Skip to main content
Fig. 6 | BMC Bioinformatics

Fig. 6

From: Interpretable prediction of necrotizing enterocolitis from machine learning analysis of premature infant stool microbiota

Fig. 6

NEC risk scores and test characteristics. Confidence scores from the growing bag were used to calculate dynamic NEC risk scores as described in the main text. Each patient in both clinical cohorts was analyzed individually, using a MIL model that was naïve to that patient’s data. Aggregate risk scores for all patients were binned by days preceding disease or study discharge, and are displayed with standard deviation. The trend lines were generated by LOWESS curve fitting (A). Separate survival curves were generated for patients in the Warner and Olm cohorts (B, C) where the outcome of interest was a risk score above 0.35 during hospitalization. Differences between affected and non-affected patient curves were statistically significant (****\(p<0.001\); Kolmogorov–Smirnov test) for both cohorts.Timelines for all patients from both cohorts are labeled with the last sample collected (D). For those who were predicted to have NEC based on a risk score >0.35, the first sample to cross the risk threshold is indicated, with dotted lines illustrating lead-time before disease onset (for true positives) or the last sample collected before aging out of the study (for false positives). Prediction outcomes for all patients used across both cohorts are displayed in the confusion matrix (E)

Back to article page