The UCSF Office of Population Health (OPH) Complex Care Management (CCM) team provides advanced care management services to adult patients with high medical and/or psychosocial complexity who are high utilizers of inpatient or emergency department (ED) services. The CCM program involves essential high-touch support such as assessing individual patients’ healthcare challenges, developing targeted care plans, providing health education and coaching, coordinating linkage to care, and connecting to other community resources. Prior analysis of this program’s outcomes showed a statistically significant decrease in ED and observation encounters for patients enrolled in the program. The impact of this program therefore has significant potential to help address UCSF Health’s current ED crowding and bed capacity challenges, reduce readmission rates, and help meet quality metrics associated with specific patient populations. 

Currently, one of the most time-consuming challenges faced by the CCM team is determining individual patient eligibility for the CCM program. Despite using a reporting workbench that identifies patients meeting initial objective criteria, the team must still manually chart review to determine eligibility, which can consume up to 30 minutes per patient. This manual process can also be inconsistent, as it has been noted that different reviewers assess medical and social complexities differently based on their training and clinical background. 

Ideally, much of the current time spent by the OPH CCM team on manual chart review could instead be spent on higher-yield patient care activities, expanding the team’s ability to manage a higher volume of patients without requiring additional staff, as well as allowing its team members to practice at the top of their licenses. Additionally, standardizing definitions across patient enrollment criteria and workflows around the patient identification process would reduce variability between team members in applying inclusion and exclusion criteria.  

We believe that AI can be used to optimize these areas of need and have done preliminary work to show how this can apply. Through the AI demonstration project, we hope to validate our initial prototype; build an automated pipeline at scale; design, implement, and study appropriate clinical workflows; and work with appropriate governance committees to ensure safe and effective use. While this proposal is initially focused on addressing the needs of the CCM program in OPH, we believe that this model can be scaled to multiple other programs within OPH who would benefit from improved efficiency in patient identification processes.  

This work is particularly important given the current environment for UCSF Health. As our health system grows, we may not be able to hire more staff, and so will need to be more efficient. If successful, this initiative will allow CCM team members to re-allocate approximately one day per week to seeing additional patients instead of performing chart reviews. 

To address these challenges, we built an initial prototype using prompt engineering in Versa to optimize the CCM patient eligibility determination process. This AI system mimics the current workflow of CCM staff, reviewing relevant notes for complex patients with high inpatient and ED utilization within a set time period for specific inclusion and exclusion criteria. Specifically, patients must meet at least one of six indicators of medical or psychosocial complexity. These indicators include:

1. Uncontrolled chronic medical conditions
2. Use of high-risk medications
3. Evidence of barriers to medical adherence or limited health literacy
4. Lack of social support
5. Older age with frailty or poor functional status or cognitive impairment
6. Social determinants of health (SDoH) challenges

We completed an initial assessment comparing LLM-generated eligibility determinations with those from manual reviews using 27 distinct patient charts. Our findings revealed 100% sensitivity and 77% specificity in eligibility determination. The lower specificity was attributed to over-inclusion, as the initial prompt lacked explicit instructions on defining medical and social complexity. Patients with lower complexity levels were incorrectly classified as eligible. We expect that further refinements will lead to better specificity. During this initial assessment, we also observed the AI system’s ability to standardize the review process, as it identified 3 patients who were initially misclassified for eligibility upon re-review by the CCM team. 

The initial results from this AI system were received positively by the CCM team and OPH leadership. Together, we believe that this can be used as an adjunct tool to significantly streamline the time spent on and improve the overall quality of eligible patient review. 

Twice a month, the CCM team reviews a list of patients who have met certain utilization thresholds and other simple inclusion criteria, and subsequently performs chart review for the preceding 6 months of data to determine whether these patients are eligible for enrollment in the CCM program. Approximately 73 patients are reviewed every month. For each patient, the AI system described above would review and synthesize relevant patient notes to summarize findings for the CCM team. It would generate the following outputs: eligibility determination, specific inclusion and exclusions that led to that determination, as well as which notes informed those inclusion and exclusion decisions (see question 4).This setup increases transparency in the AI system, as it can show how it reached its conclusions as well as where to find relevant information for further review. This output should ideally be integrated into the existing Epic Reporting Workbench report used by the CCM team, as it includes patient data and patient outreach tools. CCM team members would then use this information to assist in their review.  

Please see above for an example of what the output from the AI system might look like. This data would ideally be integrated into an Epic Reporting Workbench report. Here, we only show 1 exclusion criteria for illustrative purposes. In reality, the output would include many more criteria currently utilized by the CCM team.  

False negatives, false positives, and hallucinations are critical errors we will monitor for in this AI system. These errors can arise at multiple levels, such as when the LLM is outputting answers to individual inclusion and exclusion criteria, and when it is giving a final eligibility assignment for each patient.  As we envision this system being used to augment the review process rather than automate it, it is particularly important to monitor the false negative, false positive, and hallucination rates in its responses to individual inclusion and exclusion criteria, as we anticipate the CCM team to primarily rely on these responses to optimize their review. Large error rates in any of these categories will limit the usefulness of the AI system.  

Including notes from CareEverywhere in the AI system will also be fundamental to ensuring accuracy of the AI output. As these notes are also included in review by CCM staff in their current processes 

The only way to measure if these errors are occurring is to retain some number of complete manual chart reviews per month (i.e.10-20) and compare the AI system’s performance to this gold standard. In terms of mitigation strategies, there are a few possible strategies, including: (1) For criteria that includes math/logic/structured data, leverage traditional reporting rather than generative AI. (2) Prioritize low false negative and hallucination rates over false positive rates, if able. (3) When new patterns of error are discovered, use iterative cycles of prompt engineering to attempt to resolve the error. 

Please see below for lists of measures being tracked already, as well as measures that we want to track as part of the AI deployment. For this project to be successful, the AI system should be accurate, increase efficiency/satisfaction, increase the capacity of the CCM team, and be low cost to deploy and maintain relative to its benefits.  

Executive Sponsor: Tim Judson, MD, Interim Chief Population Health Officer 

Co-project leads: Esther Hsiang, MD, MBA, Interim Medical Director of Care Delivery Transformation, UCSF Office of Population Health and Leo Liu, MD, MS, FAMIA, Associate Program Director, Clinical Informatics Fellowship.