684 - Machine Learning Characterizes Timing-Function Phenotypes Predicting Surgical Outcomes in Infants Undergoing Kasai Procedure for Biliary Atresia

Publication Number: 4669.684

Erwin T. Cabacungan, Medical College of Wisconsin, Children's Wisconsin, Milwaukee, WI, United States; Ruby Gupta, Children's Hospital of Wisconsin, Milwaukee, WI, United States

Background: Outcomes in biliary atresia (BA) vary with surgical timing and hepatic reserve, but traditional predictors fail to explain this heterogeneity. Machine learning offers a data-driven approach to identify clinically meaningful phenotypes.

Objective: To identify distinct clusters of infants undergoing Kasai procedure based on surgical timing and hepatic function, and to evaluate how these clusters relate to perioperative complications and outcomes.

Design/Methods: This IRB-exempt study analyzed infants who underwent Kasai portoenterostomy in the 2012-2023 National Quality Improvement Program-Pediatric database (NSQIP-P) database. Infants ≥200 days at surgery, with invalid albumin values, or without postoperative BA ICD-CM 9/10 codes were excluded (Fig.1A). K-means clustering of age at surgery and preoperative serum albumin identified three phenotypes (Fig.1B): (1) Early/Decreased Function(E/D); (2) Early/Preserved Function(E/M); and (3) Late/Mixed Function(L/M). Demographic, preoperative, operative, and postoperative characteristics were compared. Multivariable logistic regression and negative binomial models, adjusted for covariates, evaluated associations between cluster membership and postoperative outcomes.

Results: Clusters differed in race, weight at surgery, sepsis status, nutritional and cardiac risk factors, and hematologic disorders (Table 1A). Laboratory profiles showed rising bilirubin and AST from E/D to L/M groups, while coagulation parameters followed a different pattern: E/P had slightly lower PT (12.3 vs 12.8 sec) and INR (1.0 vs 1.1) than E/D, whereas L/M had the highest PT (13.5 sec) and INR (1.1). Operative characteristics were similar (Table 1B). Postoperative complications, including unplanned reoperation, wound dehiscence, prolonged hospitalization, and discharge disposition, also varied by cluster. E/P infants had lower odds of bleeding or transfusion (OR 0.62, p=0.014) and were less likely to remain hospitalized at 30 days (OR 0.38, p=0.035), whereas L/M infants had higher odds of unplanned reoperation (OR 2.43, p=0.04) and longer hospital stay (IRR 1.21, p=0.012) (Table 2). Adjusted mean hospital stay increased from 12-13 days in E/D and E/P clusters to 15 days in L/M cluster (Fig.2).

Conclusion(s): Distinct Kasai phenotypes emerged when surgical timing and hepatic function were analyzed together. E/P infants had the most favorable perioperative course, while L/M infants had greater morbidity and prolonged recovery. Optimizing both surgical timing and hepatic condition may improve BA outcomes, highlighting the value of data-driven phenotyping for surgical risk stratification.

Figure 1A. Flow Diagram of Final Study Cohort from the NSQIP-P Database (2102 -2023) and Figure 1B. Heterogeneity of the Distribution of Age at Surgery and Preoperative Serum Albumin Across Three Clusters in Infants with Biliary Atresia
Fig. 1A - Illustrates the cohort selection process from the initial 1.3 million surgical records to the final 776 infants included in the analysis.
Fig. 1B - Clustering based on age at Kasai portoenterostomy surgery and preoperative serum albumin identified three distinct groups: (1) Early/Decreased Function (blue), younger infants with low albumin levels; (2) Early/Preserved Function (red), younger infants with relatively higher albumin levels; (3) Late/Preserved Function (green), older infants with variable albumin levels. Each point represents an infant. The dashed reference lines denote the median albumin value (3.5g/dl) and early surgical threshold (60 days).

Table 1A. Demographic and Preoperative Characteristics of Infants Undergoing Kasai Procedure by Cluster Group and Table 1B. Operative and Postoperative Complications/Outcomes of Infants Undergoing Kasai Procedure by Cluster Group
Table 1A & B - Comparison of baseline demographic, preoperative, operative, and postoperative characteristics across the three clusters derived from age at surgery and preoperative serum albumin levels: Cluster 1 (Early/Decreased Function), Cluster 2 (Early/Preserved Function), and Cluster 3 (Late/Mixed Function). Age at surgery and preoperative serum albumin, the clustering variables, are highlighted to demonstrate distinct timing-function phenotypes.
1Categorical variables were analyzed using Chi-square or Fisher's exact test.
2Continuous variables were analyzed using one-way ANOVA or Kruskal-Wallis test based on data distribution.
Significant differences are indicated in bold red.

Table 2. Association Between Cluster Phenotypes and Postoperative Outcomes Following Kasai Procedure and Figure 2. Adjusted Mean Hospital Stay (days) by Cluster Phenotype
Table 2 - Multivariable logistic and negative binomial regression models examined associations between clusters (Early/Decreased Function (reference group), Early/Preserved Function, and Late/Mixed Function) and key postoperative outcomes.
1Models were adjusted for race, weight at surgery, cardiac risk factors, SIRS/sepsis within 48 hours of surgery, nutritional support before surgery, and hematologic disorder before surgery.
Significant differences are indicated in bold red.
Figure 2 - Error bars represent 95% confidence intervals. Late/Mixed (L/M) function infants had significantly longer hospital stays compared to the early clusters.