Assessing the Impact of Early Albumin Administration in Patients Requiring Intravascular Resuscitation: Application of Machine Learning Techniques to Electronic Health Record (EHR) Data
Author(s)
Viayna E1, Ardiles T2, Runken MC2, Weimer I3, Zhang X3, Pathan F3, Lodaya K3
1Grifols International, Sant Cugat del Valles, B, Spain, 2Grifols SSNA, RTP, NC, USA, 3Boston Strategic Partners, Inc., Boston, MA, USA
Presentation Documents
OBJECTIVES: Exogenous human albumin administration has shown improved outcomes in patients requiring intravascular resuscitation (e.g., spontaneous bacterial peritonitis, and hepatorenal syndrome). However, its acute use in other conditions and liver disease-associated complications requiring fluid resuscitation is unclear. We applied machine learning (ML) techniques to electronic health records data from patients with anemia and gastrointestinal bleeding, alcoholic liver disease, hepatic failure, acute pancreatitis, and nephrotic syndrome, to identify subpopulations that may benefit from early albumin administration.
METHODS: Using de-identified Cerner Real World Data and ICD-10-CM codes, we identified five cohorts receiving albumin during hospitalization between 10/1/2015–03/31/2021. Each cohort was 1:1 propensity score matched (PSM) on baseline characteristics and early albumin receipt (<24 hours of admission), using logistic regression and k-nearest neighbor models. Decision tree (DT) models were used to predict four outcomes: 30-day mortality (30DM), 90-day readmission, hospital length of stay, and hospital-free days. After model visualization, we identified relevant splits in the DTs to assess subgroups of interest. Once identified, the model conditions were applied to the full dataset to extract the subgroup. Significance testing was conducted using chi-square test.
RESULTS: In total, 361,704 distinct patients were analyzed. After PSM, the most notable positive signal was observed for the 30DM outcome in the anemia with gastrointestinal bleeding cohort (N=10,388). DT models classified favorable outcomes for a subgroup (n=598) with no cirrhosis and severe anemia (baseline hemoglobin <7.0 g/dL). This prediction was validated on the full dataset where this subgroup had a 25.8% lower 30DM rate compared to those not receiving early albumin (p=0.04).
CONCLUSIONS: Among five distinct cohorts, ML techniques were used to identify patient subgroups potentially benefiting from early albumin. Our findings indicate a viable methodology for patient subgroup selection and suggest that non-cirrhotic patients with severe anemia may benefit from early albumin.
Conference/Value in Health Info
Value in Health, Volume 26, Issue 6, S2 (June 2023)
Code
MSR6
Topic
Methodological & Statistical Research
Topic Subcategory
Artificial Intelligence, Machine Learning, Predictive Analytics
Disease
Systemic Disorders/Conditions (Anesthesia, Auto-Immune Disorders (n.e.c.), Hematological Disorders (non-oncologic), Pain)