Demystifying the localized relationships between population health outcomes and multi-source determinants: A spatially varying GeoAI framework
Nduwayezu, Gilbert LU
(2025)
- Abstract
- Data-driven modeling frameworks have become essential tools for guiding surveillance strategies and informing public health policies across diverse population health challenges. Accurate, fine-scale disease estimates often lacking from direct surveys are critical for policy planning, given that spatial heterogeneity and nonlinear dynamics among determinants of health challenge classical models, limiting their utility for targeted public health interventions. Advances in geospatial artificial intelligence (GeoAI), and increased computational power, have enabled deeper insights into spatial non-stationarity, the shifting strength and direction of relationships across space. These advances enhance both the accuracy and contextual relevance of... (More)
- Data-driven modeling frameworks have become essential tools for guiding surveillance strategies and informing public health policies across diverse population health challenges. Accurate, fine-scale disease estimates often lacking from direct surveys are critical for policy planning, given that spatial heterogeneity and nonlinear dynamics among determinants of health challenge classical models, limiting their utility for targeted public health interventions. Advances in geospatial artificial intelligence (GeoAI), and increased computational power, have enabled deeper insights into spatial non-stationarity, the shifting strength and direction of relationships across space. These advances enhance both the accuracy and contextual relevance of spatial modeling, supporting localized decision-making in population health outcomes. Drawing on diverse spatial frameworks, this thesis developed, tested, and applied localized spatially varying GeoAI methodologies, offering integrated modeling approaches to address stunting among children in the complexities of such public health concern. Paper I develops localized spatially varying approaches to reveal significant intra-area variation in stunting prevalence and nonlinear relationships using cross-sectional socioeconomic and fine-scale remotely sensed climatic and agroecological data to better characterise household microenvironments. The approach provided a more detailed understanding of how local environments shape nutrition outcomes and demonstrating the importance of considering both scale and nonlinearity in stunting research. Building on this, Paper II implements a hybrid spatial machine learning (ML) framework to detect fine-scale heterogeneity in stunting prevalence, while also quantify localized disparities that national-level surveys overlook. The framework captured spatially heterogeneous outcomes across most areas, with predictors exhibiting regions-specific effects that vary according to different thresholds of influence. Paper III advances the analysis by implementing a hybrid spatially varying deep learning (DL) approach, which captured convoluted nonlinear influence of fine socio-economic determinants on child stunting outcomes. The algorithm fairly captured variability in stunting outcomes, highlighting key child, maternal, and household determinants whose contributions varied across space, though limitations in training data size constrained broader generalizability. Paper IV further refines this perspective by introducing a predictive multilevel spatial ensemble learning (SEL) framework to produce small area estimates (SAEs) of stunting risk by combining geomasked household data with agroecological and remote sensing (RS) indicators. This approach demonstrated the capacity of predictive models to generalize beyond sampled survey clusters and produce continuous prevalence surfaces at scales as fine as 1 km². Overall, these papers highlight that trade-offs between interpretability, generalizability, and spatial scale in these analytical and predictive models remain challenging to navigate and must be evaluated case by case according to research priorities. The methodologies presented in this thesis aim to generate fine-scale, interpretable risk estimates that can support targeted nutrition interventions in data-scarce settings. (Less)
Please use this url to cite or link to this publication:
https://lup.lub.lu.se/record/75434166-a0e3-46d7-b9fb-142330ac6f9e
- author
- Nduwayezu, Gilbert
LU
- supervisor
- opponent
-
- Professor Scholz, Johannes, Department of Geoinformatics – Z_GIS, Paris-Lodron University Salzburg
- organization
- publishing date
- 2025
- type
- Thesis
- publication status
- published
- subject
- keywords
- Population health, childhood stunting, multi-source data, spatially varying GeoAI, small area estimates, interpretability, localized interventions
- pages
- 78 pages
- publisher
- Lund University
- defense location
- Världen (GC1:F111). The defence will be live streamed, but part of the premises is to be excluded from the live stream: https://lu-se.zoom.us/s/64886467528
- defense date
- 2025-12-11 13:00:00
- ISBN
- 978-91-89187-65-8
- 978-91-89187-66-5
- project
- Demystifying the localized relationships between population health outcomes and multi-source determinants: A spatially varying GeoAI framework
- language
- English
- LU publication?
- yes
- id
- 75434166-a0e3-46d7-b9fb-142330ac6f9e
- date added to LUP
- 2025-11-11 14:45:51
- date last changed
- 2025-11-18 03:46:14
@phdthesis{75434166-a0e3-46d7-b9fb-142330ac6f9e,
abstract = {{Data-driven modeling frameworks have become essential tools for guiding surveillance strategies and informing public health policies across diverse population health challenges. Accurate, fine-scale disease estimates often lacking from direct surveys are critical for policy planning, given that spatial heterogeneity and nonlinear dynamics among determinants of health challenge classical models, limiting their utility for targeted public health interventions. Advances in geospatial artificial intelligence (GeoAI), and increased computational power, have enabled deeper insights into spatial non-stationarity, the shifting strength and direction of relationships across space. These advances enhance both the accuracy and contextual relevance of spatial modeling, supporting localized decision-making in population health outcomes. Drawing on diverse spatial frameworks, this thesis developed, tested, and applied localized spatially varying GeoAI methodologies, offering integrated modeling approaches to address stunting among children in the complexities of such public health concern. Paper I develops localized spatially varying approaches to reveal significant intra-area variation in stunting prevalence and nonlinear relationships using cross-sectional socioeconomic and fine-scale remotely sensed climatic and agroecological data to better characterise household microenvironments. The approach provided a more detailed understanding of how local environments shape nutrition outcomes and demonstrating the importance of considering both scale and nonlinearity in stunting research. Building on this, Paper II implements a hybrid spatial machine learning (ML) framework to detect fine-scale heterogeneity in stunting prevalence, while also quantify localized disparities that national-level surveys overlook. The framework captured spatially heterogeneous outcomes across most areas, with predictors exhibiting regions-specific effects that vary according to different thresholds of influence. Paper III advances the analysis by implementing a hybrid spatially varying deep learning (DL) approach, which captured convoluted nonlinear influence of fine socio-economic determinants on child stunting outcomes. The algorithm fairly captured variability in stunting outcomes, highlighting key child, maternal, and household determinants whose contributions varied across space, though limitations in training data size constrained broader generalizability. Paper IV further refines this perspective by introducing a predictive multilevel spatial ensemble learning (SEL) framework to produce small area estimates (SAEs) of stunting risk by combining geomasked household data with agroecological and remote sensing (RS) indicators. This approach demonstrated the capacity of predictive models to generalize beyond sampled survey clusters and produce continuous prevalence surfaces at scales as fine as 1 km². Overall, these papers highlight that trade-offs between interpretability, generalizability, and spatial scale in these analytical and predictive models remain challenging to navigate and must be evaluated case by case according to research priorities. The methodologies presented in this thesis aim to generate fine-scale, interpretable risk estimates that can support targeted nutrition interventions in data-scarce settings.}},
author = {{Nduwayezu, Gilbert}},
isbn = {{978-91-89187-65-8}},
keywords = {{Population health; childhood stunting; multi-source data; spatially varying GeoAI; small area estimates; interpretability; localized interventions}},
language = {{eng}},
publisher = {{Lund University}},
school = {{Lund University}},
title = {{Demystifying the localized relationships between population health outcomes and multi-source determinants: A spatially varying GeoAI framework}},
url = {{https://lup.lub.lu.se/search/files/232727881/e-nailing_ex_Gilbert.pdf}},
year = {{2025}},
}