Role of immature granulocyte and blood biomarkers in predicting perforated acute appendicitis using machine learning model

Abstract

BACKGROUND

Acute appendicitis (AAp) is a prevalent medical condition characterized by inflammation of the appendix that frequently necessitates urgent surgical procedures. Approximately two-thirds of patients with AAp exhibit characteristic signs and symptoms; hence, negative AAp and complicated AAp are the primary concerns in research on AAp. In other terms, further investigations and algorithms are required for at least one third of patients to predict the clinical condition and distinguish them from uncomplicated patients with AAp.

AIM

To use a Stochastic Gradient Boosting (SGB)-based machine learning (ML) algorithm to tell the difference between AAp patients who are complicated and those who are not, and to find some important biomarkers for both types of AAp by using modeling to get variable importance values.

METHODS

This study analyzed an open access data set containing 140 people, including 41 healthy controls, 65 individuals with uncomplicated AAp, and 34 individuals with complicated AAp. We analyzed some demographic data (age, sex) of the patients and the following biochemical blood parameters: White blood cell (WBC) count, neutrophils, lymphocytes, monocytes, platelet count, neutrophil-to-lymphocyte ratio, lymphocyte-to-monocyte ratio, mean platelet volume, neutrophil-to-immature granulocyte ratio, ferritin, total bilirubin, immature granulocyte count, immature granulocyte percent, and neutrophil-to-immature granulocyte ratio. We tested the SGB model using n-fold cross-validation. It was implemented with an 80-20 training-test split. We used variable importance values to identify the variables that were most effective on the target.

RESULTS

The SGB model demonstrated excellent performance in distinguishing AAp from control patients with an accuracy of 96.3%, a micro aera under the curve (AUC) of 94.7%, a sensitivity of 94.7%, and a specificity of 100%. In distinguishing complicated AAp patients from uncomplicated ones, the model achieved an accuracy of 78.9%, a micro AUC of 79%, a sensitivity of 83.3%, and a specificity of 76.9%. The most useful biomarkers for confirming the AA diagnosis were WBC (100%), neutrophils (95.14%), and the lymphocyte-monocyte ratio (76.05%). On the other hand, the most useful biomarkers for accurate diagnosis of complicated AAp were total bilirubin (100%), WBC (96.90%), and the neutrophil-immature granulocytes ratio (64.05%).

CONCLUSION

The SGB model achieved high accuracy rates in identifying AAp patients while it showed moderate performance in distinguishing complicated AAp patients from uncomplicated AAp patients. Although the model's accuracy in the classification of complicated AAp is moderate, the high variable importance obtained is clinically significant. We need further prospective validation studies, but the integration of such ML algorithms into clinical practice may improve diagnostic processes.

Key Words: Acute appendicitis; Complicated acute appendicitis; Machine learning; Stochastic gradient boosting

Core Tip: This study uses an open access database and a Stochastic Gradient Boosting (SGB) machine learning algorithm to tell the difference between acute appendicitis (AAp) patients who are complicated and those who are not complicated. It also finds important biomarkers for both groups by using variable importance values that come from the modeling process. The SGB model demonstrated excellent precision in identifying AAp patients while exhibiting average performance in differentiating complicated AAp patients from uncomplicated ones.

INTRODUCTION

Acute appendicitis (AAp) is a common medical condition characterized by inflammation of the appendix, often requiring emergency surgical intervention. It is most frequently observed between 10 and 30 years of age but can occur in all age groups[1,2]. Prompt and accurate diagnosis is essential to both reducing negative appendectomy rates and preventing complications such as perforation and diffuse peritonitis, which significantly increase morbidity and mortality[1,3-7]. The diagnosis of AAp relies on clinical examination, radiological imaging such as ultrasound, computed tomography (CT), or magnetic resonance imaging (MRI), biochemical markers, and scoring systems[8]. While ultrasound is a widely preferred non-invasive method, its effectiveness depends on operator expertise and patient-specific anatomical factors[3,9]. CT is the most definitive imaging technique in adults due to its high sensitivity and specificity, reducing unnecessary surgeries but exposing patients to radiation[8,10,11]. MRI is a preferred alternative in pregnant women, offering high sensitivity and specificity in AAp diagnosis[8]. Various scoring systems, including the Alvarado Score, Raja Isteri Pengiran Anak Saleha Appendicitis, adult appendicitis scoring system, Appendicitis Inflammatory Response Score, and Pediatric Appendicitis Score, improve diagnostic precision and help reduce negative appendectomy rates[12-15]. Several inflammatory markers such as white blood cell (WBC) count, neutrophils, lymphocytes, monocytes, platelets, C-reactive protein (CRP), bilirubin, and derived ratios like neutrophil-to-lymphocyte ratio (NLR) and platelet-to-lymphocyte ratio have also been investigated for their diagnostic utility[1,16,17].

The distinction between uncomplicated and complicated AAp is crucial for determining the appropriate treatment strategy. Uncomplicated AAp presents with localized inflammation, whereas complicated AAp includes perforation, gangrene, or abscess formation, necessitating urgent surgical or percutaneous radiological interventions[18-20]. Accurately identifying these subtypes is essential for optimal management, as uncomplicated cases may be managed with either appendectomy or nonoperative antibiotic treatment[11,21,22], while complicated cases require immediate surgery to minimize complications such as peritonitis and pelvic abscess[23,24]. Delayed treatment leads to increased morbidity, and complicated cases account for 18%-34% of all AAp patients[25,26]. To improve differentiation between these subtypes, clinicians use imaging techniques, scoring systems, and biochemical markers[24,27]. Various studies have proposed cutoff values for biomarkers to distinguish between uncomplicated and complicated AAp; however, differences in study populations and methodologies limit their widespread applicability[6,23,28-33].

Despite significant advancements, traditional diagnostic methods still have limitations in differentiating between uncomplicated and complicated AAp. Most studies have relied on classical biostatistical approaches that analyze clinical, radiological, and biochemical parameters[7,34,35]. However, the generalizability of these findings remains uncertain, as variations across patient populations and medical centers affect diagnostic accuracy. This has led to increasing interest in artificial intelligence (AI) techniques, particularly machine learning (ML), which can improve diagnostic precision by minimizing human bias and efficiently analyzing high-dimensional datasets[36-38].

ML has emerged as a valuable tool in the differential diagnosis of AAp, addressing the limitations of traditional diagnostic approaches. Unlike conventional statistical methods, ML models can efficiently analyze large, multidimensional datasets, capturing complex patterns that enhance diagnostic precision. Recent systematic reviews have demonstrated the effectiveness of ML-based models in distinguishing AAp from other abdominal pathologies and classifying cases as uncomplicated or complicated with higher accuracy than conventional methods[39-43]. These studies highlight the superior predictive power of ML algorithms trained on a combination of clinical, laboratory, and imaging data. By integrating diverse variables, ML models optimize risk stratification and facilitate timely, evidence-based treatment decisions. This capability is particularly crucial in emergency settings, where early and precise differentiation between uncomplicated and complicated cases directly influences patient outcomes and management strategies.

All these diagnostic approaches aim to minimize negative appendectomy rates and predict AAp complications preoperatively, which is essential for optimizing treatment planning. In this study, we aimed to apply ML models to differentiate AAp from control cases and to distinguish uncomplicated from complicated AAp. Additionally, we sought to determine the most relevant variables for diagnosing AAp and predicting complicated cases, thereby improving diagnostic accuracy and supporting clinical decision-making in AAp management.

MATERIALS AND METHODS

Dataset

The study uses an open-access data set that includes patients with and without AAp[24]. The study's data set encompasses information from 140 individuals. 41 of them are healthy (control) individuals, and 99 individuals are AAp patients. The study excluded patients with pregnancy, blood transfusion, immunosuppressive or steroid use, hematological malignancy, and missing data. Based on the surgical findings and histopathological reports of the individuals in the data set, the patients were divided into 3 groups: The control group, uncomplicated AAp and complicated AAp (perforated, abscess, and peritonitis). After the separation process, 41 control patients, 65 uncomplicated AAp and 34 complicated AAp were respectively in the groups. All groups were documented with demographic data such as age and sex, immature granulocyte count (n; × 10³/L), immature granulocyte percent (%), WBC, neutrophils, lymphocytes, monocytes, platelet count, NLR, lymphocyte-to-monocyte ratio (LMR), mean platelet volume (MPV), neutrophil-to-immature granulocyte ratio, and ferritin levels. For details about the data set, see the study of Turkes et al[24].

Stochastic Gradient Boosting

Stochastic Gradient Boosting (SGB) is an advanced ML methodology that integrates gradient boosting concepts with stochastic sampling methods to improve model efficacy and mitigate overfitting. This strategy is especially efficacious in high-dimensional datasets where conventional boosting techniques may falter. SGB operates by sequentially training a series of weak learners, typically decision trees (DTs), and adjusting each learner to the residuals of its predecessors. This iterative method enables the model to progressively enhance its predictions by concentrating on the faults of preceding models[44,45]. A primary advantage of SGB is its capacity to incorporate unpredictability into the training process. By randomly picking a fraction of the training data for each iteration, SGB helps alleviate the danger of overfitting, a prevalent challenge in ML models, particularly when handling intricate datasets[46]. This stochastic characteristic not only improves generalization but also facilitates quicker convergence relative to conventional gradient boosting techniques[47,48]. Optimizing the performance of SGB models necessitates the precise tweaking of hyperparameters, including the learning rate and the number of trees[49]. In summary, SGB signifies a notable progression in ensemble learning methodologies, providing improved efficacy via its stochastic sampling method. Its utilization in several domains demonstrates its versatility and efficacy in tackling intricate predictive modeling issues. The hypermarameters used in SGB are as follows: N.trees (Number of Trees): 100; interaction.depth (Maximum Depth of Trees): 1; n.minobsinnode (Minimum Observations in Node): 10; shrinkage (Learning Rate): 0.1; bag.fraction (Subsample Ratio): 0.5; train.fraction (Training Fraction): 1. These settings help prevent overfitting and optimize model performance.

Modelling phase

We partition the dataset into 80% training data and 20% testing data for modeling with SGB. This study utilized the 5-fold cross-validation procedure, a resampling method, to ensure the model's validity. The n-fold cross-validation method entails partitioning the dataset into n subsets and subsequently applying the model to each subset. In the subsequent phase, we designate one component from the total n components for testing and use the remaining n-1 components for training. We evaluate the cross-validation method in the last stage by calculating the mean of the values derived from the models. We assessed the modeling performance using various metrics such as accuracy, balanced accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1-score. Ultimately, the modeling test yielded variable importance values to identify the variables most influential on the target variable.

Study protocol and ethics committee approval

This descriptive and cross-sectional study involving human participants was in accordance with the ethical standards of the institutional and national research committee and with the Helsinki Declaration of 1964 and its later amendments or comparable ethical standards. The IRB of Inonu University for Non-Interventional Clinical Research granted ethical approval, with the approval date and number being 2024/6809. We used the Strengthening the Reporting of Observational Studies in Epidemiology standard to improve the quality of reporting for observational studies[50].

Statistical analysis

The study summarizes the data using the median (95% confidence interval). Normality was checked with Shapiro-Wilk. Mann-Whitney We used the U test to compare the groups. In statistical analyses, P < 0.05 was considered significant. We performed the analyses using IBM SPSS 25.0.

RESULTS

This study included 140 patients aged between 18 and 70. The gender distribution of the 140 participants shows that men and women are equally represented (50% female, 50% male). The median age of the patients was 37 (34-41) years.

Table 1 presents the comparison of different groups stratified according to AAp status in terms of demographic and biochemical data using pairwise analysis. Age (P = 0.003), WBC (P < 0.001), neutrophil (P < 0.001), monocyte, NLR (P < 0.001), MLR (P < 0.001), MPV (P = 0.003), immature granulocytes (IG) count (P < 0.001), IG percent (P < 0.001), and neutrophil to IG ratio (P < 0.001) were all statistically different between the control and AAp. These results shows that biomarkers indicating inflammation are higher in AAp patients compared to the control group.

Table 1 Comparison of different groups stratified by acute appendicitis status, using pairwise.

Variables (median)	Control (n = 41)	AAp (n = 99)	P value	Control (n = 41)	Uncomplicated AAp (n = 65)	P value	Control (n = 41)	Complicated (n = 34)	P value	Uncomplicated AAp (n = 65)	Complicated AAp (n = 34)	P value
Age (year)	43 (37-54)	33 (32-39)	0.003	43 (37-54)	33 (31-40)	0.001	43 (37-54)	36 (32-47)	0.131	33 (31-40)	36 (32-47)	0.091
Platelet	258 (243-291)	248 (237-263)	0.349	258 (243-291)	247 (236-264)	0.381	258 (243-291)	253 (230-276)	0.459	247 (236-264)	253 (230-276)	0.985
Total bilirubin	0.61 (0.49-0.73)	0.73 (0.63-0.87)	0.254	0.61 (0.49-0.73)	0.67 (0.62-0.82)	0.337	0.61 (0.49-0.73)	0.93 (0.61-1.13)	0.187	0.67 (0.62-0.82)	0.93 (0.61-1.13)	0.329
WBC	6.8 (6.5-7.6)	14.5 (13.5-15.3)	< 0.001	6.8 (6.5-7.6)	13.6 (12.9-14.7)	< 0.001	6.8 (6.5-7.6)	15.8 (14.6-16.8)	< 0.001	13.6 (12.9-14.7)	15.8 (14.6-16.8)	0.009
Neutrophil	3.98 (3.70-4.45)	11.5 (10.5-12)	< 0.001	3.98 (3.70-4.45)	10.7 (9.9-12.6)	< 0.001	3.98 (3.70-4.45)	12.2 (11.5-13.8)	< 0.001	10.7 (9.9-12.6)	12.2 (11.5-13.8)	0.047
Lymphocyte	2.5 (2.2-2.6)	2.1 (1.8-2.5)	0.088	2.5 (2.2-2.6)	2 (1.7-2.5)	0.057	2.5 (2.2-2.6)	2.2 (1.8-2.8)	0.400	2 (1.7-2.5)	2.2 (1.8-2.8)	0.401
Monocyte	0.42 (0.39-0.47)	0.69 (0.62-0.73)	< 0.001	0.42 (0.39-0.47)	0.6 (0.6-0.7)	< 0.001	0.42 (0.39-0.47)	0.81 (0.73-0.95)	< 0.001	0.6 (0.6-0.7)	0.81 (0.73-0.95)	0.001
NLR	1.7 (1.6-2.0)	5.81 (5-6.68)	< 0.001	1.7 (1.6-2.0)	6 (4.46-7)	< 0.001	1.7 (1.6-2.0)	5.7 (5.1-7.3)	< 0.001	6 (4.46-7)	5.7 (5.1-7.3)	0.491
LMR	5.97 (5.45-6.64)	2.81 (2.56-3.14)	< 0.001	5.97 (5.45-6.64)	2.9 (2.6-3.5)	< 0.001	5.97 (5.45-6.64)	2.53 (2.16-3.14)	< 0.001	2.9 (2.6-3.5)	2.53 (2.16-3.14)	0.038
MPV	9.6 (9.1-10)	10.3 (10.1-10.6)	0.003	9.60 (9.10-10.2)	10.3 (9.9-10.6)	0.009	9.60 (9.10-10.2)	10.2 (9.8-10.8)	0.009	10.3 (9.9-10.6)	10.2 (9.8-10.8)	0.693
IG (× 103/μL)	0.01 (0.01-0.01)	0.02 (0.02-0.03)	< 0.001	0.01 (0.01-0.01)	0.02 (0.02-0.03)	< 0.001	0.01 (0.01-0.01)	0.03 (0.03-0.06)	< 0.001	0.02 (0.02-0.03)	0.03 (0.03-0.06)	0.001
IG (%)	0.10 (0.10-0.20)	0.1 (0.1-0.2)	< 0.001	0.10 (0.10-0.20)	0.11 (0.10-0.20)	< 0.001	0.10 (0.10-0.20)	0.20 (0.20-0.40)	< 0.001	0.11 (0.10-0.20)	0.20 (0.20-0.40)	0.018
Neutrophil to IG ratio	330 (15-387)	495 (440-570)	< 0.001	330 (15-387)	513 (453-593)	< 0.001	330 (15-387)	442 (383-553)	0.001	513 (453-593)	442 (383-553)	0.182
Ferritin	34 (26-65)	71.3 (35.5-117)	0.070	34 (26-65)	84 (41-124)	0.044	34 (26-65)	30 (25-163)	0.440	84 (41-124)	30 (25-163)	0.490

AAp: Acute appendicitis; IG: Immature granulocytes; MPV: Mean platelet volume; NLR: Neutrophil-to-lymphocyte ratio; LMR: Lymphocyte-to-monocyte ratio; WBC: White blood cell.

A group of people with complicated AAp had significantly higher median age WBC (P = 0.009), neutrophil (P = 0.047), monocyte (P = 0.001), IG count (P = 0.001), and IG percent (P = 0.018) than a group of people with uncomplicated AAp. On the other hand, the median LMR (P = 0.038) was higher in the uncomplicated AAp group than in the complicated AAp group. Overall, these findings indicate a pronounced inflammatory state in patients with complicated AAp compared to the uncomplicated AAp group.

The uncomplicated AAp group had significantly higher median WBC (P < 0.001), neutrophils (P < 0.001), monocytes (P < 0.001), NLR (P < 0.001), MPV (P = 0.009), IG count (P < 0.001), IG percent (P < 0.001), neutrophil-to-IG ratio (P < 0.001), and ferritin (P = 0.004) than the control group. On the other hand, age (P = 0.001) and LMR (P < 0.001) were higher in the control group than in the uncomplicated AAp group.

The median WBC (P < 0.001), neutrophil (P < 0.001), monocyte (P < 0.001), NLR (P < 0.001), MPV (P = 0.009), IG count (P < 0.001), IG percent (P < 0.001), and neutrophil/IG ratio (P = 0.001) were significantly higher in the complicated AAp group compared to the control group. On the other hand, the median LMR (P < 0.001) was higher in the control group than in the complicated AAp group.

Table 2 presents the results of the training and test set metrics for modeling using SGB by AAp status (control vs entire AAp, uncomplicated vs complicated AAp). The results for control vs entire AAp are as follows. With an accuracy of 96.3%, the model demonstrates excellent overall effectiveness. The balanced accuracy is 97.4%, indicating it distinguishes well between the two conditions. Sensitivity is strong at 94.7%, showing the model's ability to correctly identify cases of AAp. The specificity is perfect at 100%, meaning it accurately classifies all healthy individuals. Additionally, the PPV is also 100%, confirming that all positive diagnoses are correct. However, the NPV stands at 88.9%, indicating a potential need for caution with negative results. The F1 score of 97.3% highlights a robust balance between precision and recall. The micro and macro aera under the curve (AUC) values obtained from the model are 94.7% and 97.4% respectively and the model is quite successful. Overall, this model excels in diagnosing AAp, although potential inaccuracies in negative findings warrant attention.

Table 2 Results of training and test set metrics of groups based on acute appendicitis status.

Groups, metrics (%)	Control vs entire AAp		Uncomplicated vs complicated AAp
	Training set	Test set	Training set	Test set
Accuracy	96.5	96.3	82.5	78.9
Balanced accuracy	95.4	97.4	81.1	80.1
Sensitivity	97.5	94.7	76.9	83.3
Specificity	93.9	100	85.2	76.9
Positive predictive value	97.5	100	71.4	62.5
Negative predictive value	93.9	88.9	88.5	90.9
F1 score	97.5	97.3	74.1	71.4
AUC	96.4	94.7	82.5	79.0
Accuracy	96.5	96.3	82.5	78.9

AAp: Acute appendicitis; AUC: Aera under the curve.

The variable importance values revealing the variables explaining the dependent variable obtained from the model are given in Table 3. Additionally, Figure 1A presents a graphical representation of the variable importance values. According to variable importance values, the most important variables determining the difference between AAp and control groups were WBC (100%), neutrophil (95.14%), LMR (76.05%), NLR (48.87%), and monocyte (35.46%). Less important variables include MPV (19.78%), total bilirubin (9.95%), IG percent (7.15%), lymphocyte (6.98%), age (5.89%), neutrophil to IG ratio (4.90%), platelet (3.28%), and sex (2.15%).

Figure 1 Importance of variables for differentiating. A: Importance of variables for differentiating between acute appendicitis (AAp) and control groups; B: Importance of variables for differentiating between complicated AAp and uncomplicated AAp. PLT: Platelet; IG: Immature granulocytes; MPV: Mean platelet volume; NLR: Neutrophil-to-lymphocyte ratio; LMR: Lymphocyte-to-monocyte ratio; WBC: White blood cell.

Table 3 Importance of variables in differentiating acute appendicitis from the control group.

Variables	Variables importance (%)
WBC	100
Neutrophil	95.14
LMR	76.05
NLR	48.87
Monocyte	35.46
MPV	19.78
Total bilirubin	9.95
IG (%)	7.15
Lymphocyte	6.98
Age	5.89
Neutrophil to IG ratio	4.90
Platelet	3.28
Sex	2.15

WBC: White blood cells; NLR: Neutrophil-to-lymphocyte ratio; LMR: Lymphocyte-to-monocyte ratio; MPV: Mean platelet volume; IG: Immature granulocytes.

The results for complicated AAp vs uncomplicated AAp are as follows. The accuracy is 78.9%, indicating decent overall performance, while balanced accuracy is 80.1%, reflecting good identification of both conditions. Sensitivity is high at 83.3%, meaning the model effectively detects actual AAp cases. However, the specificity is lower at 76.9%, indicating the possibility of misclassifying some healthy individuals. The PPV is 62.5%, indicating limited reliability for positive results, whereas the NPV is strong at 90.9%. The F1 score of 71.4% shows a balanced performance. The micro and macro AUC values obtained from the model are 79% and 53.2% respectively. Overall, the metrics show a promising diagnostic capability for AAp, but the lower PPV implies the need for additional testing in clinical practice to confirm positive diagnoses.

The variable importance values revealing the variables explaining the dependent variable obtained from the model are given in Table 4. Additionally, Figure 1B presents a graphical representation of the variable importance values. We found that the variables most important for differentiating between uncomplicated and complicated AAp were total bilirubin (100%), WBC (96.90%), neutrophil to IG ratio (64.05%), NLR (63.53%), ferritin (63.27%), and monocyte (62.13%). On the other hand, LMR (34.83%), MPV (34.53%), lymphocyte (34.33%), neutrophils (29.54%), age (28.75%), sex (0.80%), and IG count (0.38%) were the variables with the lowest importance. In general, total bilirubin and WBC are the most critical factors in differentiating both AAp types. Figure 2 summarizes the study's results graphically.

Figure 2 Graphical summary of the study. NLR: Neutrophil-to-lymphocyte ratio; LMR: Lymphocyte-to-monocyte ratio; WBC: White blood cell.

Table 4 Importance of variables in differentiating complicated acute appendicitis from the uncomplicated acute appendicitis.

Variables	Variables importance (%)
Total bilirubin	100
WBC	96.90
Neutrophil to IG ratio	64.05
NLR	63.53
Ferritin	63.27
Monocyte	62.13
LMR	34.83
MPV	34.53
Lymphocyte	34.33
Platelet	31.07
Neutrophil	29.54
Age	28.75
Sex	0.80
IG	0.38

WBC: White blood cells; NLR: Neutrophil-to-lymphocyte ratio; LMR: Lymphocyte-to-monocyte ratio; MPV: Mean platelet volume; IG: Immature granulocytes.

DISCUSSION

The uncomplicated and complicated AAp incidences were 84.2-102.7 and 19.4-27.2 per 100000 people for years, respectively. Lifetime incidence of AAp is around 8.6% in men and 6.7% in women, whereas epidemiological studies show that the lifetime risk of having an appendectomy is 12% for men and 23% for women[1,6].

Differences between uncomplicated and complicated AAp are essential for establishing suitable management strategies and forecasting patient outcomes. Typically, conservative management or simple appendectomy can treat uncomplicated AAp, but complicated AAp increases the risk of serious complications[8,23]. Therefore, preoperative accurate prediction is of critical importance for patient care. Research indicates that many biochemical and clinical parameters (age, sex, pain, migration of pain, anorexia, nausea, vomiting, fever, dysuria, etc.) can assist in distinguishing between uncomplicated and complicated AAp[23,32,51,52].

Numerous studies have explored the relationship between AAp and basic biochemical markers such as WBC, neutrophilia, lymphocytes, platelets, MPV, platelet distribution width (PDW), red cell distribution width, CRP, total or direct bilirubin, and ferritin. Studies often use WBC as the primary laboratory marker for AAp[31]. Complicated cases of AAp significantly increase neutrophils, as their counts rise in tandem with the severity of infections[53]. Studies show a marked elevation of neutrophil levels in patients with either uncomplicated or complicated AAp[31]. Researchers have correlated NLR with more severe manifestations of AAp, where the excessive inflammatory response may cause lymphocyte levels to diminish[54,55]. NLR has demonstrated high sensitivity and specificity in identifying both uncomplicated and severe AAp[56]. Researchers have investigated MPV, an additional hematological parameter, in relation to AAp. Increased MPV has been associated with inflammatory disorders, such as AAp, indicating that bigger platelets may signify an intensified inflammatory response[57]. Research indicates a positive correlation between MPV and WBC, supporting the notion that platelet activation is integral to the inflammatory response in AAp[58]. Monocyte counts, although less commonly emphasized, can help elucidate the inflammatory process, as they generally rise in reaction to infection[59]. Researchers have investigated ferritin in relation to AAp, but its significance is less defined than that of WBC and neutrophil counts. Increased ferritin levels may suggest a systemic inflammatory response; nevertheless, additional research is required to elucidate its diagnostic significance in AAp[18].

Researchers have also investigated bilirubin levels as potential biomarkers for predicting the severity of AAp[60-63]. Arredondo Montero et al[64] published a meta-analysis in 2022 showing that serum bilirubin increases significantly in patients with complicated AAp. The specificity of bilirubin as a diagnostic marker for complicated AAp significantly exceeds that of WBC counts, suggesting its potential utility in clinical practice[63]. Various ML models have begun to accurately predict the diagnosis of AAp in recent years, coinciding with the entry of AI into the health field. Published studies have shown that increases in total or direct bilirubin levels are the most important biomarkers in the diagnosis of AAp and especially complicated AAp[7,65].

As understood from the study of Turkes et al[24], in which we used the open access data set, there have been studies published in recent years on the use of IG count, IG percent, and neutrophil-to-IG ratio in the differential diagnosis and severity of AAp[24,66-68]. However, there is no published study on the use of IG in ML-based studies. In this study, we showed the importance of IG count, IG percent and neutrophil to IG ratio in the differential diagnosis of uncomplicated and complicated AAp using classical statistical analysis and also showed that the neutrophil to IG ratio can be used in the diagnosis of perforated AAp using the ML model. Although the variable importance rate is slightly low, we believe that it would be appropriate to use different ML models in multicentric and prospective studies. Up to this part of the discussion, data obtained from classical statistical analyses of some important biochemical markers used in the diagnosis and prediction of uncomplicated and complicated AAp have been summarized.

From this point on, we aim to discuss the usability of ML algorithms in the classification and accurate prediction of AAp. We are gradually utilizing ML methodologies to enhance the diagnostic precision of AAp in both adult and pediatric patients. Akbulut et al[7] conducted a study using explainable ML based on CatBoost, demonstrating 88.2% accuracy in distinguishing AAp from the control group and 92% accuracy in distinguishing uncomplicated AAp from complicated ones. Wei et al[65] conducted a study using nine different ML models [Random Forest (RF), Linear regression (LR), Classification and regression trees (CART), Bayes, Filexible Discriminant Analysis, Support Vector Machine (SVM), neural networks (NN), Gradient Boosting Machines (GBM), K-means Clustering (KNN)], demonstrating that the GBM achieved the highest accuracy rate in classifying AAp and recognizing complicated AAp. The study reveals that the accuracy rates ranged from 88.89% to 95.56%. Shahmoradi et al[69] conducted a study using the optimized artificial NN of the SVM, which demonstrated the model's 95% accuracy in distinguishing AAP patients from normal ones. Akmese et al[70] conducted a study using seven different ML algorithms (NN, KNN, LR, SVM, CART, RF, GBT), demonstrating that the GBT algorithm was the best at predicting AAp with 95.31% accuracy. However, the authors stated that the accuracy rates they obtained from the algorithms varied between 64.84% and 95.31%. In order to accurately predict AAp, Mijwil and Aggarwal[71] compared various ML algorithms (LR, Bayes, GLM, DTs, SVM, GBT, RF), and observed that the obtained accuracy rates varied between 64.74% and 83.75%. This study used the SGB ML model, which classified and predicted AAp and the control group with a high accuracy of 96.3%. On the other hand, it predicted the stratification of uncomplicated and complicated AAp with an accuracy rate of 78.9%. Both results are similar to the literature.

The SGB-based ML model, which this study used to compare AAp and control groups, demonstrated high performance in AAp classification and prediction. With an accuracy of 96.3%, a sensitivity of 94.7%, and a specificity of 100%, the model offers a reliable option in clinical applications. In particular, the PPV was 100%, while the NPV was 88.9%, indicating that some false negative results may occur. Furthermore, the F1 score of 97.3% shows that the model performs well in terms of both accuracy and sensitivity. In conclusion, the integration of the SGB model in the diagnosis of AAp can significantly improve the diagnostic processes; however, it is important to consider the limitations of the model and potential false negative results. Future research should emphasize the validity and reliability of the model with different data sets. In this context, the integration of the SGB model in clinical practice can significantly improve diagnostic processes.

This study showed that the SGB-based ML model, which compared uncomplicated and complicated AAp groups, performed moderately in complicated AAp stratification and prediction, achieving an accuracy rate of 78.9%, and had a moderate impact on these diagnoses[72]. The sensitivity rate of 83.3% reflects the ability to correctly identify cases of complicated AAp, while the specificity rate of 76.9% indicates the presence of false positives. This could result in the incorrect recognition of some uncomplicated AAp cases as complicated, leading to unnecessary surgical interventions. The PPV was 62.5%, and the NPV was 90.9%, indicating that the model is reliable when giving negative results, but the accuracy of positive results is low. In addition, although the F1 score of 71.4% indicates a balanced performance of the model, this value may not be sufficient in a clinical context. In conclusion, although the SGB model has some strengths in differentiating uncomplicated and complicated AAp, its overall performance is far from satisfactory. Future research should focus on the evaluation of different variables and algorithms to improve the performance of the model. These findings provide important clues for the development of a more effective model for the diagnosis of AAp. The model's variable importance values revealed the total bilirubin variable as the most crucial factor in diagnosing both uncomplicated and complicated AAp.

Examining the literature data reveals the use of numerous ML algorithms for the classification and accurate prediction of both complicated and uncomplicated AAp. However, the literature lacks sufficient data on the significance of clinical and biochemical variables. Marcinkevics et al[73] conducted a study using the RF classifier algorithm, demonstrating that the parameter with the strongest relationship between the level and the severity of AAp is CRP. Akbulut et al[7] conducted a study using ML-based explainable AI (SHAP model), demonstrating that the five most important variables in the prediction of complicated AAp are CRP, PDW, age, MPV, and bilirubin. In the same study, the authors showed that the seven most important variables in the prediction of uncomplicated AAp are bilirubin, PNR, PDW, mean corpuscular volume, WBC, CRP, and neutrophil. Males et al[74] used a ML-based explainable AI (SHAP model) to find the seven most important clinical and biochemical factors for predicting AAp. These were CRP, WBC, the length of the symptoms, neutrophils (%), Na concentration, the ratio of thrombocytes to lymphocytes, and lymphocytes (%). Aydin et al[75] prepared a study using six different ML algorithms, which revealed that PDW, WBC, neutrophils, and lymphocytes were the five most important biochemical variables in predicting uncomplicated AAp, while PDW, neutrophils, WBC, G, and lymphocytes were the most important variables in predicting complicated AAp. The most significant issue with this study is that, despite having the lowest accuracy rate (93.97%) among the six ML algorithms, the decision tree algorithm is the one that most accurately reflects the results of variable importance[75]. This study demonstrates that the SGB-based ML model's variable importance identifies WBC count, neutrophil, and LMR as the three most crucial variables in the diagnosis of AAp, highlighting the impact of the immune response throughout the disease's progression. Future research may improve diagnostic processes and patient management by further studying the association of these variables with clinical outcomes. This study also demonstrates that the SGB-based ML model's variable importance identifies total bilirubin, WBC, the neutrophil to IG ratio, NLR, and ferritin as the five most crucial variables in the diagnosis of complicated AAp.

Key points

The current study investigates the improvement of accurate AAp diagnosis through ML approaches utilizing biochemical and demographic characteristics. The SGB model effectively distinguished between AAp and control groups. The model's algorithm was only somewhat good at classifying and predicting both complicated and uncomplicated AAp cases. However, looking at the different meanings of the results showed that they were very useful for clinical purposes. If there is insufficient correlation between the model's accuracy and the importance of the variables, it is crucial to prioritize the importance of the variables from a clinical perspective. There is also a need for larger patient datasets and external validation to ensure the validity and control of the study results.

CONCLUSION

The present study demonstrates the importance of biomarkers in the diagnosis of AAp and the potential of ML models to improve diagnostic processes for AAp. Despite the model's successes suggesting its potential benefits in clinical practice, the results for patients with complicated and uncomplicated AAp indicate the need for prospective and multicentric studies with larger data sets to enhance the model's validity and reliability. Systems obtained by applying AI methods to clinical data can accelerate the diagnostic processes before surgery and make the management of patients more effective.