Scholarlink

<p>Accurate and timely diagnosis of infectious diseases is vital for effective clinical management. We present MedTransNet, a multimodal transformer framework for early detection, risk stratification, and antimicrobial guidance from electronic health records (EHR).&#xa0;Modality-specific encoders (demographics, laboratories, medications, vital signs) feed cross- and self-attention blocks with causal temporal masking. Monte Carlo dropout provides epistemic uncertainty. We evaluate on the eICU Collaborative Research Database (eICU-CRD v2.0.1), using patient-level 5-fold cross-validation; uncertainty is calibrated via temperature scaling. Early detection time is computed as the latency between the model’s alert threshold crossing and the earliest clinical recognition time.&#xa0;On infectious-disease classification, MedTransNet achieves accuracy <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(0.905 \pm 0.012\)</EquationSource> </InlineEquation>, precision <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(0.875 \pm 0.013\)</EquationSource> </InlineEquation>, recall <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(0.890 \pm 0.012\)</EquationSource> </InlineEquation>, F1 <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(0.882 \pm 0.011\)</EquationSource> </InlineEquation>, and AUC-ROC <InlineEquation ID="IEq5"> <EquationSource Format="TEX">\(0.901 \pm 0.007\)</EquationSource> </InlineEquation>. For sepsis, the model flags risk on average <InlineEquation ID="IEq6"> <EquationSource Format="TEX">\(11.8 \pm 2.3\)</EquationSource> </InlineEquation> hours before clinical recognition at the tuned operating point. Uncertainty correlates with error and improves calibration (ECE <InlineEquation ID="IEq7"> <EquationSource Format="TEX">\(\downarrow\)</EquationSource> </InlineEquation> with temperature scaling). Attention- and attribution-based analyses highlight clinically coherent predictors.&#xa0;MedTransNet integrates multimodal signals, temporal dynamics, and uncertainty to deliver accurate, interpretable, and risk-aware predictions for early infectious-disease detection and antimicrobial decision support. Findings warrant prospective and multi-center validation for stewardship impact.</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

MedTransNet: a transformer-based clinical decision support framework for early infectious disease prognosis and antimicrobial stewardship

  • 期刊论文

  • 发表日期：

    2025年12月4日

查看全文
  • Yassin Ben Youssef,
  • Asma Amdouni,
  • Ghaith Manita,
  • MD Ella Ben Hassine

摘要

Accurate and timely diagnosis of infectious diseases is vital for effective clinical management. We present MedTransNet, a multimodal transformer framework for early detection, risk stratification, and antimicrobial guidance from electronic health records (EHR). Modality-specific encoders (demographics, laboratories, medications, vital signs) feed cross- and self-attention blocks with causal temporal masking. Monte Carlo dropout provides epistemic uncertainty. We evaluate on the eICU Collaborative Research Database (eICU-CRD v2.0.1), using patient-level 5-fold cross-validation; uncertainty is calibrated via temperature scaling. Early detection time is computed as the latency between the model’s alert threshold crossing and the earliest clinical recognition time. On infectious-disease classification, MedTransNet achieves accuracy \(0.905 \pm 0.012\) , precision \(0.875 \pm 0.013\) , recall \(0.890 \pm 0.012\) , F1 \(0.882 \pm 0.011\) , and AUC-ROC \(0.901 \pm 0.007\) . For sepsis, the model flags risk on average \(11.8 \pm 2.3\) hours before clinical recognition at the tuned operating point. Uncertainty correlates with error and improves calibration (ECE \(\downarrow\) with temperature scaling). Attention- and attribution-based analyses highlight clinically coherent predictors. MedTransNet integrates multimodal signals, temporal dynamics, and uncertainty to deliver accurate, interpretable, and risk-aware predictions for early infectious-disease detection and antimicrobial decision support. Findings warrant prospective and multi-center validation for stewardship impact.