An interpretable DIC risk prediction model based on convolutional neural networks with time series data.

Area of Science:

• Critical Care Medicine
• Biomedical Informatics
• Hematology

Background:

• Disseminated intravascular coagulation (DIC) is a severe, life-threatening condition often occurring in the late stages of various coagulation disorders.
• Early identification of patients at risk for DIC is crucial for improving prognosis and reducing mortality.
• Current methods for early DIC prediction are insufficient, highlighting the need for advanced predictive models.

Purpose of the Study:

• To develop and validate a novel, interpretable deep learning model for the early prediction of DIC risk.
• To assess the model's performance in identifying high-risk patients within an intensive care unit (ICU) setting.
• To utilize explainability techniques to understand the model's decision-making process for DIC prediction.

Main Methods:

• A deep learning-based time series model was developed to predict DIC risk.
• The model was trained and validated on a cohort of ICU patients admitted between January 1, 2019, and January 1, 2022.
• Gradient-weighted Class Activation Mapping (Grad-CAM) was employed for model interpretability, visualizing prediction drivers.

Main Results:

• The deep learning model demonstrated exceptional predictive performance, achieving an Area Under the Curve (AUC) of 0.986.
• The model attained high accuracy (95.7%) and F1-score (0.935) in identifying patients at risk of DIC.
• Grad-CAM analysis provided visual heatmaps, illustrating the basis for the model's predictions and enhancing transparency.

Conclusions:

• The developed interpretable deep learning model effectively predicts the risk of DIC in ICU patients.
• Early identification of high-risk individuals facilitates timely clinical intervention, potentially improving treatment efficacy.
• This approach offers a promising tool for enhancing DIC management and patient care in critical care settings.