Performance of an artificial intelligence-based tool for cervical precancer screening in five countries in Africa: a prospective, observational, diagnostic accuracy study

Henry Phiri & Groesbeck Parham et al.

Cervical cancer kills 300 000 women annually, almost entirely in low-income and middle-income countries. Visual inspection of the cervix using acetic acid (VIA) remains a common method of screening but has suboptimal accuracy. Automated visual evaluation (AVE) is an internally validated artificial intelligence (AI)-based tool designed to assist health workers to improve VIA and to aid diagnostic accuracy. We aimed to measure sensitivity and specificity of AVE for detection of histology-confirmed cervical intraepithelial neoplasia 2 or greater (CIN2 +) in real-world clinical settings, including testing its performance relative to naked-eye VIA. This prospective diagnostic accuracy study took place in government health facilities in Malawi, Rwanda, Senegal, Zambia, and Zimbabwe convenience-sampled between March 17, 2022, and Jan 13, 2023. Eligible participants were women attending routine cervical screening. Women who were pregnant; less than 6 weeks postpartum; had a previous hysterectomy or removal of the cervix; previous treatment for cervical precancer or cancer; and impaired visualisation of the cervix due to inflammatory changes associated with acute infection were excluded. Participants formed a consecutive series. The index test was the previously validated AI-based AVE system. The reference test was histopathology assessment of cervical tissue sample. Among 24  447 eligible women, 9559 (39·1%) were women living with HIV and 11 635 (47·6%) women were positive on at least one test: 7695 (31·5%) on HPV, 5610 on AVE (22·9%), and 2314 (9·6%) on VIA. Of 18  086 women with confirmed final status, 526 (2·9%) had CIN2 +. AVE was found to have sensitivity of 60·1% (95% CI 55·5-64·5) for CIN2 + and specificity of 81·9% (81·3-82·5) for less than CIN2. VIA was found to have sensitivity of 36·6% (32·2-41·1) and specificity of 94·2% (93·8-94·5). Improved sensitivity of AVE over VIA was statistically significant (p<0·0001). AVE-assisted VIA, considering a participant positive when either VIA or AVE is positive, was found to have sensitivity of 71·8% (67·6-75·8) and specificity of 79·0% (78·3-79·6). HPV sensitivity was 90·4% (87·4-92·9) and specificity was 80·1% (79·5-80·7). AVE showed increased sensitivity compared to VIA, with moderate loss in specificity. This method could potentially increase detection of cervical precancerous lesions. High HPV and CIN2 + positivity, influenced by high HIV positivity among participants, underscore the importance of scaling up population-based screening programmes in resource-limited settings to support cervical cancer elimination. Unitaid and Global Health Labs.
Authors
Henry Phiri, Francois Uwinkindi, Mamadou Diop, Mulindi Mwanahamuntu, Lucia Gondongwe, Bothwell Takaingofa Guzha, Caroline Barrett, Jessica Trenc Joseph, Matthew P Horning, Malick Anne, Dipayan Banik, Frehiwot Birhanu, Stephen Burkot, Samson Chisele, Nang'andu Chizyuka, Marieme Diallo, Halimatou Diop-Ndiaye, Sylvie Gaju, Liming Hu, Rob Jenison, Sourabh Kulhare, Shannon Kuyper, Zohreh Laverriere, Chifundo Makwakwa, Evans Malyangu, Tatenda Maparo, Chengetai Mukotsanjera, Ishan Shah, Aaron Lunda Shibemba, Timothy Tchereni, Emily Meserve, Danny A Milner, Carlos Parra-Herran, Eric Yang, Jelena Mirkovic, Karen Hariharan, Groesbeck Parham