MD Anderson’s Oncology Expert Advisor powered by IBM Watson: A Web-based cognitive clinical decision support tool.

Health Services Research
Session Type and Session Title: 
Oral Abstract Session, Health Services Research
Abstract Number: 
J Clin Oncol 32:5s, 2014 (suppl; abstr 6506)
Koichi Takahashi, Hagop M. Kantarjian, Guillermo Garcia-Manero, Rick J Stevens, Courtney Denton Dinardo, Joshua Allen, Emily Hardeman, Scott Carrier, Cynthia Powers, Pat Keane, Sherry Pierce, Mark Routbort, Thai Nguyen, Brett Smith, Jeffery Frey, Keith Perry, John C. Frenzel, Rob High, Andrew Futreal, Lynda Chin; The University of Texas MD Anderson Cancer Center, Houston, TX; IBM, Rocester, NY; IBM, Raleigh, NC; IBM, Durham, NC; IBM, Austin, TX

Abstracts that were granted an exception in accordance with ASCO's Conflict of Interest Policy are designated with a caret symbol (^).

Abstract Disclosures


Background: Application of cognitive computing technology in cancer has potential to democratize expert knowledge and enable personalized evidenced based oncology care. Methods: We developed a web based prototype of MD Anderson’s Oncology Expert Advisor (OEA), a cognitive clinical decision support tool powered by IBM Watson. The Watson technology is IBM’s third generation cognitive computing system based on its unique capabilities in natural language processing and deep QA (question-answer). OEA is built with four core capabilities: 1) Patient Evaluation through interpretation of structured and unstructured clinical data to create a dynamic case summary with longitudinal view of the pertinent events 2) Treatment and Management suggestions based on patient profile weighed against consensus guidelines, relevant literature, and MD Anderson expertise, which will include approved therapies, genomic based therapies as well as automatic matching to eligibility criteria of clinical trials at MD Anderson, 3) Care Pathway Advisory that supports management of patients by alerting adverse events and/or suggesting proactive care support, and 4) Patient-oriented Research functionalities for identification of patient cohorts and generation of hypothesis. Results: We trained OEA by loading 400 cases of historical patient cases and assessed the accuracy of OEA treatment suggestions using MD Anderson’s physicians’ decisions as benchmark. False positive result was defined when OEA recommends a non-correct answer with high confidence, whereas false negative was defined when OEA recommends a correct answer with low confidence. When 200 leukemia cases were tested to assess accuracy of standard-of-care (SOC) treatment recommendation, false positive rate was 2.9%, whereas false negative rate was 0.4%. Overall accuracy of SOC treatment recommendation by OEA was 82.6%. Conclusions: OEA is able to generate dynamic patient case summary by interpreting structured and unstructured clinical data and suggest personalized treatment options with reasonably high accuracy. Live system evaluation of OEA is ongoing and application of OEA in clinical practice is expected to be piloted at our institution.