(148) Development and Validation of Text Mining Approach for Detecting Recurrence for Early Breast Cancer Patients: A Multi-institutional Study in Taiwan

Background: In previous breast cancer studies, the disease recurrence was always defined by proxy index like "time to next treatment" while using electronic medical records (EMRs) database. However, there remains a scarcity of employing text mining methodologies to detect disease recurrence events in Taiwan.

Objectives: Our study aimed to develop text mining techniques for detecting recurrence of early breast cancer.

Methods: We conducted a retrospective study from the Chang Gung Research Database (CGRD), the largest multi-institutional EMRs database in Taiwan. We included patients diagnosed with estrogen receptor (ER)/progesterone receptor (PR) positive and human epidermal growth factor receptor 2 (HER2) negative early breast cancer (stage 1-3). We randomly selected 10% patients from each stage. Our text mining approach involves analyzing unstructured images textual data to detect breast cancer recurrence events. First, we identified instances of no recurrence, we searched for negative keywords ("no local recurrence”, “without definite recurrence", etc.) to exclude patients without recurrence. Second, we utilized positive keywords ("metastasis", "new bilateral lung metastases", etc.) for detecting recurrence. To establish a gold standard for comparison, two clinical doctors independently reviewed the random samples.

Results: A total of 820 ER/PR positive and HER2 negative early breast cancer patients were evaluated. For all stage, the positive predictive value (PPV), negative predictive value (NPV), sensitivity, and specific were 50.6% (95% confidence interval [CI]: 39.7 - 61.5%), 97.2% (96.1 - 98.4%), 67.2% (55.4 - 78.9%), 94.7% (93.1% - 96.3%), respectively. The highest PPV was observed in stage 2 patients (n = 356, 43.4%), the PPV, NPV, sensitivity, and specific were 60.4% (46.5 - 74.2%), 97.4% (95.6% - 99.1%), 78.3% (65.1 - 91.6%), 94.0% (91.4% - 96.6%), respectively.

Conclusions: Our text mining approach to define recurrence events had low PPV and high NPV. Further investigation is warranted to explore the integration of pathology reports text mining to discern between recurrence and multiple primary cancers accurately. Combining these methodologies holds promise in improving the clarity of detection in such cases and enhancing the overall reliability of recurrence detection in early breast cancer patients.