Developing Artificial Intelligence Models for Extracting Oncologic Outcomes from Japanese Electronic Health Records

Area of Science:

• Oncology
• Medical Informatics
• Artificial Intelligence

Background:

• Establishing AI frameworks for extracting oncological outcomes from large databases is challenging.
• Electronic Health Records (EHRs) contain valuable unstructured data for cancer research.

Purpose of the Study:

• To develop and evaluate AI models for extracting treatment responses in lung cancer patients.
• To utilize unstructured EHR data for predicting oncological outcomes.

Main Methods:

• Developed and compared AI models including Bidirectional Encoder Representations from Transformers (BERT), Naïve Bayes, and Longformer.
• Trained models on University of Miyazaki Hospital (UMH) data and validated on the Life Data Initiative (LDI) dataset.
• Extracted treatment responses to calculate time to progression (TTP) and compared AI predictions with manual curation.

Main Results:

• The BERT model demonstrated superior precision, recall, and F1 scores compared to Naïve Bayes and Longformer on the UMH dataset.
• BERT model's prediction accuracy remained consistent when applied to the multi-hospital LDI dataset.
• Kaplan-Meier analysis showed similar time to progression trends between AI-predicted and manually curated data across treatment lines.

Conclusions:

• AI models, particularly BERT, can effectively extract treatment responses from unstructured EHR data for lung cancer patients.
• The developed AI framework shows potential for large-scale oncological outcome analysis.
• Further model refinement is necessary to enhance performance in clinical applications.