Construction and Utilization of a Neural Network Model to Predict Current Procedural Terminology Codes from Pathology...

Area of Science:

• Medical Informatics
• Computational Pathology
• Artificial Intelligence in Healthcare

Background:

• Pathology specimen accessioning initially involves tentative Current Procedural Terminology (CPT) code assignment.
• Subsequent verification steps by pathology staff and a keyword-based application aim to ensure CPT code accuracy.
• A high accuracy rate (>97%) in initial CPT code assignments was observed, yet opportunities for improvement exist.

Purpose of the Study:

• To develop and evaluate a neural network model for predicting CPT codes directly from pathology report texts.
• To integrate this predictive model into an existing CPT code-checking application to enhance accuracy and efficiency.
• To assess the model's performance in both accurate CPT code prediction and the detection of coding errors.

Main Methods:

• Utilized R programming language and the R Keras package for model development and prediction.
• A three-layer neural network architecture was employed, including word-embedding, bidirectional Long Short-Term Memory (LSTM), and densely connected layers.
• Input data consisted of concatenated header-diagnosis texts from finalized pathology reports; data were partitioned into training and validation sets.

Main Results:

• The neural network model achieved high prediction accuracy: 97.5% on the validation set and 97.6% on the test set.
• In cases with initially incorrect CPT codes, the model disagreed with the initial assignment in 73.6% of instances.
• The model identified nine additional specimens with CPT coding errors that were missed by all prior manual and application-based checks.

Conclusions:

• A neural network model leveraging pathology report text can accurately predict CPT codes.
• The model demonstrates moderate sensitivity in detecting CPT coding errors, complementing existing verification processes.
• Neural networks show significant potential to play an increasingly important role in optimizing CPT coding within surgical pathology.