Scholarlink

<p>Digital twins for oncology must move beyond risk forecasting to deliver actionable treatment guidance. We present an EHR-native causal policy layer for hospitalized prostate-cancer (PCa) patients that estimates drug-level counterfactual outcomes and produces safety-aware recommendations. Using a PCa cohort from MIMIC-IV, we study commonly used inpatient medications across five endpoints - Mortality, ICU transfer, ED (Emergency Department) 30-day revisit, AKI (Acute Kidney Injury), and LOS (Length of Stay). Effects are estimated with cross-fitted doubly-robust estimators (AIPW/DR) under propensity trimming, with ATE (Average Treatment Effect), ATT (Average Treatment Effect on the Treated), ATC (Average Treatment Effect on the Controls) and FDR (False Discovery Rate) control. We then learn benefit-targeted policies and enforce a conservative safety overlay using curated DDI (Drug-Drug Interaction) rules, comorbidity vetoes, and frequent-pattern mining. The drug-level estimates are clinically coherent (e.g., loop diuretics increase AKI risk; acetaminophen shortens LOS), and benefit-targeted policies outperform treat-all and risk-only baselines. Frequent-pattern analysis highlights cardio-renal and analgesia-bowel clusters, underscoring the need for explicit safety gates. Overall, the layer advances digital twins from prediction to prescription, yielding individualized, auditable treatment menus rather than risk scores alone.</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

Estimating treatment effects and recommending safe drug policies in prostate cancer inpatients: a causal inference approach towards digital twins

  • 期刊论文

  • 发表日期：

    2026年3月21日

查看全文
  • Annette John,
  • Reda Alhajj,
  • Jon Rokne

摘要

Digital twins for oncology must move beyond risk forecasting to deliver actionable treatment guidance. We present an EHR-native causal policy layer for hospitalized prostate-cancer (PCa) patients that estimates drug-level counterfactual outcomes and produces safety-aware recommendations. Using a PCa cohort from MIMIC-IV, we study commonly used inpatient medications across five endpoints - Mortality, ICU transfer, ED (Emergency Department) 30-day revisit, AKI (Acute Kidney Injury), and LOS (Length of Stay). Effects are estimated with cross-fitted doubly-robust estimators (AIPW/DR) under propensity trimming, with ATE (Average Treatment Effect), ATT (Average Treatment Effect on the Treated), ATC (Average Treatment Effect on the Controls) and FDR (False Discovery Rate) control. We then learn benefit-targeted policies and enforce a conservative safety overlay using curated DDI (Drug-Drug Interaction) rules, comorbidity vetoes, and frequent-pattern mining. The drug-level estimates are clinically coherent (e.g., loop diuretics increase AKI risk; acetaminophen shortens LOS), and benefit-targeted policies outperform treat-all and risk-only baselines. Frequent-pattern analysis highlights cardio-renal and analgesia-bowel clusters, underscoring the need for explicit safety gates. Overall, the layer advances digital twins from prediction to prescription, yielding individualized, auditable treatment menus rather than risk scores alone.