MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Evaluating the Linearity of Risk Functions Across Radiotherapy Outcomes Using Deep Learning

C Ahern1 , T Pheiffer1*, C Berlind1 , W Lindsay1 , Y Xiao2 , C Simone3 , (1) Oncora Medical, Inc., Philadelphia, PA, (2) University of Pennsylvania, Philadelphia, PA, (3) University of Maryland School of Medicine, Baltimore, Maryland

Presentations

(Sunday, 7/29/2018) 4:00 PM - 4:55 PM

Room: Davidson Ballroom A

Purpose: Many survival models, such as the Cox Proportional Hazards (CPH) model, estimate risk as a linear function of input features. We quantify the advantage of deep learning models with non-linear risk functions to model radiotherapy outcomes.

Methods: In this IRB-approved analysis, a dataset of 16,689 radiotherapy courses performed at one institution between 2008 and 2015 with over 240 clinical features was created from EMRs and treatment planning systems using automated extraction software. Five efficacy outcomes - local, nodal, and distant failure, death, and unplanned hospitalization - were also extracted. Each of 63 cancer adverse events was tracked for each patient according to CTCAEv4. Time to outcome was defined as the elapsed number of days from first fraction to outcome, grade ≥1 toxicity, or last visit. For each outcome, regularized CPH models and deep survival networks were trained using 3-fold cross-validation on an 80-20 train-test split, with performance measured by the concordance index.

Results: The deep learning models outperformed the CPH models for 43 of the 68 total outcomes, measured by difference in concordance index, with mean (mean Δ=0.011) and range (-0.073, 0.230). The outcomes exhibiting the largest improvement with deep learning were fatigue (0.260), depression (0.182), local failure (0.043), and aspiration (0.036).

Conclusion: The improved performance of deep survival networks over CPH models indicates these outcomes have a risk function that may be non-linear in the input features. Fatigue and depression are systemic outcomes with multifactorial causes. Risk of local recurrence varies by treatment site and patient characteristics. Risk of aspiration is dependent on involvement of critical structures in head and neck cancers or prior resections for lung cancer. Multi-center validation of these findings is needed and future studies should investigate potential mechanisms underlying non-linearities or interactions in risk functions.

Funding Support, Disclosures, and Conflict of Interest: CA, TP, CG, and WL are employed by and own equity in Oncora Medical, Inc.

Keywords

Not Applicable / None Entered.

Taxonomy

TH- Dataset analysis/biomathematics: Machine learning techniques

Contact Email