Constructing a software platform by harnessing emerging data science tools for improved analytics and monitoring of female genital schistosomiasis and

点击翻译按钮获取中文摘要

The goal of this project is to develop an open-source, point-of-care data platform that combines automated visual examination technology and individual spatio-temporal risk profiles to assist in female genital schistosomiasis (FGS) diagnostic decision-making. Four smaller work packages, focussing on low-resource intensive data collection/aggregation, machine learning and Bayesian networks, will all feed into a final fifth package to develop the point-of-care platform. Schistosomiasis is second only to malaria in terms for infectious disease public health impact, and FGS is a particularly neglected form of the disease. Safe, accurate, and low-cost diagnostic solutions for FGS are not readily available, and there are still gaps in understanding FGS symptoms and transmission dynamics. The work completed in this project will help improve the current FGS landscape by providing a point-of-care diagnostic software platform to be used by clinicians, including those not speciality trained to review colposcope images. Early and accurate detection of FGS not only means earlier treatment of the disease in an individual but also has positive ramifications for controlling community transmission in endemic areas. Work package one (WP1) will focus on developing and trialling the Badger Data System (BDS). This system is currently at a prototype phase, based on the credit card-sized Badger 2040W E-ink device developed by Pimoroni. This system will be programmed as a daily symptomology diary and given to women as part of a pilot longitudinal study, nested within the Schista! Study in Zambia, to investigate disease dynamics and progression of FGS throughout the menstrual cycle. Not only will the output of this pilot further the understanding of FGS, but it will also result in the development of an open-source data collection system that can be applied in a range of contexts. FGS can be identified through visual examination of the cervix using a colposcope. Work package two (WP2) and three (WP3) will involve refining techniques for colposcope image pre-processing (WP2) and then building a machine learning framework (WP3) for FGS identification and classification, essentially giving a computer the ability to 'see' FGS in a colposcope image. This process is known as automated visual examination (AVE). Various deep learning methods for AVE (convolutional neural networks (CNN), ResNET, Deep SVDD) will be trialled in WP3 and selected for the final data platform based on sensitivity and specificity performance. In regards to the pre-processing (WP2), the machine learning in WP3 will only be as good as the images used for training so both quality and volume are important. To increase the volume of training images, AI-powered image enhancement and pixel upscaling software will be assessed for their accuracy in upscaling colposcope images. If the enhancement is found to be true to life, then these images will be fed into WP3 to increase the sensitivity and specificity of the software platform. This upscaling software may also assist with accurately processing lower-resolution images so that handheld imaging devices (smartphones, digital cameras, handheld colposcopes) can be used.In work package four (WP4), to strengthen the results of the AVE, data from various sources, including the results of WP1, will be collected on variables to increase model sensitivity further. These variables will include individual demographics, temporal, spatial, and environmental data and will be assessed for their association with FGS using various regression techniques and geographic information systems (Eg. ArcGIS). To decode the complex web of disease transmission, the explanatory variables will be integrated into a Bayesian network to produce a risk model, which will be integrated into the software platform for more accurate diagnostic classifications.