First, the early detection of clinical deterioration holds the promise of reducing mortality rates and improving outcomes. However, this remains a challenging task in both clinical and ambulatory settings.
By integrating reliable machine learning and artificial intelligence models into clinical decision support systems, Stanford Health Care successfully addressed this issue. This initiative not only enhanced patient care but also streamlined clinical workflows, reducing wait times, elevating care standards, and facilitating crucial conversations.
At Stanford Health Care, Dr. Shreya Shah, a proficient scientific endocrinologist and board-certified medical informatics professional, is a leading authority on the integration of artificial intelligence in healthcare.
During the upcoming 2023 HIMSS AI in Healthcare Forum scheduled for December 14–15 in San Diego, Dr. Shah will present a groundbreaking research scenario titled “How Stanford Health Transformed Patient Care by Combining Compassion with AI-Driven Innovations.”
To delve deeper into Stanford Health Care’s utilization of AI and ML technologies, we engaged in a discussion with Dr. Shah to gain insights into her pioneering work.
Q. Why is the identification of medical deterioration in individuals still a challenging task?
A. The complexity of patient conditions in hospitals, especially in academic medical centers, has increased with the shift towards outpatient care and higher acuity levels. Diagnosing deterioration in such patients, who exhibit a wide range of symptoms, requires meticulous attention. It often involves sifting through extensive data that evolves over time and necessitates coordinated efforts from multidisciplinary care teams.
Communication breakdowns, information overload, and cognitive biases within teams can impede the timely identification of clinical deterioration, leading to adverse outcomes like unplanned ICU admissions and emergency interventions. Additionally, varying perceptions of risk among team members can further complicate the situation.
These challenges can be effectively addressed through standardized care coordination workflows that empower all team members to make informed decisions regarding patient care.
Q. How did you determine that AI and ML were the optimal solutions to address this issue?
A. Recognizing the need for a coordinated clinical response to identify high-risk patients without overburdening clinicians, we leveraged an ML model to predict patients at risk of future deterioration. The key was to provide timely predictions without overwhelming clinicians with redundant information.
Our focus was not solely on the accuracy of the predictive model but on fostering collaborative workflows where medical and non-medical teams could collectively assess risks and responses for flagged patients. This approach ensured a team-based approach driven by probabilistic modeling.
Our implementation strategy centered on integrating the ML model into the healthcare system, establishing efficient team processes for collaborative workflows, and ensuring scalable deployment of AI-integrated systems in healthcare settings. The goal was to create a comprehensive system that aligns with clinical, operational, and strategic requirements while embracing cutting-edge technology.
Q. Can you provide an example of how Stanford effectively addressed clinical deterioration by incorporating validated AI and ML models into clinical decision support systems?
A. Upon validating our clinical deterioration model on internal data, the alerts generated by the model were seamlessly integrated into our ERP system, complete with contributing factors. This integration included real-time alerts to the care team, enhancing response times and coordination.
The ML model, capable of providing predictions on patient deterioration every 15 minutes, served as a risk assessment tool, aiding in care coordination. Site-specific validation of the model’s accuracy in predicting deterioration events, such as unplanned ICU transfers, within a specific timeframe led to a 20% reduction in such events. This validation process significantly improved standardized multidisciplinary assessments.
Subjective evaluations confirmed the model’s effectiveness in guiding collaborative workflows and achieving consensus among multidisciplinary team members. By utilizing an accurate and regularly updated risk signal, we established a seamless communication process between physicians and the broader care team, enhancing patient care coordination.
