Can AI Apps Improve Skin Cancer Detection, or Just Increase Healthcare Consumption?

🔍 Key Finding An AI-based mHealth app for skin cancer detection increased claims for premalignant and malignant skin lesions by 32% compared to non-users, but also led to a three to fourfold increase in claims for benign skin tumors and nevi, likely due to the app’s suboptimal specificity. While promising for improving skin cancer detection, further improvements in diagnostic accuracy and targeted implementation in high-risk populations are needed for cost-effective and optimal use of healthcare resources.

🔬 Methodology Overview

• Design: Retrospective, population-based, pragmatic study.
• Participants: 18,960 mHealth app users (SkinVision) matched 1:3 with 56,880 controls who did not use the app. Matching criteria included age, sex, socioeconomic status, residential area, history of skin cancer, and comorbidities.
• Intervention: Free access to the SkinVision app for skin cancer detection.
• Data Collection: Dermatological healthcare claims data from a large Dutch health insurance company for the year 2019, along with app usage data (number of assessments, risk ratings by the app’s CNN and teledermatologists).
• Outcome Measures: Frequency and type of dermatological healthcare claims (premalignant and malignant skin lesions, benign skin tumors and nevi, unrelated dermatological claims), diagnostic and therapeutic interventions, and direct healthcare costs.
• Statistical Analysis: Odds ratios (OR) to compare claim proportions between mHealth users and controls, two-proportion z-tests, Fisher’s Exact Test, Benjamini-Hochberg procedure for multiple testing correction, and short-term cost-effectiveness analysis using incremental cost-effectiveness ratio (ICER).
• Subgroup Analyses: Comparisons within mHealth users based on high-risk vs. low-risk assessments and a separate analysis for users with a prior history of skin cancer.

📊 Results

• Increased (pre)malignant skin lesion claims: mHealth app users had a 32% increase in claims for (pre)malignant skin lesions compared to controls (6.0% vs. 4.6%; OR 1.3 [95% CI 1.2-1.4]).
• Higher benign lesion claims: mHealth app users had a three- to four-fold higher risk of claims for benign skin tumors and nevi (5.9% vs. 1.7%; OR 3.7 [95% CI 3.4-4.1]). This was largely driven by claims for nevi (OR 4.0 [95% CI 3.6-4.4]).
• Increased biopsies and excisions: mHealth app users had twice as many biopsies and excisions at the GP level (75 per 1000 persons vs. 34 per 1000 persons; p<0.001) and more hospital-based excisions (55 vs. 25 per 1000 persons; p<0.001).
• Higher costs: mHealth app users had higher average annual dermatological healthcare costs (€64.97 vs. €43.09; Δ €21.88 [95% CI 17.90-25.85]; p<0.001). The largest cost difference was for nevi (€11.05 vs. €2.71 per person; p<0.001).
• Cost per additional (pre)malignancy: The incremental cost-effectiveness ratio (ICER) was €2567 per additional (pre)malignant lesion detected using the app, compared to standard care. €1843 of this was attributed to the increased costs associated with benign lesions.
• Subgroup analysis (high-risk assessment): mHealth users with at least one high-risk assessment had a 9.1% claim rate for (pre)malignancy, compared to 4.3% for those with only low-risk assessments. The high-risk group also had a six-fold higher risk of claims for benign skin tumors/nevi (OR 6.7 [95% CI 5.9-7.7]).

💡 Clinical Impact AI-based mHealth app use increased detection of premalignant and malignant skin lesions by 32%, but also led to a three to fourfold increase in benign skin lesion assessment. Further improvements in diagnostic accuracy and targeted implementation for high-risk populations are needed for optimal, cost-effective mHealth app integration into dermatological care.

🤔 Limitations

• Underestimation of mHealth app impact due to reliance on claims data.
• Lack of data on the number, type, and stage of skin cancers.
• Lack of detailed primary care skin cancer data.
• Potential residual confounding due to differences between app users and non-users.
• Potential overdiagnosis and overtreatment in app users due to increased worry about skin lesions.
• Limited generalizability due to younger study population, excluding a large portion of the elderly skin cancer population.
• Cross-sectional, short-term cost-effectiveness analysis limited to dermatological healthcare costs, lacking long-term data on health and cost impact.

✨ What It Means For You This study suggests that AI-powered mHealth apps for skin cancer detection can increase detection rates of premalignant and malignant skin lesions, but also lead to a substantial increase in consultations for benign lesions, potentially straining resources. Doctors should be aware of this potential for increased workload related to benign lesions when advising patients on the use of such apps, and further research is needed to optimize the balance between improved detection and overdiagnosis.

Reference Smak Gregoor AM, Sangers TE, Bakker LJ, Hollestein L, Uyl-de Groot CA, Nijsten T, Wakkee M. An artificial intelligence based app for skin cancer detection evaluated in a population based setting. npj Digital Medicine. 2023;6:90. https://doi.org/10.1038/s41746-023-00831-w