S.P. Robertson

Creating a Data Science Platform for Developing Complication Risk Models for Personalized Treatment Planning in Radiation Oncology

The common approach to assessing risk in radiation oncology treatment uses Lyman-Kutcher-Burman (LKB) derived models to calculate normal tissue complication probability (NTCP). LKB is not sufficiently robust to capture the modern clinical reality of three-dimensional intensity modulated radiation therapy (IMRT) treatments, the approach accounts for only two factors -- Dmax and Veff. We present a data science platform designed to facilitate the rapid creation of data-derived NTCP models. The platform extracts native Philips Pinnacle data such as dose grids and contoured regions using the cross-vendor DICOM RT standard. Further, outcome data is encoded using Common Terminology Criteria for Adverse Events 4.0. Thus, the platform exploits the normal clinical workflow and information encoded with a standard ontology. Over the course of less than three weeks we used the platform to create NTCP models for two complications (xerostomia and voice dysfunction due to parotid and larynx irradiation, respectively). We assess the resulting platform with a focus on its context within a Learning Health System (LHS). We believe that the system reported can serve as a guide to the development of radiation oncology data science platforms in particular and local-level LHS components in general.

Evaluation of classification and regression tree (CART) model in weight loss prediction following head and neck cancer radiation therapy

Objective We explore whether a knowledge–discovery approach building a Classification and Regression Tree (CART) prediction model for weight loss (WL) in head and neck cancer (HNC) patients treated with radiation therapy (RT) is feasible. Methods and materials HNC patients from 2007 to 2015 were identified from a prospectively collected database Oncospace. Two prediction models at different time points were developed to predict weight loss ≥5 kg at 3 months post-RT by CART algorithm: (1) during RT planning using patient demographic, delineated dose data, planning target volume–organs at risk shape relationships data and (2) at the end of treatment (EOT) using additional on-treatment toxicities and quality of life data. Results Among 391 patients identified, WL predictors during RT planning were International Classification of Diseases diagnosis; dose to masticatory and superior constrictor muscles, larynx, and parotid; and age. At EOT, patient-reported oral intake, diagnosis, N stage, nausea, pain, dose to larynx, parotid, and low-dose planning target volume–larynx distance were significant predictive factors. The area under the curve during RT and EOT was 0.773 and 0.821, respectively. Conclusions We demonstrate the feasibility and potential value of an informatics infrastructure that has facilitated insight into the prediction of WL using the CART algorithm. The prediction accuracy significantly improved with the inclusion of additional treatment-related data and has the potential to be leveraged as a strategy to develop a learning health system.

SU-E-P-26: Oncospace: A Shared Radiation Oncology Database System Designed for Personalized Medicine, Decision Support, and Research

Purpose: Advancement in Radiation Oncology (RO) practice develops through evidence based medicine and clinical trial. Knowledge usable for treatment planning, decision support and research is contained in our clinical data, stored in an Oncospace database. This data store and the tools for populating and analyzing it are compatible with standard RO practice and are shared with collaborating institutions. The question is - what protocol for system development and data sharing within an Oncospace Consortium? We focus our example on the technology and data meaning necessary to share across the Consortium. Methods: Oncospace consists of a database schema, planning and outcome data import and web based analysis tools.1) Database: The Consortium implements a federated data store; each member collects and maintains its own data within an Oncospace schema. For privacy, PHI is contained within a single table, accessible to the database owner.2) Import: Spatial dose data from treatment plans (Pinnacle or DICOM) is imported via Oncolink. Treatment outcomes are imported from an OIS (MOSAIQ).3) Analysis: JHU has built a number of webpages to answer analysis questions. Oncospace data can also be analyzed via MATLAB or SAS queries.These materials are available to Consortium members, who contribute enhancements and improvements. Results: 1) The Oncospace Consortium now consists of RO centers at JHU, UVA, UW and the University of Toronto. These members have successfully installed and populated Oncospace databases with over 1000 patients collectively.2) Members contributing code and getting updates via SVN repository. Errors are reported and tracked via Redmine. Teleconferences include strategizing design and code reviews.3) Successfully remotely queried federated databases to combine multiple institutions’ DVH data for dose-toxicity analysis (see below – data combined from JHU and UW Oncospace). Conclusion: RO data sharing can and has been effected according to the Oncospace Consortium model: http://oncospace.radonc.jhmi.edu/. John Wong - SRA from Elekta; Todd McNutt - SRA from Elekta; Michael Bowers - funded by Elekta

Effects of perineural invasion on biochemical recurrence and prostate cancer-specific survival in patients treated with definitive external beam radiotherapy

OBJECTIVES Perineural invasion (PNI) has not yet gained universal acceptance as an independent predictor of adverse outcomes for prostate cancer treated with external beam radiotherapy (EBRT). We analyzed the prognostic influence of PNI for a large institutional cohort of prostate cancer patients who underwent EBRT with and without androgen deprivation therapy (ADT). MATERIAL AND METHODS We, retrospectively, reviewed prostate cancer patients treated with EBRT from 1993 to 2007 at our institution. The primary endpoint was biochemical failure-free survival (BFFS), with secondary endpoints of metastasis-free survival (MFS), prostate cancer-specific survival (PCSS), and overall survival (OS). Univariate and multivariable Cox proportional hazards models were constructed for all survival endpoints. Hazard ratios for PNI were analyzed for the entire cohort and for subsets defined by NCCN risk level. Additionally, Kaplan-Meier survival curves were generated for all survival endpoints after stratification by PNI status, with significant differences computed using the log-rank test. RESULTS Of 888 men included for analysis, PNI was present on biopsy specimens in 187 (21.1%). PNI was associated with clinical stage, pretreatment PSA level, biopsy Gleason score, and use of ADT (all P<0.01). Men with PNI experienced significantly inferior 10-year BFFS (40.0% vs. 57.8%, P = 0.002), 10-year MFS (79.7% vs. 89.0%, P = 0.001), and 10-year PCSS (90.9% vs. 95.9%, P = 0.009), but not 10-year OS (67.5% vs. 77.5%, P = 0.07). On multivariate analysis, PNI was independently associated with inferior BFFS (P<0.001), but not MFS, PCSS, or OS. In subset analysis, PNI was associated with inferior BFFS (P = 0.04) for high-risk patients and with both inferior BFFS (P = 0.01) and PCSS (P = 0.05) for low-risk patients. Biochemical failure occurred in 33% of low-risk men with PNI who did not receive ADT compared to 8% for low-risk men with PNI treated with ADT (P = 0.01). CONCLUSION PNI was an independently significant predictor of adverse survival outcomes in this large institutional cohort, particularly for patients with NCCN low-risk disease. PNI should be carefully considered along with other standard prognostic factors when treating these patients with EBRT. Supplementing EBRT with ADT may be beneficial for select low-risk patients with PNI though independent validation with prospective studies is recommended.

Using Big Data Analytics to Advance Precision Radiation Oncology

Big clinical data analytics as a primary component of precision medicine is discussed, identifying where these emerging tools fit in the spectrum of genomics and radiomics research. A learning health system (LHS) is conceptualized that uses clinically acquired data with machine learning to advance the initiatives of precision medicine. The LHS is comprehensive and can be used for clinical decision support, discovery, and hypothesis derivation. These developing uses can positively impact the ultimate management and therapeutic course for patients. The conceptual model for each use of clinical data, however, is different, and an overview of the implications is discussed. With advancements in technologies and culture to improve the efficiency, accuracy, and breadth of measurements of the patient condition, the concept of an LHS may be realized in precision radiation therapy.

A Shape-Based Dose Model for the Prediction of High Grade Radiation Induced Xerostomia for Head and Neck Cancer Patients

Ductal region 73.72 % 84.47 % 58.41 % A control group of organ level DVH features was tested, using D10, D20, ..., D90 for both parotids. The use of shape based dose features was able to improve predictive capability. For each parotid gland, shells were created with outer bounds defined by, 2 mm expansion, no expansion, and 5 mm contraction. Then, shells were partitioned into octants defined by x, y, and z axes, creating 24 substructures per gland (48 per patient). D90 and D20 were computed for each derived structure

SU-F-R-47: Quantitative Shape Relationship Analysis of PTV Modification for Critical Anatomy Sparing and Its Impact On Pathologic Response for Neoadjuvant Stereotactic Radiotherapy for Pancreatic Cancer

PURPOSE Stereotactic body radiation therapy (SBRT) may be used to increase surgery candidacy in borderline resectable (BRPC) and locally advanced (LAPC) pancreatic cancer. However, the planning target volume (PTV) may need to be limited to avoid toxicity when the gross tumor volume (GTV) is anatomically involved with surrounding critical structures. Our study aims to characterize the coverage of GTV and investigate the association between modified PTV and pathologic (pCR) or near pathologic (npCR) complete response rates determined from the surgical specimen. METHODS Patients treated with neoadjuvant pancreas SBRT followed by surgery from 2010-2015 were selected from Oncospace. Overlap volume histogram (OVH) analysis was performed to determine the extent of compromise of the PTV from both the GTV and a standard target (GTV+3mm). Subsequently, normalized overlap volume (%) was calculated for: (1) GTV-PTV, and (2) GTV+3mm expansion-PTV. A logistic regression model was used to identify the association between the overlap ratios and ≥ npCR(pCR/npCR) stratified by active breathing control (ABC) versus free-breathing status. RESULTS Eighty-one (BRPC: n=42, LAPC: n=39) patients were available for analysis. Nearly 40% (31/81) had ≥npCR and 75% (61/81) were able to complete ABC. Mean coverage of the GTV-PTV was 92.6% (range, 59.9%-100%, SD = 8.68) and coverage of the GTV+3mm expansion-PTV was 85. 2% (range, 59.9% -100.0%, SD= 8.67). Among the patients with ABC, every 10% increase in GTV coverage doubled the odds to have ≥npCR (OR = 1.82, p=0.06). Coverage of GTV+3mm expansion was not associated with ≥npCR regardless of ABC status. CONCLUSION Preferential sparing of critical anatomy over GTV-PTV coverage with ABC management suggests worse ≥npCR rates for neoadjuvant SBRT in BRPC and LAPC. Limiting the GTV and GTV+3mm expansion in free-breathing patients was not associated with pathologic response perhaps due to larger GTV definitions as a result of motion artifacts in free-breathing CT scans. Collaboration with Toshiba, Elekta, and Phillips.

SU-D-BRB-02: Combining a Commercial Autoplanning Engine with Database Dose Predictions to Further Improve Plan Quality

PURPOSE Database dose predictions and a commercial autoplanning engine both improve treatment plan quality in different but complimentary ways. The combination of these planning techniques is hypothesized to further improve plan quality. METHODS Four treatment plans were generated for each of 10 head and neck (HN) and 10 prostate cancer patients, including Plan_A: traditional IMRT optimization using clinically relevant default objectives; Plan_B: traditional IMRT optimization using database dose predictions; Plan_C: autoplanning using default objectives; and Plan_D: autoplanning using database dose predictions. One optimization was used for each planning method. Dose distributions were normalized to 95% of the planning target volume (prostate: 8000 cGy; HN: 7000 cGy). Objectives used in plan optimization and analysis were the larynx (25%, 50%, 90%), left and right parotid glands (50%, 85%), spinal cord (0%, 50%), rectum and bladder (0%, 20%, 50%, 80%), and left and right femoral heads (0%, 70%). RESULTS All objectives except larynx 25% and 50% resulted in statistically significant differences between plans (Friedmans χ2 ≥ 11.2; p ≤ 0.011). Maximum dose to the rectum (Plans A-D: 8328, 8395, 8489, 8537 cGy) and bladder (Plans A-D: 8403, 8448, 8527, 8569 cGy) were significantly increased. All other significant differences reflected a decrease in dose. Plans B-D were significantly different from Plan_A for 3, 17, and 19 objectives, respectively. Plans C-D were also significantly different from Plan_B for 8 and 13 objectives, respectively. In one case (cord 50%), Plan_D provided significantly lower dose than plan C (p = 0.003). CONCLUSION Combining database dose predictions with a commercial autoplanning engine resulted in significant plan quality differences for the greatest number of objectives. This translated to plan quality improvements in most cases, although special care may be needed for maximum dose constraints. Further evaluation is warranted in a larger cohort across HN, prostate, and other treatment sites. This work is supported by Philips Radiation Oncology Systems.

MO-G-304-01: FEATURED PRESENTATION: Expanding the Knowledge Base for Data-Driven Treatment Planning: Incorporating Patient Outcome Models

Purpose: To extend the capabilities of knowledge-based treatment planning beyond simple dose queries by incorporating validated patient outcome models. Methods: From an analytic, relational database of 684 head and neck cancer patients, 372 patients were identified having dose data for both left and right parotid glands as well as baseline and follow-up xerostomia assessments. For each existing patient, knowledge-based treatment planning was simulated for by querying the dose-volume histograms and geometric shape relationships (overlap volume histograms) for all other patients. Dose predictions were captured at normalized volume thresholds (NVT) of 0%, 10%, 20, 30%, 40%, 50%, and 85% and were compared with the actual achieved doses using the Wilcoxon signed-rank test. Next, a logistic regression model was used to predict the maximum severity of xerostomia up to three months following radiotherapy. Baseline xerostomia scores were subtracted from follow-up assessments and were also included in the model. The relative risks from predicted doses and actual doses were computed and compared. Results: The predicted doses for both parotid glands were significantly less than the achieved doses (p < 0.0001), with differences ranging from 830 cGy ± 1270 cGy (0% NVT) to 1673 cGy ± 1197 cGy (30% NVT). The modelled risk of xerostomia ranged from 54% to 64% for achieved doses and from 33% to 51% for the dose predictions. Relative risks varied from 1.24 to 1.87, with maximum relative risk occurring at 85% NVT. Conclusions: Data-driven generation of treatment planning objectives without consideration of the underlying normal tissue complication probability may Result in inferior plans, even if quality metrics indicate otherwise. Inclusion of complication models in knowledge-based treatment planning is necessary in order to close the feedback loop between radiotherapy treatments and patient outcomes. Future work includes advancing and validating complication models in the context of knowledge-based treatment planning. This work is supported by Philips Radiation Oncology Systems.

SU-E-T-544: A Radiation Oncology-Specific Multi-Institutional Federated Database: Initial Implementation

PURPOSE To implement a common database structure and user-friendly web-browser based data collection tools across several medical institutions to better support evidence-based clinical decision making and comparative effectiveness research through shared outcomes data. METHODS A consortium of four academic medical centers agreed to implement a federated database, known as Oncospace. Initial implementation has addressed issues of differences between institutions in workflow and types and breadth of structured information captured. This requires coordination of data collection from departmental oncology information systems (OIS), treatment planning systems, and hospital electronic medical records in order to include as much as possible the multi-disciplinary clinical data associated with a patients care. RESULTS The original database schema was well-designed and required only minor changes to meet institution-specific data requirements. Mobile browser interfaces for data entry and review for both the OIS and the Oncospace database were tailored for the workflow of individual institutions. Federation of database queries--the ultimate goal of the project--was tested using artificial patient data. The tests serve as proof-of-principle that the system as a whole--from data collection and entry to providing responses to research queries of the federated database--was viable. The resolution of inter-institutional use of patient data for research is still not completed. CONCLUSIONS The migration from unstructured data mainly in the form of notes and documents to searchable, structured data is difficult. Making the transition requires cooperation of many groups within the department and can be greatly facilitated by using the structured data to improve clinical processes and workflow. The original database schema design is critical to providing enough flexibility for multi-institutional use to improve each institution s ability to study outcomes, determine best practices, and support research. The project has demonstrated the feasibility of deploying a federated database environment for research purposes to multiple institutions.