AI-Driven Risk Modeling for Piperacillin-Tazobactam-Associated Thrombocytopenia in the ICU

Tuesday, March 24, 2026

11:15 AM - 12:15 PM Central Time

Location: Connections Central - RST 16

YZ

Yani Zhang, MD, M.Med

Disclosure information not submitted.
Imad Isaac, MD (he/him/his)

MedStar Washington Hospital Center, District of Columbia

Disclosure(s): No relevant financial relationship(s) to disclose.
LM

Lokesh Manjani, MD

Attending Physician
MedStar Washington Hospital Center

Disclosure information not submitted.
ZJ

Zeqiang Ji, MD

Fellow Physician
Beijing Tiantan Hospital

Disclosure information not submitted.
YG

Yuhao Guo, MD, PhD

Resident Physician
Danbury Hospital/Yale University School of Medicine

Disclosure information not submitted.
SM

Saad Mohammad, MD

Attending Physician
MedStar Washington Hospital Center

Disclosure information not submitted.
XW

Xiangni Wu, MD, PhD

Resident Physician
University of Missouri Kansas City

Disclosure information not submitted.
LY

Linjie Yan, MD

Fellow Physician
Buffalo General Medical Center

Disclosure information not submitted.
AF

Allan Fong, MS

Sr. AI Research Scientist
MedStar Health Research Institute

Disclosure information not submitted.

Introduction: Thrombocytopenia is common in ICU, with a reported prevalence of 8.3–67.6%. A platelet count nadir < 100,000/μL or ≥ 30% drop is associated with higher mortality. Piperacillin-Tazobactam (PIP-TAZ), a widely used empiric antibiotics, has been associated with rapid, severe thrombocytopenia. We conducted a retrospective study to examine PIP-TAZ associated thrombocytopenia (PAT) in ICU patients and developed a machine learning (ML) model for outcome prediction.

Methods: We used a 10-year retrospective dataset from regional acute care facilities to build two ML models: a binary classifier for in-hospital mortality and a regression model for hospital length of stay (hLOS) among survivors. Features included age, sex, race, baseline platelet count, and nadir platelet count. For classification, we trained an XGBoost model (1,000 estimators, max depth 3, learning rate 0.01) using an 80/20 stratified train-test split. For regression model, the same features were used. Model performance was evaluated using MAE, RMSE, and R². Model performance was evaluated using metrics including precision, recall, F1-score, PR AUC, and ROC AUC.

Results: In predicting mortality, the model performed well on the training data (precision 0.96, recall 0.393, F1-score 0.558, PR AUC 0.875, ROC AUC 0.967). The testing PR AUC was 0.169, and ROC AUC was 0.56, suggesting overfitting. Race (Black, 22.2%; White, 17.4%) and platelet counts (baseline 17.2%; nadir 16.8%) were top predictors, followed by age (16.0%) and sex (10.4%). For hLOS, training results showed moderate performance (MAE 4.55, RMSE 6.63, R² 0.62), while testing performance dropped (MAE 8.21, RMSE 16.81, R² –0.08). Nadir (32.4%) and baseline (20.3%) platelet counts were the most predictive features for hLOS.

Conclusions: Black race and nadir platelet count emerged as key predictors of mortality and prolonged hLOS in ICU patients with PAT. To our knowledge, this is the first ML study of PAT in ICU settings. These findings may aid in identifying high-risk patients, increasing clinical suspicion for PAT, and supporting predictive modeling for drug-associated adverse outcomes.