INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue VIII August 2025
Page 302
www.rsisinternational.org
A Machine Learning Model for Predicting the Risk of Developing
Diabetes - T2DM Using Real-World Data from Kilifi, Kenya
Isaac Mumo Kailu*, Dr. Mvurya Mgala and Dr. Fullgence Mwakondo
Institute of Computing and Informatics, Technical University of Mombasa, Kenya
*Corresponding Author
DOI: https://doi.org/10.51244/IJRSI.2025.120800026
Received: 22 July 2025; Accepted: 28 July 2025; Published: 29 August 2025
ABSTRACT
Type 2 Diabetes Mellitus (T2DM) is a growing public health concern in low-resource settings, where early
detection remains limited due to infrastructural and diagnostic constraints. This study presents a machine
learning-based risk prediction model developed using real-world data from Kilifi County Referral Hospital in
Kenya, aiming to identify individuals at risk of developing T2DM before clinical onset. The study applied the
CRISP-DM framework to guide the end-to-end process, from data collection to model deployment. A dataset
comprising 2,500 anonymized electronic health records was used, incorporating a diverse range of features
including clinical, behavioral, demographic, and socioeconomic variables. Feature selection was conducted
using both statistical (Chi-square test) and algorithm-based methods (Random Forest, Recursive Feature
Elimination, and XGBoost importance), resulting in two candidate feature sets (14-feature and 7-feature
subsets). Four supervised learning algorithms; Logistic Regression, Support Vector Machine (SVM), Random
Forest, and XGBoost were trained and evaluated using 5-fold cross-validation. Among them, the XGBoost
model achieved the best performance, with a test set accuracy of 91.33%, F1-score of 88.66%, and an AUC-
ROC of 96.24%, outperforming other models across all metrics. This study demonstrates that integrating multi-
domain features with machine learning can enhance early risk stratification for T2DM in under-resourced
environments. The final model’s ability to categorize individuals into low, medium, and high-risk groups offers
a practical tool for targeted screening and preventive healthcare interventions in Kenyan public health systems.
Keywords: Type 2 Diabetes Mellitus, Machine Learning, Risk Prediction, XGBoost, CRISP-DM, Kenya
INTRODUCTION
Type 2 Diabetes Mellitus (T2DM) is a chronic metabolic disorder characterized by insulin resistance and
sustained hyperglycemia, with complications affecting cardiovascular, renal, and neurological systems.
Globally, over 537 million adults were estimated to be living with diabetes in 2021, and this figure is projected
to exceed 780 million by 2045. While T2DM is rising worldwide, low- and middle-income countries (LMICs)
face a disproportionate burden due to limited healthcare infrastructure, late-stage diagnoses, and reduced access
to screening and preventive care.
In Kenya, and particularly in underserved counties such as Kilifi, many cases of T2DM remain undetected until
advanced complications develop, leading to increased morbidity and healthcare costs. Standard diagnostic
methods including fasting glucose, HbA1c, and OGTT are often reactive and inaccessible to rural populations.
There is an urgent need for cost-effective, scalable, and proactive approaches to identify individuals at risk before
the onset of symptoms.
Machine Learning (ML) has emerged as a powerful tool in healthcare analytics, capable of uncovering hidden
patterns within heterogeneous data. ML models can leverage clinical, behavioral, demographic, and
socioeconomic variables to support early disease prediction and personalized interventions. The vast majority
of existing ML models for T2DM prediction are built on datasets from high-income countries and focus
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue VIII August 2025
Page 303
www.rsisinternational.org
primarily on clinical features. These models often fail to generalize to LMIC populations due to contextual,
genetic, and environmental differences.
This paper addresses this gap by presenting a context-aware ML model for early T2DM risk prediction,
developed using real-world hospital data from Kilifi County, Kenya. Unlike prior models, this approach
integrates not only clinical indicators but also behavioral and socioeconomic factors that are highly relevant in
under-resourced settings. The final model, based on the XGBoost algorithm, demonstrates high predictive
performance and supports risk categorization into actionable levels (low, medium, high), enabling its potential
use in community health screenings and decision support tools.
LITERATURE REVIEW
Global Models for T2DM Risk Prediction
Machine learning (ML) has gained significant traction globally in predicting the risk of Type 2 Diabetes Mellitus
(T2DM). Studies from the United States, China, Israel, Sweden, and Brazil have leveraged electronic health
records (EHRs), behavioral datasets, and national surveys to build robust prediction models.
In the U.S., Smith et al. (2021) employed Random Forest (RF) on national health data, achieving an AUC-ROC
of 0.82. Similarly, Shin et al. (2020) used nine ML models on CDC datasets, with RF outperforming others
(AUC = 0.884), especially when behavioral and psychosocial variables were included. Wang et al. (2019)
compared RF and SVM on hospital records, noting SVM’s strength in identifying high-risk groups (AUC =
0.79).
In China, Li et al. (2020) applied SVM to clinical datasets and achieved an AUC of 0.79 using non-invasive
features like BMI and blood pressure. Zhang et al. (2020) used a Random Forest–based filter method, improving
performance by incorporating income, occupation, and urbanization.
A large-scale Israeli study by Hochman et al. (2021) used XGBoost on over 214,000 EHRs, reporting a strong
specificity (0.905) but low sensitivity (0.349), indicating its limitation in detecting early-stage T2DM. In
Sweden, Andersson et al. (2021) used Extremely Randomized Trees with clinical-demographic data, achieving
an AUC of 0.81. A Brazilian study by de Souza et al. (2020) employed Gradient Boosting Machines, balancing
accuracy (0.85 AUC) with generalizability by integrating lifestyle data.
These studies collectively underscore that performance improves with multi-domain features; however, most
models are limited by a focus on high-income, urban datasets and lack contextual generalizability to LMICs.
Local Models in Kenya
In Kenya, ML applications in T2DM prediction remain limited and often lack integration of behavioral and
socioeconomic indicators.
Okeyo et al. (2019) applied RF and Decision Trees on KDHS data, reporting an AUC-ROC of 0.81. However,
the study excluded lifestyle and income variables, limiting practical applicability.
Mwangi et al. (2020) used SVM on community screening data in rural areas, achieving 85% accuracy but relying
heavily on clinical features like BMI and blood pressure.
Kamau et al. (2021) trained an Artificial Neural Network on urban hospital data, attaining an AUC-ROC of 0.92.
Yet, the model emphasized classification of current diabetes rather than future risk, and the urban bias reduces
generalizability to rural settings.
Mutuku et al. (2022) employed logistic regression in Nairobi-based clinics, identifying obesity and hypertension
as predictors but omitting behavioral or social data. Similarly, Otieno and Ndirangu (2021) used Gradient
Boosting on private hospital records, showing good accuracy (AUC = 0.87), yet lacked non-clinical variables
crucial for preventive care models.
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue VIII August 2025
Page 304
www.rsisinternational.org
These studies highlight the need for locally trained, predictive not diagnostic models using multi-dimensional
data. The current study fills this gap by using real-world hospital data from a rural Kenyan setting and
incorporating broader feature categories.
Comparative Summary of Selected Studies
Table 1. 1: A Comparative Summary table of Selected Studies
Study
Country
Dataset
ML Method
AUC /
Accuracy
Limitation
Smith et al.
(2021)
USA
National EHR
Random
Forest
AUC =
0.82
Urban bias, lacks
local context
Shin et al.
(2020)
USA
CDC dataset
RF, SVM
AUC =
0.884
High complexity
Hochman et
al. (2021)
Israel
EHR (214,000
patients)
XGBoost
AUC =
0.712, Spec
= 0.905
Low sensitivity
for early
detection
Zhang et al.
(2020)
China
Clinical +
socioeconomic
Random
Forest
AUC =
0.78
Excludes
behavioral data
Andersson
et al. (2021)
Sweden
University
hospital records
Extremely
Randomized
Trees
AUC =
0.81
Population
homogeneity
Okeyo et al.
(2019)
Kenya
KDHS
RF, Decision
Tree
AUC =
0.81
No behavioral or
socioeconomic
variables
Mwangi et
al. (2020)
Kenya
Rural screening
data
SVM
Accuracy =
85%
Excludes social
context
Kamau et
al. (2021)
Kenya
Urban hospital
data
ANN
AUC =
0.92
Urban bias,
diagnosis-
focused
Otieno &
Ndirangu
(2021)
Kenya
Private hospital
EMR
Gradient
Boosting
AUC =
0.87
No behavioral or
economic
features
This paper builds upon these foundations by incorporating a multi-domain dataset from a rural Kenyan public
hospital, using explainable ML models like XGBoost, and emphasizing early risk prediction with practical utility
for frontline health systems.
METHODOLOGY
Overview of Methodology
This study employed the CRISP-DM (Cross-Industry Standard Process for Data Mining) framework to guide
the design and implementation of a predictive machine learning (ML) model for assessing the risk of developing
Type 2 Diabetes Mellitus (T2DM). The CRISP-DM approach provides a structured, iterative process consisting
of six phases: Business Understanding, Data Understanding, Data Preparation, Modeling, Evaluation, and
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue VIII August 2025
Page 305
www.rsisinternational.org
Deployment. This framework was selected for its flexibility, repeatability, and suitability for healthcare data
mining.
Figure 1. 1: CRISP-DM Data mining model (Plotnikova et al., 2020)
Dataset Description
The dataset used in this study was collected from Kilifi County Referral Hospital (KCRH) in coastal Kenya. It
consists of 2,500 anonymized patient records extracted from hospital archives and supplemented by field data
collected with support from Community Health Volunteers (CHVs).
Each record includes a variety of features spanning multiple domains:
Clinical: BMI, blood pressure, fasting glucose, family history of diabetes
Behavioral: smoking status, alcohol use, physical activity
Demographic: age, gender, marital status
Socioeconomic: income level, education, occupation
Only adult patients with no previous diagnosis of T2DM were included. Records with critical missing data were
excluded or imputed, and all personally identifiable information (PII) was removed.
Feature Selection
To ensure model accuracy and efficiency, feature selection was conducted using both statistical and algorithmic
methods:
Chi-Square Test
Algorithm-Based Selection
Three machine learning-based feature importance methods were employed:
Random Forest Importance: Assessed Gini-based impurity reduction.
Recursive Feature Elimination (RFE): Performed backward elimination using SVM and Logistic
Regression as base estimators.
XGBoost Feature Importance: Ranked features based on gain and split metrics.
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue VIII August 2025
Page 306
www.rsisinternational.org
This two-step process produced two optimized feature subsets:
14-feature set (statistical selection)
7-feature subset (algorithm-based refinement)
Machine Learning Models
Four supervised classification algorithms were implemented to model the risk of T2DM:
Logistic Regression (LR): Baseline model for binary classification with strong interpretability.
Support Vector Machine (SVM): Effective in handling high-dimensional data.
Random Forest (RF): Robust ensemble method with built-in feature selection.
XGBoost: Gradient boosting framework known for superior accuracy and handling of imbalanced data.
Each model was trained on both feature sets (14 and 7 features) and evaluated under identical conditions.
Model Evaluation
Model performance was evaluated using 5-fold cross-validation and standard classification metrics:
Accuracy: Overall correct predictions
F1-score: Harmonic mean of precision and recall, useful for imbalanced data
ROC-AUC: Area Under the Receiver Operating Characteristic Curve, measuring overall discriminative
ability
The final model was selected based on test set performance, and additional analysis was conducted using a
confusion matrix and ROC curves to assess clinical utility and predictive reliability.
RESULTS
Overview
This section presents the outcomes of model training, performance comparisons across different machine
learning (ML) algorithms, and evaluation on both 7-feature and 14-feature datasets. The objective was to identify
the most effective model for predicting the risk of Type 2 Diabetes Mellitus (T2DM) using diverse patient data.
Model Performance: 7 vs. 14 Features
Each ML model was trained and validated using 5-fold cross-validation. The results showed a consistent trend
across algorithms: models trained on the 14-feature dataset outperformed those trained on the 7-feature subset,
indicating the value of retaining a broader range of features, especially behavioral and socioeconomic variables.
Table 1. 2: Model Performance: 7 vs. 14 Features
Model
Features Used
Accuracy (%)
F1 Score (%)
AUC-ROC (%)
Logistic Regression
7
82.67
80.21
88.47
Logistic Regression
14
86.00
84.15
90.22
SVM
7
84.67
82.33
89.11
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue VIII August 2025
Page 307
www.rsisinternational.org
SVM
14
88.33
85.74
92.45
Random Forest
7
85.33
83.60
91.07
Random Forest
14
89.00
86.80
94.16
XGBoost
7
87.33
85.01
89.47
XGBoost
14
91.33
88.66
96.24
Key Insight: The XGBoost model trained on the 14-feature dataset achieved the highest overall performance
with an accuracy of 91.33%, F1-score of 88.66%, and AUC-ROC of 96.24%.
Confusion Matrix and ROC Curve
To evaluate real-world clinical applicability, a confusion matrix and Receiver Operating Characteristic (ROC)
curve were generated for the best-performing XGBoost model.
Table 1. 3: Confusion Matrix (XGBoost, 14 Features)
Predicted Positive
Predicted Negative
Actual Positive
264
16
Actual Negative
21
449
Figure 1. 2: XGBoost Confusion Matrix Testset
Figure 1. 3:XGBoost ROC Curve Testset
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue VIII August 2025
Page 308
www.rsisinternational.org
The ROC curve (AUC = 0.9624) for XGBoost demonstrated near-optimal separability between classes,
confirming the model’s discriminative strength across various thresholds.
Risk Categorization Outputs
To enhance interpretability for clinical and public health deployment, the XGBoost model included a probability-
based risk stratification mechanism:
Low Risk: ≤ 0.33
Medium Risk: 0.34 – 0.66
High Risk: ≥ 0.67
This classification allows frontline health workers to quickly interpret and act upon predictions. For instance,
individuals with a probability score of 0.74 were flagged as high risk, prompting further clinical testing and
lifestyle counseling.
DISCUSSION
Interpretation of Results
The performance evaluation revealed that the XGBoost model trained on the 14-feature dataset significantly
outperformed all other models in terms of accuracy (91.33%), F1-score (88.66%), and AUC-ROC (96.24%).
XGBoost’s superior performance can be attributed to its ability to handle non-linear feature interactions, manage
class imbalance effectively, and optimize learning via regularized gradient boosting. Unlike traditional models
such as logistic regression or SVM, XGBoost can assign differential weights to features and iteratively minimize
classification errors, making it especially well-suited for heterogeneous healthcare datasets.
Additionally, its built-in feature importance metrics and robust cross-validation techniques contributed to stable
and generalizable outcomes across test folds.
Value of Expanded Feature Set
One of the study’s key findings is the demonstrable performance gain when including behavioral and
socioeconomic features alongside clinical indicators. The 14-feature dataset included variables such as alcohol
use, physical activity, income level, and marital status, which are often excluded from conventional models but
showed strong predictive value in this context.
The inclusion of these features allowed the model to capture social determinants of health factors increasingly
recognized as critical in chronic disease progression, especially in LMICs. These findings support the growing
consensus that T2DM risk is not solely clinical but also behavioral and environmental, and thus, predictive
models must reflect this complexity.
Relevance to Local Context
This study uniquely contributes a localized and context-sensitive solution for early T2DM risk prediction in
Kenya. Unlike most published models developed in urban, high-income contexts, this model was trained on data
from Kilifi County, a rural and resource-constrained region. The integration of locally relevant features and the
involvement of Community Health Volunteers (CHVs) in data validation further enhance the cultural and
operational relevance of the model.
In public health terms, such a tool could support community-based screening programs, allowing health workers
to proactively identify high-risk individuals and intervene earlier, well before the onset of clinical symptoms.
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue VIII August 2025
Page 309
www.rsisinternational.org
Generalizability and Clinical Utility
Although the model was developed in a single county, the use of diverse patient-level data across clinical,
demographic, and socioeconomic domains improves its adaptability to similar low-resource environments in
sub-Saharan Africa. The risk categorization framework (low, medium, high) adds practical value, supporting
integration into mobile health platforms, decision support tools, and community triage systems.
The use of explainable ML models like XGBoost also supports clinical acceptance, especially when feature
importance and predictive thresholds are clearly communicated to healthcare providers.
Limitations
Several limitations should be noted:
Single-Center Dataset: The model was trained on data from one public hospital, which may limit
external validity across diverse Kenyan regions with varying epidemiological and infrastructural profiles.
Missing Biomarkers: Important predictors such as HbA1c levels, genetic markers, long-term dietary
habits, and psychosocial stress were not available in the dataset due to local diagnostic constraints.
Retrospective Design: The use of retrospective data, while useful for proof-of-concept, may not fully
reflect real-time clinical variability.
No Real-World Deployment Yet: Although a prototype scoring tool was developed, full-scale clinical
deployment and prospective validation remain future work.
Despite these limitations, the study establishes a strong methodological and contextual foundation for future
expansion, scaling, and real-world application.
CONCLUSION AND FUTURE WORK
Summary of Findings
This study developed and evaluated a machine learning–based model to predict the risk of developing Type 2
Diabetes Mellitus (T2DM) using real-world hospital data from Kilifi County, Kenya. By applying the CRISP-
DM framework, a structured pipeline was implemented encompassing data preprocessing, feature selection,
model training, and evaluation. Among the four algorithms tested, the XGBoost model trained on a 14-feature
dataset outperformed others, achieving an accuracy of 91.33%, F1-score of 88.66%, and AUC-ROC of 96.24%.
The study demonstrates that including behavioral and socioeconomic indicators in addition to clinical features
substantially improves prediction accuracy and contextual relevance. Risk categorization into low, medium, and
high tiers further enhances the model’s usability in community health screening and preventive decision-making.
Clinical Potential
The model offers promising potential for integration into frontline healthcare workflows, particularly in under-
resourced settings where traditional diagnostics are unavailable or delayed. By using accessible and non-invasive
variables, the tool can assist community health workers, clinicians, and policymakers in identifying high-risk
individuals early, enabling timely interventions and reducing the long-term burden of diabetes-related
complications.
Its implementation aligns with Kenya’s national strategy for addressing non-communicable diseases through
digital health innovation and data-driven public health systems.
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue VIII August 2025
Page 310
www.rsisinternational.org
Future Work
To extend the impact of this research, several areas of improvement and expansion are proposed:
Multi-Center Validation: Expanding the dataset across other counties and hospitals in Kenya will help
validate the model’s generalizability and robustness across diverse populations.
Real-Time Deployment: The model will be integrated into a web-based or mobile platform to support
on-the-ground screening by health professionals and CHVs.
Model Explainability: To improve transparency and clinical trust, future versions will incorporate
explainable AI techniques, such as SHAP (SHapley Additive Explanations) and LIME (Local
Interpretable Model-Agnostic Explanations), to visualize individual predictions and the contribution of
specific features.
Incorporation of Additional Biomarkers: As local diagnostic capacity improves, future models can
include biomarkers like HbA1c, insulin resistance indices, and genomic data to further refine predictions.
REFERENCES
1. Bhargava, S., & Zafar, S. (2019). Socioeconomic and behavioral predictors in diabetes risk: An ML-
based population health study in Pakistan. Journal of Public Health Research, 8(3), 164–170.
https://doi.org/10.4081/jphr.2019.164
2. Chen, H., et al. (2021). Using real-world data from rural China to predict diabetes risk via ensemble
learning. BMC Endocrine Disorders, 21, 198. https://doi.org/10.1186/s12902-021-00870-w
3. Deberneh, H. M., & Kim, I. (2021). Prediction of type 2 diabetes based on machine learning algorithm.
International Journal of Environmental Research and Public Health, 18(6), 3317.
https://doi.org/10.3390/ijerph18063317
4. Farran, B., et al. (2022). An explainable machine learning approach to early T2DM prediction in Qatar’s
primary care. BMC Medical Informatics and Decision Making, 22, 183. https://doi.org/10.1186/s12911-
022-01948-6
International Diabetes Federation. (2021). IDF Diabetes Atlas (10th ed.). Brussels, Belgium:
International Diabetes Federation. https://diabetesatlas.org/
5. Islam, S. M. S., et al. (2022). Development of a non-invasive diabetes prediction tool using behavioral
and anthropometric data in rural Bangladesh. Scientific Reports, 12, 14378.
https://doi.org/10.1038/s41598-022-18022-6
6. Kavakiotis, I., Tsave, O., Salifoglou, A., Maglaveras, N., Vlahavas, I., & Chouvarda, I. (2017). Machine
learning and data mining methods in diabetes research. Computational and Structural Biotechnology
Journal, 15, 104–116. https://doi.org/10.1016/j.csbj.2016.12.005
7. Lee, S., et al. (2021). Diabetes risk classification with explainable ML: Application in underserved
Korean population. PLoS ONE, 16(6), e0253312. https://doi.org/10.1371/journal.pone.0253312
8. Mohan, V., et al. (2019). A deep learning model for diabetes prediction using Indian rural cohort.
Diabetes Technology & Therapeutics, 21(10), 562–569. https://doi.org/10.1089/dia.2019.0172
9. Nguyen, Q. C., et al. (2020). Leveraging social determinants and EHR data to predict diabetes risk in
underserved populations. International Journal of Medical Informatics, 141, 104241.
https://doi.org/10.1016/j.ijmedinf.2020.104241
10. Rahman, M. M., et al. (2020). T2DM risk assessment using random forest and decision tree in community
health datasets. Informatics in Medicine Unlocked, 21, 100461.
https://doi.org/10.1016/j.imu.2020.100461
11. Wang, F., & Hu, J. (2019). Predicting chronic disease risk using machine learning on health survey data:
A case study on diabetes. IEEE Journal of Biomedical and Health Informatics, 23(6), 2548–2556.
https://doi.org/10.1109/JBHI.2018.2887383