| dc.description.abstract | Monitoring quality metrics continuously in breast cancer care can help healthcare providers and organizations engage with patients, advance the delivery of care, and tackle financial challenges. In this work, we develop the Pathfinder method to identify treatment event patterns that relate to quality metrics in breast cancer clinical data. We used manually curated cancer registry data and administrative data to evaluate our method and compare the data sources. We aim to demonstrate that we can effectively track quality metrics by abstracting raw data to the proper level of granularity.
We developed the Pathfinder method that consists of six subtasks: data extraction, data standardization, vertical abstraction, horizontal abstraction, quality metric querying, and quality metric generation. We used cancer registry treatment event and administrative CPT data from Vanderbilt University Medical Center from 2000-2012. We characterized the data sources, assessed the abstraction process, and measured the quality of the CPT codes. We then used the data to evaluate three quality metrics: rate of re-excision after initial breast conserving surgery, radiation therapy after breast conserving surgery, and chemotherapy usage in early stage disease. Finally, we used an event sequence mining method to identify common treatment event patterns to characterize and compare our data sources.
Cancer registry and CPT data for 2679 breast cancer patients were used in our study. The application of our variable abstraction process produced an approximately 12-fold reduction in the number of unique treatment event sequences from the raw sequence to most abstracted state. The quality metrics developed from the cancer registry data matched expected national rates. The CPT data was often aligned with the curated cancer registry events, but did show gaps with 37% of patients missing at least one CPT code. Despite this, the CPT data still produced similar surgical quality metrics to the cancer registry data. The CPT and cancer registry data did occasionally have different rates of frequent event patterns occurring in the patient population.
This work demonstrates how our temporal abstraction method can enable us to transform raw clinical data to the level of abstraction necessary to generate the desired quality metric. We were able to measure a set of quality metrics over a 12-year period with both manually curated cancer registry data and administrative CPT data that matched expected national rates. Despite this, near real-time metrics are difficult to achieve given the manual nature of cancer registry curation and the high frequency of missing CPT data from care delivered outside the organization. Future work is needed to develop data driven methods that fit the abstraction framework and can utilize clinical data that is generated during the course of care. This work can help healthcare providers, organizations, and patients make better healthcare decisions and assess performance. | |
