Screening programs (like all healthcare services) not only have to prove medical effectiveness, but also cost-effectiveness. The test itself must be relatively low-cost if it is to be deployed on a large scale. False positives must be minimized not only to avoid additional costs of diagnostic workup, but also to prevent health risks of unnecessary interventions and the psychological strain induced by positive test results.
Several imaging-based screening tests exist currently, such as breast cancer screening with mammography, lung cancer screening with chest CT, and colon cancer screening with CT colonography. The convoluted reimbursement for these three is a case study in the challenges of payment policy for screening tests. Screening mammography and lung cancer screening are covered by both Medicare and commercial payers. CT colonography is covered by the major commercial payers but not Medicare. A complete discussion of the economic evaluation of screening tests is beyond the scope of this article. However, payment policy is an important component of screening programs as reimbursement is necessary to incentivize adoption. In addition, lack of payment policy is a known barrier to screening implementation.
To detect diseases earlier, we need to predict who is going to be diagnosed in the future. The prevalence of a disease is often more important for costs and outcome of a screening program than the test validity. Current screening programs are not suitable for early detection of rare diseases with low prevalence. This concept is explained by Bayes’ Theorem, which predicts potentially higher-than-tolerable false positive rates for disease detection in the setting of rare diseases, even when using a screening test with high sensitivity and specificity. The complex task of forecasting risk could be bolstered by AI tools, which have potential to refine screening guidelines based on a person’s level of risk for developing a certain type of cancer. A shift from mass screening to selective screening could alter the cost-benefit equation to make screening feasible for cancers of the bladder, pancreas, kidney, and others. AI’s potential ability to decrease the cost of screening tests (through technical efficiencies and targeted risk modeling) and increase the quality of screening tests (through reduction in overdiagnosis) increases the overall value of screening in general and could negate many of the criticisms of current screening.
One example of an AI tool targeting population screening is a vertebral compression fracture algorithm. The algorithm uses deep learning to identify incidental osteoporotic compression fractures on chest CT performed for other reasons. This information could be used to assist healthcare providers in accurately identifying patients at risk and placing them under supervision or in fracture-prevention programs to reduce the risks of subsequent osteoporotic fractures. Unlike many of the currently-marketed AI tools focused on triage, this type of algorithm shifts emphasis to population health in ways that potentially foreshadow the future of screening.
CMS recently approved a Category III CPT® code for the vertebral compression fracture AI tool, largely predicated on the potential impact on population health. It is important to note that Category III codes are a set of temporary codes assigned to emerging technologies, services, and procedures. Unlike Category I codes, these codes are typically not reimbursed by Medicare or commercial payers. The lack of payment is certainly tied to the currently sparse data for true outcome advantages of using the tool, as well as well-defined cost savings. Nonetheless, this is the first AI code specific to radiology and provides a glimpse of how this type of technology may fit into the fee-for-service system.
With or without AI, screening tools cannot increase downstream costs in healthcare. Not only must false positives be minimized, but there must also be safeguards in place against fraud and abuse. The potential for this type of abuse is arguably greater with AI-augmented screening, as larger populations can be screened in less time. Widespread screening tools used by companies to market a product or procedure to a targeted population would present an ethical dilemma, as well as an economic one. AI screening tools that serve as hypersensitive detection algorithms may find “disease” that would have otherwise never impacted a patient’s longevity or quality of life. In screening, as in all of our profession, we must strive to provide all of the care that is necessary and none that is not.