Artificial intelligence-powered prediction of diabetic complications: from clinical data to molecular omics

May combine two claims.

Physiochemical indicator-based DR models were predominantly single-center (75%, 6/8), with only 2 multicenter studies conducted in the United States and Northern Ireland.

“In contrast, physiochemical indicator-based models were predominantly single-center (75%, 6/8), with only 2 multicenter studies conducted in the United States and Northern Ireland [39, 40], reflecting a limited degree of external validation across populations.”

• cites [39] Tarasewicz, 2023, Development and validation of a diabetic retinopathy risk stratification algorithm, Diabetes Care sourcePartly supported

  Data Collection describes a U.S. diabetic-retinopathy prediction study using Kaiser Permanente Northern California records. The Conclusions say validation occurred within the same single health care delivery system and that external validation elsewhere is needed. Our copy does not establish that this was a multicenter study or substantiate the citing sentence’s count of two multicenter studies among eight models. (LLM, conf=0.83)

  Experimental — accuracy not yet measured
• cites [40] Wright, 2023, Identifying the severity of diabetic retinopathy by visual function measures using both traditional statistical methods and interpretable machine learning: A cross-sectional study, Diabetologia sourcePartly supported

  The Methods (Data collection; page unavailable) identify this as a Northern Ireland study that recruited from a population cohort and diabetes clinics. The paper does not establish that it was multicenter or report the claimed tally of six single-center and two multicenter studies. The Discussion (Strengths and limitations; page unavailable) says its models lacked external validation. (LLM, conf=0.82)

  Experimental — accuracy not yet measured

May combine two claims.

Most molecular-omics-driven DN prediction studies were single-center; only two transcriptomics studies utilized public multi-cohort integrative to enable external validation.

“most studies were single-center; only two transcriptomics studies utilized public multi-cohort integrative to enable external validation [76, 77].”

• cites [76] Zhu, 2023, Machine-learning algorithm-based prediction of a diagnostic model based on oxidative stress-related genes involved in immune infiltration in diabetic nephropathy patients, Front Immunol sourcePartly supported

  The Materials and methods section (no page given) says the study downloaded two public GEO expression datasets, GSE30528 and GSE96804. The Results section reports testing the diagnostic model on GSE96804 and then on GSE30528, supporting its use of a second cohort for validation. The paper does not establish that most other studies were single-center or that only two transcriptomics studies used this approach. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [77] Han, 2022, Identification and verification of diagnostic biomarkers for glomerular injury in diabetic nephropathy based on machine learning algorithms, Front Endocrinol (Lausanne) sourceSupported

May combine two claims.

Most DN clinical feature-based studies did not specify the prediction horizon, with only a few targeting new-onset DN over 1- or 3-year periods.

“Notably, most studies did not specify the prediction horizon, with only a few targeting new-onset DN over 1- or 3-year periods [65, 70].”

• cites [65] Zhu, 2024, Machine learning-based predictive Modeling of diabetic nephropathy in type 2 diabetes using integrated biomarkers: A single-Center retrospective study, Diabetes Metab Syndr Obes sourcePartly supported

  The Abstract Methods and Methods > Data Collection and Variables describe patients initially free of DN who were followed for 36 months to identify new cases. This supports the paper as an example of a three-year prediction study, but it does not establish that most other studies omitted a prediction horizon or address the one-year studies. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [70] Song, 2020, Longitudinal risk prediction of chronic kidney disease in diabetic patients using a temporal-enhanced gradient boosting machine: Retrospective cohort study, JMIR Med Inform sourcePartly supported

  The Methods > Experimental Design section describes this study’s rolling prediction of one-year diabetic kidney disease risk, supporting it as an example of a study with a specified horizon. The passages found do not establish that most other studies omitted a prediction horizon or that only a few targeted one- or three-year periods. (LLM, conf=0.82)

  Experimental — accuracy not yet measured

May combine two claims.

Proteomics-based models enabled early discrimination among CKD etiologies, suggesting utility for non-invasive DN diagnosis.

“Proteomics-based models further enabled early discrimination among CKD etiologies, suggesting utility for non-invasive DN diagnosis [80].”

• cites [80] Glazyrin, 2020, Proteomics-based machine learning approach as an alternative to conventional biomarkers for differential diagnosis of chronic kidney diseases, Int J Mol Sci sourcePartly supported

  The Discussion reports that plasma-proteomic models distinguished diabetic nephropathy from glomerulonephritis, and the Conclusions suggest that proteomics-based machine learning may help diagnose diabetic nephropathy at an early stage. But the Discussion describes the tested patients as having previously diagnosed renal disease; the study does not establish early-stage discrimination among CKD etiologies. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

May combine two claims.

Zhu et al. analyzed historical EHR data from 2184 Chinese T2D patients and developed an SVM-based model for DN prediction.

“Zhu et al. [65] analyzed historical EHR data from 2184 Chinese T2D patients and developed an SVM-based model for DN prediction.”

• cites [65] Zhu, 2024, Machine learning-based predictive Modeling of diabetic nephropathy in type 2 diabetes using integrated biomarkers: A single-Center retrospective study, Diabetes Metab Syndr Obes sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Han et al. highlighted PRKAR2B and TGFBI as potential markers for glomerular injury.

“Han et al. [77] highlighted PRKAR2B and TGFBI as potential markers for glomerular injury.”

• cites [77] Han, 2022, Identification and verification of diagnostic biomarkers for glomerular injury in diabetic nephropathy based on machine learning algorithms, Front Endocrinol (Lausanne) sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Diabetic patients face a 2- to 4-fold higher risk of developing CVD compared to the general population, with earlier onset and poorer prognosis.

“Epidemiological studies reveal that diabetic patients face a 2- to 4-fold higher risk of developing CVD compared to the general population, with earlier onset and poorer prognosis [83, 84].”

• cites [83] Vilanova, 2020, Prediabetes is independently associated with subclinical carotid atherosclerosis: An observational study in a non-urban Mediterranean population, J Clin Med sourcePartly supported

  The Introduction (page not provided) reports a two- to four-fold higher risk of atherosclerotic CVD and poorer cardiovascular prognosis in people with type 2 diabetes than in people without diabetes. It discusses atherosclerosis beginning during prediabetes, but does not establish that CVD begins earlier in diabetic patients than in the general population. (LLM, conf=0.88)

  Experimental — accuracy not yet measured
• cites [84] Gu, 2024, The long-term cost-effectiveness of once-weekly semaglutide versus sitagliptin for the treatment of type 2 diabetes in China, Heal Econ Rev sourcePartly supported

  The Discussion states that people with type 2 diabetes have a 2- to 4-fold higher risk of cardiovascular disease than nondiabetic individuals. It also describes cardiovascular complications as a leading cause of death in this group, but does not establish earlier onset or a poorer prognosis relative to nondiabetic people. The claim also broadens the paper’s type 2 diabetes population to diabetic patients generally. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

CVD remains the leading cause of mortality among diabetic patients.

“CVD, one of the most severe systemic complications of diabetes, remains the leading cause of mortality among diabetic patients [9, 81, 82].”

• cites [9] Einarson, 2018, Prevalence of cardiovascular disease in type 2 diabetes: A systematic literature review of scientific evidence from across the world in 2007-2017, Cardiovasc Diabetol sourcePartly supported

  The Background calls CVD the main cause of death in patients with type 2 diabetes, and the Results report that it accounted for 50.3% of deaths in that group. The citing sentence says “diabetic patients” generally, while the review’s mortality finding is specific to type 2 diabetes. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [81] Strain, 2018, Diabetes, cardiovascular disease and the microcirculation, Cardiovasc Diabetol sourcePartly supported

  The Abstract states that CVD is the leading cause of mortality in people with type 2 diabetes. The citing sentence says this of diabetic patients generally, extending the claim beyond the population specified in the source. (LLM, conf=0.85)

  Experimental — accuracy not yet measured
• cites [82] Isath, 2023, Exercise and cardiovascular health: A state-of-the-art review, Prog Cardiovasc Dis sourceNot checked — no full text of the source

May combine two claims.

Yin et al. developed an XGBoost-based DN prediction model with serum metabolomics data, achieving an AUC of 0.966 and identifying eight potential metabolic biomarkers including acetylcarnitine, glutarylcarnitine, tyrosine, serine, and methionine.

“Yin et al. [79] developed an XGBoost-based DN prediction model with serum metabolomics data, achieving an AUC of 0.966 and identifying eight potential metabolic biomarkers including acetylcarnitine, glutarylcarnitine, tyrosine, serine, and methionine.”

• cites [79] Yin, 2024, Explainable machine learning-based prediction model for diabetic nephropathy, J Diabetes Res sourcePartly supported

  Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured

May combine two claims.

Huang et al. reported that a multimodal model combining two metabolites, sphingomyelin C18:1 and phosphatidylcholine diacyl C38:0, with clinical variables achieved optimal prediction of DN, yielding an AUC of 0.825, which outperformed the model based solely on established clinical predictors (AUC = 0.800).

“Huang et al. [103] reported that a multimodal model combining two metabolites, sphingomyelin C18:1 and phosphatidylcholine diacyl C38:0, with clinical variables achieved optimal prediction of DN, yielding an AUC of 0.825, which outperformed the model based solely on established clinical predictors (AUC = 0.800).”

• cites [103] Huang, 2020, Machine learning approaches reveal metabolic signatures of incident chronic kidney disease in individuals with prediabetes and type 2 diabetes, Diabetes sourcePartly supported

  Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured

May combine two claims.

Ravaut et al. applied ML to Canadian administrative health data, achieving satisfactory performance for adverse CVD events.

“Ravaut et al. [88] applied ML to Canadian administrative health data, achieving satisfactory performance for adverse CVD events.”

• cites [88] Ravaut, 2021, Predicting adverse outcomes due to diabetes complications with machine learning using administrative health data, NPJ Digit Med sourceSupported

May combine two claims.

Chu et al. incorporated psychosocial determinants into a DNN framework, achieving AUC = 0.91, suggesting that psychosocial and behavioral variables provide additional predictive value beyond traditional biomarkers.

“Chu et al. [97] further incorporated psychosocial determinants into a DNN framework, achieving AUC = 0.91, suggesting that psychosocial and behavioral variables provide additional predictive value beyond traditional biomarkers.”

• cites [97] Chu, 2021, Roles of anxiety and depression in predicting cardiovascular disease among patients with type 2 diabetes mellitus: A machine learning approach, Front Psychol sourcePartly supported

  The Results (CVD Risk Prediction Model in Patients With T2DM; page unavailable) report a DNN AUC of 0.91 and rank anxiety and depression among its top five predictors. The paper does not report a comparison with a biomarkers-only model, so it does not establish additional predictive value beyond traditional biomarkers; the cited results also do not establish that for behavioral variables. (LLM, conf=0.87)

  Experimental — accuracy not yet measured

May combine two claims.

None of the included multimodal studies performed comprehensive clinical cost-benefit analysis, and only one reported clinical utility evaluations using decision-curve analysis (DCA).

“None of the included studies performed comprehensive clinical cost–benefit analysis, and only one reported clinical utility evaluations using decision-curve analysis (DCA) [104].”

• cites [104] Shen, 2023, Metabolite-assisted models improve risk prediction of coronary heart disease in patients with diabetes, Front Pharmacol sourcePartly supported

  The Methods and Results report decision-curve analysis (DCA) to evaluate the models’ clinical net benefit. This supports identifying the cited study as one that performed a clinical utility evaluation. The paper does not establish that it was the only included study to do so, or that none of the included studies performed a comprehensive clinical cost–benefit analysis. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

May combine two claims.

Chu et al. applied XGBoost framework with psychosocial variables, significantly improving CVD risk prediction.

“Chu et al. [97] applied XGBoost framework with psychosocial variables, significantly improving CVD risk prediction.”

• cites [97] Chu, 2021, Roles of anxiety and depression in predicting cardiovascular disease among patients with type 2 diabetes mellitus: A machine learning approach, Front Psychol sourcePartly supported

  The Methods and Results describe a CVD prediction model for patients with type 2 diabetes that includes anxiety and depression, but identify it as a deep neural network, not XGBoost. The Results report its performance (AUC 0.91), but do not establish a statistically significant improvement in prediction from adding psychosocial variables. The indexed passages provide section labels but no page numbers. (LLM, conf=0.91)

  Experimental — accuracy not yet measured

May combine two claims.

Shen et al. reported that a metabolite-assisted model incorporating four key metabolites and nine clinical factors achieved an AUC of 0.962-0.979, significantly outperforming clinical-only models (AUC 0.576-0.745).

“Shen et al. [104] reported that a metabolite-assisted model incorporating four key metabolites and nine clinical factors achieved an AUC of 0.962–0.979, significantly outperforming clinical-only models (AUC 0.576–0.745).”

• cites [104] Shen, 2023, Metabolite-assisted models improve risk prediction of coronary heart disease in patients with diabetes, Front Pharmacol sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Shen et al. used a RF-based model to evaluate CVD risk in diabetic patients assisted by metabolic features, selecting metabolic biomarkers with a mean decrease in Gini >1.0 for model construction and training.

“Shen et al. [104] used a RF-based model to evaluate CVD risk in diabetic patients assisted by metabolic features, selecting metabolic biomarkers with a mean decrease in Gini >1.0 for model construction and training.”

• cites [104] Shen, 2023, Metabolite-assisted models improve risk prediction of coronary heart disease in patients with diabetes, Front Pharmacol sourcePartly supported

  The Results (page not provided) say random forest selected metabolites with mean decrease Gini >1.0 in both training and test sets for further model development. But the Methods and Results describe logistic regression for the CVD risk-prediction models; random forest was used for biomarker selection, not as the prediction model. (LLM, conf=0.87)

  Experimental — accuracy not yet measured

May combine two claims.

RF can effectively handle high-dimensional data while offering high predictive accuracy and strong parallel computing capability.

“RF can effectively handle high-dimensional data while offering high predictive accuracy and strong parallel computing capability [115].”

• cites [115] Hu, 2023, A review on longitudinal data analysis with random forest, Brief Bioinform sourcePartly supported

  The Introduction (page not provided) says RF is suited to high-dimensional settings, and the Standard RF algorithm section describes its predictive power. The passages do not establish the claimed strong parallel computing capability. (LLM, conf=0.85)

  Experimental — accuracy not yet measured

May combine two claims.

An analysis compared eight ML algorithms, with RF performing best for early CVD detection.

“an analysis compared eight ML algorithms, with RF performing best for early CVD detection [87].”

• cites [87] Das, 2024, Determinants of developing cardiovascular disease risk with emphasis on type-2 diabetes and predictive modeling utilizing machine learning algorithms, Medicine (Baltimore) sourceSupported

May combine two claims.

Li et al. built a DR risk prediction model with XGBoost, identifying key features such as age, FBG, and HbA1c.

“Li et al. [15] built a DR risk prediction model with XGBoost, identifying key features such as age, FBG, and HbA1c.”

• cites [15] Li, 2021, Predictive model and risk analysis for diabetic retinopathy using machine learning: A retrospective cohort study in China, BMJ Open sourceSupported

Probably not a substantial claim: may not pass S2 (checkable).

Professional oversight and validation are needed in clinical deployment of LLMs for cardiovascular prevention information.

“The model provided accurate and comprehensive answers for 75% of items, though readability was limited for highly specialized clinical topics, underscoring the need for professional oversight and validation in clinical deployment.”

• cites [138] Lorenzoni, 2025, Large language models to develop evidence-based strategies for primary and secondary cardiovascular prevention, Eur Heart J Digit Health sourceSupported

May combine two claims.

The BERT-based model demonstrated the scalability and accuracy of LLMs for pathological report structuring and clinical data mining.

“In renal function research, Bobart et al. [136] developed a biomedical BERT-based natural language processing model to automatically extract diagnostic information from kidney biopsy reports, achieving an average AUC of 0.95 across reports for DN, lupus nephritis, and other diseases, reducing manual annotation workload and demonstrating the scalability and accuracy of LLMs for pathological report”

• cites [136] Bobart, 2025, Development of a natural language processing model for extracting kidney biopsy pathology diagnoses, Kidney Med sourcePartly supported

  The Methods describe a PubMed BERT model that extracts diagnoses from kidney biopsy reports, and the Results report a mean AUROC of 0.95 across diagnoses. The Conclusions call the system accurate and scalable, but the Discussion describes large-scale use and testing in external cohorts as future work. The paper supports this specific model’s potential for clinical data extraction, not a demonstrated result for LLMs generally; it also does not measure a reduction in manual annotation workload. (LLM, conf=0.82)

  Experimental — accuracy not yet measured

May combine two claims.

RAG provides external, verifiable knowledge support, thereby significantly reducing hallucinations and factual errors.

“By dynamically retrieving authoritative medical literature, clinical guidelines, or EHR knowledge bases before text generation, RAG provides external, verifiable knowledge support, thereby significantly reducing hallucinations and factual errors.”

• cites [141] Sharma, 2025, Retrieval-augmented generation: A comprehensive survey of architectures, enhancements, and robustness Frontiers sourceCould not verify

  The Introduction (pp. 1–2) and Conclusions (p. 9) discuss CNF-guided Verilog generation, not retrieval-augmented generation. Searches for RAG, hallucinations, medical literature, clinical guidelines, and EHR knowledge bases found no passage addressing the claimed reduction in factual errors. (LLM, conf=0.78)

  Experimental — accuracy not yet measured

May combine two claims.

Li et al. developed DeepDR-LLM, a multimodal vision-language system for DR diagnosis and management, enabling multimodal generation of diagnostic and treatment recommendations and providing support for personalized diabetes care.

“Li et al. [135] developed DeepDR-LLM, a multimodal vision-language system for DR diagnosis and management, enabling multimodal generation of diagnostic and treatment recommendations and providing support for personalized diabetes care.”

• cites [135] Li, 2024, Integrated image-based deep learning and language models for primary diabetes care, Nat Med sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Zhu et al. developed the HaluEval-Wild benchmark based on RAG to evaluate hallucination rates in LLMs.

“For instance, Zhu et al. [142] developed the HaluEval-Wild benchmark based on RAG to evaluate hallucination rates in LLMs.”

• cites [142] Zhu, 2024, Halueval-wild: Evaluating hallucinations of language models in the wild sourceSupported

May combine two claims.

In diabetic retinopathy grading tasks, ChatGPT-4 Omni performs well for proliferative diabetic retinopathy detection but shows markedly lower accuracy for mild or early-stage lesions.

“For instance, in DR grading tasks, ChatGPT-4 Omni performs well for PDR detection but shows markedly lower accuracy for mild or early-stage lesions [131], highlighting insufficient robustness under domain shift [140].”

• cites [131] Chetla, 2025, Assessing the diagnostic capabilities of ChatGPT-4 omni in grading diabetic retinopathy fundoscopy using color fundus photographs, Clin Ophthalmol sourceSupported
• cites [140] Sopchoke, 2018 sourceCould not verify

  The Abstract and Research Methodology (p. 1) describe a cross-domain music-and-movie recommendation system, while Results and Discussion (p. 2) report recommendation accuracy. Searches for ChatGPT-4 Omni and diabetic retinopathy grading found no passage about PDR detection or mild or early-stage lesions. The claim is therefore not verifiable from this source. (LLM, conf=0.78)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Retrieval-Augmented Generation (RAG) has emerged as an effective paradigm for medical applications.

“First, Retrieval-Augmented Generation (RAG) has emerged as an effective paradigm [141].”

• cites [141] Sharma, 2025, Retrieval-augmented generation: A comprehensive survey of architectures, enhancements, and robustness Frontiers sourceCould not verify

  The indexed copy discusses language models for Verilog code generation, including in Background and Motivation (p. 2) and Methodology (p. 4), rather than RAG. Searches for retrieval-augmented generation and medical or clinical applications found no passage supporting the claim. The printed citing sentence does not itself mention medical applications. (LLM, conf=0.76)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Readability of ChatGPT-4o outputs was limited for highly specialized clinical topics.

“The model provided accurate and comprehensive answers for 75% of items, though readability was limited for highly specialized clinical topics, underscoring the need for professional oversight and validation in clinical deployment.”

• cites [138] Lorenzoni, 2025, Large language models to develop evidence-based strategies for primary and secondary cardiovascular prevention, Eur Heart J Digit Health sourceSupported

May combine two claims.

Ayhan et al. employed the multimodal Gemini 1.5 Pro model with contextual learning to perform binary DR classification, achieving an accuracy of 0.841 (95% CI: 0.803-0.879) and an F1 score of 0.88 (95% CI: 0.844-0.909), comparable to RETFound, without requiring retraining or large-scale labeled datasets.

“Ayhan et al. [130] employed the multimodal Gemini 1.5 Pro model with contextual learning to perform binary DR classification, achieving an accuracy of 0.841 (95% CI: 0.803–0.879) and an F1 score of 0.88 (95% CI: 0.844–0.909), comparable to RETFound, without requiring retraining or large-scale labeled datasets.”

• cites [130] Ayhan, 2026, In-context learning for data-efficient diabetic retinopathy detection via multimodal foundation models, Ophthalmol Sci sourceSupported

May combine two claims.

GPT-4 exhibited a 20% hallucination rate without RAG, which dropped to 5% when RAG was incorporated.

“Their study showed that GPT-4 exhibited a 20% hallucination rate without RAG, which dropped to 5% when RAG was incorporated.”

• cites [142] Zhu, 2024, Halueval-wild: Evaluating hallucinations of language models in the wild sourceSupported

May combine two claims.

Factuality bias and hallucination are among the most critical limitations of LLMs.

“First, factuality bias and hallucination are among the most critical limitations [139].”

• cites [139] Kalai, 2025, Why language models hallucinate sourcePartly supported

  The Conclusions (p. 16) treat hallucination as a consequential language-model shortcoming and argue that mainstream evaluations reward it. The paper does not identify a separate limitation called “factuality bias,” nor does it rank that limitation among the most critical. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

• AssumptionRests on an assumption: LLMs are primarily trained on general-purpose corpora rather than rigorously validated medical knowledge bases.

May combine two claims.

Chetla et al. utilized ChatGPT-4 Omni to grade DR severity on the EyePACS dataset, achieving high precision for PDR detection (accuracy 75.6%, precision 92.2%), demonstrating the feasibility of general-purpose LLMs for low-cost DR screening.

“Chetla et al. [131] utilized ChatGPT-4 Omni to grade DR severity on the EyePACS dataset, achieving high precision for PDR detection (accuracy 75.6%, precision 92.2%), demonstrating the feasibility of general-purpose LLMs for low-cost DR screening.”

• cites [131] Chetla, 2025, Assessing the diagnostic capabilities of ChatGPT-4 omni in grading diabetic retinopathy fundoscopy using color fundus photographs, Clin Ophthalmol sourcePartly supported

  Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Ding et al. introduced the 3DS framework, which adopts a model-centric approach to align data selection with the model's knowledge distribution, thereby improving adaptability and robustness in medical tasks.

“For example, Ding et al. [144] introduced the 3DS (Decomposed Difficulty-based Data Selection) framework, which adopts a model-centric approach to align data selection with the model’s knowledge distribution, thereby improving”

• cites [144] Ding, 2024, 3DS: Decomposed difficulty data Selection's case study on LLM medical domain adaptation sourceSupported

May combine two claims.

Omics approaches are particularly valuable for identifying early biomarkers and disease subtypes, potentially enabling targeted intervention before irreversible damage occurs.

“These approaches are particularly valuable for identifying early biomarkers and disease subtypes, potentially enabling targeted intervention before irreversible damage occurs [161, 162].”

• cites [161] Kim, 2015, Integrative phenotyping framework (iPF): Integrative clustering of multiple omics data identifies novel lung disease subphenotypes, BMC Genomics sourcePartly supported

  The Methods section describes detecting biomarkers associated with lung-disease clusters, and the Discussion says multi-omics integration may identify new subphenotypes and drug targets (page numbers unavailable). The paper does not establish that these are early biomarkers or that targeted intervention could occur before irreversible damage. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [162] Fechete, 2011, Mapping of molecular pathways, biomarkers and drug targets for diabetic nephropathy, Proteomics Clin Appl sourceNot checked — no full text of the source

May combine two claims.

The CSEDB study demonstrated an average 13.3% performance drop in high-risk scenarios, with general-purpose models underperforming domain-specific medical LLMs in both safety and effectiveness metrics.

“Their study demonstrated an average 13.3% performance drop in high-risk scenarios, with general-purpose models underperforming domain-specific medical LLMs in both safety and effectiveness metrics.”

• cites [146] Wang, 2025, A novel evaluation benchmark for medical LLMs: Illuminating safety and effectiveness in clinical domains sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Expert feedback indicated that RAG-enhanced GPT-4 reduced hallucinations and offered comprehensive contextual information, thorough consideration of parameter settings, and clear explanations of underlying assumptions.

“Expert feedback further indicated that RAG-enhanced GPT-4 not only reduced hallucinations but also offered comprehensive contextual information, thorough consideration of parameter settings, and clear explanations of underlying assumptions.”

• cites [142] Zhu, 2024, Halueval-wild: Evaluating hallucinations of language models in the wild sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Through optimized prompts and workflows, LLMs could reduce critical error rates below human note-taking levels in medical text summarization.

“Similarly, Asgari et al. [147] developed a framework to evaluate hallucination and clinical safety in medical text summarization, showing that through optimized prompts and workflows, LLMs could reduce critical error rates below human note-taking levels,”

• cites [147] Asgari, 2025, A framework to assess clinical safety and hallucination rates of LLMs for medical text summarisation, NPJ Digit Med sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Asgari et al. developed a framework to evaluate hallucination and clinical safety in medical text summarization.

“Similarly, Asgari et al. [147] developed a framework to evaluate hallucination and clinical safety in medical text summarization, showing that through optimized prompts and workflows, LLMs could reduce critical error rates below human note-taking levels,”

• cites [147] Asgari, 2025, A framework to assess clinical safety and hallucination rates of LLMs for medical text summarisation, NPJ Digit Med sourceSupported

May combine two claims.

Wang et al. proposed the Clinical Safety-Effectiveness Dual-Track Benchmark (CSEDB), designed by 32 specialist physicians, covering 26 clinical departments and 30 key indicators.

“Wang et al. [146] proposed the Clinical Safety-Effectiveness Dual-Track Benchmark (CSEDB), designed by 32 specialist physicians, covering 26 clinical departments and 30 key indicators, including critical illness recognition, medication safety, and guideline adherence.”

• cites [146] Wang, 2025, A novel evaluation benchmark for medical LLMs: Illuminating safety and effectiveness in clinical domains sourceSupported

May combine two claims.

The 'black-box' nature, dependency on large labeled datasets, and high computational cost hinder widespread clinical implementation of deep learning.

“Nevertheless, the ‘black-box’ nature [168], dependency on large labeled datasets, and high computational cost hinder their widespread clinical implementation.”

• cites [168] Aggarwal, 2022, COVID-19 image classification using deep learning: Advances, challenges and opportunities, Comput Biol Med sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

LLMs represent a rapidly emerging research paradigm in diabetic complication prediction.

“Currently, LLMs represent a rapidly emerging research paradigm in diabetic complication prediction, with growing studies exploring their potential lies in processing unstructured clinical text [169], enabling personalized health management via conversational agents [170], and supporting biomedical knowledge discovery [171].”

• cites [169] Maarseveen, 2025, Improving musculoskeletal care with AI enhanced triage through data driven screening of referral letters, NPJ Digit Med sourcePartly supported

  The Introduction (page not available) says that NLP and LLMs have potential for analyzing unstructured referral letters, supporting the point attached to [169] about processing clinical text. But the study concerns musculoskeletal triage, not diabetic complication prediction; its Methods describe an NLP pipeline and XGBoost prediction models rather than an LLM-based prediction study. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [170] Liu, 2024, Global trends and hotspots of ChatGPT in medical research: A bibliometric and visualized study, Front Med (Lausanne) sourcePartly supported

  The Introduction (page unavailable) says ChatGPT can analyze medical records and clinical notes and extract insights from unstructured text, supporting the application nearest citation [169]. The Discussion describes rapid uptake of LLMs in medicine, but the paper does not establish an emerging research paradigm specifically in diabetic complication prediction. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [171] Cooper, 2015, The center for causal discovery of biomedical knowledge from big data, J Am Med Inform Assoc sourcePartly supported

  Citation [171] is attached to the claim that LLMs support biomedical knowledge discovery. The Introduction describes computational methods for discovering causal knowledge, but it does not identify those methods as LLMs or address diabetic complication prediction. (LLM, conf=0.85)

  Experimental — accuracy not yet measured

May combine two claims.

Most omics-based DR cohorts were derived from Asian populations (primarily China and Turkey), while only a few transcriptomic studies utilized publicly available multi-regional datasets.

“Regarding population composition, most cohorts were derived from Asian populations (primarily China and Turkey), while only a few transcriptomic studies utilized publicly available multi-regional datasets.”

May combine two claims.

Among clinical feature-based DN models, all were retrospective, 45.5% were multicenter, cohort sizes ranged from 133 to 605,328 participants, external validation was performed in 36.4%, and AUCs ranged from 0.747 to 0.989.

“Among clinical feature-based models, all (100%) were retrospective, 45.5% were multicenter studies, and cohort sizes ranged from 133 to 605 328 participants.”

• Also stated in: § Diabetic nephropathy · plain

  “External validation was performed in 36.4% of studies, with reported AUCs ranging from 0.747 to 0.989.”

May combine two claims.

All omics-based DR models were retrospective and single-center designed.

“All (100%) were retrospective and single-center designed, with cohort sizes of 50–317 participants, and 28.6% had external validation.”

May combine two claims.

The review identified 23 studies on AI-based DR prediction, comprising 8 fundus image-based, 8 clinical physiological feature-based, and 7 molecular omics-based studies.

“Summary of 23 studies on AI-based DR prediction categorized by input modality: fundus images (n = 8 studies), clinical physiological features (n = 8), and molecular omics data (n = 7).”

May combine two claims.

Among clinical feature-based DR studies, seven were multicenter, five of which were fundus image-based.

“Among these, seven studies were multicenter, five of which were fundus image-based, suggesting that imaging models have been more frequently validated across sites.”

May combine two claims.

Among clinical feature-based DR models, 87.5% were retrospective and 25% multicenter.

“For clinical feature-based models, 87.5% were retrospective and 25% multicenter, including 260–276 794 participants.”

May combine two claims.

Tree-based algorithms, especially RF and XGBoost, dominate physiochemical-based DR prediction, consistently identifying renal function, long-term glycemic control, BP, lipid levels, and inflammation as major risk factors.

“Overall, tree-based algorithms, especially RF and XGBoost, dominate physiochemical-based DR prediction, consistently identifying renal function, long-term glycemic control, BP, lipid levels, and inflammation as major risk factors.”

May combine two claims.

Among omics-based DR models, datasets included metabolomics (n=3), transcriptomics (n=2), and metagenomics (n=2).

“For omics-based models, datasets included metabolomics (n = 3 studies), transcriptomics (n = 2), and metagenomics (n = 2).”

May combine two claims.

The review identified 16 studies on AI-based DN prediction, including 11 clinical feature-based and five molecular omics-based studies.

“A total of 16 studies on AI-based DN prediction were reviewed, including 11 using clinical feature-based models and five using molecular omics-based models.”

May combine two claims.

Only one of the 16 clinical feature-based DR studies was prospective; all others were retrospective.

“Only one was prospective, while all others adopted a retrospective design.”

May combine two claims.

Only studies using public resources provided accessible datasets, while none of the self-collected cohort studies shared raw data or model code.

“Only studies using public resources provided accessible datasets, while none of the self-collected cohort studies shared raw data or model code.”

May combine two claims.

Model inputs for physiochemical-based DR prediction are predominantly structured variables, while subjective and unstructured features such as lifestyle behaviors and functional assessments are underutilized.

“Additionally, model inputs are predominantly structured variables, while subjective and unstructured features such as lifestyle behaviors and functional assessments are underutilized.”

May combine two claims.

Reported AUCs for DN clinical feature-based models ranged from 0.75 to 0.99, with most models achieving 0.80-0.90.

“In terms of performance, reported AUCs ranged from 0.75 to 0.99, with most models achieving 0.80–0.90, indicating generally good discriminative ability.”

May combine two claims.

DN incidence in included clinical feature-based studies was ~30%-50%, indicating a generally balanced dataset.

“Regarding sample balance, DN incidence was ~30%–50%, indicating a generally balanced dataset.”

May combine two claims.

Most DN clinical feature-based studies did not provide code or data, with only a few based on public datasets offering partially reproducible resources, restricting algorithm transparency and reproducibility.

“Most studies did not provide code or data, with only a few based on public datasets offering partially reproducible resources, which restricts algorithm transparency and reproducibility.”

May combine two claims.

The study designs of molecular-omics-driven DN prediction studies were predominantly retrospective, with only one prospective cohort.

“The study designs were predominantly retrospective, with only one prospective cohort.”

May combine two claims.

Common predictive features for DN clinical models included HbA1c, serum creatinine, eGFR, diabetes duration, BP, and presence of retinopathy or CVD history; these were consistently identified as important determinants of DN onset.

“Common predictive features included HbA1c, serum creatinine, eGFR, diabetes duration, BP, and the presence of retinopathy or CVD history; these features were consistently identified as important determinants of DN onset.”

May combine two claims.

In the review, 11 studies employing clinical features for DN prediction were included, all of which were retrospective and lacked prospective validation.

“In this review, 11 studies employing clinical features for DN prediction were included (Table S3), all of which were retrospective and lacked prospective validation.”

May combine two claims.

Among omics-based DN models, datasets included proteomics (n=1), transcriptomics (n=2), and metabolomics (n=2); most (80%) were retrospective, 20% multicenter, cohort sizes between 48 and 562 participants, 40% conducted external validation, and AUCs ranged from 0.874 to 0.966.

“For omics-based models, datasets included proteomics (n = 1 studies), transcriptomics (n = 2 studies), and metabolomics (n = 2 studies).”

• Also stated in: § Diabetic nephropathy · plain

  “Most (80%) were retrospective, 20% multicenter, with cohort sizes between 48 and 562 participants; 40% conducted external validation. Model performance (AUC) ranged from 0.874 to 0.966.”

Probably not a substantial claim: may not pass S3 (load-bearing).

DN clinical feature-based studies had broad geographical distribution, covering China, the USA, Iran, Singapore, Australia, and Northern Ireland.

“The geographical distribution of DN studies is broad, covering China, the USA, Iran, Singapore, Australia, and Northern Ireland, reflecting active global exploration in this field.”

May combine two claims.

External validation was conducted in only 4 of 11 DN clinical feature-based studies, suggesting limited generalizability.

“However, high performance was often based on internal validation, with external validation conducted in only 4 (4/11) studies, suggesting limited generalizability.”

May combine two claims.

The included multimodal studies mainly employed feature-level fusion, while decision-level or representation-level integration frameworks remain rare.

“The included studies mainly employed feature-level fusion, while decision-level or representation-level integration frameworks remain rare.”

May combine two claims.

All 14 CVD prediction studies were retrospective, with 64.3% (9/14) being single-center studies.

“All studies were retrospective, with single-center designs predominating, of which 64.3% (9/14) were single-center studies.”

May combine two claims.

Only two CVD prediction studies conducted independent external validation, while most relied on internal validation.

“Only two studies conducted independent external validation, while most relied on internal validation.”

May combine two claims.

Reported AUC values for CVD prediction models ranged from 0.701 to 0.910, with most models achieving 0.700-0.800.

“Reported AUC values ranged from 0.701 to 0.910, with most models achieving 0.700–0.800”

May combine two claims.

In total, five multimodal modeling studies were reviewed, covering all three major complications, with most focused on DN (n = 3) and CVD (n = 2), while DR was represented by a single study.

“In total, five multimodal modeling studies were reviewed, covering all three major complications, with most focused on DN (n = 3) and CVD (n = 2), while DR was represented by a single study.”

Probably not a substantial claim: may not pass S4 (based).

External validation is seldom implemented in current AI research.

“External validation is the gold standard for assessing model generalizability, yet it is seldom implemented in current AI research.”

May combine two claims.

Prediction timeframes vary widely from 1 to 5 years, with some studies incorporating observation and buffer periods, while others rely on cross-sectional data only.

“For instance, prediction timeframes vary widely from 1 to 5 years, with some studies incorporating observation and buffer periods, while others rely on cross-sectional data only.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The limited availability of source data and model codes undermines the reproducibility of current AI models.

“Moreover, the limited availability of source data and model codes further undermines the reproducibility of current AI models.”

May combine two claims.

AUC, accuracy, or F1-score are often derived from distinct prediction horizons, population characteristics, and endpoint definitions, and some studies report ranges while others omit certain metrics.

“Although AUC, accuracy, or F1-score are commonly reported, these metrics are often derived from distinct prediction horizons, population characteristics, and endpoint definitions, and some studies report ranges while others omit certain metrics, reflecting incomplete and heterogeneous reporting of evaluation results.”

May combine two claims.

Omics-driven models spanning metabolites, transcripts, and microbial profiles improve DR risk prediction and uncover mechanistic insights, moving beyond phenotypic screening toward precision diagnostics.

“Omics-driven models spanning metabolites, transcripts, and microbial profiles improve DR risk prediction and uncover mechanistic insights, moving beyond phenotypic screening toward precision diagnostics.”

May combine two claims.

Gut-retina axis perturbations may contribute to DR and offer opportunities for microbiota-based interventions.

“These findings suggest gut-retina axis perturbations may contribute to DR and offer opportunities for microbiota-based interventions.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Standardizing image quality variations and ensuring compatibility across different fundus camera devices remain unaddressed in current DR models.

“Standardizing image quality variations and ensuring compatibility across different fundus camera devices remain unaddressed in current models.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The literature selection process followed the PRISMA framework.

“The entire process followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) framework and is illustrated in Fig. 1.”

May combine two claims.

Targeted omics techniques deliver clinically translatable signatures yet restrict biomarker innovation through predefined analyte panels.

“Nevertheless, the targeted techniques deliver clinically translatable signatures yet restrict biomarker innovation through predefined analyte panels.”

May combine two claims.

The Du et al. model achieved robust cross-center prediction and supported early risk assessment.

“achieving robust cross-center prediction and supporting early risk assessment.”

• cites [66] Du, 2024, Development and external validation of a machine learning model to predict diabetic nephropathy in T1DM patients in the real-world, Acta Diabetol sourceNot checked — no full text of the source

May combine two claims.

A separate study involving 133 diabetic patients combined PCA with SVM, achieving 88.7% accuracy, demonstrating the effectiveness of feature selection combined with ML.

“A separate study involving 133 diabetic patients combined principal component analysis (PCA) with SVM, achieving 88.7% accuracy, demonstrating the effectiveness of feature selection combined with ML [72].”

• cites [72] Maniruzzaman, 2021, Risk prediction of diabetic nephropathy using machine learning techniques: A pilot study with secondary data, Diabetes Metab Syndr sourceNot checked — no full text of the source

May combine two claims.

Integrating multi-omics data has markedly improved DN prediction and mechanistic understanding.

“Integrating multi-omics data has markedly improved DN prediction and mechanistic understanding.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Challenges such as inconsistent data coding, missing records, and variations in standardization associated with large-scale data integration continue to constrain model accuracy and generalizability.

“However, challenges such as inconsistent data coding, missing records, and variations in standardization associated with large-scale data integration continue to constrain model accuracy and generalizability.”

Probably not a substantial claim: may not pass S4 (based).

Molecular-omics-driven AI models represent a transformative step in DN prediction, shifting the focus from conventional clinical indicators to molecular-level biomarkers that reveal early disease processes.

“In summary, molecular-omics-driven AI models represent a transformative step in DN prediction, shifting the focus from conventional clinical indicators to molecular-level biomarkers that reveal early disease processes.”

May combine two claims.

A study involving 9059 participants developed an XGBoost model for MACE prediction in T2D, attaining an AUC of 0.82 and identifying phosphate and blood urea nitrogen as novel predictors.

“A study involving 9059 participants developed an XGBoost model for MACE prediction in T2D, attaining an AUC of 0.82 and identifying phosphate and blood urea nitrogen as novel predictors [90].”

• cites [90] Abegaz, 2023, Machine learning algorithms to predict major adverse cardiovascular events in patients with diabetes, Comput Biol Med sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Tree-based models, particularly RF and XGBoost, dominate the field due to their ability to capture complex nonlinear interactions.

“Tree-based models, particularly RF and XGBoost, dominate the field due to their ability to capture complex nonlinear interactions.”

May combine two claims.

High performance in CVD prediction models typically came from internal validation, indicating that generalizability remains a challenge.

“High performance typically came from internal validation, indicating that generalizability remains a challenge.”

Probably not a substantial claim: may not pass S4 (based).

AI has played a significant role in the development of diagnostic models for diabetic complications, driving the evolution of predictive approaches from relying on single clinical features to integrating multimodal data.

“AI has played a significant role in the development of diagnostic models for diabetic complications, driving the evolution of predictive approaches from relying on single clinical features to integrating multimodal data.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Tree-based algorithms are preferred in clinical applications due to their strong feature selection capability, robustness to missing data, and high predictive performance.

“Tree-based algorithms are preferred in clinical applications due to their strong feature selection capability, robustness to missing data, and high predictive performance.”

May combine two claims.

DTs are prone to overfitting and sensitive to data perturbations, limiting generalizability.

“However, DTs are prone to overfitting and sensitive to data perturbations, limiting generalizability [113].”

• cites [113] Jiang, 2020, Deep decision tree transfer boosting, IEEE Trans Neural Netw Learn Syst sourceNot checked — no full text of the source

May combine two claims.

External validation and standardized benchmarking are limited in multimodal studies, impeding cross-study comparability.

“Moreover, external validation and standardized benchmarking are limited, impeding cross-study comparability.”

Probably not a substantial claim: may not pass S4 (based).

Clinical feature-based AI models for CVD risk prediction in diabetes have evolved from early, small-scale studies to large cohort-based approaches leveraging advanced algorithms.

“In conclusion, clinical feature-based AI models for CVD risk prediction in diabetes have evolved from early, small-scale studies to large cohort-based approaches leveraging advanced algorithms.”

Classified as probably not a substantial claim.

LLM applications offer new technological pathways for building intelligent systems for diabetes complication diagnosis and care.

“Overall, LLM applications in diabetes-related complication management are expanding from data structuring and clinical text understanding to intelligent interaction and personalized patient education, offering new technological pathways for building intelligent systems for”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The integration of LLMs and their agent frameworks has emerged as a research hotspot in medical AI, showing considerable potential in the intelligent diagnosis and management of diabetes-related complications.

“The integration of LLMs and their agent frameworks has emerged as a research hotspot in medical AI, showing considerable potential in the intelligent diagnosis and management of diabetes-related complications.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Aftab et al. reported that ChatGPT-4 exhibited high sensitivity for referable DR, reaching 96.2% in high-risk populations, but its limited specificity and tendency to overcall disease reduce its practicality as a general screening tool, while supporting its use as an auxiliary triage system for high-risk patients.

“Aftab et al. [134] reported that ChatGPT-4 exhibited high sensitivity for referable DR, reaching 96.2% in high-risk populations, but its limited specificity and tendency to overcall disease reduce its practicality as a general screening tool, while supporting its use as an auxiliary triage system for high-risk patients.”

• cites [134] Aftab, 2025, Evaluation of ChatGPT-4 in detecting referable diabetic retinopathy using single fundus images, AJO International sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

DL is particularly effective at capturing nonlinear relationships and latent feature structures within data compared with traditional ML.

“Compared with traditional ML, DL is particularly effective at capturing nonlinear relationships and latent feature structures within data.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Non-tree-based models are particularly suitable for small to medium-sized datasets or scenarios with relatively limited feature dimensions.

“Non-tree-based models are particularly suitable for small to medium-sized datasets or scenarios with relatively limited feature dimensions.”

Probably not a substantial claim: may not pass S2 (checkable).

These observations underscore the potential of multimodal fusion in advancing precision prediction and management of diabetic complications.

“These observations underscore the potential of multimodal fusion in advancing precision prediction and management of diabetic complications.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based). May combine two claims.

In the management of diabetic complications, LLMs and their agents are particularly suited for integrating multimodal information, providing auxiliary personalized clinical decision support, and facilitating continuous patient engagement through conversational interaction.

“In the management of diabetic complications, LLMs and their agents are particularly suited for integrating multimodal information, providing auxiliary personalized clinical decision support, and facilitating continuous patient engagement through conversational interaction.”

Probably not a substantial claim: may not pass S2 (checkable).

Current research and practice in multimodal fusion still face critical technical hurdles.

“However, current research and practice in multimodal fusion still face critical technical hurdles.”

Probably not a substantial claim: may not pass S4 (based).

Clinical and omics data differ substantially in scale, distribution, and noise patterns.

“Clinical and omics data differ substantially in scale, distribution, and noise patterns.”

May combine two claims.

Clinical-data models are effective in capturing phenotypic characteristics and disease progression, showing robust performance in short-term risk stratification.

“These models are effective in capturing phenotypic characteristics and disease progression, showing robust performance in short-term risk stratification.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Multimodal models tend to exacerbate the 'black-box' nature of AI systems, further hindering clinical interpretability.

“Moreover, multimodal models tend to exacerbate the ‘black-box’ nature of AI systems, further hindering clinical interpretability.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

Rigorous methodology in data splitting and statistical power analysis is crucial to ensure unbiased and reliable model evaluation.

“In the stage of data splitting and statistical power analysis, rigorous methodology is crucial to ensure unbiased and reliable model evaluation.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

These measures help identify potential regulatory risks and support the safe integration and continuous optimization of models in real-world clinical settings.

“These measures not only help identify potential regulatory risks, such as hallucinations, domain shifts, or inappropriate treatment suggestions, but also support the safe integration and continuous optimization of models in real-world clinical settings.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Prospective clinical validation represents the key step in bridging experimental success and clinical translation.

“Prospective clinical validation represents the key step in bridging experimental success and clinical translation.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

Regulatory clearance does not mark the end of the model lifecycle but rather the beginning of its long-term management.

“Finally, regulatory clearance does not mark the end of the model lifecycle but rather the beginning of its long-term management.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

Establishing a 'clinician-in-the-loop' validation system can enhance interaction between algorithms and clinical experts, improving clinical credibility and acceptance.

“Establishing a ‘clinician-in-the-loop’ validation system can enhance interaction between algorithms and clinical experts, improving clinical credibility and acceptance.”

Probably not a substantial claim: may not pass S2 (checkable).

Only within a framework that combines scientific validation, ethical compliance, and policy coordination can LLMs safely transition from research prototypes to clinical practice.

“Only within a framework that combines scientific validation, ethical compliance, and policy coordination can LLMs safely transition from research prototypes to clinical practice, advancing intelligent management of diabetes complications toward explainable, regulatable, and sustainable”

Probably not a substantial claim: may not pass S3 (load-bearing).

Establishing robust evaluation standards and regulatory frameworks has become a critical challenge as LLMs transition into clinical decision support systems.

“As LLMs increasingly transition into clinical decision support systems, establishing robust evaluation standards and regulatory frameworks has become a critical challenge.”

Probably not a substantial claim: may not pass S2 (checkable).

Traditional natural language processing metrics such as BLEU or ROUGE fail to fully capture the practical value of medical applications.

“Traditional natural language processing metrics, such as Bilingual Evaluation Understudy (BLEU) or Recall-Oriented Understudy for Gisting Evaluation (ROUGE), fail to fully capture the practical value of medical applications, highlighting the need for multidimensional evaluation criteria encompassing factuality, safety, and clinical reproducibility.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

AI has demonstrated great potential in predicting diabetic complications, but the pathway from research prototype to a clinically reliable tool remains full of challenges.

“Although AI has demonstrated great potential in predicting diabetic complications, the pathway from research prototype to a clinically reliable tool remains full of challenges.”

May combine two claims.

Traditional machine learning methods offer strong interpretability and low computational demands, making them more feasible for clinical adoption.

“They offer strong interpretability and low computational demands, making them more feasible for clinical adoption.”

Classified as probably not a substantial claim. May combine two claims.

The proposed framework emphasizes reproducibility, robust validation, clinical utility assessment, and regulatory compliance, aiming to provide a feasible pathway for moving AI innovations from laboratory research into clinical practice.

“Moreover, the framework emphasizes reproducibility, robust validation, clinical utility assessment, and regulatory compliance, aiming to provide a feasible pathway for moving AI innovations from laboratory research into clinical practice and ultimately enhancing patient management and decision-making.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Such practices would improve transparency, comparability, and the interpretability of AI model evidence in diabetic complication prediction.

“Such practices would improve transparency, comparability, and the interpretability of AI model evidence in diabetic complication prediction.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Translating high-performing AI models into clinically actionable tools remains a major challenge.

“Translating high-performing AI models into clinically actionable tools remains a major challenge.”

Classified as probably not a substantial claim.

The performance of AI models depends on multiple factors including task specificity, data modality, and sample size.

“In conclusion, the performance of AI models depends on multiple factors including task specificity, data modality, and sample size.”

Probably not a substantial claim: may not pass S2 (checkable).

Readers should interpret reported results cautiously, as performance metrics may not be directly comparable across studies.

“These sources of heterogeneity underscore that readers should interpret reported results cautiously, as performance metrics may not be directly comparable across studies.”