305 - From Data to Detection Through TransfusaScope: Leveraging a Novel Artificial Intelligence Tool for Data Analytics and Detection of Patterns Around Blood Transfusion

Publication Number: 1291.305

Emilee Greager, Johns Hopkins All Children's Hospital, St. Petersburg, FL, United States; Dina Ashour, Johns Hopkins All Children's Hospital, St. Petersburg, FL, United States; Luis Ahumada, Johns Hopkins, St. Petersburg, FL, United States; Georgia Spanos, Johns Hopkins All Children's Hospital, St. Petersburg, FL, United States; Jenn Jones, Johns Hopkins All Children's Hospital, St Petersburg, FL, United States; Dominique Tran, Johns Hopkins All Children's Hospital, St Petersburg, FL, United States; Maha Al-Ghafry, Johns Hopkins University School of Medicine, St Petersburg, FL, United States; Ernest Amankwah, Johns Hopkins University School of Medicine, St Petersburg, FL, United States; Khyzer B. Aziz, Johns Hopkins University School of Medicine, Baltimore, MD, United States; Cassandra Josephson, Medtronic, St. Petersburg, FL, United States

Background: Children are vulnerable to transfusion-related complications. Investigating transfusion impact is hindered by manual, labor-intensive chart reviews. Pediatric transfusions are common and continuous physiologic data are available but underutilized due to difficulty of extraction. TransfusaScope, a novel clinical intelligence tool, addresses the issue by capturing and displaying peri-transfusion physiologic data.

Objective: To apply a novel bioinformatics tool in extraction and visualization of pediatric peri-transfusion physiologic data enabling detection of transfusion-related patterns.

Design/Methods: The validated tool extracted and integrated data from the electronic medical record and continuous vital sign monitors. Variables were displayed on an interactive dashboard with synchronized timelines encompassing the duration of admission. Transfusion duration, product, hemoglobin, respiratory rate, heart rate, mean arterial pressure, blood pressure, and oxygen saturation (SpO2) around a single transfusion event were also extracted. Median vital signs peri-transfusion were compared (Wilcoxon signed rank test; p-value < 0.05). Vital signs (median and range between patients and mean within patients at 2-, 4-, and 12-hours pre and post transfusion) were analyzed. Non-random patterns were determined with Shewhart's charts. With the McNemar's test, 12-hour pre and 48-hour post transfusion patterns were compared. Mean and standard deviation were calculated using data within 3 months of the transfusion.

Results: The clinician-facing interactive TransfusaScope dashboard displays a synchronized event timeline (Figure 1) and reveals patterns in a 2-year-old male with leukemia. The SpO2 panel depicts a dip coinciding with a fluctuating respiratory rate, confirmed by manual chart review as a period of respiratory distress, and the visible recovery of SpO2 occurred after administration of oxygen. The Shewhart's charts found four distinct patterns emerge with analysis of single transfusion events for three patients aged 2-, 3-, and 18-years (Table 1). The patterns corresponding with the aforementioned 2-year-old are displayed below (Figure 2). Specific non-random patterns emerge pre-transfusion, while different non-random patterns emerge post-transfusion, which was consistent among all three patients, further demonstrating the complexity of transfusion administration on the physiologic measurements of pediatric patients.

Conclusion(s): The TransfusaScope may advance pediatric transfusion analytics by unifying multimodal, disparate data into a single, visual interface that reveals transfusion-related physiologic patterns.

Figure 1:
TransfusaScope's clinician-facing interactive dashboard including vital signs from the top down: heart rate in beats per minute, hemoglobin in g/dL, percent oxygen saturation, and respiratory rate in breaths per minute. Vertical dashed lines represent a transfusion event. Horizontal dashed lines represent a predetermined threshold of abnormality.

Table 1:
Patient Demographics, Collective Pre and Post Transfusion Data, and Non-Random Patterns on Individual Shewhart's Charts

Figure 2:
A Shewhart control chart (Nelson, L., 2018) displaying MAP, HR, and SpO2. σ =standard deviation, CL=centerline (mean). Red corresponds with measurements during active transfusion and green corresponds with outside of active transfusion. Patterns: a=One point is greater than 3 standard deviations from the center line. b= Two out of three consecutive points fall more than 2 standard deviations from the center line. c= Four out of five consecutive points fall more than 1 standard deviation from the centerline. d=Nine or more consecutive points fall on the same side of the centerline.