Neural network-assisted ("NNA") analysis of cervical smears: pooled effectiveness results and economic analysis

Objective: To determine the sensitivity of cervical cancer smear screening with neural network-assisted ("NNA") rescreening.Methods: The Papnet system of NNA analysis of cervical smears has been in clinical use worldwide for over 3 years and has been the subject of over 22 published manuscripts reporting on data from over 202,000 smears. This investigation reviewed the results of these studies and classified each study according to study design using a systematic protocol based on reference validation, diagnostic threshold for abnormal, and outcome metric. This classification taxonomy allowed for weighted (based on number of cases in each study) pooling of studies for each study design class. The pooled effectiveness metrics were used to derive the sensitivity of cervical cancer screening with NNA rescreening, using a baseline unassisted screening sensitivity of 85%. Other effectiveness metrics determined by this analysis include NNA's sensitivity as a primary screener, comparisons with primary unassisted screening, and comparisons of NNA rescreening and unassisted rescreening.Results: Analyses of the weighted, pooled mean estimates for each of the principal outcome metrics indicate the sensitivity of cervical cancer screening with NNA ranges from 90% to 99%; most pooled estimates fall in the range of 97-99%. An economic analysis using the APL-based "Cervical Cancer Screen" computer model developed by Eddy (Eddy DM. Screening for cervical cancer. Ann Intern Med 1990;113:214-26) and these effectiveness estimates as inputs showed that NNA analysis involves an accepted level of resource expenditure (approximately $40,000 per life year saved) when added to unassisted screening on a triennial basis.Conclusion: The sensitivity of cervical cancer screening with NNA rescreening using the Papnet system yields sensitivities in excess of 90% and approaching 99%.