Risk prediction
Harnessing the power of artificial intelligence: predicting type 2 diabetes onset
PhD research showing that combining retinal images, ECG features and clinical risk factors improves prediction of type 2 diabetes in Qatar, and introducing new open-source image methods.
Project leadDr. Farida Mohsen
- 8
- Publications
- 0
- Talks
- 4
- Team members
- 1
- Partner
Summary
This research taught AI to combine eye photographs, heart-tracing (ECG) data and routine clinical information to better predict who will develop type 2 diabetes.
This doctoral research began by mapping the field: scoping reviews of AI for diabetes risk prediction and of methods that fuse health records with medical images.
It then built multimodal deep-learning models on longitudinal Qatar Biobank cohorts. ECG-DiaNet combined ECG-derived features with clinical risk factors and improved prediction over standard risk scores. The research also introduced RadEx, a nonlinear Radon-based image transform that captures subtle retinal patterns, released as open-source software, and used it to grade diabetic retinopathy.
Key findings
- AUROC for ECG-DiaNet, against 0.822 for standard risk scores (p < 0.001) (Scientific Reports, 2025)
- 0.845
- Quadratic kappa for five-class diabetic retinopathy grading with RadFuse
- 0.932
- Open-source Python package for nonlinear Radon transformation
- RadEx
OverviewWhy it matters · Research question and aims · Impact
Why it matters
- Standard risk scores use a few clinical factors and miss many people at risk.
- Eye and heart measurements carry early signs of changes in blood vessels that come with diabetes.
Research question and aims
Does combining retinal images, ECG features and clinical risk factors improve prediction of type 2 diabetes onset in a Middle Eastern population?
- Review AI methods for diabetes risk prediction and for fusing health records with imaging.
- Develop multimodal deep-learning models for incident type 2 diabetes in Qatar Biobank cohorts.
- Create new image-feature methods for retinal images.
- Release methods as reusable open-source software.
Impact
Who benefits and how: patients, clinicians, health services, policy, training (e.g. "Model shared with HMC diabetes clinic")
MethodsDesign, data and methods · Timeline and milestones · Ethics and data protection
Design, data and methods
- Study design
- Retrospective longitudinal analysis of cohort data
- Setting
- Qatar Biobank
- Participants or sample
- Who is included and how many
Data
- Qatar Biobank · retinal fundus images, ECG and clinical risk factors
Analytics and AI methods
How the study works
- Scoping and systematic reviews to set the research agenda.
- Multimodal fusion of retinal fundus images, ECG-derived features and clinical data.
- Nonlinear Radon transformation (RadEx) and sinogram fusion (RadFuse) for retinal analysis.
Timeline and milestones
- 2022–2023Reviews of AI for diabetes prediction and of EHR–imaging fusion (npj Digital Medicine).
- 2024RadEx and retinopathy grading papers; PhD completed in December.
- 2025ECG-DiaNet and RadEx software papers published.
Ethics and data protection
Analysis of de-identified Qatar Biobank data under an approved data-access project.
- Approving body
- e.g. HBKU IRB, HMC MRC
- Approval number
- Approval or reference number
- Data protection
- How data are stored and protected
Team & partnersTeam · Partners and collaborators
Team
Dr. Zubair Shah
PhD supervisor
Dr. Hazrat Ali
Co-author
Lecturer in Artificial Intelligence and Data Science, University of Stirling, UK
Dr. Md Rafiul Biswas
Co-author
Postdoctoral Researcher, College of Humanities and Social Sciences, HBKU
Partners and collaborators
OutputsOutputs
Outputs
Publications 8
-
ECG features improve multimodal deep learning prediction of incident T2DM in a Middle Eastern cohort
Mohsen F, Safa A, Shah Z
Scientific Reports 15: 27164, 2025
Journal articleRead the paper Free version
-
Integrating non-linear Radon transformation for diabetic retinopathy grading
Mohsen F, Belhaouari S, Shah Z
Scientific Reports 15: 30706, 2025
Journal articleRead the paper Free version
-
RadEx: an open source Python package for nonlinear Radon transformation
Mohsen F, Islam A, Shah Z, Belhaouari SB
Software Impacts: 100779, 2025
Journal articleRead the paper
-
Introducing RadEx: adaptive parameterized feature extraction from medical images
Islam A, Mohsen F, Shah Z, Belhaouari SB
Computer Graphics International Conference (CGI 2024): 278–294, Springer, 2024
Conference paperRead the paper Conference talk
-
Mohsen F, Shah U, Islam A, Shah Z, Belhaouari SB
IEEE 26th International Workshop on Multimedia Signal Processing (MMSP 2024): 1–5, 2024
Conference paperRead the paper Conference talk
-
A scoping review of artificial intelligence-based methods for diabetes risk prediction
Mohsen F, Al-Absi HRH, Yousri NA, El Hajj N, Shah Z
npj Digital Medicine 6: 197, 2023
Journal articleRead the paper Free version
-
Artificial intelligence-based methods for fusion of electronic health records and imaging data
Mohsen F, Ali H, El Hajj N, Shah Z
Scientific Reports 12: 17981, 2022
Journal articleRead the paper Free version
-
Customized and automated machine learning-based models for diabetes type 2 classification
Mohsen F, Biswas MR, Ali H, Alam T, Househ M, Shah Z
Advances in Informatics, Management and Technology in Healthcare: 517–520, IOS Press, 2022
Short paperRead the paper
Show these in the publications list
Software, datasets and tools
Theses
Keywords
Get involved
Use RadEx or collaborate
RadEx is open-source. Contact us to collaborate on multimodal diabetes prediction.