Hospital systems face an increasingly complex operational environment. Inpatient utilization days are projected to grow 9% over the next decade1. With hospitals operating under razor thin margins, maximizing operational efficiency without additional capital investments is imperative.
GE HealthCare is bringing over 125 years of healthcare expertise to this challenge, leveraging deep domain knowledge developed through decades of partnership with clinical providers and patients worldwide. This extensive experience in medical technology and hospital operations helps inform ongoing research into new AI-powered features that will support hospital systems by simplifying the orchestration of patient care, including balancing the competing demands from all service points across the hospital system.
GE HealthCare has demonstrated measurable success in developing offerings to improve operational efficiencies for leading health systems. At Oregon Health & Science University, Command Center enabled the health system to reduce emergency department walkouts by 10%, decrease surgical length of stay by 0.5 days, and improve operating room utilization by 4.5% in its first year. The same offering allowed Duke Health to create capacity for 500 additional patients annually and achieve a 50% reduction in temporary labor requirements.23
At HLTH 2025, GE HealthCare is demonstrating ongoing research into AI-powered models and a conversational interface that will allow clinicians to interact with a conversational agent using a natural language interface, and in real time get answers to questions.
AI Pressure Forecast Model: Predicting hospital-wide pressure points
Hospital departments experience surges that cascade through the entire facility. An influx of emergency patients on a Monday morning can lead to imaging backlogs by noon, delayed admissions by evening, and surgical postponements the next day. Without advance warning, staff scramble to respond, but often this activity can unfold even as patients are already experiencing extended wait times and delayed care.
GE HealthCare is conducting research on a new AI Pressure Forecast model to address this problem. The model uses N-BEATS, a neural network architecture designed for time-series forecasting, to detect patterns in hospital data and predict pressure points up to 72 hours ahead. The model simultaneously processes multiple pressure factors across the hospital ecosystem. These can include demographic criteria like census data, emergency department metrics like boarding length of stay, imaging turnaround times, therapy consultation delays, and staffing levels. The system predicts and normalizes each factor to a score between 1 and 100, where higher values indicate greater operational stress.
The model also incorporates threshold-based override mechanisms where certain critical metrics can trigger immediate alerts when they exceed configured limits. These individual scores can be combined through weighted aggregation methods that hospitals can customize based on their operational priorities, producing department- and facility-level pressure predictions adapted to each institution’s unique context.
The model is both predictive and prescriptive. Based on its findings, it surfaces custom, dynamic action recommendations to the user, aimed at addressing specific high-pressure points projected across departments, allowing hospitals to respond to fast-paced changes as they unfold throughout the day.
Expected Day of Discharge (EDD) Model: Predicting patient discharge timing
Care teams can often face uncertainty about which patients will go home and when. One care team member thinks Mrs. Johnson might be ready tomorrow, but another believes she needs two more days to ensure all discharge needs are coordinated. This lack of coordination has downstream effects, as the case manager can’t arrange home health without knowing discharge dates. As these dynamics play out, the ED has admitted patients waiting for beds. This uncertainty causes suboptimal prioritization of resources and unnecessarily extended patient length of stay.
GE HealthCare is demonstrating research on a new EDD model. The model utilizes a fine-tuned version of BERT (which can understand context by analyzing surrounding text) to predict discharge timing. The model ingests patient demographics, admission information, and medical event sequences as inputs. It processes both structured and unstructured clinical data through a unified framework to identify patterns and predict discharge timing across multiple time windows.
Patients can often have lengthy histories in their records. To overcome the challenge of small context windows—the amount of information that a model can process—GE HealthCare’s research team developed a sliding window approach, ensuring recent events receive priority while maintaining awareness of overall trajectory.
Conversational assistant: Natural language interface for operational queries4
It’s 7 AM and the operations director needs answers fast. Are the medical-surgical units adequately staffed for the next shift? Has bed 314 been cleaned yet? Did Mr. Chen’s cardiology consult happen? Getting these answers means logging into multiple systems, navigating menus, and mentally piecing together data from different sources. By the time she has the information needed for the morning huddle, thirty precious minutes have passed.
GE HealthCare is conducting research on a conversational assistant developed to transform this daily struggle into simple conversation. Hospital leaders can now ask questions in plain English and receive immediate answers—”Show me the staffing trend across all medical-surgical units over the next 24 hours” or “Which patients are likely to discharge this afternoon and what barriers are they waiting on?” The system employs a question-answering architecture using large language models trained specifically for healthcare operations, translating natural language queries into structured database queries. Responses maintain traceability back to source data for audit purposes, with hyperlinks to corresponding views where applicable.
The technical architecture connects several components to make this natural interaction possible. The large language model—an AI system trained on vast amounts of text to understand and generate human-like language—converts user queries from plain English into structured query language. The query execution engine uses this to retrieve data from patient-level records, departmental metrics, and system-wide operational indicators. Finally, a response generation module synthesizes this data into coherent natural language responses delivered as text or tabular outputs within the conversational dialogue box.
Beyond simple query-response, the assistant maintains conversational context across interactions. After asking about emergency department pressure, a user can simply ask “What about tomorrow?” and the system understands this refers to emergency department pressure predictions for the following day. The system also adapts to each hospital’s unique vocabulary—learning institution-specific abbreviations, unit names, and operational terms through a dynamic terminology mapping system that continuously updates based on user interactions.
Together, these innovations show how GE HealthCare researchers are leveraging deep learning and generative AI to help hospitals thrive in an era characterized by a growing influx of patients, persistent labor shortages, and tight margins.
- Technology in development. Not for sale. Not cleared or approved by the U.S. FDA or any other global regulator for commercial availability. ↩︎
- https://www.vizientinc.com/insights/all/2024/10-year-forecast-recommends-innovative-models-right-sites-of-care ↩︎
- https://www.gehccommandcenter.com/2025-outcomes-source-data ↩︎
- Concept only. May never become a product. Not for Sale. Not cleared or approved by the U.S. FDA or any other global regulator for commercial availability. ↩︎
