open new doors for monitoring children
after surgery from the comfort of home and revolutionize hospitalization management.
In summary, AI and ML applications in pediatric appendicitis show promising results,
offering improved diagnostic support, management guidance, and complication prediction
for this common condition.
Necrotizing Enterocolitis
Necrotizing Enterocolitis
Necrotizing enterocolitis (NEC) remains a devastating disease primarily affecting
preterm neonates, with high morbidity and mortality. Early prediction, diagnosis,
and effective management are of utmost importance and demand innovative solutions.[29] In the future, AI may contribute to the early prediction and diagnosis of this lethal
gastrointestinal disease seen in neonatal intensive care units.
Recently, the performance of ML models using both birth characteristics and continuous
vital signs from the first days after birth in the early prediction of NEC in preterm
infants has been examined.[30] Verhoeven et al compared three ML models—including LR, support vector machine (SVM),
and eXtreme gradient boosting (XGBoost)—reporting F1 scores ranging from 0.76 to 0.82
and AUPR values between 0.77 and 0.83. Notably, splanchnic and cerebral oxygenation
were the strongest predictors. This study is especially relevant as a minimum gap
of 48 hours was maintained between the last included data point and the clinical onset
of NEC, emphasizing true prediction rather than diagnosis of early-onset cases.[30] In a nationwide Korean study, Kim et al used 38 variables—including maternal, prenatal,
and postnatal factors obtained within 1 week of birth—to train an ML algorithm to
predict surgical NEC, achieving an AUC of 0.721.[31] ML algorithms have also been applied to stool samples; in one study, longitudinal
stool microbiota profiles from preterm infants at risk for NEC enabled prediction
approximately 8 days before onset.[32]
While NEC can be identifiable using ultrasound examinations,[33] abdominal radiographs remain the diagnostic gold standard.[34] Deep neural networks have been optimized to perform comparably to surgical residents
in identifying pneumatosis on radiographic images.[35] Furthermore, deep learning models can achieve excellent diagnostic performance in
predicting the need for surgical intervention based on abdominal X-rays.[36] Lure et al demonstrated that random forest and ridge logistic regression models
can differentiate NEC from spontaneous intestinal perforation.[37]
To enhance NEC diagnosis on radiographs, Lu and coworkers analyzed X-ray images from
484 patients (262 NEC patients and 222 non-NEC) and developed a radiomics model.[38] The combination of their radiomics models and radiologists in parallel diagnosis
increased the sensitivity for diagnosing NEC and reduced the rate of missed diagnoses,
highlighting AI's value as a diagnostic aid. Other studies using Gradient-weighted
Class Activation Mapping have shown that the model's attention was focused on the
fixed dilatation of the intestinal folds, intestinal wall edema, interintestinal gas,
and portal venous gas on abdominal radiographs.[39]
In patients suffering from surgical NEC, the choice between laparotomy (LAP) or comfort
care (CC) represents a complex, ethical dilemma. In a recent retrospective study,
Verhoeven and colleagues trained a behavioral artificial intelligence technology (BAIT)
decision aid on expert knowledge.[40] BAIT is a novel technology that allows for developing decision aids. The authors
have tested this technology on 40 patients (20 LAP and 20 CC) and have found that
treatment choices by AI aligned with clinical practice in at least 80% of cases.
AI has already demonstrated promise as a supportive tool in the prediction and diagnosis
of NEC. However, most of the datasets for training, validation, and testing are relatively
small and often originate from a single institution. Given the complexity and rarity
of NEC, larger prospective, multicenter studies are needed to fully evaluate the potential
of AI in timely prediction, accurate diagnosis, and clinical decision-making. Additionally,
external validation sets should be established to enable independent assessment of
AI tools for NEC.
Hirschsprung's Disease
Contrast enemas (CEs) are widely used as an initial diagnostic tool in the evaluation
of patients with suspected Hirschsprung's disease (HD). However, their interpretation
is known to vary significantly across institutions.[41] One of the first studies applying ML to quantitative imaging features extracted
from barium enema for distinguishing HD from non-HD cases was published in 2021.[42] The analysis included 54 neonates with biopsy-confirmed short-segment HD and 59
neonates without HD. By adding five clinical features—including vomiting, abdominal
distention, meconium extracted within 24 hours, and findings on rectal palpation—to
the radiological features, SVM, and L2-regularized LR achieved an accuracy of 86.4%,
a sensitivity of 81.8%, and a specificity of 90.9%.[42] More recently, a follow-up study evaluated the interobserver agreement and diagnostic
performance of pediatric surgeons, pediatric radiologists, and a deep neural network
in the interpretation of CE for HD.[43] The deep neural network achieved high diagnostic accuracy (area under the receiver
operating characteristic curve (AUC-ROC = 0.87)), which further improved when combining
anteroposterior and lateral views (AUC-ROC = 0.92). Integration of clinical data also
enhanced sensitivity and negative predictive value. These studies underscore the potential
of AI—particularly ML—in improving the diagnostic utility of CEs in HD.
The second diagnostic cornerstone in HD is rectal biopsy—which is either performed
as rectal suction or full-thickness biopsy. Histologically, HD is characterized by
aganglionosis and hypertrophic preganglionic nerve fibers with increased acetylcholine
activity. In addition to standard hematoxylin–eosin (H&E) and acetylcholinesterase
staining, various immunohistochemical markers—including S100, β-tubulin, glial fibrillary
acidic protein, nerve growth factor receptor, and cell body markers such as calretinin,
microtubule-associated protein 2 (MAP2), and peripherin—can aid in histopathologic
assessment.[44] Most of these methods involve simplifying the detection of ganglionic cells or hypertrophic
nerve fibers. These markers help simplify the detection of ganglion cells or hypertrophic
nerves, yet histologic diagnosis still largely depends on expert interpretation. In
2019, Schilling et al demonstrated that by applying ML via an ensemble voting classifier,
an accuracy of 87.5% could be achieved. Automated diagnosis using digital pathology
with immunohistochemical panels showed promising results for markers like calretinin
and MAP2.[44] Braun and colleagues confirmed these findings, showing that an AI-based analysis
of parasympathetic hyperinnervation could identify HD with high precision.[45] Similarly, Greenberg et al developed an AI algorithm capable of accurately detecting
ganglion cells in H&E-stained colon specimens.[46] Duci et al applied a U-net classifier to resected, fixed tissue samples and demonstrated
reliable detection of ganglion cells and hypertrophic nerve fibers in HD specimens.[47] Trained on over 19,000 images, the models achieved accuracies of 92.3% and 91.5%,
respectively. Another study revealed that a deep learning approach achieved accuracy
values of over 90% in detecting ganglion cells in whole slide images of 366 frozen
and 302 formalin-fixed paraffin-embedded H&E-stained slides obtained from 164 patients
from three centers.[48] Notably, the use of heatmap overlays improved the diagnostic accuracy of pathologists
from 77% to 85.8% and halved the average diagnostic time—from 139.7 to 70.5 seconds.[48] These developments highlight the potential of AI-based tools to assist in the histopathological
diagnosis of HD and suggest a role in accelerating intraoperative frozen section analysis.
Congenital Diaphragmatic Hernia
Congenital Diaphragmatic Hernia
AI applications in congenital diaphragmatic hernia (CDH) are limited but emerging.
Nevertheless, automatic segmentation models offer the potential to facilitate and
standardize lung volume measurements, enhance data collection accuracy, and support
the development of robust AI algorithms for predicting postnatal outcomes.[49] In 2024, Conte et al explored the feasibility of using a publicly available deep
learning (DL)-based automatic segmentation system (nnUNet) for automatic MRI contouring
of the lungs and liver of fetuses with CDH.[49] The performance of this system was compared with that of a human rater with 15 years
of experience in fetal MRI. In a cohort of 39 CDH cases, the authors demonstrated
that automatic segmentation of the fetal lung and liver is feasible and shows high
accordance with manual segmentation. In a follow-up retrospective observational cohort
study including 50 infants with isolated left-sided CDH, ML algorithms were evaluated
for their ability to predict clinical outcomes.[50] The goal was to develop predictive models for mortality and persistent neonatal
pulmonary hypertension based on integrated prenatal and early postnatal data. Three
different classification algorithms were tested: XGBoost, SVM, and K-Nearest Neighbors
(KNN). The best performing model, XGBoost, achieved 88% accuracy and 95% sensitivity
for predicting mortality using 10 features and 82% accuracy for pulmonary hypertension
severity with 14 features. The area under the ROC curve was 0.87 for mortality and
0.82 for pulmonary hypertension severity.[50]
These promising results pave the way for further research into the use of AI and ML
models to support clinical decision-making and outcome prediction in CDH patients.
Biliary Atresia
Biliary atresia (BA) is a rare but severe neonatal cholangiopathy characterized by
progressive fibroinflammatory obliteration of the extrahepatic bile ducts. If not
detected or treated promptly, it can quickly result in liver cirrhosis and eventually
liver failure. Early diagnosis and timely surgical intervention—typically via Kasai
portoenterostomy—are critical for improving long-term outcomes and delaying the need
for liver transplantation.[51] However, early identification remains challenging as clinical features such as persistent
jaundice, acholic stools, and elevated liver enzymes are non-specific and can overlap
with other causes of neonatal cholestasis.
AI-based diagnostic models have shown significant potential in improving the early
and accurate identification of BA. Zhao et al developed a multimodal AI model incorporating
serum biomarkers—most notably matrix metalloproteinase-7 (MMP-7)—along with additional
laboratory and clinical variables, demonstrating excellent diagnostic performance.[52] In another study, ML techniques were applied to assess the diagnostic utility of
gamma-glutamyl transferase (GGT) across various clinical settings, revealing that
GGT's effectiveness is influenced by patient demographics.[52] Xu et al used bioinformatics and ML to analyze mRNA expression data and identified
C-X-C motif chemokine ligand 8 (CXCL8) and thymosin beta-10 (TMSB10) as key diagnostic
biomarkers for BA. CXCL8 has been implicated as a therapeutic target due to its proinflammatory
role, while TMSB10 is associated with cell polarity, a factor potentially involved
in disease progression. These findings were validated through immunohistochemistry
and qRT-PCR, confirming upregulation of both genes in BA liver samples.[53]
Ultrasound continues to be a cornerstone in diagnosing BA; however, its accuracy highly
depends on the skill of the operator performing and interpreting the scan. Recent
advances in deep learning have enabled automated analysis of ultrasound images with
high diagnostic accuracy, aiding in the differentiation of BA from other causes of
neonatal cholestasis.[54] These AI-driven tools offer consistent and reproducible results, which are particularly
valuable in settings with limited access to experienced pediatric radiologists.
Beyond diagnosis, AI models have been utilized to stratify prognosis in BA. A study
from China developed a novel, survival-based clustering model to classify BA patients
into prognostically distinct subgroups, enabling early identification of candidates
for liver transplantation.[55] This data-driven classification system provides a potential roadmap for personalized
treatment planning and long-term management. In a related study, ML models trained
on early clinical and laboratory data were developed to predict the likelihood of
BA in infants presenting with cholestasis. These models demonstrated improved predictive
accuracy and have the potential to reduce the reliance on invasive diagnostic procedures
such as liver biopsy.[56]
To improve generalizability, a multicenter cross-sectional study developed a robust
ML-based diagnostic tool applicable across various clinical environments.[57] This approach highlights the scalability of AI-driven tools and their potential
integration into routine clinical workflows. Moreover, predictive models have also
been used to forecast postoperative complications such as adhesive small bowel obstruction
following BA surgery, adding another layer of clinical utility.[58]
Conclusion
AI and ML hold immense potential to transform pediatric surgery, offering improvements
in diagnostics, surgical precision, as well as personalized care. However, several
challenges must be addressed, particularly within the pediatric context. These include
the limited availability of large datasets due to the rarity of many pediatric surgical
conditions. This fact underscores the need for data collection within international,
multicenter collaboration. Furthermore, the heterogeneity of cases and variability
in clinical practices across institutions require the development and optimization
of robust AI/ML models. Additionally, deployment of such robust models should not
be restricted to the developing institutions, but rather made available to anybody
at any time and from anywhere via cloud hosting to support health and well-being anywhere
and consequently reduce health care inequalities. On the other hand, ethical concerns
represent another major hurdle—especially regarding decision-making involving neonates
and children, where accountability and transparency are paramount. Finally, external
validation of AI/ML models on independent datasets is essential to ensure their reliability
and applicability in real-world settings. Collaborative, prospective data sharing
will be key to refining and optimizing these promising technologies.
By understanding the fundamentals of these technologies and staying informed about
their applications and developments, pediatric surgeons can harness their full potential
to ultimately enhance patient outcomes.
