Mortality prediction in patients with hyperglycaemic crisis using explainable machine learning: a prospective, multicentre study based on tertiary hospitals

Abstract

Background: Experiencing a hyperglycaemic crisis is associated with a short- and long-term increased risk of mortality. We aimed to develop an explainable machine learning model for predicting 3-year mortality and providing individualized risk factor assessment of patients with hyperglycaemic crisis after admission.

Methods: Based on five representative machine learning algorithms, we trained prediction models on data from patients with hyperglycaemic crisis admitted to two tertiary hospitals between 2016 and 2020. The models were internally validated by tenfold cross-validation and externally validated using previously unseen data from two other tertiary hospitals. A SHapley Additive exPlanations algorithm was used to interpret the predictions of the best performing model, and the relative importance of the features in the model was compared with the traditional statistical test results.

Results: A total of 337 patients with hyperglycaemic crisis were enrolled in the study, 3-year mortality was 13.6% (46 patients). 257 patients were used to train the models, and 80 patients were used for model validation. The Light Gradient Boosting Machine model performed best across testing cohorts (area under the ROC curve 0.89 [95% CI 0.77-0.97]). Advanced age, higher blood glucose and blood urea nitrogen were the three most important predictors for increased mortality.

Conclusion: The developed explainable model can provide estimates of the mortality and visual contribution of the features to the prediction for an individual patient with hyperglycaemic crisis. Advanced age, metabolic disorders, and impaired renal and cardiac function were important factors that predicted non-survival.

Trial registration number: ChiCTR1800015981, 2018/05/04.

Keywords: Explainable model; Hyperglycaemic crisis; Machine learning; Mortality.

Fig. 1

Discrimination and calibration performance of the models. A Receiver operating characteristic curves for the LR, SVM, RF, LightGBM, and DNN models. B Calibration curve for the LightGBM model

Fig. 2

The impact of the input features on predictions. Each dot represents the effect of a feature on the prediction for one patient. The redder the colour of the dots, the higher the value of the features, and the bluer the colour of the dots, the lower the value of the features. Dots to the left x-axis represent patients with values of the features decreasing mortality prediction, and dots to the right x-axis represent patients with values of the features increasing mortality prediction

Fig. 3

Examples of personalized risk factors. A An example of personalized risk factor analysis for a patient in the test set (clinical outcome was death). B An example of personalized risk factor analysis for a patient in the test set (actual clinical outcome was survival)