Integrating a Shareable Artificial Intelligence Model Into Clinical Research for Cancer Recurrence in Patients With...

Area of Science:

• Oncology
• Medical Informatics
• Natural Language Processing

Background:

• Extracting cancer recurrence data from clinical text is labor-intensive.
• Automated methods are needed to efficiently analyze radiology reports for cancer outcomes.
• The DFCI-imaging-student model was developed to extract cancer outcomes from radiology reports.

Purpose of the Study:

• To apply the DFCI-imaging-student model in a community oncology setting.
• To evaluate the model's performance in determining recurrence and time-to-recurrence for breast cancer (BC) and colorectal cancer (CRC).
• To aggregate report-level predictions to derive patient-level outcomes.

Main Methods:

• Randomly sampled 200 BC and 200 CRC patients diagnosed with stage III disease.
• Manually reviewed recurrence, recurrence date, and sites of recurrence from electronic health records.
• Applied the DFCI-imaging-student model to radiology reports and compared outcomes against manual review.

Main Results:

• Processed 7,195 radiology reports with a median follow-up of 8.4 years (BC) and 6.8 years (CRC).
• The model showed high sensitivity and specificity for recurrence detection (BC: 92.3%/92.6%, CRC: 94.3%/86.9%).
• Moderate-to-high accuracy in identifying distant metastasis sites, with low median error in time-to-recurrence.

Conclusions:

• The DFCI-imaging-student model accurately determines cancer recurrence and time-to-recurrence.
• This automated approach offers an efficient method for large-scale research.
• The model can improve recurrence surveillance and facilitate collaborative cancer research.