Abstract 5299: Machine learning approach to personalized medicine in breast cancer patients: Development of data-driven, personalized, causal modeling through identification and understanding of optimal treatments for predicting better disease outcomes

Background: In the era of personalized medicine, a major challenge is harnessing longitudinal data across the cancer care continuum, which includes multimodal data sets of biologic, molecular, and clinical information about patients (pts) and their tumors. There is a growing need for new computing analytics, such as machine learning–an important tool in healthcare bio-informatics. We report our approach to building cancer disease models in an unbiased manner through utilization of a causal machine learning and simulation platform.

Methods: The Swedish Cancer Institute (SCI) Personalized Medicine Research Program (PMRP) is a prospective registration protocol with the objective of establishing a centralized longitudinal, molecular, and phenotypic data repository. Since 2014, over 1,030 pts have been enrolled, having undergone next-generation sequencing (NGS) profiling of their tumors. Of these pts, we identified 100 breast cancer pts who also have detailed longitudinal clinical annotation within our SCI Breast Cancer Registry. All de-identified data, variables, and data points in the multimodal data types are integrated into normalized data frames to include demographics, cancer risks, tumor specifications, tumor sequencing, initial and subsequent cancer treatments, and outcomes data. A reverse engineering approach, via the Reverse Engineering and Forward Simulation (REFS) platform, is being utilized, focusing on discovering the complex causal mechanisms that determine which therapies will produce the best outcomes for an individual pt. This method goes beyond traditional approaches that rely on data correlations to match treatments to pts. The breast cancer causal model uncovers many of the possible combinations of causal relationships that drive outcomes and enables “what if?” simulations of a variety of interventions, across pts, to determine optimal therapies. Performance metrics and model robustness will be explored using a stratified, n-fold (e.g., 10-fold) cross-validation procedure, which is designed to provide an unbiased estimate of model generalization to new observations.

Results: The causal model and simulations can elevate the providers' abilities to better understand treatment responses based on pts' unique clinical data and mutational statuses; study different treatment options to optimize management; and understand the complex interactions among variables that lead to a range of treatment outcomes.

Conclusions: Knowledge generated from the simulations of the disease model can potentially streamline and support the clinical decision-making process, to include molecular tumor board deliberations, and ultimately assist providers in arriving at optimal treatment recommendations for pts.

Citation Format: Henry Kaplan, Anna Berry, Kristine Rinn, Erin Ellis, George Birchfield, Tanya Wahl, Xiaoyu Liu, Mariko Tameishi, J D. Beatty, Patricia Dawson, Vivek Mehta, Anna Holman, Mary Atwood, Shlece Alexander, Candy Bonham, Lauren Summers, Iya Khalil, Boris Hayete, Diane Wuest, Wei Zheng, Yuhang Liu, Xulong Wang, Thomas David Brown. Machine learning approach to personalized medicine in breast cancer patients: Development of data-driven, personalized, causal modeling through identification and understanding of optimal treatments for predicting better disease outcomes [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2018; 2018 Apr 14-18; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2018;78(13 Suppl):Abstract nr 5299.

Abstract 998: OPeN: the oncology precision network data sharing consortium

Background: The advent of Next-Generation Sequencing (NGS), and other molecular diagnostic technologies, has enabled the use of genomic information to guide targeted treatment in cancer patients. While this precision oncology approach can yield exciting clinical outcomes, the innumerable genomic variants identified in individual tumors effectively establishes each case as a unique N=1 clinical presentation. This scenario is contrary to a basic dogma of medical practice where historical cases and treatment outcomes guide future management and therapeutic decisions. Aggregation of large data sets, on a multi-institutional basis, has the potential to overcome the N=1 paradox and yield management insights in the implementation of precision oncology.

Methods: We have formed the Oncology Precision Network (OPeN), an oncology data sharing consortium, to aggregate big data sets consisting of clinical, genomic, pharmacological, and treatment response data from diverse patient cases. Data from Intermountain Healthcare, Stanford University, and Swedish Cancer Institute-Providence St. Joseph Health, as well as other institutions, comprises the database and is derived from 79 hospitals, over 800 physician clinics and more than 50,000 annual cases.

Results: The OPeN database can be interrogated by variant type, specific therapeutics, clinical outcomes, and by grouped variables, in a structured data format. The overarching IT platform is a cloud based, open source, triple store precision oncology solution, Syapse. These data are yielding valuable insights, including tumor mutational burden (TMB) scores and their correlation to immunotherapy response, clinical response in various drug-gene combinations, and therapy-specific adverse events.

Conclusions: We anticipate this resource will be used by the Molecular Tumor Boards of contributing institutions for clinical interpretation, and by treating providers to overcome the N=1 challenge associated with precision oncology.

Citation Format: Lincoln Nadauld, Derrick Haslem, Paul D. Tittel, Mariko Tameishi, Thomas Brown, James Ford. OPeN: the oncology precision network data sharing consortium [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2017; 2017 Apr 1-5; Washington, DC. Philadelphia (PA): AACR; Cancer Res 2017;77(13 Suppl):Abstract nr 998. doi:10.1158/1538-7445.AM2017-998

Abstract 997: Impact of a personalized medicine research program (PMRP), using targeted tumor profiling and a cloud based clinical trials matching platform, on clinical decision-making

Background: Cancer care is evolving to a model of precision medicine where genomic changes in a patient's tumor are used to inform individualized management (mgmt). The optimal approach and impact of tumor profiling on cancer care remain important research questions. We report the impact on clinical decision-making by results from a PMRP in a research practice.

Methods: A custom designed next generation sequencing (NGS) 68 gene alteration (GA) panel, covering clinically relevant genes and regions was developed in 2014. The NGS results were used to: 1) prioritize standard therapies; 2) match patients (pts) with clinical trials (CT); and 3) serve as a data mining resource. NGS testing was offered early in the course of mgmt. An Institutional Review Board approved prospective registration protocol (PMRP) was activated in 2014, with the objective of establishing a centralized longitudinal, molecular phenotypic, and research data repository. Primary endpoints include proportion of pts where NGS impacted mgmt, to include enrollment onto CT. A cloud-based informatics platform was developed to: manage PMRP; facilitate CT matching; perform quality assurance/quality improvement; pursue research initiatives.

Results: As of 11/15/2016, 869 pts gave informed consent, with 844 pts enrolled. The top primary sites included: breast (115); colorectal (111); central nervous system (103); lung (91); ovary (49); hematologic malignancies (46); pancreas (37); uterus (28); esophagus (25); skin (21). Of solid tumor pts with documented clinical stage, 130 (40%) pts had early stage cancer (I, II and III), and 193 (60%) pts had advanced stage (IV) cancer. NGS results: 739 (88%) pts with GA found; 27 (3%) pts without GA. Of pts with GA, 178 (24%) pts had actionable (on-label drugs) GA and 476 (64%) pts had applicable (off-label or CT) GA, for a total of 546 (74%) pts with actionable and/or applicable GA. The top actionable GA were: KRAS (125); PIK3CA (17); BRAF (13); EGFR (12); NRAS (11); AKT1 (3); TET2 (2); ERBB2 (2); HRAS (2). The top applicable GA, included: TP53 (225); TPMT (78); TYMS (78); PIK3CA (77); APC (56); PTEN (52); IDH1 (34); CDNK2A (22); CTNNB1 (18); TET2 (16). Care mgmt impact was reported by physicians for 508 pts with actionable/applicable GA. Physicians reported mgmt impact, at time of reporting, for 105 (21%) pts, to include: new treatment (Tx) in 30 (6%) pts; no Tx given in 18 (4%) pts; Tx changed in 12 (2%) pts; Tx stopped in 1 (<1%) pt; 6 (1%) pts enrolled onto CT. In 403 (79%) pts, physicians reported no mgmt impact, to include: insufficient evidence in 294 (58%) pts; drugs/CT access in 91 (18%) pts; refused Tx in 17 (3%) pts.

Conclusions: NGS profiling of tumors with this 68 GA panel has an impact on clinical decision- making in a minority, though substantial number, of pts. Impact on CT participation remains modest. Access to drugs and CT remains an important barrier.

Citation Format: Thomas D. Brown, Paul D. Tittel, Philip J. Gold, Charles W. Drescher, John M. Pagel, J D. Beatty, Patra Grevstad, Desiree Iriarte, Shlece Alexander, Madeleine Brindle, Xiaoyu Liu, Donielle O'connor, Mariko Tameishi, Danbin Xu, Anna B. Berry. Impact of a personalized medicine research program (PMRP), using targeted tumor profiling and a cloud based clinical trials matching platform, on clinical decision-making [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2017; 2017 Apr 1-5; Washington, DC. Philadelphia (PA): AACR; Cancer Res 2017;77(13 Suppl):Abstract nr 997. doi:10.1158/1538-7445.AM2017-997

Analysis of reimbursement (R) for next generation sequencing (NGS) on patients' tumors in the context of a personalized medicine program

6506

Background: R policies for NGS testing vary widely among private and public insurers. While drug costs are the greatest challenge in personalized or precision medicine, cost and R are substantial barriers to genomic profiling with NGS. We examined variation in coverage and R for a cohort of cancer patients (pts) treated at a tertiary oncology center. Methods: An Institutional Review Board approved prospective registration protocol was activated with the objective of establishing a centralized longitudinal clinical, molecular phenotypic, and research data repository for pts diagnosed with cancer. Based on provider assessment of medical necessity, mutations in 68 cancer associated genes were analyzed. Evaluation of R for NGS was performed from Sept, 2014 through Jan, 2017, with use of CPT code 81455. R was analyzed based on: payer type; pt age; localized vs. metastatic disease; and actionability of data. Results: 588 pts with evaluable analytic cases, and NGS testing, with R results shown in the table below. For groups with >= 10 pts: R frequency was highest in managed care programs, either private or Medicare, and least frequent in non-HMO Medicare (p<.001). In pts receiving R, payments by private HMOs were highest (p<.02). NGS results with labelled drug indications were associated with less frequent R (26% vs. 35%; p<.05), and lower payments (mean of $358 vs. $567; p<.02) compared to other NGS results. Younger age was associated with more frequent R (38% in pts <60 years, 24% in pts >= 60 years; p<.005). Neither cancer diagnosis nor stage were significantly associated with R. Conclusions: One third of pts received some R for NGS testing. R was more frequent and higher in managed care programs, both private and Medicare. R was more likely for younger age pts, while actionable NGS results were associated with lower R. These data demonstrate the need for rational, transparent, and consistent R policies, along with a value-based R model for NGS across all payer groups. [Table: see text]