Write a comment

This review is the result of a virtual, collaborative Live Review discussion organized by one of PREreview Champions on January 16, 2026. The discussion was joined by 5 people: 1 facilitator and 4 live review participants. The authors of this review have dedicated additional asynchronous time over the course of 2 weeks to help compose this final report using the notes from the Live Review. Special thanks to all participants who contributed to the discussion and made it possible to provide feedback on this preprint.

Summary:

In this research, the author investigates how Artificial Intelligence models, like Machine Learning and Deep Learning, can assist in the early diagnosis of diabetes in healthcare. The study uses three classical Machine Learning (ML) models; Decision Tree, Random Forest, and Support Vector Machine (SVM), alongside a Deep Learning (DL) model which was implemented using a Neural Network.

The results show that the Deep Learning model (Neural Network) achieved the highest accuracy, while the Random Forest model also performed strongly but required less computational resources. One interesting part of the study is the use of Neural Networks, which are well suited for learning complex, non-linear patterns hidden in datasets.

The study emphasizes the importance of early diabetes detection, as this can lead to better prognosis. This aligns with existing literature and supports the idea that AI-generated data outputs can be integrated into medical-decision support systems. The speed and precision with which the AI models employed analyze large diabetes prediction datasets, thus reducing diagnosis waiting time, represents a key strength of the study.

However, a major weakness is the lack of validation by medical experts. In addition, the paper does not make a case for why better performance could not be achieved using deeper or more advanced Neural Networks.

List of major concerns and feedback:

• The statistical methods used to evaluate model performance were not clearly explained. The author should explain any statistical tests applied, and the reason for choosing such methods.

• The dataset used in the study is private and not openly available.The provided link (https://www.kaggle.com/code/khaledalrantisi1/ml-and-dl) leads only to the source code notebook, which limits reproducibility and independent validation. Access should be provided to the sample dataset to improve transparency and reproducibility.

• Readers may find it difficult to fully agree with the conclusions because the experiment cannot be easily replicated due to the lack of accessible data.

• In figure 1, the variables are not sufficiently explained, making it difficult for readers to understand the figure and draw meaningful inferences. The author should add a brief and simple explanation of the standardization method in the figure legend.

• All figures, including confusion matrices, lack legends to explain the meaning of colors, making interpretation difficult. All figures should include clear legends and color scales in confusion matrices should be explicitly explained to improve interpretability.

• The manuscript does not clearly report the gender distribution (male vs. female) of the study population, which is an important demographic factor in diabetes research. The author should explicitly mention the exact percentage of the gender, and if possible, discuss how gender distribution may influence model performance.

List of minor concerns and feedback

• The presentation of the confusion matrix is commendable. However, the accuracy curve would benefit from the inclusion of grid lines to improve readability and ease of interpretation.

• The figures are not coordinated and properly referenced in the main text, which may confuse readers.

• The interpretation of “0” and “1” used in figure 1 should be included in the footnote for quick and clear interpretation.

• Abbreviations such as “BMI” and “HbA1c” are not defined in the figure 1 footnote.

• The numbers 0–4 in Figure 1 are not explained. Clarify whether these numbers represent serial numbers.

• Accuracy and Learning Curve in section 5.1 should be titled as a table; likewise the Results Comparison section in 6.

• The manuscript does not address potential bias or fairness issues across gender and age groups, nor does it include external clinical validation.

Additional comments

• The author could specify further features about their neural network training, like the choice of loss function, optimizer, etc.

• In the literature review section, the author should ensure that it is concurrent and well referenced. The author should try and avoid repetition.

• Although the conclusions are supported by the data, they are not written in complete sentence form. Rewrite the conclusions in clear, well-structured sentences to improve readability.

• Although the author did not include a limitations section, the discussion on computational complexity and resource demands of models such as SVMs and neural networks is insightful and relevant for practical deployment.

• Interpretation of results appears in the discussion section, where some of it would be better placed in the results section. Move direct interpretations of figures and tables to the results section and reserve the discussion section for broader implications and comparisons with existing studies.

• The highlighted points for the models in the conclusion section contain information that is better suited for the discussion.

• Conclusion needs to summarise the overall findings of the research in few sentences, therefore it should be reworded.

• Recommendations for future work should be written in sentence format, not in highlights. It should also come before the conclusion section.

• The study highlights the potential of ensemble learning, which could be explored further in future research.

• The manuscript is suitable for publication only after major revisions, especially the inclusion of external validation and clearer methodological explanations.

Concluding remarks

We thank the authors of the preprint for posting their work openly for feedback. Many thanks also to all participants of the Live Review call for their time and for engaging in the lively discussion that generated this review.

Toba Olatoye was a facilitator of this call and a PREreview champion.

Competing interests

The authors declare that they have no competing interests.

Use of Artificial Intelligence (AI)

The authors declare that they did not use generative AI to come up with new ideas for their review.