Effective Identification of Similar Patients Through Sequential Matching over ICD Code Embedding.

Area of Science:

• Medical Informatics
• Computational Health
• Clinical Data Science

Background:

• Evidence-based medicine relies on identifying similar patient conditions, often represented by International Classification of Diseases (ICD) code sequences.
• Existing methods for matching ICD-10 code sequences lack effectiveness, particularly for patients with multiple comorbidities and complex care needs.

Purpose of the Study:

• To present a novel method for matching ICD-10 code sequences that effectively captures clinical similarity.
• To improve patient matching for individuals with complex health conditions and multiple comorbidities.

Main Methods:

• Leveraging representation learning for individual ICD-10 codes.
• Explicitly incorporating the sequential order of codes in the matching process.
• Evaluating the method on a state-wide cancer data collection.

Main Results:

• The proposed method significantly outperforms state-of-the-art approaches that ignore sequential code order.
• The method demonstrates superior performance in identifying similar patients based on clinical outcomes.
• Improved identification of patients with similar readmission and mortality outlooks was observed.

Conclusions:

• The developed method offers an effective solution for matching ICD-10 code sequences, enhancing clinical similarity assessment.
• This approach is valuable for identifying similar patients, particularly those with complex care needs.
• The method is adaptable to other types of codified sequence data beyond ICD-10 codes.