Key points are not available for this paper at this time.
Watson for Oncology (WFO), an artificial intelligence platform, demonstrated a strong concordance with recommendations from a panel of oncologists in a double-blind validation study presented at the San Antonio Breast Cancer Symposium, which was held December 6–10, 2016. WFO was developed by IBM Corporation (Armonk, NY) in collaboration with Memorial Sloan Kettering Cancer Center in New York City. The computing system can extract and assess large amounts of structured and unstructured data from medical records. The system then uses natural language processing and machine learning to present cancer treatment options. The study's lead author, S. P. Somashekhar, MBBS, MS, Mch, chairman of the Manipal Comprehensive Cancer Center in Bangalore, India, says that WFO provides treatment recommendations for patients with breast, lung, and colorectal cancer. His facility recently adopted the system to support oncologists in making quality, evidenced-based decisions, he says. Dr. Somashekhar and his colleagues wanted to assess and compare the agreement between WFO and the Manipal Comprehensive Cancer Center's multidisciplinary tumor board, a group of 12 to 15 oncologists who meet weekly to review cases. The group studied the cases of 638 patients who had been treated at the hospital. They then analyzed the degree of concordance between the tumor board and the recommendations of WFO, as well as how much time it took for each method to issue recommendations. The WFO recommendations were divided into 3 categories: those recommended for standard treatment, those recommended for consideration, and those not recommended. Approximately 90% of WFO's recommendations for standard treatment or for consideration were concordant with those of the tumor board. The extent of similarity in recommendations varied with the type of breast cancer. Nearly 80% of the recommendations were concordant in patients with nonmetastatic disease, compared with only 45% in metastatic cases. In patients with triple-negative cancer, WFO agreed with the physicians 68% of the time. However, in patients with human epidermal growth factor receptor 2 (HER2)/neu-negative breast cancer, WFO and physicians were in agreement only 35% of the time. Dr. Somashekhar says the difference is not surprising because patients with triple-negative breast cancer have fewer treatment options than those with HER2/neu-negative breast cancer. More complicated cases lead to more divergent options regarding treatment, he notes. Researchers found that it took an average of 20 minutes to manually capture and analyze the data and to generate recommendations; however, after oncologists gained more familiarity with the cases, the mean time decreased to approximately 12 minutes. By comparison, WFO took 40 seconds to capture and analyze the data and then make a recommendation. However, Dr. Somashekar cautions that although artificial intelligence is a helpful step toward personalized medicine, it should be viewed as a complement to the physician's work, not a replacement. When dealing with humans, many factors cannot be addressed by a machine, such as the context and preferences of each patient, the patient-physician relationship, human touch, and empathy, he adds.
Carrie Printz (Mon,) studied this question.