top of page

4 November 2026

12

min read

Conventional Statistical and Machine-Learning Evaluation of a Public Gastric Cancer Dataset: A Data-Quality and Predictive-Modeling Cautionary Study

A comprehensive evaluation of a public gastric cancer dataset reveals major data-quality concerns and highlights why reliable clinical labels and validated data are essential for developing effective AI-based cancer prediction models.

A comprehensive evaluation of a public gastric cancer dataset reveals major data-quality concerns and highlights why reliable clinical labels and validated data are essential for developing effective AI-based cancer prediction models.

Updated: 

4 November 2026

Abstract


Background: Artificial intelligence applications are ever more prevalent in cancer-risk prediction diagnostic classification and biomarker discovery for increasingly frequent global malignancy gastric cancer. The quality of datasets used to create predictors and outcome markings meant they are coherent and clinically logical.


Objective: The aim of the present study was to analyze a public gastric cancer dataset using various statistical models.


Methods: There were a total of 212,354 observations that were assessed for their secondary analysis. Using descriptive statistics, bivariate testing, odds ratios, and multivariable logistic regression, conventional analysis. The stratified five-fold cross-validation was utilized to evaluate the machine-learning models.


Results: Positive gastric cancer labels were 20,959 thereby a prevalence of 9.87%. There was minimal association of cancer status with clinical, diagnostic variables and molecular score. Clinical models and the broader machine learning models stayed at the same level as chance, while decision-curve analysis indicated that there was no clinical net benefit.


Conclusion: Although the dataset is large and complete, it does not carry a meaningful predictive signal for the classification of gastric cancer.  Without independent verification of labels, definitions of variables, and provenance of the data, it is unsuitable for clinically valid modelling for gastric cancer prediction.


Keywords: Gastric cancer; machine learning; logistic regression; prediction model; data quality; decision-curve analysis; artificial intelligence


Introduction


Gastric cancer is one of the important gastrointestinal cancers (1). The global cancer statistics commonly demonstrate significant incidence and mortality of gastric cancer in identifying regions (2). These regions are also with major Helicobacter pylori, undue delay in diagnosis, dietary risk factors and unavailability of early screening (3). Endoscopy, histopathological confirmation, and imaging-based staging are the cornerstones of current clinical practice (4). As a result, in a clinically credible gastric cancer dataset, diagnostic variables as in a biopsy, endoscopy and CT scan findings should clearly correlate with cancer status (4).


Machine learning is being increasingly used in many areas of oncology including diagnosis, prognosis, molecular classification in gastric cancer (5-9). Research has also been conducted into microRNAs and other molecular markers in terms of their status as candidate diagnostic (10-14). Yet, large sample size alone does not suffice for prediction modeling. ...these are essential: transparent data provenance, valid outcome definition, appropriate validation, calibration, and assessment of clinical usefulness (15-25).


The present study analyzed a public gastric cancer dataset using a stepwise approach beginning with conventional statistics and progressing toward machine-learning validation. The aim was to determine whether the dataset contains clinically meaningful predictive signal for gastric cancer classification.


Methods


Study Design and Dataset


Involving a public gastric cancer dataset, it was a secondary analytical study. The sequence of analysis was data audit, descriptive analysis, bivariate analysis, multivariable logistic regression, machine-learning modelling, subgroup validation and clinical-utility analysis.


There were 212354 observations and 29 variables in the data set. Included were demographic factors, clinical risk factors, diagnostic indicators, molecular prediction scores, molecular identifiers, and a gastric cancer binary label.


Classification of Predictors and Outcome Variables in the Gastric Cancer Dataset


Table (1) shows the configuration of the dataset used for the analysis of gastric cancer. The categories of the various variables were demographic, clinical, diagnostic, molecular, derived molecular, and outcome. The study population was characterized using demographic and clinical variables and an assessment of conventional risk-factor associations. Diagnostic variables such as endoscopy, biopsy and CT scan findings were judged to be clinically important because, as a group, they are expected to relate directly to cancer status. The predicted sum and all sum two derived molecular summary variable while miRNA-related identifiers and predicted score include molecular variable. Regression and machine-learning analyses employed binary gastric cancer label as the primary outcome.


Table 1: Dataset structure

Domain

Variables

Demographic

Age, gender, ethnicity, geographical location

Clinical

Family history, smoking, alcohol, H. pylori infection, dietary habits, existing conditions

Diagnostic

Endoscopic images, biopsy results, CT scan

Molecular

miRNA accession, miRNA ID, target symbol, DIANA-microT, EIMMo, MicroCosm, miRanda, miRDB, PicTar, PITA, TargetScan

Derived molecular

predicted.sum, all.sum

Outcome

Binary gastric cancer label

 

Statistical Equations


·       Prevalence = (Positive cases / Total observations) x 100

·       Log [p / (1 - p)] = beta0 + beta1X1 + beta2X2 + ... + betakXk

·       Adjusted odds ratio = exp(beta)

·       Balanced accuracy = (Sensitivity + Specificity) / 2

·       Net benefit = TP/N - (FP/N) x [pt / (1 - pt)]


Machine-Learning Analysis


Categorical predictors were one-hot encoded. Standardization of continuous variables within training folds. Five-fold cross-validation stratified. The following models were included: logistic regression, Gaussian naive Bayes, shrinkage linear discriminant analysis, Extra Trees and a negative-class baseline. The performance was assessed via measures like ROC-AUC, PR-AUC, balanced accuracy, calibration, and clinical net benefit measures (24-31).


Results


Data Quality and Class Distribution


The first audit of data quality and their outcomes is shown in Table (2) and Figure (1). The dataset consists of 212,354 observations and 29 variables. No missing values. No exact duplicate records. The label for gastric cancer in a binary way was imbalanced due to 191,395 and 20,959 negative and positive observations in the dataset respectively which resulted in a positive prevalence of 9.87%. This imbalance suggests that crude accuracy is not a good measure of model performance because a majority-class classifier would achieve high apparent accuracy and fail to detect cancer positives. For this reason, the balanced accuracy, sensitivity, ROC AUC, PR AUC and calibration metrics in the subsequent analysis was examined.


Table 2: Overview of Dataset Completeness and Class Distribution

Characteristic

Result

Total observations

212,354

Total variables

29

Missing values

0

Exact duplicates

0

Negative labels

191,395

Positive labels

20,959

Positive prevalence

9.87%



Figure 1: Class distribution showing marked imbalance, with positive gastric cancer labels representing approximately 9.87% of observations.


Descriptive characteristics of participants


Table (3) gives a summary of descriptive statistics of the dataset. The average age of the subjects was 53.26 ± 18.98 years. The gender was distributed as males (70.03%) and females (29.97%). The subgroup of ethnicity A was the most frequently represented followed by ethnicity B and ethnicity C. Most observations were geographically coded California, while about one-fifth were coded other. The positive outcome for gastric cancer was 9.87% confirming a class-imbalanced outcome structure.


Table 3: Selected descriptive characteristics.


Variable

Main finding

Age

Mean 53.26 +/- 18.98 years

Gender

Males: 70.03%; females: 29.97%

Ethnicity

A: 60.08%; B: 29.93%; C: 9.99%

Location

California: 79.87%; Other: 20.13%

Positive cancer label

9.87%

 

Comparison of Age Distribution Between Gastric Cancer Label Groups


The age-density curves for cases with and without cancer are nearly fully overlapping for all ages.  It meant that age did not significantly separate the two label groups and was unlikely to contribute useful predictive information in regression or machine learning models.



Figure 2: Age distributions by cancer label, demonstrating near-complete overlap between positive and negative cases.


Bivariate Clinical Analysis

The unadjusted clinical associations with gastric cancer is estimated in Table (4). The positive records and the negative records were of almost the identical age, and the difference was not significant. Same, gender, smoking, and alcohol consumption. The odds ratios for H. pylori infection, dietary habits, and family history were near 1.0, indicating little association with cancer status. While family history was close to having a statistically significant impact, its actual effect was minor in size and not consistently heading in the expected risk direction. Across the bivariate clinical analysis, any interpretation of clinically meaningful separation on cancer-positive and cancer-negative records would support an outcome label that is poorly informed by established gastric cancer risk factors.

Table 4. Bivariate clinical associations.

Predictor

Comparison

OR or difference

p-value

Interpretation

Age

Positive vs negative

53.14 vs 53.27 years

0.346

No difference

Gender

Males vs females

OR 0.992

0.634

No association

Family history

Yes or no

OR 0.970

0.058

Very weak; not significant

Smoking

Yes vs no

OR 1.018

0.217

No association

Alcohol

yes vs no

OR 0.993

0.626

No association

H. pylori

yes vs no

OR 0.980

0.222

No association

Dietary habits

Yes vs no

OR 1.009

0.630

No association

 

Clinical Category-Specific Cancer Prevalence in the Dataset

Across all investigated clinical and demographic categories, the positive cancer-label prevalence remained near the overall prevalence of 9.87%. The lack of any substantial variation indicates that these predictors could not differentiate between cancer-positive and cancer-negative records and therefore offer little information for posterior regression or machine-learning models (Figure 3).




Figure 3: Cancer-label prevalence across clinical categories, showing minimal variation around the overall prevalence.


Diagnostic and Molecular Analysis


Table (5) exhibits the non-adjusted associations between the diagnostic variables and the gastric cancer label. There were no associations of interest between endoscopic image category, biopsy result, and CT scan status with cancer status as odds ratios were close to 1.0 with non-significant p-values. The study’s finding is clinically unexpected because endoscopy and biopsy are the cornerstones of diagnosis and would be expected to be strongly associated with confirmed cancer status. Lack of a diagnostic signal thus raises serious concerns regarding label validity, variable coding, or data provenance.


Table 5: Diagnostic variable associations


Variable

Comparison

OR

p-value

Interpretation

Endoscopic image category

yes vs no

1.004

0.812

No association

Biopsy result

yes vs no

1.009

0.723

No association

CT scan

yes vs no

0.982

0.326

No association

 

Standardized Mean Differences in miRNA-Related Molecular Scores by Cancer Status


Table (6) summarizes standardized mean differences in molecular prediction scores, grouped according to gastric cancer label. The SMD values for all the molecular variables were nearly zero, indicating identical distributions for cancer-positive and cancer-negative records. There was the greatest difference in PITA, but the effect size is trivial and not likely a biological signal. This evidence indicates that the classification of gastric cancer in the present dataset was not enriched by miRNA-related molecular scores and derived summary variables.


Table 6: Molecular score findings.


Variable

Main result

Interpretation

DIANA-microT

SMD 0.006

Negligible difference

EIMMo

SMD 0.000

No difference

MicroCosm

SMD 0.004

Negligible difference

miRanda

SMD 0.001

No difference

miRDB

SMD -0.008

Negligible difference

PicTar

SMD -0.001

No difference

PITA

SMD 0.017

Trivial nominal signal

Target Scan

SMD -0.003

No difference

Predicted sum

SMD -0.003

No difference

All sum

SMD -0.008

No difference

 

Multivariable Logistic Regression


The multivariable clinical logistic regression model predicting gastric cancer label is provided in Table (7). After adjustment, all displayed odds ratios were imprecise. All 95% confidence intervals crossed the null value. H. pylori and smoking, gender, age, alcohol consumption wee not associated with gastric cancer. The effect of family history was very small and unexpectedly had the wrong direction. According to these findings, the clinical variables did not yield meaningful independent predicting information in relation to gastric cancer class.

Table 7: Clinical logistic regression

Predictor

Adjusted OR

95% CI

p-value

Age per 10 years

0.996

0.989-1.004

0.343

Gender

0.993

0.962-1.024

0.651

Family history

0.970

0.940-1.001

0.058

Smoking

1.018

0.989-1.048

0.221

Alcohol

0.993

0.965-1.022

0.632

H. pylori

0.980

0.948-1.013

0.224

 

Clinical Logistic Regression Model Fit and Predictive Performance


Table (8) provides a summary of the clinical logistic regression model fit and predictive performance. The statistical analysis of the likelihood-ratio test indicated it was not statistically significant. This means that clinical predictors do not appear to improve model fit when compared with the null model. The pseudo-R2 from Nagelkerke is near zero which indicates that the model explains essentially no variation in the gastric cancer. Cross-validated ROC-AUC was 0.501, while the Brier score showed no contribution beyond the baseline outcome prevalence. According to these findings, the clinical logistic regression model in this study did not yield meaningful predictive value for gastric cancer classification.


Table 8: Clinical logistic model-level performance


Metric

Result

Likelihood-ratio p-value

0.358

Nagelkerke pseudo-R2

0.00013

Cross-validated ROC-AUC

0.501

Brier score

0.0890

 

Multivariable Clinical Predictor Effects for Gastric Cancer Classification


Figure (5) shows adjusted clinical odds ratios from the multivariable logistic regression model. Blue points represent adjusted odds ratios, and horizontal lines indicate 95% confidence intervals. The red vertical line represents the null value of OR = 1.0. All estimates were close to the null value and their confidence intervals overlapped 1.0, indicating that the clinical variables did not provide meaningful independent predictive information for gastric cancer label classification.



Figure 4: Adjusted clinical odds ratios from multivariable logistic regression; all estimates are close to the null value.


Machine-Learning Performance


Table (9) summarizes the performance of machine-learning five-fold cross-validated gastric cancer label prediction. All models demonstrated a chance-level ability to discriminate between the two groups since their ROC-AUC values were near 0.50 and balanced accuracy values were near 50%. Even though the negative-class baseline achieved high crude accuracy, its sensitivity was zero, thus not clinically useful. The model's metrics of Precision and PR-AUC were also near the prevalence of positive labels suggesting none of the models enriched for cases positive for cancer. These results show that neither linear nor non-linear machine-learning methods were able to extract a meaningful predictive signal from available clinical, diagnostic or molecular variables.


Table 9: Five-fold cross-validated machine-learning performance.


Model

Accuracy

Balanced accuracy

Sensitivity

Specificity

Precision

F1

ROC-AUC

PR-AUC

Logistic regression, clinical

64.63%

50.15%

32.10%

68.19%

9.97%

15.10%

50.16%

9.95%

Logistic regression, expanded

50.95%

49.80%

48.37%

51.23%

9.76%

16.00%

49.85%

9.85%

Gaussian naive Bayes

46.82%

49.67%

53.23%

46.12%

9.74%

16.32%

49.73%

9.78%

Shrinkage LDA

47.97%

49.88%

52.25%

47.50%

9.79%

16.23%

49.86%

9.85%

Extra Trees

54.32%

50.16%

44.98%

55.34%

9.93%

16.23%

50.15%

9.91%

Negative-class baseline

90.13%

50.00%

0.00%

100.00%

0.00%

0.00%

50.00%

9.87%

 

Comparison of ROC-AUC, Balanced Accuracy, and PR-AUC Across Models


ROC-AUC and balanced accuracy were nearly 50% for all models evaluated, indicative of chance level. PR-AUC values were nearly the same as the prevalence of positive-label (9.87%), which indicates that the models did not meaningfully enrich cancer-positive cases. This indicates that machine-learning algorithms did not detect useful predictive structure in clinically, diagnostically or molecularly available variables (Figure 5).


 

Figure 5: Five-fold cross-validated machine-learning performance; ROC-AUC and balanced accuracy remain near chance.


Model Robustness, Subgroups, and Clinical Utility


As can be seen in Table (10), advanced validation results for the clinical prediction model are mentioned. The out-of-fold ROC-AUC was 0.5013 (95% CI crossing 0.50), indicating chance-level discrimination. The PR-AUC approximated the genuine prevalence of cancer labels, suggesting no enrichment of positives. Permutation testing found that model performance did not differ significantly from random-label performance. The Brier score was like that from a prevalence-only model, and the calibration slope was far below the ideal of 1.0, suggesting uninformative probability estimates. Also, it may have been that there was no predictive signal present, and not an insufficient sample size, that is, larger training samples led to no improvement on learning-curve analysis.


Table 10: Advanced validation results.


Validation method

Result

Interpretation

Out-of-fold clinical ROC-AUC

0.5013

Chance-level

95% CI

0.4972-0.5054

Includes 0.50

PR-AUC

0.0991

Same as prevalence

Permutation p-value

0.271

Not better than random labels

Brier score

0.0890

Similar to prevalence-only model

Calibration slope

0.127

Non-informative prediction

Learning curve

No improvement with larger n

No hidden signal

 

ROC-AUC Across Increasing Training Sample Sizes


At greater numbers of training samples, ROC-AUC remained close to the level of chance-performance, or around 50%. A lack of performance gain indicates that weak discrimination of the model didn’t occur due to insufficient sample size, but because clinical variables, at least those available, have no strong predictive signal (Figure 6).



Figure 6: Learning-curve analysis; increasing training size does not improve ROC-AUC.

Subgroup Validation of the Gastric Cancer Prediction Model


Table (11) shows subgroup ROC-AUC analysis of clinical prediction model. Its ROC-AUC values were close to 50% for all gender, ethnicity, age-group, and geographic subgroups. No subgroup demonstrated an enhancement in model performance or a clinically useful distinction between cancer-positive and cancer-negative records. The absence of a predictive signal across multiple major demographic and geographic categories rather than just the overall sample is suggested by these results.


Table 11: Subgroup ROC-AUC summary

Subgroup

ROC-AUC

Interpretation

Gender 1

50.34%

Chance-level

Gender 2

49.60%

Chance-level

Ethnicity A

50.17%

Chance-level

Ethnicity B

50.20%

Chance-level

Ethnicity C

48.98%

Chance-level

Age 20-39

49.59%

Chance-level

Age 40-59

50.29%

Chance-level

Age 60-89

50.28%

Chance-level

California

50.15%

Chance-level

Other location

49.95%

Chance-level

 

Forest Plot of Subgroup ROC-AUC Values with 95% Confidence Intervals


As shown in Figure (7), the blue points reflect the subgroup ROC-AUC estimates and the horizontal lines represent the 95% confidence intervals. The red vertical line indicates discrimination by chance (50%). The confidence intervals for each subgroup superseded the chance line, indicating that the model did not display meaningful discriminatory performance for any demographic, age, ethnic, or geographic subgroup.

 


Figure 7: Subgroup ROC-AUC estimates with 95% confidence intervals; all subgroups overlap chance-level discrimination


Clinical Utility Assessment Using Decision-Curve Analysis


Table )12) gives an overview of the decision-curve analysis of the clinical prediction model at selected risk thresholds. At thresholds of 5.0% and 7.5%, net benefit from the model was exactly similar to the treat-all strategy, indicating no value added. When evaluating the threshold of 10.0%, the model net benefit was slightly negative, and at higher thresholds, it became similar to treat-none. The results suggest that the model was not clinically useful in guiding gastric cancer classification or risk-based decision-making in this cohort.


Table 12: Decision-curve analysis


Risk threshold

Model net benefit

Treat-all net benefit

Treat-none net benefit

5.0%

0.0513

0.0513

0

7.5%

0.0256

0.0256

0

10.0%

-0.0005

-0.0014

0

12.5%

0.0000

-0.0301

0

15.0%

0.0000

-0.0604

0

20.0%

0.0000

-0.1266

0

 

Decision-Curve Evaluation of Gastric Cancer Model Usefulness


As shown in Figure (8), the prediction model was not associated with a meaningful net benefit across risk thresholds. When the thresholds were low the model performance was similar to that of the treat-all. However, when the thresholds were high the performance was similar to that of the treat-none. The research results indicate the model does not provide any clinically useful decision-making value to gastric cancer risk classification.



Figure 8: Decision-curve analysis; the model provides no meaningful clinical net benefit across evaluated thresholds.


Molecular Sum Audit


The expected molecular sum was defined as the sum of DIANA-microT, EIMMo, MicroCosm, miRanda, miRDB, PicTar, PITA, and TargetScan. Table (13) shows an internal consistency check of the derived molecular variables If calculation of predicted.sum and all.sum is done from the visible molecular prediction algorithm the summed component molecular scores would be expected to correlate with them. Nevertheless, neither showed a significant link with component-score sum. It can be concluded that the obtained molecularly derived variables had little internal consistency and therefore, may not represent a linear combination of the components of the molecular score. Thus, they should be interpreted with caution. The results lead to a more general concern that there is no reliable molecular predictive structure with the dataset for gastric cancer classification.


Table 13: Molecular sum audit.


Variable

Expected relationship

Observed interpretation

predicted.sum

Should correlate with component-score sum

No meaningful correlation

all.sum

Should correlate with component-score sum

No meaningful correlation

 

Molecular Sum-Variable Validation in the Gastric Cancer Dataset

 

As shown in Figure (9), scatterplots are a comparison of the expected sum of the component scores with the reported predicted.sum and all.sum variables. The anticipated alignment of the reported sum variables if they yielded the sum of eight scores by molecular algorithm is shown in a red diagonal line. The wide dispersion as well as the near-zero correlation for both panels suggests that the reported molecular summary variables do not align meaningfully with the component scores.



Figure 9: Reported molecular sum variables are poorly aligned with the sum of component molecular scores.


Discussion


The present analysis shows a critical difference between a large dataset and a clinically useful dataset. With 212,354 observations and no missing values, gastric cancer dataset found that the predictors had an almost zero association with the gastric cancer. All of the descriptive statistics, bivariate analysis, multivariable logistic regression analyses, machine learning models, subgroup validation analyses, learning curve analysis, calibration checks, and decision-curve analyses corroborated the finding. Thus, the key takeaway is to understand that machine learning did not fail in a technical sense; it was the data that was not valid.


A major concern is the lack of association between diagnostic variables and the cancer label. In real-world gastric cancer diagnosis, endoscopy and biopsy are central diagnostic procedures (5-8). Contemporary AI studies in gastric cancer usually focus on endoscopic images, histopathology, or clinically verified diagnostic pathways, and many report strong diagnostic performance when valid image data and confirmed labels are used (9-14). In contrast, the present dataset showed no meaningful relationship between endoscopic image category, biopsy result, CT scan category, and cancer label. This discrepancy strongly suggests a problem with variable coding, label validity, or dataset generation rather than a true absence of diagnostic relevance.


This result is notably different from studies that used image-based or histopathology confirmed gastric cancer datasets. In those studies, the predictive signals are often integrated into the target itself such as the pattern of the endoscopic mucosa, shape of the lesion, architecture of the histology or molecular features. The current data set does not contain image files; it contains variables that provide codes, endoscopic_images, biopsy_results, and ct_scan. These coded variables were not the label for cancer, which means they are not diagnostic predictors.


The clinical-risk-factor findings are also inconsistent with established gastric cancer epidemiology. Previous literature consistently identifies H. pylori infection, age, sex, smoking, diet, family history, and premalignant gastric conditions as relevant gastric cancer risk factors (1-8). In the present dataset, however, H. pylori infection had an adjusted OR close to 1, smoking had an adjusted OR close to 1, and family history showed a weak inverse tendency rather than a biologically expected increased risk. Such findings are unlikely in a clinically valid gastric cancer risk dataset unless the variables are incorrectly coded, randomly generated, or disconnected from the outcome label.


The molecular findings also require caution. MicroRNAs have been widely investigated as gastric cancer biomarkers, including circulating miRNAs and tissue-based miRNA signatures (15-20). However, molecular biomarkers usually require biologically meaningful measurement, quality control, normalization, and validation. In the present dataset, molecular scores such as DIANA-microT, EIMMo, MicroCosm, miRanda, miRDB, PicTar, PITA, and TargetScan showed almost identical distributions between positive and negative labels. Although PITA showed a nominally small association, the effect size was trivial and not convincing as a biological signal.

 

The results of machine learning are therefore to be expected. The ROC-AUC values of logistic regression, Gaussian naive Bayes, shrinkage LDA, and Extra Trees are 0.50. Thus, the models were virtually unable to rank the positive cases above the negative cases. The high crude accuracy of the negative-class baseline was due to the dominance of negative observations, but it proved to be totally insensitive and clinically useless. This shows that accuracy alone can be misleading in medical datasets (24-26, 32, 33).


The results are consistent with contemporary guidance of prediction-models. TRIPOD, TRIPOD+AI, PROBAST, and PROBAST-AI state prediction models should be assessed according to transparent data sources, valid outcome definition, appropriate predictors, complete reporting, calibration, validation, and clinical usefulness (21-24). The current data set does not pass the test of clinical plausibility, as diagnostic and clinical variables do not behave as expected. So, more algorithmic optimizing won’t solve the core problem.


Table (14) compares the current findings with published AI, clinical, molecular and prediction-model literature on gastric cancer. Several studies of endoscopy powered by AI such as Hirasawa et al (14) and Luo et al (25) have documented that deep-learning models can distinguish non-pathological conditions from pathological ones above chance level when trained on real endoscopic images labelled with clinical accuracy. Likewise, clinical guidelines and reviews assert that endoscopy and biopsy are essential for gastric cancer diagnosis. In contrast, the current dataset did not reveal any significant association between endoscopic image category biopsy result CT scan status and the gastric cancer label.

 

The table also contrasts current clinicopathological findings with established epidemiology for gastric cancer. The previous studies recognized age, sex and H. pylori. Gastric cancer risk factors include pylori infection, smoking, diet, family history, and premalignant lesions. Again, these variables did not contribute to the present dataset. Research on molecular-biomarkers has shown some miRNAs are promising diagnostic/prognostic markers for gastric cancer. However, in this dataset, the miRNA-related scores were almost identical, regardless of whether labelled positive or negative. In the end, prediction-model guidance, including TRIPOD, PROBAST, and decision-curve methodology, makes the case that useful models should display valid discrimination, calibration, and clinical net benefit. ROC-AUC of the dataset remained constant at roughly 0.50 with no net clinical benefit.


Table 14: Comparison with published gastric cancer AI and prediction studies.


Area

Findings in previous studies

Findings in present dataset

Interpretation

AI-assisted endoscopy

Deep-learning models using real endoscopic images often report performance clearly above chance (13, 14).

Endoscopic image category was not associated with the cancer label.

Dataset does not behave like clinically valid endoscopy data.

Biopsy/pathology

Histopathology is central for confirming gastric cancer diagnosis (5-8).

Biopsy result had OR close to 1 and p>0.05.

Clinically suspicious; suggests label or coding problem.

Clinical risk factors

Age, sex, H. pylori, smoking, family history, and premalignant lesions are established or commonly studied risk factors (1-4, 8, 11, 12).

These variables showed negligible ar scores were nearly associations.

Clinical signal is absent or labels are unreliable.

Molecular biomarkers

miRNAs have been reported as candidate diagnostic and prognostic biomarkers (15-20).

Molecul identical between positive and negative labels.

Molecular fields do not carry meaningful cancer signal.

Machine-learning performance

Valid models usually exceed baseline when predictors are informative (21-30, 34).

ROC-AUC remained around 0.50 across models.

Machine learning could not learn because signal was absent.

Clinical utility

Useful models should improve decision-making beyond treat-all or treat-none strategies ([31,36,37).

Decision-curve analysis showed no net benefit.

Model is not clinically useful.


Strengths and Limitations


A thorough analyzed step-wise analytical framework was used for the study. This work merged ordinary statistics, logistic regression, machine-learning models, calibration assessment, permutation testing, subgroup analysis, learning-curve analysis, molecular consistency check and decision-curve analysis.


Due to non-availablity of external dataset dictionary, original clinical records, pathology reports, imaging files and independent outcome verification, the analysis was limited. Many variables were coded without clinical definitions. As a result, this study assesses the internal statistical structure of the dataset, not the epidemiology of gastric cancer.


Conclusion


Although the public gastric cancer dataset that we analyzed is large and complete, it does not contain any useful signal.  Diagnostic variables like biopsy, CT imaging, and endoscope were not associated with the cancer label and all machine-learning models performed near chance. Unless independently verified, the dataset should not be used for clinically valid gastric cancer prediction modeling with respect to label quality, variable coding and data provenance. The key takeaway from this case study is the importance of assessing data quality before developing the machine-learning model.


References


  1. Bray F, Laversanne M, Sung H, Ferlay J, Siegel RL, Soerjomataram I, Jemal A. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA: a cancer journal for clinicians. 2024 May;74(3):229-63.

  2. Sung H, Ferlay J, Siegel RL, Laversanne M, Soerjomataram I, Jemal A, Bray F. Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA: a cancer journal for clinicians. 2021 May;71(3):209-49.

  3. Ferlay J, Ervik M, Lam F, Colombet M, Mery L, Piñeros M, Znaor A, Soerjomataram I, Bray F. Global cancer observatory: cancer today. Lyon: International agency for research on cancer. 2020 Nov 25;20182020:557-67.

  4. De Martel C, Georges D, Bray F, Ferlay J, Clifford GM. Global burden of cancer attributable to infections in 2018: a worldwide incidence analysis. The Lancet global health. 2020 Feb 1;8(2):e180-90.

  5. Lordick F, Carneiro F, Cascinu S, Fleitas T, Haustermans K, Piessen G, Vogel A, Smyth EC, ESMO Guidelines Committee. Gastric cancer: ESMO Clinical Practice Guideline for diagnosis, treatment and follow-up. Annals of Oncology. 2022 Oct 1;33(10):1005-20.

  6. Ajani JA, D’Amico TA, Bentrem DJ, Chao J, Corvera C, Das P. NCCN Clinical practice guidelines in oncology: Gastric Cancer. National Comprehensive Cancer Network. 2015.

  7. Japanese Gastric Cancer Association jgca@ koto. kpu-m. ac. jp. Japanese gastric cancer treatment guidelines 2021. Gastric cancer. 2023 Jan;26(1):1-25.

  8. Smyth EC, Nilsson M, Grabsch HI, van Grieken NC, Lordick F. Gastric cancer. The Lancet. 2020 Aug 29;396(10251):635-48.

  9. Hu B, El Hajj N, Sittler S, Lammert N, Barnes R, Meloni-Ehrig A. Gastric cancer: Classification, histology and application of molecular pathology. Journal of gastrointestinal oncology. 2012 Sep;3(3):251.

  10. Joshi SS, Badgwell BD. Current treatment and recent progress in gastric cancer. CA: a cancer journal for clinicians. 2021 May;71(3):264-79.

  11. Machlowska J, Baj J, Sitarz M, Maciejewski R, Sitarz R. Gastric cancer: epidemiology, risk factors, classification, genomic characteristics and treatment strategies. International journal of molecular sciences. 2020 Jun 4;21(11):4012.

  12. Sexton RE, Al Hallak MN, Diab M, Azmi AS. Gastric cancer: a comprehensive review of current and future treatment strategies. Cancer and Metastasis Reviews. 2020 Dec;39(4):1179-203.

  13. Luo H, Xu G, Li C, He L, Luo L, Wang Z, Jing B, Deng Y, Jin Y, Li Y, Li B. Real-time artificial intelligence for detection of upper gastrointestinal cancer by endoscopy: a multicentre, case-control, diagnostic study. The Lancet Oncology. 2019 Dec 1;20(12):1645-54.

  14. Hirasawa T, Aoyama K, Tanimoto T, Ishihara S, Shichijo S, Ozawa T, Ohnishi T, Fujishiro M, Matsuo K, Fujisaki J, Tada T. Application of artificial intelligence using a convolutional neural network for detecting gastric cancer in endoscopic images. Gastric cancer. 2018 Jul;21(4):653-60.

  15. Link A, Kupcinskas J. MicroRNAs as non-invasive diagnostic biomarkers for gastric cancer: Current insights and future perspectives. World Journal of Gastroenterology. 2018 Aug 14;24(30):3313.

  16. Shin VY, Chu KM. MiRNA as potential biomarkers and therapeutic targets for gastric cancer. World journal of gastroenterology: WJG. 2014 Aug 14;20(30):10432.

  17. Chen X, Ba Y, Ma L, Cai X, Yin Y, Wang K, Guo J, Zhang Y, Chen J, Guo X, Li Q. Characterization of microRNAs in serum: a novel class of biomarkers for diagnosis of cancer and other diseases. Cell research. 2008 Oct;18(10):997-1006.

  18. Bartel DP. MicroRNAs: target recognition and regulatory functions. cell. 2009 Jan 23;136(2):215-33.

  19. Wang J, Chen J, Sen S. MicroRNA as biomarkers and diagnostics. Journal of cellular physiology. 2016 Jan;231(1):25-30.

  20. Wei H, Pu K, Liu XG, Li BX, Zhang HS, Wang H, Wang H, Sun WM, Wang YP. The diagnostic value of circulating microRNAs as a biomarker for gastric cancer: a meta-analysis. Oncology reports. 2019 Jan;41(1):87-102.

  21. Moons KG, Altman DG, Reitsma JB, Ioannidis JP, Macaskill P, Steyerberg EW, Vickers AJ, Ransohoff DF, Collins GS. Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD): explanation and elaboration. Annals of internal medicine. 2015 Jan 6;162(1):W1-73.

  22. Collins GS, Dhiman P, Andaur Navarro CL, Ma J, Hooft L, Reitsma JB, Logullo P, Beam AL, Peng L, Van Calster B, van Smeden M. Protocol for development of a reporting guideline (TRIPOD-AI) and risk of bias tool (PROBAST-AI) for diagnostic and prognostic prediction model studies based on artificial intelligence. BMJ open. 2021 Jul;11(7):e048008.

  23. Wolff RF, Moons KG, Riley RD, Whiting PF, Westwood M, Collins GS, Reitsma JB, Kleijnen J, Mallett S, PROBAST Group†. PROBAST: a tool to assess the risk of bias and applicability of prediction model studies. Annals of internal medicine. 2019 Jan 1;170(1):51-8.

  24. Moons KG, Damen JA, Kaul T, Hooft L, Navarro CA, Dhiman P, Beam AL, Van Calster B, Celi LA, Denaxas S, Denniston AK. PROBAST+ AI: an updated quality, risk of bias, and applicability assessment tool for prediction models using regression or artificial intelligence methods. bmj. 2025 Mar 24;388.

  25. Luo W, Phung D, Tran T, Gupta S, Rana S, Karmakar C, Shilton A, Yearwood J, Dimitrova N, Ho TB, Venkatesh S. Guidelines for developing and reporting machine learning predictive models in biomedical research: a multidisciplinary view. Journal of medical Internet research. 2016 Dec 16;18(12):e323.

  26. Liu X, Rivera SC, Moher D, Calvert MJ, Denniston AK, Ashrafian H, Beam AL, Chan AW, Collins GS, Deeks AD, ElZarrad MK. Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: the CONSORT-AI extension. The Lancet Digital Health. 2020 Oct 1;2(10):e537-48.

  27. Rivera SC, Liu X, Chan AW, Denniston AK, Calvert MJ, Ashrafian H, Beam AL, Collins GS, Darzi A, Deeks JJ, ElZarrad MK. Guidelines for clinical trial protocols for interventions involving artificial intelligence: the SPIRIT-AI extension. The Lancet Digital Health. 2020 Oct 1;2(10):e549-60.

  28. Vasey B, Nagendran M, Campbell B, Clifton DA, Collins GS, Denaxas S, Denniston AK, Faes L, Geerts B, Ibrahim M, Liu X. Reporting guideline for the early stage clinical evaluation of decision support systems driven by artificial intelligence: DECIDE-AI. bmj. 2022 May 18;377.

  29. Mongan J, Moy L, Kahn Jr CE. Checklist for artificial intelligence in medical imaging (CLAIM): a guide for authors and reviewers. Radiology: Artificial Intelligence. 2020 Mar 25;2(2):e200029.

  30. Sounderajah V, Ashrafian H, Golub RM, Shetty S, De Fauw J, Hooft L, Moons K, Collins G, Moher D, Bossuyt PM, Darzi A. Developing a reporting guideline for artificial intelligence-centred diagnostic test accuracy studies: the STARD-AI protocol. BMJ open. 2021 Jun 1;11(6):e047709.

  31. Vickers AJ, Elkin EB. Decision curve analysis: a novel method for evaluating prediction models. Medical Decision Making. 2006 Nov;26(6):565-74.

  32. Saito, T., & Rehmsmeier, M. (2015). The precision-recall plot is more informative than the ROC plot when evaluating binary classifiers on imbalanced datasets. PloS one, 10(3), e0118432.

  33. Davis J, Goadrich M. The relationship between Precision-Recall and ROC curves. InProceedings of the 23rd international conference on Machine learning 2006 Jun 25 (pp. 233-240).

  34. Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, Blondel M, Prettenhofer P, Weiss R, Dubourg V, Vanderplas J. Scikit-learn: Machine learning in Python. the Journal of machine Learning research. 2011 Nov 1;12:2825-30.

  35. Glenn WB. Verification of forecasts expressed in terms of probability. Monthly weather review. 1950 Jan;78(1):1-3.

  36. Steyerberg EW. Clinical Prediction Models. 2nd edn. Springer Nature Switzerland, 2019.

  37. Van Calster B, McLernon DJ, Van Smeden M, Wynants L, Steyerberg EW. Calibration: the Achilles heel of predictive analytics. BMC medicine. 2019 Dec;17(1):230.

  38. Riley RD, Ensor J, Snell KI, Harrell FE, Martin GP, Reitsma JB, Moons KG, Collins G, Van Smeden M. Calculating the sample size required for developing a clinical prediction model. Bmj. 2020 Mar 18;368.

Tags

Share This Page  on:  

​

​

ba5ce40f2f63d62c0d1604efd8628dcaa1910072.png
51cffa6ef17e6d092f78200435a6055df6b758c8.png
f264ab5cc2757f8fb5b333dcb8cd42905db961aa.png
1770449288b6b323310c7fc549b511399421d785.png

​

Our Authors

Get the Latest Healthcare
Stories in Your Inbox.

Subscribe to our newsletter to receive new articles, case insights, and thought leadership from clinicians and researchers worldwide.

bottom of page