Maria Athanasiou, Greece

National Technical University of Athens School of Electrical and Computer Engineering

Presenter of 1 Presentation

ORAL PRESENTATION SESSION

AN INTERPRETABLE LSTM-BASED PREDICTION MODEL FOR ASSESSING THE RISK OF HOSPITALIZATION AND RE-HOSPITALIZATION IN YOUTH WITH TYPE 1 DIABETES MELLITUS 

Abstract

Background and Aims

Diabetic Ketoacidosis (DKA) and hyperglycemia with ketosis in the absence of acidosis constitute major causes of hospital admission and morbidity in children and adolescents with Type 1 Diabetes Mellitus (T1DM). This study aims at the development of an interpretable prediction model for the risk assessment of hospitalization and re-hospitalization in children and adolescents with T1DM.

Methods

Data collected from a two-year follow-up of 127 T1DM patients at the “Agia Sofia” Children’s Hospital, within the framework of the “SWEET” Initiative, were used for development and evaluation purposes. Frequently identified risk factors for recurrent DKA admissions were considered to compose the input space.

The model was based on Long Short-Term Memory Neural Networks (LSTM) in order to leverage LSTM’s efficiency in handling sequential data. The unbalanced nature of the dataset was addressed by applying an ensemble learning method, based on a sub-sampling approach. Interpretation of the model’s decisions was achieved by deploying the Local Interpretable Model-agnostic Explanations (LIME) technique.

data_table.png

Results

The 3-fold cross-validation criterion was applied to assess the model’s generalization ability. The obtained results in terms of C-statistic (71.34 ± 0.05%) indicate acceptable discrimination ability. Meaningful insights about the influence of the considered risk factors on the model’s decisions were provided by applying the LIME technique.

lime_image_abstract.png

Conclusions

The obtained results demonstrate the potential of the proposed method to achieve high accuracy while ensuring user’s trust by providing appropriate explanations on the produced decisions. Acknowledgment: Supported within the framework of the ENDORSE project, which is funded by the NSRF (Grant agreement: Τ1ΕΔΚ-03695).

Hide