Generalisability through local validation: overcoming barriers due to data disparity in healthcare

Abstract

Cho et al. report deep learning model accuracy for tilted myopic disc detection in a South Korean population. Here we explore the importance of generalisability of machine learning (ML) in healthcare, and we emphasise that recurrent underrepresentation of data-poor regions may inadvertently perpetuate global health inequity.Creating meaningful ML systems is contingent on understanding how, when, and why different ML models work in different settings. While we echo the need for the diversification of ML datasets, such a worthy effort would take time and does not obviate uses of presently available datasets if conclusions are validated and re-calibrated for different groups prior to implementation.The importance of external ML model validation on diverse populations should be highlighted where possible – especially for models built with single-centre data.

Optimizing human-centered AI for healthcare in the Global South

“…emphasizes the need for ML models, especially those trained on data from a single source, to be validated on local, representative datasets. 33 This will pave the way for generalizability in ML applications for healthcare, while ensuring that the Global South is equitably represented in these solutions.…”

Sensor Data Integration: A New Cross-Industry Collaboration to Articulate Value, Define Needs, and Advance a Framework for Best Practices

“…Similar data sources combined across different studies create meta-studies to increase statistical power to detect differences and improve the generalizability of results [ 14 , 15 ];…”

Global disparity bias in ophthalmology artificial intelligence applications

“…Although data from all world countries are a distant goal, equivalent representation of all continents, ethnicities and the maximum number of countries is desired to reduce ML bias. Demographic information and other social determinants of health are typically not contained in these datasets, making it challenging to interrogate algorithms for bias 7 8. High-quality data are also fundamental for environmental-specific algorithm validation, which is essential before AI implementation.…”

“…Available automated DR algorithm performance varies considerably in performance in the real world due to limited training data, including heterogeneity in disease presentations and suboptimal image quality 9. In addition, diverse sociodemographic and ethnic representation are necessary if generalisability is a goal 8…”

“… 9 In addition, diverse sociodemographic and ethnic representation are necessary if generalisability is a goal. 8 …”