ML4Labs · Joud El-Shawa

Accepted and presented at the AAAI SecureAI4Health 2025 Symposium. A multi-modal deep learning system using Clinical BioBERT and LSTM models to predict laboratory test ordering across multiple hospitals.

Excessive laboratory testing contributes to significant healthcare costs and patient burden. This project develops a deep learning framework that predicts whether a glucose test will be ordered in the next AM/PM time bin, using the GEMINI dataset.

The model integrates Clinical BioBERT embeddings for unstructured notes with Long Short-Term Memory (LSTM) networks for temporal modeling, achieving ROC-AUC 0.92, PR-AUC 0.67, and cross-hospital generalization 0.84 ROC-AUC. Temporal recency cues further enhanced predictive stability.

The project was presented at the AAAI SecureAI4Health Symposium 2025 (poster + lightning talk) and showcased at Vector Remarkable 2025, marking progress toward real-time decision support for reducing unnecessary lab tests.

Built with Python, Keras, Tensorflow, SQLAlchemy, Clinical BioBERT, GEMINI Dataset, LSTM