Diagnosing Ovarian and Other Cancers With Human and Artificial Intelligence

Kenny Lin, MD, MPH
Posted on February 6, 2023

The Ovarian Cancer Research Alliance recently released a consensus statement that encourages women who have completed childbearing to consider having their fallopian tubes removed if they are having pelvic surgery for benign conditions. This surgical strategy for reducing ovarian cancer mortality acknowledges the negative results of the UK Collaborative Trial of Ovarian Cancer Screening, which Dr. Jennifer Middleton previously discussed on the blog. Studies have found that early ovarian cancer frequently causes constitutional, abdominal, urinary, and pelvic symptoms, which can be used in combination with a cancer antigen 125 level to estimate ovarian cancer risk. However, the low sensitivity and specificity of this “symptom index” means that it misses early cancers and leads to unnecessary cancer evaluations for benign conditions.

A population-based UK cohort study assessed the underlying cancer risk of patients 30 years and older who presented to primary care with new-onset fatigue and co-occurring vague symptoms. After excluding persons with anemia or alarm symptoms (e.g., dysphagia), researchers followed patients for up to nine months for a diagnosis of cancer. Cancer risk increased with age, reaching 3% or more in patients in their mid-60s. For all age groups combined, cancer risk was highest for women with fatigue and abdominal bloating and men with fatigue and weight loss, constipation, dyspnea, or abdominal pain.

A Danish research group developed an artificial intelligence (AI) based model that used results from common blood tests (complete blood count, electrolytes, and/or liver function tests) to generate a risk score that predicted cancer within 90 days. This  laboratory data is often readily available; another UK study of patients who were diagnosed with cancer found that 41% received common blood tests in primary care as part of their diagnostic process. However, relying solely on blood tests neglects the value of primary care physicians’ nonanalytical “gut feelings” that the patient has a benign or serious condition.

In an observational study of 155 general practitioners in Spain, a “sense of alarm,” present in 22% of consultations for new symptoms, had a sensitivity of 59% for cancer and other serious diseases and a negative predictive value of 98%. Thus, AI may also assist in cancer diagnosis by imitating the intuitive behavior of groups of family physicians. In an editorial in Annals of Internal Medicine, Dr. Gary Weissman and colleagues proposed that AI clinical decision support systems (CDSSs) utilize a “wisdom of crowds” approach that, like the best chess-playing AI systems, “reli[es] on imitation learning and collective intelligence” rather than set rules:

Averaging the judgments of many clinicians may outperform even the best clinician in the group. Training models to learn these consensus behaviors could lead to clinically significant improvements in accuracy. Furthermore, most diagnostic errors are the result of overlooking common diagnoses rather than very rare ones.… An AI CDSS that offers human-like suggestions may improve the reliability of clinical care by helping to avoid these clinical blunders.… Having an AI system that acts more like a thoughtful human guide rather than a black-box arbiter of truth may be the best next move.