Disease classification via interpretable machine learning based on multi-center routine coagulation test.

Area of Science:

• Clinical diagnostics
• Machine learning in healthcare
• Coagulation testing

Background:

• Routine coagulation tests are crucial for diagnosing various conditions.
• Developing interpretable models for disease classification is essential for clinical application.
• Multi-center data collection ensures generalizability of findings.

Purpose of the Study:

• To develop an interpretable machine learning model for disease classification using routine coagulation tests.
• To identify key features associated with specific diseases for improved clinical diagnosis.
• To leverage multi-center data for robust model development and validation.

Main Methods:

• Unsupervised clustering and supervised machine learning models (LightGBM, Random Forest, SVM, etc.) were employed.
• Data from 11 hospitals were used, with 10-fold cross-validation and external validation on 2 hospitals.
• SHAP and Decision Tree analyses were utilized for model interpretability.

Main Results:

• The LightGBM model demonstrated superior performance in classifying valvular heart disease (VHD) and pulmonary infection (PI).
• High F1-scores (0.8890 for VHD, 0.7233 for PI in cross-validation) and AUCs (0.9500 for VHD, 0.8023 for PI) were achieved.
• External validation confirmed strong generalization (F1-scores: 0.9259 VHD, 0.7464 PI; AUCs: 0.9493 VHD, 0.8297 PI).
• International normalized ratio (INR) was identified as a key feature for VHD, and age for PI.

Conclusions:

• Machine learning models utilizing multi-center coagulation test data offer effective and interpretable disease classification.
• These models support the automation of clinical diagnosis.
• The identified key features provide valuable insights for clinicians.