Interpretation of clinical variables in the use of artificial intelligence-based decision support for surgical necrotizing enterocolitis: A prospective study.

PURPOSE: End-of-life decision-making in neonates with surgical necrotizing enterocolitis (NEC) is complex and varies between clinicians depending on their interpretation of clinical variables. This study aimed to create transparency in this variability and assess its impact on the use of Behavioral Artificial Intelligence Technology (BAIT) as a tool to inform clinical decision making.

METHODS

Attendees of multidisciplinary team meetings where the decision between surgery and comfort care was made for a neonate with surgical NEC were asked to interpret clinical variables required for BAIT. Variability in interpretation was analyzed using standard deviations, and assessed for correlation with variability in BAIT predictions using Spearman's rank correlation. Linear mixed effects models were used to assess the influence of demographics on these predictions and Spearman's rank correlation evaluated association with personal treatment preference, aiming to identify factors contributing to inter-rater differences.

RESULTS

The largest variations were observed in the interpretation of growth since birth (SD = 0.29), congenital comorbidities (SD = 0.26), and hemodynamics (SD = 0.23). Variability in variable interpretation correlated to variability in BAIT predictions (ρ = 0.782, p < 0.001). Clinicians' specialty was the primary factor contributing to these variations, and there was a significant association between treatment preference and model prediction (ρ = 0.552, p < 0.001).

IMPLICATIONS: The decision-making process is highly complex and influenced by personal factors. BAIT can offer transparency into these variations and serve as a reflective tool for clinicians, enabling more informed and consistent decision-making for surgical NEC.