Machine learning-based diagnosis for disseminated intravascular coagulation (DIC): Development, external validation,...

Area of Science:

• Hematology
• Medical Informatics
• Artificial Intelligence in Medicine

Background:

• Disseminated intravascular coagulation (DIC) diagnosis faces challenges due to a lack of specific biomarkers.
• Current composite scoring systems for DIC are rapid but require optimization of fibrin-related markers and cut-off values.

Purpose of the Study:

• To optimize the diagnostic utility of DIC-related parameters using a machine learning (ML) approach.
• To evaluate the diagnostic value of an ML-based approach for DIC diagnosis.

Main Methods:

• Retrospective review of 656 DIC-suspected cases (Set 1) and external validation with 217 cases (Set 2).
• Investigation of 46 DIC-related parameters, including clinical findings and laboratory results.
• Development and testing of an artificial neural network (ANN) model using 32 selected parameters.

Main Results:

• The ANN model achieved higher AUC values than traditional scoring systems (ISTH, JMHW, JAAM) in both datasets.
• Set 1: ANN AUC 0.981 vs. ISTH 0.945, JMHW 0.943, JAAM 0.928.
• Set 2: ANN AUC 0.968 vs. ISTH 0.946.
• Evaluation of parameter importance revealed differences compared to traditional scoring systems.

Conclusions:

• Machine learning effectively optimizes the use of clinical parameters for robust DIC diagnosis.
• The ML approach shows potential as a supportive tool for clinical decision-making, possibly integrated into electronic health records.
• Further prospective validation is needed to confirm the clinical benefits of this ML-based approach.