The UCSF Health Problem: Hospital Acquired Pressure Injury (HAPI) is a preventable injury to skin or soft tissue that is acquired during a patient’s hospital stay. Reducing HAPI rates is a top priority for UCSF Health leadership as the occurrence of HAPI is detrimental to patient experience and outcomes, results in significant costs (estimated cost to the health system for 1 HAPI is $18,000-$27,000), and is a critical quality measure in the evaluation of hospital performance. The creation of interdisciplinary workflows, utilization of structured problem-solving, and electronic dashboards to monitor HAPI rates and risk have reduced HAPI rates at Benioff Children’s Hospital (BCH) by 64% over the last 15 months. However, poor HAPI bundle compliance continues to be a driver of HAPI as current HAPI rates are 27% higher than national benchmarks. Critical care units account for 95% of HAPI at BCH.

The Gap: There is a lack of real-time electronic medical record (EMR) tools that assist interdisciplinary bedside teams to more effectively support efforts to prevent HAPI.

The objective of this proposal is to develop an AI report to reduce HAPI in critical care units at BCH (Oakland and SF) by improving bundle compliance through an interdisciplinary approach.

Generalizability: The AI report can be leveraged to support HAPI prevention and management across UCSF Hospitals (including adult patients) and could also be adapted to prevent other top harms (e.g Catheter Associated Blood Stream Infections, etc).  

How might AI help? Bedside clinical teams document hundreds of clinical observations related to evidence-based risk factors for HAPI prevention daily in flowsheet rows and notes for each patient, making it challenging to assess if care provided meets expected standards. For example, routine aspects of patient care (such as nutrition delivery) are often disrupted due to procedures, feeding intolerance, or other factors, leading to gaps in knowledge of actual vs ideal state of delivered nutritional goals. Lastly, there are knowledge gaps regarding how to modify care goals when patients with risk factors for developing HAPI are identified.

Multimodal generative AI is ideal for creating a report that can summarize clinical data because it can seamlessly process and integrate diverse data types—such as structured data and free-text notes—into a unified model. By combining techniques like natural language processing for unstructured text and machine learning for structured data, the AI report can extract meaningful insights across all data formats. Importantly, its generative capabilities enable it to summarize clinical data and propose actionable content (such as tailored recommendations) in real-time.

Data Sources: All data elements requested are either documented in flowsheet rows or templated notes in Apex at the individual patient level. Data requested for this report are based on published, validated risk factors for HAPI development. Structured Data: Nutrition (e.g. formula or TPN prescribed, rate of formula/TPN delivery, hours over which nutrition delivered, regular diet order, or percent of regular diet finished, HAPI bundle elements (e.g. repositioning (time of turn and position in which patient is repositioned), perfusion, skin hygiene, mobility promotion, device rotation, application of barrier creams, etc), Medication (e.g. medication administration of vasoactive medication, medications administered through central lines), Medical Devices/LDA (e.g. endotracheal tube, noninvasive positive pressure device, etc).  Unstructured Data: Nutrition: Goal nutritional intake (found in templated registered dietician note), HAPI injury status: Found in templated wound care notes.

How would an end-user find and use it?

Location: The AI report will be found in the “Summary Tab” as nursing, respiratory therapy, and physicians/advanced practice providers (APP) routinely access this tab when viewing patient charts in the inpatient setting. The proposed alert to remind bedside to staff to perform key HAPI bundle components will trigger when a bundle element is overdue when a patient’s chart is opened.

Timing of Support: Support will be most effective in real-time, as HAPI prevention efforts are carried out as frequently as every two hours in critical care units. Reminders to comply with unit policies (bundle assessment/documentation) and suggested recommendations will be made available as the need for improvement is identified.

What end users may see? The AI tool will generate a report summarizing compliance with best practices for the current day and the past week. It will also provide recommendations to improve bundle compliance. This alert can be temporarily silenced if clinical stability prevents compliance. For physicians and APPs, a suggestion box with action items will appear when they click on the hyperlink under the suggested interventions section. For immediate action items (e.g., placing a wound care consult), an alert will appear as soon as the chart is opened.

How are recommendations explained: The non-compliant bundle element and the reason for non-compliance will be visually highlighted, along with a brief suggestion for getting back into compliance. Some interventions need only brief explanations (e.g unit policy is to turn patient every 2 hours, but it has been 4 hours since last turn). For more nuanced suggestions (such as optimizing nutrition), the report will describe why and length of time nutrition is not at goal and propose actionable suggestions (such as possible means for nutritional delivery).

Picture of Embedded AI Tool:

What are the risks of AI error? 

AI errors could be caused by missing data and hallucinations. Missing data could lead to both under and overestimation of HAPI risk. However, because a key component of this report is to increase bundle compliance (and by association documentation of bundle compliance), we plan to mitigate errors from missing data by improving data entry. To validate the AI model, domain experts will regularly review its outputs and compare them with existing data reports (HAPI dashboard, nurse audit data) and analysis of component data (e.g., skin assessment, patient positing, etc) through clarity queries. Additionally, accuracy of information extracted from note text will be assessed by comparing AI output of notes with text matching (text mined from templated notes [existing standard]). The project lead has extensive experience analyzing structured and unstructured EMR data (Mahendra et al, Pulm Circ, 2023; Mahendra et al, Crit Care Explor, 2021). A feedback button will also be available for end-users to report inaccuracies. 

Feedback: A feedback button will be available for end-users to report inaccuracies at any time in the EMR.  Additionally, we will build an active user-centered design process working group to seek and incorporate iterative input from end users as the report is developed and implemented in the clinical setting. Feedback will be used to 1. Continuously retrain and refine the model for improved accuracy and reliability 2. Optimize bedside use of this tool 3. Identify other factors (staffing, equipment availability, etc) that may be identified as important factors to monitor to improve HAPI prevention effectiveness. 

How will success be measured? Success will be defined by both outcome and process measures. The primary outcome measure will be BCH Stage 2 and greater HAPI rates/1,000 patient days. The goal would be to realize a sustained reduction of HAPI rates to below national benchmark data across BCH. HAPI rates are readily available as they are closely monitored across UCSF Health in the zero-harm dashboard. The secondary outcome is improved adherence to the HAPI prevention bundle (a key measure of adoption and utilization of the tool). Documentation of bundle compliance will be assessed through query of structured flowsheet row data and in person audits (routinely performed by nursing and documented in the digital rounding tool). An automated dashboard is also being built to report bundle compliance by UCSF Health IT.  

Qualifications: Malini Mahendra MD (project lead) is a pediatric intensivist and data scientist. As a certified clarity data analyst with expertise in use of machine learning and natural language processing algorithms, she has extensive experience analyzing UCSF EMR data for quality improvement.  She is also the local quality improvement and informatics lead for the mission bay pediatric intensive care unit.

Deborah Franzon MD, MHA (collaborator) is the Executive Medical Director for Quality and Safety at BCH with over 20 years of clinical expertise and implementation science experience.   Dr. Franzon has been the recipient of prior UCSF open proposal awards including: Learning Health System Innovations (2017-2019) building a machine learning predictive model for extubation in PICU patients, soon to be integrated into Epic; and Caring Wisely (2021) to reduce delirium in PICU patients across BCH SF and Oakland.  She developed, implemented and published on the impact of an EHR enhanced dashboard in reducing CLABSI rates in the PICU. (Pediatrics, 2014) She has a proven track reducing harm, improving clinical outcomes, and driving data-informed, innovative change through collaborative, patient-centered leadership.

This project is supported by BCH Leadership: Nicholas Holmes MD MBA (President of BCH), Joan Zoltanski MD MBA (Chief Medical Officer of BCH), Judie Boehmer RN MN (Chief Nursing Officer BCH), Michael Lang MD (Chief Medical Information Officer for Children's Services), Jeff Fineman MD (Pediatric Critical Care Division Chief), Shan Ward MD and Loren Sacks MD (MB Pediatric ICU and Pediatric Cardiac ICU Medical Directors), Mary Nottingham RN and Lori Fineman RN (MB Clinical Nurse Specialists, Pediatric Critical Care and Pediatric Cardiac Critical Care), Mandeep Chadha MD (BCH-Oakland Pediatric Critical Care Quality Lead).

Selected References:

Mahendra M, Chu P, Amin EK, Nawaytou H, Duncan JR, Fineman JR, Smith-Bindman R. Associated radiation exposure from medical imaging and excess lifetime risk of developing cancer in pediatric patients with pulmonary hypertension. Pulm Circ. 2023 Aug 21;13(3):e12282. doi: 10.1002/pul2.12282. PMID: 37614831; PMCID: PMC10442605.

Mahendra M, Luo Y, Mills H, Schenk G, Butte AJ, Dudley RA. Impact of Different Approaches to Preparing Notes for Analysis With Natural Language Processing on the Performance of Prediction Models in Intensive Care. Crit Care Explor. 2021 Jun 11;3(6):e0450. doi: 10.1097/CCE.0000000000000450. PMID: 34136824; PMCID: PMC8202578.

Pageler NM, Longhurst CA, Wood M, Cornfield D, Suermondt J, Sharek P, Franzon D. Use of electronic medical record-enhanced checklist and electronic dashboard to decrease CLABSIs. Pediatrics. 2014;133(3):e738-e746. doi:10.1542/peds.2013-2249

Solutions for Patient Safety Operational Definition and Prevention Bundle. (rev. 2020). https://static1.squarespace.com/static/62e034b9d0f5c64ade74e385/t/636177d9b97cd87674b58f18/1667332057726/PI-Bundle-Op+Def.pdf

Padula WV, Delarmente BA. The national cost of hospital-acquired pressure injuries in the United States. Int Wound J. 2019;16(3):634-640. doi:10.1111/iwj.13071

Johnson AK, Kruger JF, Ferrari S, et al. Key Drivers in Reducing Hospital-acquired Pressure Injury at a Quaternary Children's Hospital. Pediatr Qual Saf. 2020;5(2):e289. Published 2020 Apr 7. doi:10.1097/pq9.0000000000000289