- Home
- Projects
- Longitudinal analysis of risk factors associated with diabetes using multimodal data and machine learning
Longitudinal analysis of risk factors associated with diabetes using multimodal data and machine learning
PhD research that fused bone scans, retinal images, ECG, genetic, laboratory and lifestyle data from Qatar Biobank to predict who will develop type 2 diabetes, and to study its complications.
Part of DAQIQ · دقيق: Diabetes Analytics for Quality of life through Individualised care in Qatar.
- Status
- Completed
- Duration
- 2022–2026
- Funding
- PhD research, College of Science and Engineering, HBKU
- Team
- 3 people
- Publications
- 5
In plain language
Many people develop type 2 diabetes years after their first health check. This research used data that people in Qatar had already given in routine health checks, such as bone scans and eye photographs, to predict who is likely to develop diabetes years ahead.
Why it matters
- Diabetes can already be diagnosed cheaply with blood tests such as HbA1c. The real challenge is knowing who will develop it next, so prevention can start early.
- Most AI studies look at one type of data at one point in time. This work followed the same people over time and combined many kinds of data.
Overview
Using longitudinal data from Qatar Biobank, this doctoral research linked each participant’s baseline health check with their status at follow-up, and trained AI models to predict future type 2 diabetes from the baseline data alone.
Separate studies examined DXA bone measurements, retinal (ophthalmoscopic) images, and respiratory and lung-function data, each fused with clinical records. The models were interpreted to find risk factors specific to the Qatari population, and to probe complication pathways such as respiratory health.
Research question
Can routine multimodal health-check data from Qatar Biobank predict who will develop type 2 diabetes years later, and which factors drive that risk?
Objectives
- Predict incident type 2 diabetes in Qatari adults from longitudinal multimodal data.
- Combine imaging, bone, ECG, genetic, laboratory and lifestyle data and test which add predictive value.
- Compare deep learning, tabular transformers, classical machine learning and large language models on the same tasks.
- Identify and explain population-specific risk factors, including differences by age and sex.
Approach and data
- Longitudinal Qatar Biobank data, with baseline records labelled by diabetes status at follow-up.
- DXA study: 1,382 participants; feature engineering on 263 bone measurements and a tabular transformer fused with clinical data.
- Retinal study (DiabEye-Q): 2,043 participants; fractal features from retinal images combined with clinical data using XGBoost and a Vision Transformer.
- Interpretation of model predictions and analysis by age group and sex.
Key results
- AUROC for predicting future type 2 diabetes from longitudinal DXA bone scans, with accuracy above 91% (Scientific Reports, 2025)
- 0.96
- AUROC for DiabEye-Q, combining retinal-image features with clinical data, with 88% accuracy (AIiH 2025)
- 93%
- study to use longitudinal DXA bone scans to predict future diabetes in Qatar
- First
Timeline
- 2022PhD started; systematic review of genetic biomarkers and machine learning.
- 2025DXA, retinal and respiratory studies published.
- 2026PhD completed; Dr. Khan continues with DAQIQ as a Research Assistant.
Team
Partners
- Qatar Biobank (data access, project QF-QBB-RES-ACC-00181)
Publications 5
-
Artificial intelligence–based diabetes risk prediction from longitudinal DXA bone measurements
Khan S, Shah Z
Scientific Reports 15: 25706, 2025
Journal articleLongitudinal analysis of risk factors associated with diabetes using multimodal data and machine learningDr. Sulaiman KhanDr. Zubair Shah
-
Khan S, Biswas MR, Shah Z
Artificial Intelligence in Healthcare (AIiH 2025), Lecture Notes in Computer Science 16038, Springer, 2025
Conference paperTalk and slidesLongitudinal analysis of risk factors associated with diabetes using multimodal data and machine learningDr. Sulaiman KhanDr. Md Rafiul BiswasDr. Zubair Shah
-
Early prediction of type 2 diabetes using multimodal data and tabular transformers
Khan S, Biswas MR, Shah Z
3rd International Conference on Foundation and Large Language Models (FLLM 2025): 592–599, IEEE, 2025
Conference paperFree versionTalk and slidesLongitudinal analysis of risk factors associated with diabetes using multimodal data and machine learningDr. Sulaiman KhanDr. Md Rafiul BiswasDr. Zubair Shah
-
Khan S, Shah Z
IEEE International Conference on Information Reuse and Integration for Data Science (IRI 2025): 289–294, 2025
Conference paperTalk and slidesLongitudinal analysis of risk factors associated with diabetes using multimodal data and machine learningDr. Sulaiman KhanDr. Zubair Shah
-
Genetic biomarkers and machine learning techniques for predicting diabetes: systematic review
Khan S, Mohsen F, Shah Z
Artificial Intelligence Review 58(2): 41, 2024
Journal articleLongitudinal analysis of risk factors associated with diabetes using multimodal data and machine learningDr. Sulaiman KhanDr. Farida MohsenDr. Zubair Shah
Talks and media
Next steps
- Extend to wearable sensors, continuous glucose monitoring and electronic health records.
- Validate the models in other Gulf cohorts.
Ethics
Analysis of de-identified Qatar Biobank data under an approved data-access project.
Get involved
Work with us on diabetes prediction
We welcome collaborators with longitudinal cohorts or clinical data who want to validate or extend these models.