Deep self-organizing competitive neural network for magnetically controlled capsule endoscopy-based auxiliary diagnosis of clinically significant gastric lesions

Abstract

BACKGROUND

Magnetically controlled capsule endoscopy (MCCE) is a well-tolerated alternative to invasive gastroscopy, but its diagnostic efficiency and consistency are limited by large data volume and physician-dependent diagnosis.

AIM

To construct a deep self-organizing competitive neural network (DSomNN) to enhance MCCE-based auxiliary diagnosis of clinically significant gastric lesions (CSGL).

METHODS

A total of 7425 MCCE patients were initially enrolled; after exclusion of 35 patients according to predefined inclusion/exclusion criteria, 7390 patients were analyzed. Demographic data, chief complaints, MCCE findings, and five key gastric lesions annotated by gastroenterologists were used for model training. The primary prediction endpoint is CSGL, defined as gastric cancer, intestinal metaplasia, atrophic gastritis, gastric ulcer, or gastric polyps. The DSomNN extracts variable associations and establishes CSGL risk criteria.

RESULTS

Results show close correlations between CSGL and clinical data. In the validation/test set, the error rate is < 5%. In conclusion, the DSomNN performs well in CSGL risk prediction, addresses MCCE limitations, and serves as a valuable auxiliary tool for clinical gastroenterology decision-making.

CONCLUSION

This study developed a DSomNN, integrating global vectors for word representation, Text convolutional neural network-based deep feature extraction, and self-organizing competitive learning for the auxiliary identification of CSGL using demographic characteristics, clinical symptoms, and MCCE findings.

Key Words: Auxiliary diagnosis; Magnetically controlled capsule endoscopy; Deep self-organizing competitive neural network; Text convolutional neural network; Clinically significant gastric lesions

Core Tip: Magnetically controlled capsule endoscopy (MCCE) is a well-tolerated alternative to invasive gastroscopy, but its diagnostic efficiency and consistency are limited by large data volume and physician-dependent diagnosis. This study constructs a deep self-organizing competitive neural network (DSomNN) to enhance MCCE-based diagnostic accuracy and efficiency, and the DSomNN performs well in gastric cancer risk prediction, addresses MCCE limitations, and serves as a valuable auxiliary tool for clinical gastroenterology decision-making.

INTRODUCTION

Gastric cancer (GC) imposes a substantial disease burden across the Americas and Eastern Europe, with a male incidence rate twice that of females[1,2]. Although early-stage GC is highly curable, most cases are detected at advanced stages, resulting in poor patient prognosis. In China, nearly 20% of the population lives with chronic gastric disorders, including gastritis, gastric ulcers, and GC[3], and GC ranks as the second leading cause of cancer-related mortality nationwide[4]. Notably, China’s GC early diagnosis rate remains below 10%, substantially lower than the 70% reported in Japan and 50% in South Korea. Early detection, timely diagnosis, and standardized treatment of GC, alongside targeted intervention for precancerous lesions, such as atrophic gastritis, gastric ulcers, and gastric polyps, are essential for reducing GC incidence and mortality.

Conventional esophagogastroduodenoscopy (C-EGD) with endoscopic biopsy serves as the diagnostic gold standard for gastric diseases. However, C-EGD causes procedural discomfort and carries inherent contraindications, restricting its applicability in large-scale population screening. Additionally, widespread population coverage, shortages of specialized endoscopists, and poor patient tolerance collectively limit the detection efficacy of gastric precancerous lesions via C-EGD[5].

Magnetically controlled capsule endoscopy (MCCE) has emerged as a simple, non-invasive alternative to C-EGD, with comparable diagnostic accuracy. Featuring high-resolution imaging, favorable safety profiles, minimal cross-infection risk, and superior patient comfort, MCCE is suitable for all age groups, particularly cardiovascular patients receiving antithrombotic therapy[6]. Accordingly, MCCE has become a critical tool for the primary screening and routine evaluation of gastric and esophageal diseases[7,8]. Our prior work has systematically validated MCCE clinical utility in small intestinal diseases[9] and reflux esophagitis[10]. At our institution, approximately 2000 patients with suspected gastric disease undergo annual MCCE screening, representing a leading clinical practice globally.

Nevertheless, MCCE has notable limitations: It cannot perform in vivo lesion biopsy or pathological assessment. Each MCCE examination generates tens of thousands of gastric images, placing a heavy interpretive burden on clinicians. This workload compromises diagnostic sensitivity and specificity and increases the risk of missed lesions. To date, no intelligent analysis system has integrated artificial intelligence (AI), particularly convolutional neural networks (CNNs), with MCCE images and patients’ clinical characteristics for rapid diagnosis and clinically significant gastric lesions (CSGL) risk prediction, nor have studies used patient symptoms and MCCE reports to analyze and predict disease outcomes.

With advancements in AI and deep learning, their application in healthcare has gained increasing attention[11]. Deep convolutional neural networks have been successfully used to diagnose diseases by extracting hierarchical image features from large datasets[12-14]. Given the shortage of clinical doctors, AI-assisted systems can alleviate clinicians’ workload, shorten diagnostic time, and reduce manual diagnostic errors.

To address the aforementioned limitations of MCCE and improve gastric disease diagnosis, we compile data from 7425 patients who underwent MCCE, including 496 who received C-EGD and biopsy pathological examinations, from 2016 to 2022, aiming to explore hidden information using AI methods and provide a new approach for the diagnosis of CSGL.

MATERIALS AND METHODS

Data collection, dataset construction and analytical framework

Clinical reports from a total of 7390 patients with MCCE were retrospectively enrolled to construct the analytical dataset. A total of 7425 patients who underwent MCCE were initially screened. After applying predefined inclusion/exclusion criteria—complete demographic data, symptoms, and endoscopic reports; written informed consent; no decline to participate; and no multiple severe comorbidities—35 patients were excluded. The final analytical cohort comprised 7390 patients. The dataset was split into a training set (n = 5543) and a test set (n = 1847) at a ratio of approximately 3:1.

Machine learning algorithms were applied for in-depth data analysis to assist clinical decision-making regarding the necessity of further diagnostic examination for enrolled patients. The methodological pipeline established in this study consisted of four sequential components: Data source screening, standardized dataset construction, development of a deep self-organizing competitive network model, and comprehensive data analysis (Figure 1).

Figure 1 The process of study: Data collection primarily considers the richness of the data. Dataset construction can achieve data standardization. Neural networks include text vectorization, deep feature mining, and clustering. Data analysis examines the correlation and clustering effects. IM: Intestinal metaplasia.

MCCE: Equipment, image acquisition, and standardized report generation

MCCE examinations were performed using a commercial system provided by Ankon Medical Technologies Co., Ltd. (Wuhan, China). The capsule endoscope measured 27 mm in length and 11.8 mm in diameter, with a weight of 4.8 g. It was equipped with a single camera offering a field of view of 140° ± 10°. The capsule was actively controlled by an external magnetic field and traversed the stomach under magnetic guidance. The dynamic frame rate ranged from 0.5 to 2 frames per second (fps). Images were captured and recorded at an average rate of 0.8 fps with a resolution of 480 × 480 pixels. Each video frame was sequentially numbered and stored in a dedicated directory.

Diagnostic reports were subsequently generated from the capsule endoscopy images in accordance with the standards specified in the Chinese Expert Consensus on Clinical Applications of MCCE (2021, Shanghai)[15]. The structured reports were organized by anatomical gastric segments, including the cardia and fundus, gastric body, antrum, angularis, and pylorus. As depicted in Figure 2, representative images from each gastric region were analyzed based on clinical indications. Each lesion was documented by its precise anatomical location, and the morphological features were individually described to compile the final endoscopic report.

Figure 2  Diagnostic results generated from magnetically controlled capsule endoscopy images.

Gastric disease manifestations are closely associated with patient demographics (e.g., gender, age) and clinical symptomatology, in addition to findings from capsule endoscopy. For instance, patients with chronic gastric disorders often present with persistent symptoms such as epigastric discomfort. Integrating patient-reported symptoms into the analytical dataset enhances the comprehensiveness of feature information. The constructed machine learning model was then applied to deeply mine the inherent correlations within this integrated dataset, thereby extracting detailed feature insights to facilitate more accurate recognition and differential diagnosis of gastric conditions.

Dataset construction, sample characterization, and clinical labeling for multimodal gastric disease analysis

Dataset construction was based on a real-world cohort of 7390 patients who underwent MCCE between 2016 and 2022 at Shanghai Sixth People’s Hospital. Among these patients, 496 underwent standard upper endoscopy with biopsy sampling. All participants provided written informed consent, and all personal identifying information was anonymized to protect privacy. The study was approved by the Ethics Committee of Shanghai Sixth People’s Hospital.

Sample selection criteria: To optimize data quality and enable robust exploration of latent clinical associations, each sample was characterized by clinicians using consensus-based definitions. The model used 20 clinical features: 2 demographic variables, 7 symptoms, and 11 endoscopic findings. Physical characteristics: Gender, Age. Presenting symptoms: Abdominal pain, abdominal distension, abdominal discomfort, heartburn, acid regurgitation, gastroesophageal reflux disease (GERD), bleeding. Endoscopic outcomes: GC, atrophic gastritis, gastric ulcer, gastric polyps, non-atrophic gastritis, erosive gastritis, abnormal gastric mucosa, yellow-white macules, varices, bleeding with scabs, abnormal gastric antrum.

Inclusion and exclusion criteria: To ensure dataset integrity and analytical validity, only patients with complete data (demographics, symptoms, and endoscopic reports) were included. Patients with incomplete datasets, those who declined study participation, and individuals with multiple severe comorbidities were excluded.

Clinical significance labeling: The primary endpoint was CSGL, defined as GC, intestinal metaplasia (IM), atrophic gastritis, gastric ulcer, or gastric polyps. Among the 7390 patients, 496 underwent conventional endoscopy with biopsy. Pathological results were used as the reference standard when available. For patients without pathological examination, clinical significance labels were determined by gastroenterologists after integrating demographic characteristics, symptoms, and MCCE reports. Disagreements were resolved by a third senior specialist. IM was not included as an input feature in the primary 20-feature deep self-organizing competitive neural network (DSomNN) model; it was used only as a reference label in the pathological subgroup. This labeling strategy may introduce verification and label bias, which is discussed in the limitations section.

Final clinical classification was determined by specialist physicians, who integrated patient characteristics, symptoms, and capsule endoscopy reports. Each sample was ultimately encoded with the 20 clinical features listed above for statistical analysis.

Following rigorous curation, data from 7390 patients met the final dataset criteria. The dataset was split into a training set (n = 5543) and a test set (n = 1847) at a ratio of approximately 3:1. The internal test set was used for label-quality assessment and internal validation and was not an independent external validation cohort. Two subsets of the internal test set were used for focused validation: Test Set a (n = 500) and Test Set b (n = 500). The overall dataset construction workflow is illustrated in Figure 3.

Figure 3  Schematic illustration of the dataset construction pipeline.

Development and validation of a DSomNN

A novel DSomNN was developed to discriminate expression patterns between clinically significant and non-significant gastric lesions using the curated dataset, thereby enabling automated data clustering and binary classification. As illustrated in Figure 4, the DSomNN architecture was constructed in three sequential stages.

Figure 4 Framework of the deep self-organizing competitive neural network. This figure presents the end-to-end architecture and operational principles of the proposed deep self-organizing competitive neural network for clinical feature and magnetically controlled capsule endoscopy report analysis. A: Overview of the deep self-organizing competitive neural network workflow, illustrating the integration of feature extraction, competitive learning, and diagnostic prediction, along with the selected network architecture and key components; B: Detailed feature extraction process via text convolutional neural network, including text vectorization, multi-kernel convolution, rectified linear unit activation, and average pooling to capture hierarchical clinical semantic features; C: Schematic of the self-organizing competitive clustering mechanism, where neurons compete for input patterns to learn topological feature distributions and enable unsupervised clustering of high-dimensional data for disease classification. TextCNN: Text convolutional neural network; GloVe: Global vectors for word representation.

Word vectorization: The 20 standardized clinical features were converted into dense high-dimensional vectors using global vectors for word representation (GloVe). This embedding technique effectively captures semantic relationships and contextual similarities among clinical features.

Deep feature extraction: A text convolutional neural network (TextCNN) was employed to perform hierarchical deep feature mining on the vectorized input data. TextCNN is particularly suited for extracting local and global patterns from structured clinical text and categorical feature data.

Self-organizing competitive clustering: A competitive layer with two neurons classified the TextCNN-derived feature representations into clinically significant vs non-significant gastric lesions. The output was a binary vector, with 1 denoting abnormal samples requiring further examination and treatment.

Feature correlation analysis and model validation

Feature correlation analysis (Figure 5) was performed to characterize interdependencies among variables: Figure 5A: Correlation matrix between clinically significant features and other endoscopic findings. Figure 5B: Associations between clinical symptoms and endoscopic features. Figure 5C: Age-related distribution patterns of gastric diseases. Heatmap visualization revealed distinct feature-feature correlations, providing critical latent information to guide deep learning analysis.

Figure 5 Correlation analysis of key features in the magnetically controlled capsule endoscopy dataset. Comprehensive correlation analyses were performed to investigate the relationships among multimodal features derived from magnetically controlled capsule endoscopy (MCCE) reports. A: Correlation matrix illustrating pairwise associations between core clinical features of high diagnostic significance and other quantitative/qualitative features extracted from MCCE reports, revealing heterogeneous correlation strengths across feature pairs; B: Heatmap depicting the correlation patterns between clinical manifestations/symptoms and endoscopic features in the MCCE dataset; C: Age-related distribution patterns of gastric diseases, demonstrating distinct correlation patterns across different age groups. GC: Gastric cancer; AG: Atrophic gastritis; GU: Gastric ulcer; GP: Gastric polyps; SG: Erosive gastritis; YWM: Yellow-white macules; VV: Varices; BS: Bleeding and scabs; AGA: Abnormal gastric antrum; OGA: Other gastric abnormalities; BR: Bile reflux; AP: Abdominal pain; AD: Abdominal distension; ADis: Abdominal discomfort; H: Heartburn; AR: Acid regurgitation; GERD: Gastroesophageal reflux; Hem: Bleeding.

Model performance was validated using the internal test set and two validation subsets derived from MCCE and standard gastroscopy with biopsy, respectively. Test Set a comprised 500 patients who underwent conventional endoscopy with available pathology/reference data. Among them, 98 patients (19.60%) were identified by gastroenterologists as requiring further evaluation; the DSomNN correctly identified 92, corresponding to a sensitivity of 93.88%. Test Set b comprised 500 patients who underwent MCCE, for whom IM data were unavailable. Specialists determined that 101 patients (20.20%) needed additional evaluation; the model correctly identified 94, corresponding to a sensitivity of 93.07%. The overall classification accuracy for Test Set b was 94.95%. Full confusion matrices and additional metrics were not available in the original dataset.

GloVe-based word vectorization for clinical feature embedding

Global Vectors for GloVe is an unsupervised natural language processing algorithm developed by Pennington, Socher, and Manning (Stanford University, 2014) that transforms discrete textual terms into continuous dense vector representations. This embedding framework enables quantitative measurement of semantic similarities and relational patterns between terms, making it suitable for modeling clinical feature relationships. In this study, each data point contains information on 20 distinct vocabulary terms, with a focus on capturing the global information of the corpus. Therefore, we employed the GloVe model, which combines both local statistical information and global context to generate word vectors.

First, a co-occurrence matrix based on the 20 feature attributes from the 7390 patient records was constructed. The dimensions of the co-occurrence matrix were defined as 20 × 20. Next, a minimum loss function Jloss was defined to calculate the co-occurrence frequencies. This function helps optimize the vector representations, ensuring that words appearing in similar contexts are closer together in the vector space.

(1)Jloss=∑i,j‍f(pi,j)(logpi,j-wi,wj)2.

where, f(pi,j) represents the weight function used to balance the different co-occurrence frequencies of word pairs, wi and wj denote the word vectors for word i and j, respectively.

Equation (1) was optimized via mini-batch stochastic gradient descent to learn high-quality clinical feature embeddings. The GloVe model was trained in a fully unsupervised manner without requiring manual labels, leveraging global statistical patterns inherent in the corpus. By minimizing the weighted least-squares objective, the algorithm effectively captures semantic similarities and relational structures among clinical features, ensuring that terms with similar clinical contexts are mapped to proximal locations in the high-dimensional vector space.

TextCNN-based deep neural network for clinical feature analysis

Following GloVe-based word vectorization, a TextCNN was employed to perform deep feature analysis on the 20 predefined clinical features, with the goal of uncovering intricate interdependencies and patterns within the dataset. The TextCNN architecture is composed of three core modules: Convolutional layers, pooling layers, and fully connected layers, which are responsible for local feature extraction, dimensionality reduction, and final classification mapping, respectively[16].

The workflow of the TextCNN model is schematically illustrated in Figure 4B. The input layer consists of the vectorized clinical feature matrix, with a defined dimension of 20 × 30 (20 clinical features × 30-dimensional GloVe embeddings). Subsequently, 1-dimensional convolutional layers were utilized to extract hierarchical deep features. Multiple convolutional kernels with varying window sizes (2, 4, and 5) were applied to scan vertically across the word-vector sequence, enabling the capture of local contextual relationships of different lengths. This convolution operation generated feature maps with dimensions of 19, 17, and 16 for kernel sizes 2, 4, and 5, respectively. The rectified linear unit activation function was then applied to introduce non-linearity, thereby enhancing the model’s capacity to learn complex feature representations.

Average pooling was adopted to down-sample the high-dimensional feature maps, retaining the most salient and representative features while reducing computational complexity. After sequential processing through the convolutional, pooling, and fully connected layers, the model generated compact feature representations and quantified the underlying correlations between the input word vectors. Finally, a fully connected layer was constructed based on the pooled output and the number of target classification categories. The softmax activation function was employed to compute the normalized probability distribution across output classes, facilitating the final binary classification of clinical significance.

Construction of the self-organizing competitive neural network

Competitive neural networks represent a class of unsupervised learning models capable of autonomously extracting intrinsic patterns from unlabeled input data and performing data classification or clustering via competitive learning mechanisms[17]. Through this process, the network identifies salient feature distributions within the dataset, enabling robust pattern recognition and classification. The output of the competitive neural network is derived from the output of TextCNN, consisting of 20 neurons. We defined the competitive layer with 2 neurons, resulting in output samples represented as a binary vector, where values of 0 and 1 indicate the classification, with 1 denoting abnormal samples that require further examination and treatment. The calculations for the neurons in the competitive layer are performed as follows: sn=∑m=120‍ωm,nxm,n, (2), where, sn represents the n-th element of the input sample. According to the competitive mechanism, sk denotes the neuron in the competitive layer with the maximum value, and it can be defined as sk=1sk>sm0else.

The weight update mechanism is

(3),ωm,n=ωm,n+γxmΔ-ωm,n,

where, γ represents learn rate. Δ is a weight gain.

Baseline models and statistical comparison

Baseline comparisons with logistic regression, random forest, a traditional artificial neural network, and a standalone TextCNN were not completed in the original study. Therefore, the incremental performance of the hybrid DSomNN architecture cannot be definitively established from the available data. These baseline comparisons require re-training and verification and will be addressed in future work.

RESULTS

Analysis of the constructed dataset

In this section, we analyzed the construction of the dataset and the performance of the DSomNN algorithm. We focused on correlations among symptoms, signs, and endoscopy reports, as well as interrelationships among the 20 clinical features. For validation of the DSomNN algorithm, we assessed diagnostic accuracy and clinical effectiveness. Table 1 summarizes the baseline characteristics, including clinical signs, main symptoms, and capsule endoscopy findings, for all 7390 enrolled patients. Regarding primary symptoms, abdominal pain (26.87%) and abdominal bloating (22.15%) were the most prevalent complaints. On capsule endoscopy, the most frequent diagnoses were superficial gastritis (n = 4229, 57.23%), bile reflux (n = 2295, 31.06%), and non-atrophic gastritis (n = 635, 8.59%).

Table 1 Statistical analysis of the constructed dataset.

		Symptoms
		AP	AD	ADis	H	AR	GERD	Hem
Feature	Mean	810	341	714	60	193	113	121
	Women	1176	417	923	74	265	134	94
	Total	1986	758	1637	134	458	247	215
	Age interval	9-99	14-92	12-92	23-87	13-88	13-92	22-93
Report	GC	11	4	11	1	1	3	0
	AG	35	21	53	9	16	9	7
	GU	146	41	89	6	26	10	30
	GP	217	111	204	15	45	42	28
	NAG	203	77	202	23	61	57	12
	SG	1618	580	1242	103	347	155	184
	AGM	83	28	65	8	25	12	15
	YWM	60	20	47	8	15	10	11
	VV	4	1	0	1	1	0	3
	BS	51	19	34	4	11	3	40
	AGA	157	85	191	22	66	27	30
	OGA	20	11	11	5	7	3	2
	BR	820	328	721	53	207	103	63

AP: Abdominal pain; AD: Abdominal distension; ADis: Abdominal discomfort; H: Heartburn; AR: Acid regurgitation; GERD: Gastroesophageal reflux; Hem: Bleeding; GC: Gastric cancer; AG: Atrophic gastritis; GU: Gastric ulcer; GP: Gastric polyps; NAG: Non-atrophic gastritis; SG: Erosive gastritis; AGM: Abnormal gastric mucosa; YWM: Yellow-white macules; VV: Varices; AGA: Abnormal gastric antrum; OGA: Other gastric abnormalities; BR: Bile reflux.

We further analyzed the symptom profiles across specific gastric pathologies. Among 35 patients with GC, the predominant symptoms were abdominal pain (n = 11, 31.43%), abdominal discomfort (n = 11, 31.43%), bloating (n = 4, 11.43%), and GERD (n = 3, 8.57%). Notably, nearly one-third of patients with GC presented with abdominal pain.

For atrophic gastritis (n = 149), the leading symptoms were discomfort (n = 53, 35.57%), abdominal pain (n = 35, 23.49%), bloating (n = 21, 14.09%), and acid regurgitation (n = 16, 10.74%).

In patients with gastric ulcers (n = 384), the primary symptoms included abdominal pain (n = 146, 38.02%), discomfort (n = 89, 23.18%), bloating (n = 41, 10.68%), and gastrointestinal bleeding (n = 30, 7.81%).

Performance validation of the DSomNN algorithm on distinct test datasets

To validate diagnostic performance, we assessed classification accuracy and clinical utility using two validation subsets, as shown in Tables 2 and 3. In Table 2, the classification distributions obtained by the conventional clinical assessment and the neural network were generally comparable in both datasets. In dataset a*, which consisted of 500 patients undergoing conventional gastroscopy, the conventional assessment classified 402 cases as normal and 98 cases as requiring further evaluation, corresponding to a detection rate of 19.60%. The network classified 410 cases as normal and 90 cases as requiring further evaluation. In dataset b*, which consisted of 500 patients undergoing MCCE, the conventional assessment identified 101 cases (20.20%) requiring further evaluation, whereas the network identified 99 cases (19.80%). Overall, the proportions of patients requiring further evaluation identified by the network were close to those obtained by the conventional assessment, suggesting good consistency in the overall classification distribution between the two approaches.

Table 2 Accuracy analysis in different dataset.

		Conventional			DSomNN
Type	Number	Normal	Need	Rate, %	Normal	Need	Rate, %
a∗	n = 500	402	98	19.60	410	90	18.00
b∗	n = 500	399	101	20.20	401	99	19.80

The data of a∗ samples from intubation gastroscope, and the data of b∗ samples from magnetically controlled capsule endoscopy. DSomNN: Deep self-organizing competitive neural network.

Table 3 Accuracy and available performance metrics in different datasets.

Dataset	Reference standard	n	Expert-identified significant cases, %	Correctly identified	Sensitivity
Test Set a	Conventional endoscopy/pathology	500	98 (19.60)	92	93.88
Test Set b	MCCE	500	101 (20.20)	94	93.07

MCCE: Magnetically controlled capsule endoscopy.

In Table 3 Test Set a comprised patients who underwent conventional endoscopy with available pathology/reference data. Of 500 patients, gastroenterologists identified 98 (19.60%) requiring further evaluation and targeted treatment. The DSomNN flagged 92 cases as high-risk. Among the 98 clinically indicated patients, the algorithm correctly identified 92, achieving a sensitivity of 93.88%. Test Set b consisted of 500 patients who underwent MCCE, for whom IM data were unavailable. Specialists determined that 101 patients (20.20%) needed additional evaluation; the model correctly identified 94. The overall classification accuracy for Test Set b was 94.95%.

Longitudinal predictive analysis and clinical feedback

Following the initial validation, we applied the trained DSomNN model to perform retrospective predictive analysis and clinical feedback evaluation on patient data collected consecutively from 2016 to 2022. The annual patient cohorts from 2017 to 2022 were as follows: 732 (2017), 1439 (2018), 1714 (2019), 1310 (2020), 1275 (2021), and 700 (2022). The remaining 220 patients enrolled between July and December 2016 were included in the overall dataset but were not included in the annual trend analysis because they represented a partial year. The algorithm identified patients requiring further diagnostic workup as: 146 (19.95%), 275 (19.11%), 334 (19.49%), 256 (19.54%), 247 (19.37%), and 134 (19.14%), respectively (Table 4). Across all six full years, the proportion of patients flagged for further investigation remained stable and consistent (ranging from 19.11% to 19.95%), demonstrating the robustness and high predictive accuracy of the DSomNN algorithm in a real-world, longitudinal clinical setting.

Table 4 Regression testing.

Year	Test samples	Detection number	Detection rate, %
2017	732	146	19.95
2018	1439	275	19.11
2019	1714	334	19.49
2020	1310	256	19.54
2021	1275	247	19.37
2022	700	134	19.14

DISCUSSION

In recent years, machine learning has been extensively applied in endoscopic diagnostics, demonstrating remarkable performance in assisting the detection of esophageal cancer, GC, and other gastrointestinal disorders[18,19]. However, prevailing approaches predominantly rely on the analysis of endoscopic images, where compromised image quality and localized sampling errors may substantially compromise diagnostic accuracy. Although specialist physicians can render more holistic clinical judgments, their assessments remain heavily reliant on specialized expertise and subjective experience. Accordingly, the present study proposes an innovative framework that integrates patient signs, primary complaints, and endoscopic report data to construct a comprehensive multi-dimensional dataset. We further develop a novel DSomNN and train it on data from 7390 patients, encompassing 147800 data points across 20 clinical features collected between 2016 and 2022. By deciphering deep feature representations and inherent inter-feature correlations, our model classifies patients according to their need for further diagnostic evaluation or therapeutic intervention. The primary objective is to classify patients according to the presence of CSGL and to support clinical decision-making regarding further diagnostic evaluation or therapeutic intervention.

Several novel aspects of this investigation merit emphasis. First, to the best of our knowledge, this is the first study to concurrently incorporate patient signs, chief complaints, and endoscopic report features into a unified dataset. The construction of an auxiliary diagnostic model using 7390 patient records with a total of 147800 data points provides a rich, multi-modal data resource for machine learning-based gastric disease analysis. Second, this research represents the initial application of a hybrid deep learning model, combining TextCNN with a self-organizing competitive network, for the auxiliary diagnosis of gastric diseases. This integrated architecture is designed to enhance both the efficiency and accuracy of specialist evaluations. Third, the established analytical model achieves robust predictive performance, with a sensitivity of 93.88% in Test Set a and 93.07% in Test Set b, and an overall classification accuracy of 94.95% in Test Set b, signifying a substantial improvement in automated diagnostic efficiency.

The feature composition of the dataset is a critical determinant of model learning efficacy. Existing clinical evidence identifies several significant risk factors for GC[20]. Gastric ulcers carry a malignant transformation rate of 1%-2%, while the rate for atrophic gastritis in China ranges from 1.2% to 7.1%. Furthermore, adenomatous polyps are well-established precancerous lesions of the stomach. Chronic inflammatory conditions, if left untreated or inadequately managed, can progressively escalate the risk of malignant transformation[21]. Therefore, during dataset construction, we intentionally include five key categories of clinically significant pathological indicators: GC, IM, atrophic gastritis, gastric ulcers, and gastric polyps. Moreover, distinct gastric pathologies manifest with characteristic symptom profiles. For example, approximately 38.0% of patients with gastric ulcers presented with abdominal pain, whereas only 6 patients (1.56%) reported heartburn. This underscores a strong correlation between primary patient complaints and specific gastric diseases, justifying the integration of symptomatic data into the predictive model. Additionally, epidemiological studies confirm that patient age and gender are closely associated with the incidence and progression of gastric diseases[21], prompting us to incorporate these demographic variables to ensure a more comprehensive representation of each patient’s clinical profile.

In the validation phase, we initially perform a descriptive statistical analysis of data from 7390 patients collected from 2016 to 2022. 3623 patients (49.03%) reported abdominal pain or discomfort, representing the most prevalent symptom complex. Patients with clinically significant lesions, including GC, frequently exhibit symptoms such as abdominal pain and bloating, confirming a clear clinical correlation between symptomatology and pathology. Sex differences in symptom presentation and clinical outcomes warrant further prospective investigation.

Limitations

First, only 496 patients (approximately 6.7%) in the final cohort had pathological gold-standard information; for the remaining patients, CSGL labels relied on specialist judgment. Moreover, because MCCE reports were used both for labeling and as model inputs, the estimated performance may be affected by label, verification, and incorporation bias. Therefore, the model should be regarded as an auxiliary tool rather than a substitute for pathological diagnosis. Second, this was a single-center study without independent external validation, which may limit generalizability. Third, baseline models such as logistic regression, random forest, traditional neural networks, and standalone TextCNN were not fully evaluated, and the incremental performance of DSomNN cannot yet be definitively established. Fourth, complete performance metrics, including specificity, positive predictive value, negative predictive value, f1 score, area under the curve, and 95%CI, require further verification. Future multicenter studies with broader pathological coverage, comprehensive model evaluation, and external validation are needed.

CONCLUSION

This study developed a DSomNN integrating GloVe-based feature representation, TextCNN-based deep feature extraction, and self-organizing competitive learning for the auxiliary identification of CSGL using demographic characteristics, clinical symptoms, and MCCE findings. The model demonstrated high sensitivity in two internal test sets, achieving 93.88% in Test Set a and 93.07% in Test Set b, indicating its potential to support the identification of patients requiring further diagnostic evaluation. These findings suggest that integrating multidimensional clinical information with deep learning may provide a useful approach for MCCE-assisted clinical decision-making. However, further validation using baseline model comparisons, complete diagnostic performance metrics, larger pathological reference cohorts, and independent multicenter datasets is required before clinical implementation.