(607) Machine learning versus binary scoring to predict adverse outcomes of pregnant patients with diabetic ketoacidosis

To identify factors on admission that are associated with a composite adverse outcome in pregnant patients with diabetic ketoacidosis (DKA), and to determine the predictive performance of a binary scoring system versus a machine learning (ML) model.

Study Design:

This was a retrospective cohort study of all DKA admissions from 2014-2024. The primary outcome was a composite of: ICU admission, mechanical ventilation, or a prolonged hospital stay ( >14 days). We used mixed-effects logistic regression to identify characteristics present on admission that are independently associated with the composite outcome. We then developed and compared two predictive models: 1) simple binary score based on admission glucose ( >350 mg/dL), bicarbonate (< 10 mEq/L), and anion gap ( >20 mEq/L), and 2) ML model using restricted cubic splines. Both models were validated over 1000 repetitions on a held-out test set, and their performance was assessed using the Area Under the Curve (AUC).

Results:

158 admissions from 107 pregnancies were identified. In the multivariable analysis, admission glucose (aOR 1.72, 95%CI 1.03–2.84; p=0.03), anion gap (aOR 0.29, 95%CI 0.10–0.69; p=0.01), and bicarbonate (aOR 0.18, 95%CI 0.07–0.41; p< 0.001) were independently associated with the composite outcome. Conditional density plots revealed a linear relationship for glucose, while anion gap and bicarbonate had non-linear relationships with the outcome. The ML-based scoring model, which incorporated these non-linearities, had better predictive performance than the simple binary scoring system, and a risk calculator based on this model was developed. The median AUC for the ML model was 0.75 (IQR 0.69–0.81), compared to 0.70 (IQR 0.65–0.74) for the binary score, with a mean AUC difference of 0.044 (95%CI 0.039-0.049, p< 0.001).

Conclusion:

Admission glucose, anion gap, and bicarbonate levels are independent predictors of adverse outcomes in pregnant patients with DKA. A ML model incorporating non-linear relationships of these predictors has improved accuracy for risk stratification compared to a simple threshold-based scoring system.