Intelligent prediagnosis for nontraumatic acute abdomen with surface-level information using machine learning

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Diagnostic Systems

Background:

• Prediagnosis is crucial for medical triage but often limited by available information.
• Nontraumatic acute abdomen (NTAA) presents a diagnostic challenge with superficial data.
• An intelligent framework was developed to address NTAA prediagnosis limitations.

Purpose of the Study:

• To develop and evaluate an intelligent framework for prediagnosing nontraumatic acute abdomen (NTAA) using limited information.
• To recursively infer disease information across hierarchical levels (I-level, II-level, III-level).
• To compare the performance of various machine learning algorithms for NTAA prediagnosis.

Main Methods:

• A retrospective dataset of NTAA patients was utilized.
• A machine learning framework with combined binary classifiers was designed.
• Recursive Feature Elimination with Cross-Validation (REFCV) was used for feature refinement.
• Five algorithms (Logistic Regression, DNN, SVM, RF, XGBoost) were assessed with five-fold cross-validation and grid search.

Main Results:

• I-Level disease identification metrics exceeded 0.90.
• II-Level classification metrics generally surpassed 0.80.
• High recognition rates were achieved for common III-level NTAA conditions.
• Logistic Regression demonstrated performance comparable to more complex algorithms.

Conclusions:

• The developed framework effectively discerns primary NTAA disease categories and subtypes.
• Prediagnosis of NTAA using superficial information is achievable with the proposed framework.
• Logistic Regression is a sufficient algorithm for this task, without significant advantages from more complex models.