In 1991-1992, Eric Horvitz and Mike Shwe created a handsfree AI-powered trauma care system. The system demonstrated the power of delivering diagnostic assessments and recommendations for guiding observations in time-critical settings. The prototype employed speech understanding, a heads-up display, and a Bayesian-network--based sequential diagnosis system that reasoned about the criticality of injuries from observations. The system generated and continued to update differential diagnoses and employed value-of-information to recommend the next best observations to make about the patient. Speech recognition was enhanced by providing the recognizer with context based on real-time inferences about likelihoods of different utterances. The system was implemented on an early portable PC that was attached to the caregiver's belt. The project was partly funded by DARPA.
The work was done within Knowledge Industries, an AI startup founded in 1988 by Jack Breese, David Heckerman, and Eric Horvitz (later acquired by Microsoft). The system employed a Private Eye personal display and a pocket-sized PC running speech recognition, calling Bayesian inference using a Bayesian network, in a sequential-diagnosis architecture, employing value-of-information computations to compute and display recommendations for best next observations to make. A command and control speech system was used to make selections and input findings about a patient. The system was designed for exploration of the feasibility of providing paramedics with expert diagnostic decision support in time-critical settings. Brad Cushing (U Maryland Shock Trauma) and, later, Adam Seiver (Stanford Trauma) helped to assess the Bayesian trauma care model. Later refinements were supported by an SBIR grant to Knowledge Industries from DARPA.
Handsfree AI-powered trauma decision-support (1992). Eric Horvitz, playing role of paramedic, and Jim White, starring as injured cyclist, demonstrate handsfree trauma decision support system, employing sequential diagnosis.
Handsfree trauma decision-support system (1992). Eric Horvitz with handsfree trauma decision support system. Pocket-sized PC loaned from DARPA ran speech recognition and sequential diagnostic reasoning about patient status using Bayesian network assessed from experts Brad Cushing (Baltimore Shock Trauma) and Adam Seiver (Stanford University Trauma). System continues to listen for observations by voice, updates differential diagnosis and recommends next best evaluation actions.
Short video showing use of the system.
Longer video with additional background and details on Bayesian-network--powered sequential diagnosis with an abdominal pain knowledge base.
AI modeling: E. Horvitz and A. Seiver. Time-Critical Action: Representations and Application. Proceedings of the Thirteenth Conference on Uncertainty in Artificial Intelligence, August 1997.