INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XI, November 2025  
Predictive Modelling of Health Risks Using Logistic Regression for  
Personalized Dietary Recommendations  
Chandana H M, Darshan S, Deeksha D K, G Jayakrishna Reddy, Dr Athi Narayanan  
Dept. of Computer Science and Engineering, DSATM, Bengaluru, India  
DOI : https://doi.org/10.51583/IJLTEMAS.2025.1411000110  
Received: 08 December 2025; Accepted: 15 December 2025; Published: 23 December 2025  
ABSTRACT  
The prevalence of diet-related chronic diseases such as Type 2 Diabetes Mellitus (T2DM) and hypertension  
necessitates the development of intelligent nutritional systems that go beyond user preference to prioritize  
clinical safety. Traditional food recommender systems often suffer from a "health-blind" bias, optimizing  
primarily for taste or popularity. This paper proposes a novel Risk-Aware Diet Recommendation System  
(RADRS) that integrates Logistic Regression (LR) for interpretable health risk assessment with Linear  
Programming (LP) for dietary optimization. By training LR models on the NHANES and Pima Indians  
Diabetes datasets, the system calculates individual disease probabilities. These probabilities dynamically  
configure nutritional constraints—specifically modifying sodium, carbohydrate, and saturated fat limits—  
within an LP solver. The proposed architecture bridges the gap between predictive health analytics and  
personalized meal planning, ensuring that recommendations are both palatable and medically compliant.  
Keywords—Health Informatics, Logistic Regression, Recommender Systems, Linear Programming,  
Personalized Nutrition, Chronic Disease Management.  
INTRODUCTION  
The global burden of non-communicable diseases (NCDs) is largely driven by modifiable behavioral risk  
factors, with unhealthy diet being a primary contributor. Despite the proliferation of health applications, a  
significant gap remains between clinical nutritional guidelines (e.g., DASH, ADA) and the daily food choices of  
individuals. Existing food recommendation systems typically rely on Collaborative Filtering (CF) or Content-  
Based Filtering (CBF), which prioritize user satisfaction and serendipity but often neglect physiological  
constraints.  
Recent advancements in "health-aware" recommender systems have attempted to incorporate nutritional data.  
However, many utilize "black-box" deep learning models that lack the interpretability required for clinical  
trust. Furthermore, risk is often treated as a binary attribute (diagnosed vs. healthy), failing to capture pre-  
clinical states where dietary intervention is most effective.  
This paper introduces a Risk-Aware Diet Recommendation System (RADRS) that employs Logistic Regression  
(LR) to predict the probability of specific metabolic conditions. Unlike complex ensemble methods, LR  
provides interpretable Odds Ratios (OR), allowing the system to justify restrictions (e.g., "Sodium is limited  
because your calculated hypertension risk is in the 85th percentile"). These risk probabilities are mathematically  
mapped to specific nutrient constraints in a Linear Programming optimization model, ensuring that the  
generated meal plans strictly adhere to medical standards such as the Dietary Approaches to Stop Hypertension  
(DASH) or Therapeutic Lifestyle Changes (TLC) diets.  
LITERATURE REVIEW  
Recommender Systems in Health  
Early diet systems utilized Knowledge-Based (KB) approaches, relying on static rule sets (ontologies) to filter  
foods. While safe, these systems often suffered from the "cold start" problem and low user engagement due to  
www.ijltemas.in  
Page 1160  
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XI, November 2025  
repetitive suggestions. Modern approaches have shifted toward hybrid systems. For instance, the FKGM model  
uses Knowledge Graph Convolutional Networks to embed health information into the recommendation process,  
yet it optimizes primarily for semantic relevance rather than strict nutrient bounding.  
Predictive Modelling in Healthcare  
Machine Learning (ML) is widely used for disease prediction. Studies using the NHANES and Pima Indians  
Diabetes datasets have benchmarked various algorithms. While Random Forests (RF) and Gradient Boosting  
often achieve marginally higher accuracy. Regression remains the gold standard in clinical settings due to its  
transparency and the direct mapping of feature weights to risk magnitude. Deep learning models, while  
powerful, often obscure the specific drivers of risk, making them less suitable for explaining why a specific  
dietary constraint (e.g., low saturated fat) was triggered.  
Dietary optimization  
The "Diet Problem," first formulated by Stigler, utilizes Linear Programming (LP) to find a low-cost diet  
meeting nutritional. Modern variations use LP to maximize Healthy Eating Index (HEI) scores or user  
preference curves subject to constraints. However, few systems dynamically adjust these constraints based on  
real-time ML risk predictions.  
PROPOSED SYSTEM ARCHITECTURE  
The RADRS architecture consists of four sequential processing modules: Data Acquisition, Risk Assessment,  
Constraint Generation, and Diet Optimization.  
Data Acquisition and Preprocessing  
The system accepts user demographic (age, gender) and anthropometric (BMI, waist circumference) data. For  
model training, we utilize the National Health and Nutrition Examination Survey (NHANES) dataset, widely  
regarded for its combination of interview and physical examination data.  
Preprocessing: Missing values in biochemical features (e.g., fasting glucose) are handled via k-Nearest  
Neighbors (KNN) imputation to preserve data structure.  
Balancing: To address class imbalance (e.g., fewer diabetic vs. non-diabetic cases), we apply  
the Synthetic Minority Over-sampling Technique (SMOTE), ensuring the model remains sensitive to at-  
risk minority classes.  
Predictive Risk Engine (Logistic Regression)  
We employ binary Logistic Regression to estimate the probability P(Y=1|X) of a user developing condition Y  
(e.g., T2DM), given feature vector X. The probability is given by the sigmoid function:  
Where  
i represents the coefficient for feature xi. The model outputs a probability score (0 ≤ P ≤ 1) for  
multiple conditions: Pdiabetes, Phypertension, and PCVD  
.
Interpretability: The coefficients provide the log-odds. For instance, a high positive  
for BMI indicates that  
weight management is a critical lever for risk reduction in that specific user, prioritizing caloric deficit  
constraints.  
www.ijltemas.in  
Page 1161  
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XI, November 2025  
Dynamic Constraint Generation  
This module translates probabilistic risk scores into deterministic mathematical constraints (C) for the  
optimization engine. Let  
Hypertension Logic:  
be the clinical risk threshold (e.g., 0.5).  
If Phyper  
>
Sodiummax = 1500 mg/day  
Potassiummin= 4700 mg/day  
Else:  
Sodiummax=2300 mg/day  
Diabetics Logic:  
If Pdiabetes  
>
Carbohydratesmax = 45% of total calories  
Fibermin= 30g/da  
Glycemic Loadmax=100  
General Health Logic:  
All plans must meet the Healthy Eating Index-2020 (HEI-2020) standards for micronutrient density.  
Diet Optimization (Linear programming)  
We formulate the meal planning task as a Constraint Satisfaction Problem (CSP) solvable via Linear  
Programming.  
Objective Function: Minimize the deviation from user preference (Cost Z).  
Where xj is the quantity of food item j, and pj is the normalized user preference score (0-1) for item j.  
Subject to:  
Where nij is the content of nutrient i in food j, and Li, Ui are the bounds derived from Module 3.  
www.ijltemas.in  
Page 1162  
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XI, November 2025  
Figure 1 System Architecture of the Personalized Dietary Recommendation Pipeline.  
EXPERIMENTAL METHODOLOGY  
Dataset Description  
Training Data:  
Pima Indians Diabetes Dataset: Used for initial diabetes model validation (n=768). Features: Glucose,  
BP, BMI, Age.  
Framingham Heart Study: Used for cardiovascular risk modeling. Features: Systolic BP, Cholesterol,  
Smoking status.  
Food Database: USDA FoodData Central database serves as the source for nutrient composition (Foundation  
Foods), mapped to the FoodOn ontology to categorize ingredients.  
Performance Metrics  
Predictive Accuracy: Evaluated using Area Under the Curve (AUC-ROC) and Recall (Sensitivity). High  
sensitivity is prioritized to minimize False Negatives (missing an at-risk user).  
Nutritional Adherence: Measured by the Constraint Satisfaction Rate (CSR)—the percentage of generated  
meal plans that successfully meet all defined safety constraints (e.g., Sodium < 1500mg).  
Diet Quality: The generated plans are scored using the HEI-2020 algorithm (score 0–100), assessing alignment  
with Dietary Guidelines for Americans.  
www.ijltemas.in  
Page 1163  
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XI, November 2025  
RESULTS AND DISCUSSIONS  
Risk Model Performance  
Preliminary validation on the Pima dataset indicates that Logistic Regression achieves an Accuracy of ~77.6%  
with an AUC of 0.83 after SMOTE balancing.7 While Random Forest classifiers achieved slightly higher  
accuracy (79%), the LR model's ability to output interpretable Odds Ratios (e.g., "Age increases risk by factor  
1.04 per year") is critical for the explanation interface of the recommendation system.  
Table 1 Performance comparison of classifiers  
Metric  
Accuracy  
Logistic Regression  
Random Forest  
79%  
77.6%  
0.83  
AUC-ROC  
Recall  
0.86  
0.81  
0.77  
Low  
0.88  
Precision  
0.74  
Interpretability  
High  
Optimization Efficacy  
Simulations using the PuLP optimization library in Python demonstrate a 100% Constraint Satisfaction Rate  
for feasible solution spaces. For users flagged as "High Risk" for hypertension (P > 0.7), the system  
successfully filtered out high-sodium processed foods, replacing them with potassium-rich alternatives (leafy  
greens, bananas) to satisfy the dual constraints of Sodium ≤ 1500mg and Potassium ≥ 4700mg.  
Table 2 Optimization Efficiency  
User Risk group  
Sodium Constraint Met  
Potassium Constraint Met  
Overall CSR  
(%)  
Low Risk  
100%  
100%  
100%  
100%  
98%  
100%  
99%  
Moderate Risk  
High Hypertension Risk  
(P>0.7)  
100%  
100%  
Comparison with Baseline  
Compared to a standard Collaborative Filtering baseline (User-KNN), the RADRS produced meal plans with  
significantly higher HEI-2020 scores (Mean: 85 vs. 62). While the baseline system recommended popular but  
nutrient-poor items (e.g., pizza, soda) based on peer similarity, the RADRS effectively penalized these items  
via the optimization constraints.  
www.ijltemas.in  
Page 1164  
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XI, November 2025  
Figure 2 Constraint Satisfaction Rate (CSR) across different user risk groups (Low, Moderate, High). The  
optimization engine consistently satisfies sodium, potassium, and overall nutrient constraints with CSR ≥ 0.99  
for all feasible solution spaces.  
CONCLUSION  
This paper presented the design of a Risk-Aware Diet Recommendation System that uniquely couples  
interpretable machine learning with mathematical optimization. By using Logistic Regression to quantify  
health risks and Linear Programming to solve the nutritional selection problem, the system ensures that dietary  
proposals are not just personalized to taste, but calibrated to physiological needs. Future work will focus on  
integrating Large Language Models (LLMs) to generate natural language explanations for the constraints and  
incorporating Feedback Loops to refine user preference vectors over time.  
REFERENCES  
1. M. Elsweiler, et al., “Bringing the ‘healthy’ into food recommenders,” CEUR Workshop Proceedings,  
2015.  
2. S. Tarima, “Logistic Regression: A Mathematical Approach,” CTSI Biostatistics, 2011.  
3. Y. Chen, et al., “Personalized Food Recommendation as Constrained Question Answering over a  
Large-scale Food Knowledge Graph,” in Proc. WSDM ’21, 2021.  
4. G. Adomavicius and A. Tuzhilin, “Toward the next generation of recommender systems: A survey,”  
IEEE Transactions on Knowledge and Data Engineering, 2005.  
5. L. Li, et al., “Health-Aware Food Recommendation Based on Knowledge Graph and Multi-Task  
Learning,” IEEE Access, 2023.  
6. A. S. Abdalrada, et al., “Logistic Regression Model for Predicting the Progression of Diabetes,” Iraqi  
Journal for Computer Science and Mathematics, 2024.  
7. Z. Xie, et al., “Building Risk Prediction Models for Type 2 Diabetes Using Machine Learning  
Techniques,” Preventing Chronic Disease, 2019.  
8. G. Stigler, “The Cost of Subsistence,” Journal of Farm Economics, 1945.  
9. Centers for Disease Control and Prevention (CDC), “National Health and Nutrition Examination  
Survey Data,” National Center for Health Statistics, 2017–2020.  
10. N. V. Chawla, et al., “SMOTE: Synthetic Minority Over-sampling Technique,” Journal of Artificial  
Intelligence Research, 2002.  
11. National Cancer Institute (NCI), “Healthy Eating Index–2020,” 2023.  
12. UCI Machine Learning Repository, “Pima Indians Diabetes Database,” 2016.  
13. National Heart, Lung, and Blood Institute (NHLBI), “Framingham Heart Study,” 2023.  
14. D. Dooley, et al., “FoodOn: A harmonized food ontology,” NPJ Science of Food, 2018.  
15. T. Fawcett, “An introduction to ROC analysis,” Pattern Recognition Letters, 2006.  
www.ijltemas.in  
Page 1165