P0993 - Comparison of Logistic Regression Model With Machine Learning Models to Predict Acute Liver Injury in Patients Hospitalized With COVID-19

Shivabalan Kathavarayan Ramu, MD¹, Anjali Byale, MD^2
1Cleveland Clinic, Cleveland, OH; ²UPMC, Pittsburgh, PA

Introduction: We aimed to compare Logistic Regression with machine learning to predict Acute Liver Injury (ALI) in patients hospitalized with COVID-19. We also developed a nomogram to predict ALI during hospitalization in COVID-19 patients with demographic, clinical and laboratory findings.

Methods: We analyzed data from 1384 COVID-19 patients from Stony Brook University. Out of 1384, 48 (3.5%) sustained ALI during hospitalization. Out of the original 123 variables, we selected 58 variables to build a multivariable logistic regression model (MLR) and train machine learning models. All missing values were imputed using random forest approach in R. The data was randomly split into training (80%) and testing (20%) and a MLR model was developed from the training dataset. Following machine learning models were trained: Random Forest with 10 trees, Support Vector Machine with Radial Basis Function kernel, Stochastic Gradient Descent, Decision tree and MLR with Lasso regularization (C=.16). The predictions were used to construct Kaplan-Meir (K-M) curves to compare the length of hospital stay and duration of mechanical ventilation between ALI and non-ALI patients. A nomogram was constructed using the variables selected by MLR that could predict ALI. All analyses were performed in RStudio and Orange (Data mining Toolbox in Python). K-M graphs were generated in GraphPad Prism9

Results: MLR showed an accuracy of 87.6% on the validation dataset. Random Forest and Support Vector Machine each showed accuracy of 100%. Decision tree performed the lowest (accuracy: 50%). The top ranked variable of importance as selected by the MLR model for predicting ALI was Neutrophil/Lymphocyte ratio followed by N-Terminal B natriuretic peptide, Triglyceride/Lymphocyte ratio, days-prior-symptoms, and lactate. The overall length of hospital stay was higher in patients who had ALI (median, 55 days, and 6 days, respectively; hazard ratio, 0.38; 95% CI, 0.25 to 0.56) (P< .001 for interaction). Whereas median duration of mechanical ventilation in patients was also higher in patients with ALI (median, 40 days, and 10 days, respectively; hazard ratio, 0.33; 95% CI, 0.22 to 0.49) (P< .001 for interaction).

Discussion: Machine learning models outperformed MLR to predict ALI during hospitalization in COVID-19 patients. We also developed a nomogram for easy prediction of COVID-19 related ALI using simple bedside variables. This model can help physicians stratify COVID-19 patients with a high probability of ALI and manage preemptively.




Disclosures:


Shivabalan Kathavarayan Ramu, MD¹, Anjali Byale, MD². P0993 - Comparison of Logistic Regression Model With Machine Learning Models to Predict Acute Liver Injury in Patients Hospitalized With COVID-19, ACG 2023 Annual Scientific Meeting Abstracts. Vancouver, BC, Canada: American College of Gastroenterology.