Encrypted login | home

Program Information

The Relative Impact of Clinical Variables On Radiotherapy Outcome Predictions

no image available
C Berlind

C Ahern1 , C Berlind1*, W Lindsay1 , P Gabriel2 , C Simone3 , (1) Oncora Medical, Inc., Philadelphia, PA, (2) University of Pennsylvania, Philadelphia, PA, (3) University of Maryland, Baltimore, MD


SU-E-FS4-4 (Sunday, July 30, 2017) 1:00 PM - 1:55 PM Room: Four Seasons 4

Purpose: Integrating and extracting data from heterogeneous sources is a major hurdle to applying predictive analytics in oncology. To guide the prioritization of integration efforts, we quantify the relative impact of different types of clinical variables on the performance of radiotherapy outcome predictions.

Methods: In this IRB-approved analysis, a dataset of 16,689 consecutive radiotherapy courses performed at one institution from 2008-2015 was created from EMRs and treatment planning systems using automated extraction software. Over 230 variables were extracted for each patient, along with 76 outcomes including survival, recurrence, hospitalization, and occurrence of 64 adverse events scored per CTCAEv4.0. Variables were split into six segments: demographics, diagnosis history, tumor information, surgical history, chemotherapy history, and radiation treatment data. Random forests to predict each of the 76 outcomes were trained with the entire set of variables using 3-fold cross-validation on an 80-20 train-test split of the data, with performance measured by area under the ROC curve (AUC). This procedure was repeated on training sets excluding each one of the variable segments in turn. The impact of each variable segment was measured by the change in AUC (ΔAUC) from the full model to the model trained with that segment excluded.

Results: Excluding tumor information led to the largest average drop in AUC across outcomes (Average ΔAUC = -0.022), followed by diagnosis history (-0.008), demographics (-0.005), radiation treatment data (-0.005), chemotherapy history (-0.003), and surgical history (-0.001).

Conclusion: The large average ΔAUC for tumor information (size, site, staging) suggests that tumor registries are a high-priority data source. Other segments have lower impacts on average, but are each very important for predicting specific outcomes, suggesting that prioritizing efforts should be driven by the desired clinical outcomes. Further research on improving data extraction methods for these segments is needed.

Funding Support, Disclosures, and Conflict of Interest: Christopher Ahern, Christopher Berlind, and William Lindsay are employees of and hold securities in Oncora Medical, Inc., a software company in Philadelphia, PA.

Contact Email: