International Journal of Clinical Case Reports, 2025, Vol.15, No.6, 293-302 http://medscipublisher.com/index.php/ijccr 298 However, at present, many studies only focus on whether the prediction is accurate or not, and pay little attention to whether the model can explain and has clinical value. This can easily turn the model into an "incomprehensible black box", affecting doctors' trust and use of it (Mohsen et al., 2023). Therefore, in recent years, the research direction has shifted from "pursuing only the highest accuracy rate" to "finding a balance between effectiveness and interpretability", enabling the model to predict accurately and provide clear and useful evidence for clinical work (Hasan et al., 2024; Kiran et al, 2025; Khokhar et al., 2025). 5.2 Integration with clinical processes and information systems For AI predictive models to truly take root, it is necessary to smoothly integrate them into existing clinical processes and information systems. A common practice is to incorporate AI tools into electronic health records (EHR) or clinical decision support systems (CDSS), so that risks can be evaluated in real time based on data and alerts can be issued, helping doctors intervene as early as possible (Hasan et al., 2024). Some AI-driven CDSS have been connected through compatible standards and electronic medical records. When doctors review medical records and determine treatment plans, they can directly see the risk classification results given by the model, which is beneficial for resource allocation and process optimization (Nimri and Phillip, 2025). However, in actual use, problems such as incompatible interfaces, inconsistent data, and inconvenient operation may still be encountered. Therefore, it is necessary to develop highly interactive applications to make viewing and operation more convenient and reduce interference with daily work (Hasan et al., 2024). During the development and promotion process, listening more to the opinions of frontline medical staff and adjusting the functions and display methods according to actual needs is the key to improving the model usage rate and avoiding additional workload (Tarumi et al., 2021; Khokhar et al., 2025). 5.3 Cost-effectiveness, ethical issues and data privacy Whether the AI diabetes screening tool is cost-effective or not is an important factor in determining whether it can be widely promoted. Existing studies have shown that when using AI to screen diabetes and its complications (such as diabetic retinopathy), it can not only maintain or even improve the diagnostic efficiency, but also often achieve a good incremental cost-benefit ratio and bring more quality-adjusted life years, especially in national physical examinations or chronic disease management programs (Rajesh et al., 2023). However, the initial investments such as system construction, personnel training and information integration still need to be comprehensively considered in combination with the long-term saved medical expenses and improved treatment effects (Khalifa and Albadawy, 2024; Akune et al., 2025). Meanwhile, ethical and privacy issues cannot be ignored either: The extensive use of health data raises concerns about patient information confidentiality, informed consent, and algorithmic bias (Khokhar et al., 2025). Therefore, it is necessary to ensure data security and fair usage through strict security measures, a transparent model development process, and regular bias checks to win public trust. Healthcare workers, data experts, ethicologists and policymakers working together can help formulate relevant norms that not only protect patients' rights and interests, but also promote the application of technological innovation, and continuously assess its ethical, legal and social impacts to ensure that the use of AI in the medical field is responsible and fair (Khalifa and Albadawy, 2024; Wang et al., 2024). 6 Limitations of Current Research and Problems to be Solved 6.1 Limitations of research design and evidence level Although early diabetes screening models based on artificial intelligence have developed rapidly, a large number of studies rely on retrospective data collected for other purposes, which is prone to selection bias and weakens the applicability of conclusions in real clinical scenarios (Chun and Kim, 2023). Many studies have relatively small sample sizes (some less than 1 000 cases), and lack sample size demonstrations that match the number of predictors, increasing the risk of overfitting and reducing statistical power and model stability (Mohsen et al., 2023; Mittal et al., 2025; Huang et al., 2025).
RkJQdWJsaXNoZXIy MjQ4ODYzNA==