Scholarlink

Background <p>Diabetes is a chronic illness that requires early stratification and consistent management to prevent serious complications. Despite advancements in AI, existing tools often focus on prediction in isolation, lacking an integrated link to post-diagnosis management.</p> Objective <p>This study aims to develop and evaluate a hybrid feasibility-stage system that integrates a Decision Support System (DSS) for diabetes prediction with a Recommender System (RS) for guideline-based lifestyle management.</p> Methods <p>A hybrid framework was proposed combining Deep Learning (DL) and rule-based techniques. The DSS employs astacked Bidirectional Long Short-Term Memory (BiLSTM) model to capture complex relationships in clinical data. A rule-based RS, anchored in the ADA 2023 and PES 2022 guidelines, generates non-prescriptive recommendations for diet, exercise, and mental wellness to support clinical triage and patient education. The framework was evaluated using a three-tiered protocol: stratified tenfold cross-validation (CV) for internal stability, an independent hold-outtest set (N = 84), and a preliminary external validation cohort (N = 15).</p> Results <p>The proposed model achieved a mean Accuracy of 94.58% ± 2.70 and a mean Sensitivity of 97.51% ± 3.90, significantly outperforming traditional ML and DL baselines (McNemar p = 0.0063). On the independent hold-out set(N = 84), the model achieved 96.0% accuracy (95% CI: 89.3–99.2%) and 100% sensitivity (95% CI: 92.7–100%), witha well-calibrated Brier score of 0.072. External validation demonstrated 93.3% concordance with expert assessments.</p> Conclusion <p>These findings highlight the potential of hybrid intelligent systems to support preliminary risk stratification and clinician-mediated lifestyle guidance. Future research will focus on large-scale, multi-center validation and the adaptation of the framework to handle missing clinical features, ensuring practical utility beyond this preliminary feasibility-stage prototype.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A hybrid deep learning approach for diabetes prediction and personalized recommendations

  • 期刊论文

  • 发表日期：

    2026年5月2日

查看全文
  • Ayesha Shahid,
  • Muskan Zahra,
  • Saiqa Adrees,
  • Muhammad Hamid,
  • Fahima Hajjej,
  • Tagrid Abdullah N. Alshalali

摘要

Background

Diabetes is a chronic illness that requires early stratification and consistent management to prevent serious complications. Despite advancements in AI, existing tools often focus on prediction in isolation, lacking an integrated link to post-diagnosis management.

Objective

This study aims to develop and evaluate a hybrid feasibility-stage system that integrates a Decision Support System (DSS) for diabetes prediction with a Recommender System (RS) for guideline-based lifestyle management.

Methods

A hybrid framework was proposed combining Deep Learning (DL) and rule-based techniques. The DSS employs astacked Bidirectional Long Short-Term Memory (BiLSTM) model to capture complex relationships in clinical data. A rule-based RS, anchored in the ADA 2023 and PES 2022 guidelines, generates non-prescriptive recommendations for diet, exercise, and mental wellness to support clinical triage and patient education. The framework was evaluated using a three-tiered protocol: stratified tenfold cross-validation (CV) for internal stability, an independent hold-outtest set (N = 84), and a preliminary external validation cohort (N = 15).

Results

The proposed model achieved a mean Accuracy of 94.58% ± 2.70 and a mean Sensitivity of 97.51% ± 3.90, significantly outperforming traditional ML and DL baselines (McNemar p = 0.0063). On the independent hold-out set(N = 84), the model achieved 96.0% accuracy (95% CI: 89.3–99.2%) and 100% sensitivity (95% CI: 92.7–100%), witha well-calibrated Brier score of 0.072. External validation demonstrated 93.3% concordance with expert assessments.

Conclusion

These findings highlight the potential of hybrid intelligent systems to support preliminary risk stratification and clinician-mediated lifestyle guidance. Future research will focus on large-scale, multi-center validation and the adaptation of the framework to handle missing clinical features, ensuring practical utility beyond this preliminary feasibility-stage prototype.