Scholarlink

This study investigates bias in artificial intelligence (AI) systems for hospitalization risk prediction using a national dataset of 886,814 individuals. The research evaluates models trained on electronic health records, focusing on demographics, comorbidities, and prior hospitalizations. Logistic regression, neural networks, and random forests were employed with techniques like downsampling and SMOTE to address class imbalance. Results showed poor model performance on original data (AUC 0.506–0.512). Performance bias was observed in 63.33% models trained with datasets created through downsampling and SMOTE, with highest AUC in Black Americans (0.712), followed by White (0.673) and Asian Americans (0.671). Subpopulation-specific models performed worse. Performance biases were evident, but not statistically significant.

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

Prevalences of AI Bias in Adolescent Hospitalization Risk Prediction

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Ryan Wu,
  • Azadeh Miran,
  • Yan Cheng,
  • Yijun Shao,
  • Adrienne N. Poon,
  • Philip Candilis,
  • Andrew Robie,
  • Meghan Davies,
  • T. Sean Vasaitis,
  • LaQuandra S. Nesbitt,
  • Qing Zeng-Treitler

摘要

This study investigates bias in artificial intelligence (AI) systems for hospitalization risk prediction using a national dataset of 886,814 individuals. The research evaluates models trained on electronic health records, focusing on demographics, comorbidities, and prior hospitalizations. Logistic regression, neural networks, and random forests were employed with techniques like downsampling and SMOTE to address class imbalance. Results showed poor model performance on original data (AUC 0.506–0.512). Performance bias was observed in 63.33% models trained with datasets created through downsampling and SMOTE, with highest AUC in Black Americans (0.712), followed by White (0.673) and Asian Americans (0.671). Subpopulation-specific models performed worse. Performance biases were evident, but not statistically significant.