Scholarlink

<p>Postoperative complications following complex brain tumor surgery remain a significant cause of morbidity. This study aimed to develop and independently validate a multi-modal artificial intelligence model integrating preoperative imaging, intraoperative physiological time-series, and biochemical data for complication prediction following glioma and skull base tumor surgery. This retrospective multi-center study included 15,000 patients from 7 centers across 5 countries. A Transformer-based architecture with cross-modal fusion was developed. Missing data were handled using multiple imputation. Model performance was evaluated using AUC, calibration metrics, decision curve analysis, SHAP interpretability, SHAP interaction analysis, and temporal attention visualization. The model achieved AUC 0.91 (95% CI: 0.89–0.93) in derivation (<i>n</i> = 8,300) and 0.87 (95% CI: 0.85–0.89) in independent multi-center validation (<i>n</i> = 6,700). Calibration was moderate (Brier score 0.112, 41.7% improvement over null; ECE = 0.048; H-L <i>p</i> = 0.017). Tumor midline invasion (OR = 3.2, 95% CI: 2.1–4.9, <i>p</i> = 0.001) and postoperative hemoglobin nadir &lt; 100&#xa0;g/L (OR = 2.8, 95% CI: 1.8–4.2, <i>p</i> = 0.003) were the strongest predictors. SHAP interaction analysis revealed a synergistic effect between midline invasion and hypotension duration (interaction SHAP = 0.062). Temporal attention visualization showed highest attention to MAP during induction/emergence and to BIS during tumor resection. Decision curve analysis demonstrated positive net benefit at clinically relevant thresholds (15%-35%). This multi-modal Transformer model showed good discrimination and moderate calibration across multi-ethnic populations, with imperfect calibration at high predicted probabilities. Key risk factors and their interactions were identified through SHAP analysis. Prospective validation against expert clinical judgment is needed before clinical deployment.</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

Multimodal artificial intelligence predicts postoperative complications following complex brain tumor surgery in an international multicenter retrospective cohort

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月1日

查看全文
  • Ping Zhang,
  • Hao Lin,
  • Sharma Madhusudan,
  • Seidu A. Richard,
  • Min Li,
  • Jun Liu,
  • Igor F. Sikorsky,
  • Pedro Machado,
  • Zhigang Lan

摘要

Postoperative complications following complex brain tumor surgery remain a significant cause of morbidity. This study aimed to develop and independently validate a multi-modal artificial intelligence model integrating preoperative imaging, intraoperative physiological time-series, and biochemical data for complication prediction following glioma and skull base tumor surgery. This retrospective multi-center study included 15,000 patients from 7 centers across 5 countries. A Transformer-based architecture with cross-modal fusion was developed. Missing data were handled using multiple imputation. Model performance was evaluated using AUC, calibration metrics, decision curve analysis, SHAP interpretability, SHAP interaction analysis, and temporal attention visualization. The model achieved AUC 0.91 (95% CI: 0.89–0.93) in derivation (n = 8,300) and 0.87 (95% CI: 0.85–0.89) in independent multi-center validation (n = 6,700). Calibration was moderate (Brier score 0.112, 41.7% improvement over null; ECE = 0.048; H-L p = 0.017). Tumor midline invasion (OR = 3.2, 95% CI: 2.1–4.9, p = 0.001) and postoperative hemoglobin nadir < 100 g/L (OR = 2.8, 95% CI: 1.8–4.2, p = 0.003) were the strongest predictors. SHAP interaction analysis revealed a synergistic effect between midline invasion and hypotension duration (interaction SHAP = 0.062). Temporal attention visualization showed highest attention to MAP during induction/emergence and to BIS during tumor resection. Decision curve analysis demonstrated positive net benefit at clinically relevant thresholds (15%-35%). This multi-modal Transformer model showed good discrimination and moderate calibration across multi-ethnic populations, with imperfect calibration at high predicted probabilities. Key risk factors and their interactions were identified through SHAP analysis. Prospective validation against expert clinical judgment is needed before clinical deployment.