Natural language processing with machine learning to predict outcomes after ovarian cancer surgery

Area of Science:

• Oncology
• Medical Informatics
• Radiology

Background:

• Ovarian cancer surgery often involves complex patient data.
• Predicting postoperative complications and readmissions is crucial for patient management.
• Current prediction models may not fully utilize unstructured data from imaging reports.

Purpose of the Study:

• To evaluate if natural language processing (NLP) of preoperative CT scans enhances prediction of postoperative complications and hospital readmission in ovarian cancer patients.
• To compare NLP-augmented machine learning models with discrete data predictors alone.

Main Methods:

• A cohort of 291 women with ovarian cancer undergoing debulking surgery was identified.
• Machine learning models incorporated discrete data (age, comorbidities, labs) and NLP of unstructured CT scan reports.
• Prediction accuracy was measured using the area under the receiver operating characteristic curve (AUC).

Main Results:

• NLP of CT scans significantly improved the prediction of 30-day hospital readmission (AUC 0.70) compared to discrete data alone (AUC 0.56).
• The model incorporating NLP also showed improved prediction for postoperative complications.
• The study included patients with a mean age of 59 and specific preoperative lab values.

Conclusions:

• Natural language processing combined with machine learning significantly enhances the prediction of postoperative complications and hospital readmissions in ovarian cancer surgery.
• Integrating unstructured text data from preoperative CT scans offers a valuable improvement over traditional discrete data predictors.