A patient-similarity-based model for diagnostic prediction

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Decision Support Systems
• Computational Psychology

Background:

• Simulating clinical reasoning is crucial for accurate medical diagnoses.
• Existing diagnostic prediction methods often lack generalizability and interpretability.
• Leveraging patient similarity for diagnostic prediction is an underexplored area.

Purpose of the Study:

• To develop and evaluate a novel patient-similarity-based framework for automated diagnostic prediction.
• To simulate clinical reasoning by retrieving analogous patients and predicting diagnoses.
• To improve the accuracy, generalizability, and interpretability of diagnostic decision support.

Main Methods:

• A patient-similarity framework inspired by structure-mapping theory was proposed.
• Patient similarity was defined using diagnosis sets, converting multilabel classification to regression.
• Both similar (positive analogy) and dissimilar (negative analogy) patients were utilized for hypothesis generation and rejection.

Main Results:

• The patient-similarity model significantly outperformed one-vs-all and traditional k-NN baselines.
• The f-1 score for positive-analogy prediction was 0.698, increasing to 0.703 with negative analogy.
• The model demonstrates promising performance for larger, heterogeneous, and incomplete clinical datasets.

Conclusions:

• The proposed model offers more accurate, generalizable, and interpretable diagnostic decision support.
• This framework represents a novel application of clinical big data and artificial intelligence.
• The approach holds significant potential for advancing AI-driven medical diagnostics.