An End-to-End Natural Language Processing Application for Prediction of Medical Case Coding Complexity: Algorithm...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare

Background:

• Medical coding converts clinical documentation to standard codes for reimbursement and analysis.
• Current AI solutions for medical coding have limited effectiveness, especially for complex cases.
• Optimizing medical coding is crucial for hospital operations.

Purpose of the Study:

• To enhance medical coding efficiency and accuracy by improving code selection.
• To develop a multimodal machine learning solution to predict coding complexity before the coding process.
• To utilize coding complexity to optimize work distribution among medical coders, minimizing errors and improving throughput.

Main Methods:

• Collected 2060 cases rated for coding complexity (1-4) by human coders.
• Used 2 expert coders to establish a gold standard for 3.01% of cases.
• Extracted and concatenated text and metadata features from electronic health records for machine learning models.
• Trained and evaluated two models: one for predictive power and generalizability, and another against the human expert benchmark.

Main Results:

• The first model achieved a macro-F1-score of 0.51 and 0.59 accuracy for 4-scale complexity classification.
• The model effectively distinguished between simple (1-2) and complex (3-4) cases with a macro-F1-score of 0.65 and 0.71 accuracy.
• The second model achieved 61% agreement with expert ratings on the gold standard, with a macro-F1-score of 0.62, compared to 66% agreement between experts.

Conclusions:

• A multimodal machine learning approach effectively predicts coding complexity using clinical text and patient metadata.
• Integrating this model into hospital systems allows automatic case distribution with performance comparable to human experts.
• The proposed solution has the potential to significantly improve coding efficiency and accuracy at scale.