Stanford Health uses AI to reduce clinical deterioration events

Detecting patient clinical deterioration early on holds the promise to decrease mortality and improve outcomes. However, it remains a challenge in both hospital and ambulatory settings.

Stanford Health Care addressed this challenge by incorporating validated models of artificial intelligence and machine learning into clinical decision support systems. They also integrated AI into clinical workflows and improved the patient experience – including reducing wait times, improving quality of care and facilitating critical conversations.

Dr. Shreya Shah is a practicing academic internist, board certified practitioner in clinical informatics and expert in healthcare integration of artificial intelligence at Stanford Health Care. 

She will be speaking on the health system’s AI efforts at the 2023 HIMSS AI in Healthcare Forum, scheduled for December 14-15 in San Diego, offering a case study titled, “How Stanford Health Transformed Patient Care by Combining Compassion with AI-Driven Innovations.”

We spoke with Shah to get a sneak preview of her session and a deeper understanding of how Stanford Health Care is using AI and ML.

Q. Why does detecting clinical deterioration in patients remain a challenge?

A. Patients in hospitals are of increasing complexity and severity of illness while lesser acuity care moves to the home, outpatient care or subacute level of management. Within an academic medical center, this is even more profound with patients at high risk for clinical deterioration.

Early signs may be subtle and vary widely between patients. Identifying which patients need the closest attention is needle-in-a-haystack activity. Moreover, these patients are cared for by multi-person care teams and require assessments of large amounts of data that change over time.

Teams can experience communication gaps, information overload and cognitive biases leading to unanticipated clinical deterioration with major consequences such as emergency resuscitation efforts and unplanned transfers to ICU care. There may also be varying degrees of alignment among team members about perceptions of risk.

Standardized workflows for care coordination that empower all care team members in patient care decisions could help overcome these challenges.

Q. How did you decide that AI and ML was the way to go to help with this challenge?

A. We needed to identify patients at increased risk and align the care team around a collaborative, standardized clinical response. We determined that an ML model can identify patients with a high probability of a future clinical deterioration event without additional tasks for our working clinicians.

The predictions would have to be performed early enough to allow for enough time for the care team to respond. Accuracy is always a concern, and clinicians often believe that the AI system will not tell them something they do not already know.

In our implementations, the emphasis was not whether the model predictions were correct. Rather, for any given patient flagged by the model, physician and nonphysician care team members had to carry out a structured collaborative workflow to assess risk and response. Thus, a probabilistic model creates a team-based trigger.

Our implementation effort focused on these priority areas: 1) Designing a system that would integrate the ML model into a complex healthcare system, 2) Building effective teams and processes to enable the collaborative workflows required for successful implementation, and 3) Deployment of these AI-integrated systems in a way that is both sustainable and scalable for the healthcare enterprise.

The focus was on creating a holistic system that not only incorporates advanced technology but also aligns with the clinical, operational and strategic needs.

Q. What is one example of how incorporating validated models of AI and ML into clinical decision support systems helped Stanford with the clinical deterioration challenge?

A. Our clinical deterioration model was validated on our data to assure model performance; then, the signals were integrated into our EHR with full transparency, including contributing factors and augmented with a mobile alert to the care team.

The ML model is able to update predictions on hospitalized patients every 15 minutes and was used to act as an objective assessor of risk and helped to facilitate alignment and coordination in patient care as an AI-integrated system.

The model underwent site-specific validation to ensure its effectiveness in predicting clinical deterioration events like unplanned ICU transfers within a 6- to 18-hour window. This workflow led to significant increases in multidisciplinary standardized patient assessments and a resulting 20% reduction in clinical deterioration events.

Qualitative evaluation results identified that the model was useful in aligning mental models and driving the necessary workflows for patients flagged by the model with consensus across multidisciplinary team members. By using a reliably and continuously updated risk signal, we aligned the physicians with the rest of the care team to enact a consistent workflow.

Follow Bill’s HIT coverage on LinkedIn: Bill Siwicki
Email him: bsiwicki@himss.org
Healthcare IT News is a HIMSS Media publication.