基于算法的智能手机应用程序评估成人皮肤癌风险:诊断准确性研究的系统综述
该系统综述评估了基于算法的智能手机应用程序在可疑皮肤病变中评估皮肤癌风险的诊断准确性。纳入9项研究、6款可识别的应用程序,多数研究样本量小、方法学质量差,存在选择性招募、图像不可评估率高及差异验证等问题。SkinVision在两项研究中检测恶性或癌前病变的敏感度为80%(95%CI 63%至92%)、特异度为78%(95%CI 67%至87%),但当前应用程序不能可靠检测所有黑色素瘤或其他皮肤癌。
要点
- 当前基于算法的智能手机应用程序不能可靠检测所有黑色素瘤或其他皮肤癌。
- SkinVision检测恶性或癌前病变敏感度80%(95%CI 63%至92%),特异度78%(95%CI 67%至87%)。
- SkinScan在单研究中检测黑色素瘤敏感度0%、特异度100%(n=15,5例黑色素瘤)。
核心结论
- 当前基于算法的智能手机应用程序不能可靠检测所有黑色素瘤或其他皮肤癌。
- 在临床相关人群和应用程序预期用户中使用时,检测性能可能比本文报告更差。
- 当前授予基于算法应用程序CE标志的监管流程不能为公众提供充分保护。
研究设计与样本
- 系统综述,检索Cochrane Central Register of Controlled Trials、MEDLINE、Embase、CINAHL、CPCI、Zetoc、Science Citation Index及在线试验注册库,从建库至2019年4月10日。
- 纳入任何设计、评估基于算法的智能手机应用程序对可疑皮肤癌皮肤病变图像进行评估的研究。
- 参考标准包括组织学诊断或随访,以及专家建议进一步检查或干预。
- 由两名作者独立提取数据并使用QUADAS-2评估有效性。
- 共纳入9项研究,评估6款不同的可识别智能手机应用程序。
主要发现
- 6项研究通过组织学或随访验证结果(n=725个病变),3项研究通过专家建议验证结果(n=407个病变)。
- 研究样本量小且方法学质量差,存在选择性招募、图像不可评估率高及差异验证。
- 病变选择和图像采集由临床医生而非智能手机用户完成。
- SkinScan在单研究中评估(n=15,5例黑色素瘤),检测黑色素瘤敏感度0%、特异度100%。
- SkinVision在两项研究中评估(n=252,61个恶性或癌前病变),检测恶性或癌前病变敏感度80%(95%CI 63%至92%)、特异度78%(95%CI 67%至87%)。
关键数值与推荐
- SkinScan检测黑色素瘤:敏感度0%、特异度100%(单研究,n=15,5例黑色素瘤)。
- SkinVision检测恶性或癌前病变:敏感度80%(95%CI 63%至92%)、特异度78%(95%CI 67%至87%)(两项研究,n=252,61个恶性或癌前病变)。
- SkinVision对照专家建议的准确性差(三项研究)。
- 两款CE标志应用程序可供下载。
临床怎么用
- 不应依赖当前基于算法的智能手机应用程序检测所有黑色素瘤或其他皮肤癌。
- 临床相关人群和应用程序预期用户使用时,检测性能可能比研究报告更差。
- 当前CE标志监管流程不能为公众提供充分保护。
局限性
- 纳入研究样本量小且方法学质量差。
- 存在选择性招募、图像不可评估率高及差异验证。
- 病变选择和图像采集由临床医生而非智能手机用户完成。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
OBJECTIVE: To examine the validity and findings of studies that examine the accuracy of algorithm based smartphone applications ("apps") to assess risk of skin cancer in suspicious skin lesions. DESIGN: Systematic review of diagnostic accuracy studies. DATA SOURCES: Cochrane Central Register of Controlled Trials, MEDLINE, Embase, CINAHL, CPCI, Zetoc, Science Citation Index, and online trial registers (from database inception to 10 April 2019). ELIGIBILITY CRITERIA FOR SELECTING STUDIES: Studies of any design that evaluated algorithm based smartphone apps to assess images of skin lesions suspicious for skin cancer. Reference standards included histological diagnosis or follow-up, and expert recommendation for further investigation or intervention. Two authors independently extracted data and assessed validity using QUADAS-2 (Quality Assessment of Diagnostic Accuracy Studies 2 tool). Estimates of sensitivity and specificity were reported for each app. RESULTS: Nine studies that evaluated six different identifiable smartphone apps were included. Six verified results by using histology or follow-up (n=725 lesions), and three verified results by using expert recommendations (n=407 lesions). Studies were small and of poor methodological quality, with selective recruitment, high rates of unevaluable images, and differential verification. Lesion selection and image acquisition were performed by clinicians rather than smartphone users. Two CE (Conformit Europenne) marked apps are available for download. SkinScan was evaluated in a single study (n=15, five melanomas) with 0% sensitivity and 100% specificity for the detection of melanoma. SkinVision was evaluated in two studies (n=252, 61 malignant or premalignant lesions) and achieved a sensitivity of 80% (95% confidence interval 63% to 92%) and a specificity of 78% (67% to 87%) for the detection of malignant or premalignant lesions. Accuracy of the SkinVision app verified against expert recommendations was poor (three studies). CONCLUSIONS: Current algorithm based smartphone apps cannot be relied on to detect all cases of melanoma or other skin cancers. Test performance is likely to be poorer than reported here when used in clinically relevant populations and by the intended users of the apps. The current regulatory process for awarding the CE marking for algorithm based apps does not provide adequate protection to the public. SYSTEMATIC REVIEW REGISTRATION: PROSPERO CRD42016033595.
引用
- Freeman K, Dinnes J, Chuchu N, et al. Algorithm based smartphone apps to assess risk of skin cancer in adults: systematic review of diagnostic accuracy studies. BMJ. 2020. PMID: 32041693. doi:10.1136/bmj.m127
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。