Operating Point Optimization for Efficient Mammogram Triage Using Only Highly Elevated Probability scores

Using AI breast cancer detection scores to triage screening mammogram patients has been described in several studies. This type of triage could expedite a cancer diagnosis and significantly shorten the time from screening mammogram to treatment. However, the optimal operating point, or score value, above which patients should be triaged, has not been well defined. Identifying practical operating points is important because the Positive Predictive Value (PPV), sensitivity, specificity, and false-negative (FN) rate at different operating points can vary greatly.

Using patient datasets constructed from a clinical practice, the performance metrics of various AI scoring models were calculated at different operating point levels.

Mathematical analysis of the very highest cancer detection case scores revealed that they are highly specific for the presence of breast cancer with few false positives. Therefore, if only the few highest case scores are used for clinical triage, the expectation is that most patients triaged will have cancer.

In clinical practice, this approach could be used to identify patients who would benefit from immediate diagnostic workup without overwhelming the clinic logistics with false-positive cases.