Scholarlink

<p>Advances in artificial intelligence are reshaping neurocritical care by enabling predictive models that support timely and informed clinical decisions. This paper introduces a hybrid multimodal Seq2Seq-NLP architecture that integrates structured time-series data with unstructured clinical narratives to forecast cranial lesion progression. The model leverages a hierarchical attention mechanism to dynamically fuse temporal and semantic features, enhancing both predictive performance and interpretability. Evaluated on four datasets—MIMIC-IV, eICU, TBIcare, and a real-world cohort from Charles Nicolle Hospital of Tunis (CNHT)—the model outperformed conventional baselines, achieving MAE scores between 0.24 and 0.27, RMSE from 0.31 to 0.35, and AUC values ranging from 0.86 to 0.89. An F1-score improvement of 14% and a Cohen’s Kappa of 0.82 on CNHT data confirm its robustness and clinical agreement. Furthermore, attention-based analysis highlights the interpretability of predictions, emphasizing the role of radiology reports and physiological signals as key indicators. Fairness analysis across demographic subgroups revealed potential disparities in predictive performance, which were addressed through stratified sampling and subgroup-aware loss reweighting to promote more equitable model behavior. Overall, this work demonstrates the potential of interpretable multimodal AI systems to support risk stratification and personalized care in neurocritical settings.</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

Predicting Cranial Lesion Progression with a Multimodal Seq2Seq Attention Model

  • 期刊论文

  • 发表日期：

    2026年6月15日

查看全文
  • Mariem Medini,
  • Riadh Bouslimi,
  • Kaouthar Nouira

摘要

Advances in artificial intelligence are reshaping neurocritical care by enabling predictive models that support timely and informed clinical decisions. This paper introduces a hybrid multimodal Seq2Seq-NLP architecture that integrates structured time-series data with unstructured clinical narratives to forecast cranial lesion progression. The model leverages a hierarchical attention mechanism to dynamically fuse temporal and semantic features, enhancing both predictive performance and interpretability. Evaluated on four datasets—MIMIC-IV, eICU, TBIcare, and a real-world cohort from Charles Nicolle Hospital of Tunis (CNHT)—the model outperformed conventional baselines, achieving MAE scores between 0.24 and 0.27, RMSE from 0.31 to 0.35, and AUC values ranging from 0.86 to 0.89. An F1-score improvement of 14% and a Cohen’s Kappa of 0.82 on CNHT data confirm its robustness and clinical agreement. Furthermore, attention-based analysis highlights the interpretability of predictions, emphasizing the role of radiology reports and physiological signals as key indicators. Fairness analysis across demographic subgroups revealed potential disparities in predictive performance, which were addressed through stratified sampling and subgroup-aware loss reweighting to promote more equitable model behavior. Overall, this work demonstrates the potential of interpretable multimodal AI systems to support risk stratification and personalized care in neurocritical settings.