The UCSF Health Problem

Hepatic encephalopathy (HE) is a common complication in patients with cirrhosis and is associated with significantly increased morbidity, mortality, and healthcare utilization.  Early-stage or covert hepatic encephalopathy (CHE), which can be present in up to 60% of patients with cirrhosis, often goes unrecognized by both patients and clinicians because its symptoms are subtle.  Without timely identification and intervention, such as initiating or titrating lactulose or rifaximin, patients can rapidly progress to overt HE – a  complication of cirrhosis that is associated with poor clinical outcomes.  Early intervention in the subclinical stage of the disease (CHE) could potentially prevent complications, improve quality of life, and reduce the burden on healthcare systems.  Current clinical methods for detecting CHE, however, are resource-intensive (due to the use of validated psychometric testing) and therefore not done in routine clinical practice.   This underscores the need for an efficient, potentially AI-driven approach, to screen and detect CHE in routine clinical practice. 

How Might AI Help?

Our group is currently exploring novel methodologies for detection of CHE.  This is strong emerging evidence that suggest a patient’s voice can serve as a biomarker for CHE (PMID 35861546, PMID 39264936).  By leveraging AI-driven analysis of speech characteristics, it may be possible to detect subtle changes, such as altered speed, pitch, or articulation patterns, that correlate with CHE.  At UCSF, the implementation of ambient AI scribes/recordings in clinical settings provides a potentially rich source of audio data.  Our proposed AI solution seeks to analyze these ambient recordings obtained through routine clinical care, applying algorithms (including large language models and specialized speech-processing models) to detect early signals of hepatic encephalopathy.  This approach allows for scalable and efficient implementation, encourage early intervention by flagging potential CHE prior to overt symptoms, and reduce burdens on clinicians as it would be implemented as a part of routine clinical care.

How Would an End-User Find and Use It?

The AI tool would ideally be integrated within existing tabs/dashboards for ambient AI recording and documentation systems (e.g. Abridge and Ambience).  The system would either analyze the recording in the background either in real-time or after the encounter has ended.  If the recording has patterns (either speed, pitch, or articulation) that are suggestive of CHE, an alert would be generated and sent to the treating clinician (envisioned to be hepatologists and advanced practice providers in the liver diseases clinic).  A summary report with insights summarizing the speech abnormalities and potential clinical significance would then be sent as an inbox result to the treating clinician. 

Example of AI Output

What Are the Risks of AI Errors?

AI-based detection of covert hepatic encephalopathy introduces several potential pitfalls:

• False Positives: Overestimating risk could lead to unnecessary patient anxiety.
• False Negatives: Missing early HE signs could delay treatment, leading to worse outcomes.
• Bias and Variations: Voice and speech patterns can vary by accent, language, and comorbid conditions, necessitating robust training sets.
• Data Privacy Concerns: Data use and protection policies for ambient recordings to be used in a research or quality improvement context are still in active development.

To mitigate these risks, we will conduct rigorous validation using retrospective and prospective data in limited pilot studies within the UCSF Hepatology and Liver Diseases outpatient clinics.  Continuous performance monitoring and iterative model refinement will ensure high sensitivity, specificity, and generalizability.  Given that this is an assistive AI tool, final clinical decision-making powers resides with the treating provider, therefore providing a “human-in-the-loop” check for deployment.

How Will We Measure Success?

We will assess the efficacy and impact through quantitative and qualitative metrics:

• Clinical Impact and Detection Accuracy

• Sensitivity / Specificity: Proportion of true CHE cases correctly identified versus psychometric testing in pilot studies.
• Positive Predictive Value (PPV) / Negative Predictive Value (NPV): Assessing how often the tool is correct in its detections or ruling out CHE.
• Change in Overt HE Rates: Monitoring whether early identification reduces the incidence of overt HE among the population of patients with cirrhosis seen at UCSF.

• Workflow Efficiency and Adoption

• Time to Diagnosis: Tracking time from first clinical visit to diagnosis and treatment pre- and post-implementation.
• Provider Utilization: Assess EHR usage time related to summary statements/results generated by the AI tool.

• Equity and Bias Analysis

• Demographic Subgroup Performance: Ensuring consistent tool accuracy across different demographic, language, and cultural groups.

Describe Your Qualifications and Commitment

• This project is led by Dr. Jin Ge, MD, MBA.  Dr. Ge is a transplant hepatologist and a data science and AI researcher within the Division of Gastroenterology and DoC-IT.  He serves as the Director of Clinical AI for the Division of Gastroenterology.  He is experiences in developing, testing, and deploying digital technologies for liver disease care.  He is currently running a pilot study to detect covert hepatic encephalopathy by using chatbots – this study is being extended to include ambient voice recordings in support of this proposal.  If selected, he will commit at least 10% effort for 1 year towards this project to ensure its success.

• Dr. Irene Y. Chen, SM, PhD is an Assistant Professor of Computational Precision Health at UC Berkeley and UCSF.  Her research is focused on safe clinical deployments of artificial intelligence and machine learning models in clinical settings.  She is a close collaborator of Dr. Ge’s and is particularly interested in the potential of utilizing voice biomarkers as a clinical diagnostic tool.