Improving AI/ML-readiness of Synthetic Data in a Resource-Constrained Setting

点击翻译按钮获取中文摘要

PROJECT SUMMARY / ABSTRACT The parent project, UZIMA-DS (UtiliZing Health Information for Meaningful Impact in East Africa through Data Science), aims to create a scalable and sustainable platform to apply novel approaches to data assimilation and advanced artificial intelligence (AI)/machine learning (ML)-based methods to improve health outcomes in two health domains: maternal, newborn and child health; and mental health. Led by the Aga Khan University in East Africa (AKU) and the University of Michigan, UZIMA-DS is a U54 Research Hub funded under the NIH Data Sci- ence for Health Discovery and Innovation in Africa Initiative. During these first two years, UZIMA-DS has focused on acquiring and harmonizing multimodal data sources. However, we and many other DS-I Africa awardees have encountered several barriers to efficiently and effectively creating AI-ready data sets, which include: 1) regulatory concerns around privacy and confidentiality, 2) heterogeneity in data laws across countries limiting the accessibil- ity of data, and 3) a lack of sufficient datasets not only for training ML models and validation but also for training students and early career investigators for capacity building. Synthetic data, or data that is generated artificially using computational techniques such as AI, is a promising technique that could address these barriers and ena- ble the broad sharing of AI-ready data sets. As part of this administrative supplement, we propose to create an AI-ready synthetic data set using one of our real UZIMA-DS data sets from Kenya: the Kaloleni-Rabai Health and Demographic Surveillance Systems (KRHDSS). KRHDSS is a population-based demographic and health surveil- lance system established in 2017 by AKU. Information is collected at least annually on ~40 demographic, health, social determinants of disease, and vital events from a resident population of about 99,000 individuals. Leverag- ing our preliminary work using a Microsoft Azure instance, we will create AI-ready synthetic datasets for research and training and evaluate whether causal relationships in real data are preserved in synthetic datasets. The overarching goal of this proposal is to “put data to work” by developing a roadmap for the curation and use of AI-ready synthetic data using FAIR principles (findable, accessible, interoperable, and re‑usable) that can be eas- ily accessed and shared for research and training purposes across the globe. Ultimately, this work has the poten- tial to promote more effective and efficient sharing of AI-ready data globally. Using cloud infrastructure and Health and Demographic Surveillance Systems data from rural Kenya as a use case, this work has immediate implications for how AI-ready data can be leveraged in resource-constrained settings to improve data driven health policy decisions for traditionally disadvantaged and marginalized groups.