Costs, benefits and diagnostic accuracy of machine learning applications for the autonomous detection of melanoma in high-risk individuals.

Abstract

Melanoma of the skin is one of the most frequently diagnosed cancers in Australia, with an estimated 18,257 new cases and over 1300 deaths in 2023. Australia continues to record the highest global incidence. Early detection of invasive melanoma is critical for survival, and timely diagnosis remains a priority. While population-level screening is not recommended, research has demonstrated that screening for melanoma in a high-risk clinic in Australia can be cost-effective.

Methods to improve melanoma detection include the use of artificial intelligence, or the practical application, machine learning. Initial trials suggested dermatologists outperformed machine learning algorithms; however, more recent studies demonstrate machine learning achieving, and in some cases surpassing, dermatologist-level diagnostic accuracy.

This thesis examines the existing body of literature on the use of machine learning to detect melanoma in adults at high-risk of developing melanoma of the skin, focusing on diagnostic accuracy, costs and benefits. Chapter 1 provides an overview of melanoma and non-melanoma skin cancer; Chapter 2 is a narrative review of machine learning in healthcare. Chapter 3 outlines the protocol for the scoping review and Chapter 4 is the scoping review. 9,188 records were screened, of which 55 studies met the inclusion criteria. Machine learning demonstrated high diagnostic accuracy in controlled settings with 78% of reader studies reporting ROAUC >0.8, compared to 60% of studies in a clinical setting. Only five studies reported on costs and benefits in high-risk populations, two conducted formal economic evaluations. One found no significant cost-effectiveness advantage, while the other reported cost savings and reduced clinician workload with machine learning. Evidence supporting the use of machine learning in high-risk populations remains limited, particularly regarding cost-effectiveness and studies in real-world clinical settings.