Diet Recommendation Using Health Data: A

Personalized Nutrition Recommendation Framework

Sahil Gaba, Reejul Bhunesh, Abhay Singh, Gaurav Garg
Department of Computer Science Engineering
University Instituet of Engineering
Chandigarh University, Mohali- 140413, Punjab, India
sahilgaba8968@[Link], reejul4thakur@[Link], kashyapabhay745@[Link], [Link]@[Link]

Abstract—Preventive healthcare depends on personalised nu- 2) A multi-objective constraint optimisation model (linear
trition, however the majority of dietary recommendations now in and mixed-integer variants) that generates meal plans
use are population-level rather than customised to each person’s that maximise preference ratings and minimise target
unique metabolic and clinical traits. This work offers a thorough
framework for automated personalised diet recommendations deviance.
that combines limited optimisation, micronutrient adequacy 3) An eight-week human research with 120 individuals that
evaluation, and physiological modelling into a pipeline that is compares the suggested approach to baseline recom-
ready for production. In order to generate daily and weekly menders and generic diets.
meal plans, our system consumes user health parameters (age, 4)
gender, height, weight, activity level, known medical conditions,
and preferences), calculates customised caloric and nutrient II. R ELATED W ORK
targets, and solves a multi-objective diet optimisation problem.
The solution minimises aim deviation and maximises preference Recommender systems for food and nutrition have grown
alignment while enforcing therapeutic limitations (such as low- quickly. Min et al. provide a comprehensive review of food
sodium for hypertension and carbohydrate limits for diabetes).We recommendation approaches, modalities, and challenges [1].
compare the framework to two baselines (rule-based and ML-
Health-aware recommenders that incorporate nutritional cri-
based recommender) and generic diet regimens in an 8-week
research with 120 participants. The findings demonstrate a teria into ranking algorithms were proposed early in the
significant decrease in nutrient deviation (from 22.4 to 5.9), field [2], [3], and later systems increasingly combine ma-
enhanced clinical indicators (mean fasting glucose reduction for chine learning with nutritional constraints [4], [5]. Recent
the diabetic cohort: 7.3 mg/dL), and good user satisfaction (USI = work explores graph-based heterogeneous models [6], LLM-
4.3/5). We offer complete mathematical formulation, algorithmic
augmented recommender chatbots [7], and deep generative
design, sensitivity analysis, complexity analysis, and privacy,
ethical, and deployment considerations. approaches for nutrition planning [8]. Systematic reviews
Index Terms—Precision nutrition, linear programming, re- and surveys summarize evaluation criteria and limitations of
stricted optimisation, dietary recommendation, personalised nu- current NRS solutions [9], [10].
trition, and health analytics. Unlike pure ML recommenders, our method tightly couples
clinical nutrient targets with optimization, similar in spirit to
I. I NTRODUCTION sequential or bilevel optimization approaches for diet plan-
ning [11], [12]. We extend these by modeling micronutrient
Long-term health outcomes are significantly influenced by adequacy, multi-day meal variety constraints, and practical
dietary behaviour; poor diets are associated with obesity, preference scoring.
type-2 diabetes, cardiovascular disease, and numerous other
disorders. Despite a wealth of public health recommendations, III. P ROBLEM D EFINITION
the most of them are still population-level and do not signif- Given a user profile U containing demographic data, an-
icantly account for individual characteristics (metabolic rate, thropometrics, medical conditions, dietary restrictions, and
clinical history, allergies, and lifestyle). There is a chance to preference signals, compute a meal plan P (daily or weekly)
automate precision nutrition, creating customised meal plans that:
that meet nutrient goals, adhere to medical restrictions, and suit • Meets caloric needs (TDEE) and macro/micronutrient
individual preferences, thanks to the growth of food metadata, targets,
wearable sensors, and computer power. A modular architecture • Observes clinical restrictions (e.g., sodium, added sug-
and optimisation framework for tailored diet recommendations ars),
are presented in this work. What we have contributed is: • Aligns with stated preferences (likes, cultural/ethical con-
1) A meticulous metabolic and micronutrient modelling straints),
pipeline that translates user health data into targeted • Minimizes deviation from nutritional targets and abrupt
nutrient levels. behavioral changes (for adherence).
 50
Percentage of Calories 50 80

Objective Value
40 60

30 40
30

20 20
20
1 2 3 4 5 6
Carbohydrates Protein Fat
Optimization Iteration
Macronutrient
Fig. 2. Convergence behaviour of the diet optimization objective during
Fig. 1. Recommended macronutrient distribution used by the diet optimization solution search.
model.

V. O PTIMIZATION F RAMEWORK
Formally, let x represent quantities of foods selected from
A. Variables and Parameters
a curated database of m food items. Each item j has nutrient
vector pj and a preference score sj (U). The optimization goal • xj : servings of food item j (continuous or integer)
is to compute x ∈ Rm • pij : amount of nutrient i per serving of j
≥0 that satisfies constraints and optimizes
the multi-objective criterion described next. • cj : calories per serving of j
• sj : preference score (0–1) computed from user profile
IV. M ATHEMATICAL M ODELING and historical behavior
• Ri : required daily amount for nutrient i
A. Metabolic Targets
We compute BMR via the Mifflin–St Jeor equations [13]: B. Single-day Linear Objective
We minimize weighted sum of (absolute) nutrient deviations
BM Rmale = 10W + 6.25H − 5A + 5, (1) and negative preference score:
BM Rfemale = 10W + 6.25H − 5A − 161, (2) X X X
min α wi |Ri − pij xj | − β sj xj , (5)
x≥0
where W (kg), H (cm), A (years). Total Daily Energy Expen- i∈N j j
diture (TDEE) is: subject to:
T DEE = BM R × AF,
X
(3) cj xj = T DEE, (6)
j
with AF = activity factor (e.g., 1.2–1.9). X
Macronutrient calorie targets follow standard ranges (ad- Li ≤ pij xj ≤ Ui , ∀i, (7)
justable clinically): j
xj ≤ Mj , ∀j (portion limits), (8)
Carbcal ∈ [0.45, 0.60] × T DEE,
xj = 0, ∀j ∈ Fexclude . (9)
P roteincal ∈ [0.15, 0.25] × T DEE,
F atcal ∈ [0.20, 0.35] × T DEE. The absolute values are linearized by introducing auxiliary
variables, producing a linear program (LP). Nutrition accuracy
Convert calories to grams using energy densities: 4 kcal/g and preference are balanced by α, β. Clinical ranges and
for protein/carbs, 9 kcal/g for fat. acceptable margins are captured by Bounds Li , Ui .

B. Micronutrient Adequacy C. Multi-day and Variety Constraints

For each micronutrient i (iron, calcium, vit D, vit C, sodium, Add restrictions to weekly plans to guarantee maximum
potassium, etc.) we define Required Daily Allowance Ri from repetition and variety:
authoritative sources [14], [15]. Adequacy ratio: 7
X
P 1[xj,d > 0] ≤ Vj ,
j pij xj d=1
M ARi = . (4)
Ri
or a mild punishment for repetition to prevent boredom, which
An acceptable interval is [0.9, 1.1]; outside this, we penalize is crucial for compliance. The MILP formulation is applicable
deviation. when discrete servings (such as integer numbers) are needed
D. Multi-objective Extension
Additionally, we are in favour of an explicit multi-objective
User Data
formulation that includes cost minimisation, choice utility, and (Age, Weight, Medical History)
nutrient deviation. Depending on the deployment preference,
we solve using either lexicographic optimisation or weighted
sum.

VI. P ERSONALIZATION S TRATEGY Metabolic Computation

Creating diet programs that are both nutritionally balanced (BMI, BMR, TDEE)
and tailored to the unique qualities of each user is one of the
main goals of the suggested framework.
First, the distribution of nutrients and caloric needs are in-
fluenced by demographic characteristics including age, gender,
and body composition. For instance, people who are more
Optimization Engine
active need to consume more calories, whereas inactive users (Linear Programming)
could need to consume fewer calories.
Second, extra dietary restrictions are brought about by
medical issues. For example, those with diabetes need to main-
tain their blood sugar levels and moderate their carbohydrate
intake, while people with hypertension need to consume less
sodium.
Personalized Diet Plan
Third, a preference scoring technique is used by the system
to include user preferences. Based on the user’s past dietary Fig. 3.
decisions, cultural eating habits, and dietary limitations like
vegetarian or vegan preferences, each food item in the database
is given a score. The user’s basal metabolic rate (BMR) and total daily
By maximising preference alignment and minimising nutri- energy expenditure (TDEE) are then estimated using the
tional divergence, the optimisation method strikes a balance metabolic computation module. Depending on the user’s ob-
between these variables. Because users are more likely to jectives, these values establish the number of calories needed
follow advice that fit their nutritional preferences and tastes, to maintain, gain, or lose weight.
this strategy increases long-term adherence. Nutrient objectives for macronutrients and micronutrients
are calculated using dietary reference standards following
VII. S YSTEM A RCHITECTURE AND I MPLEMENTATION calorie estimation. In the optimisation model, these goals serve
as constraints.
A. High-level Architecture
Next, in order to maximise preference scores and min-
The complete system is shown in Figure 3: data intake (user imise divergence from nutrient targets, the optimisation en-
profiles, wearables, lab reports) → preprocessing validation gine chooses food combinations from the well curated food
→ target computation → optimisation engine → meal plan database.
generator → monitoring and feedback loop. Lastly, the system records adherence data and user com-
B. Implementation Choices ments, allowing recommendations to be improved iteratively
in subsequent sessions.
If discrete servings are imposed, we implemented the op-
timiser utilising Gurobi/CPLEX compatibility for MILP and
CVXOPT (Python) for LP. For mobile frontend integration, VIII. A LGORITHM
the backend makes REST APIs available. The food nutrition The pipeline is summarised by the algorithm.
database is indexed for speedy retrieval and normalised (per
100 g/per serving). Algorithm 1 Personalized Diet Planner
C. System Workflow 1: Input: user profile U, food DB
2: Determine macronutrient/micronutrient targets Ri
There are multiple sequential steps in the suggested diet
3: Compute preference scores sj (U)
recommendation framework’s operational workflow. First, a
4: Compute BMR, TDEE using Mifflin–St Jeor
user interface or integrated health monitoring devices are
5: Optimise x∗
used to gather user health data, including age, gender, height,
6: Post-process meal plan (meal assembly, timing)
weight, activity level, and medical issues. Inconsistencies and
7: Return meal plan and schedule monitoring tasks
missing values are eliminated by preprocessing and validation
of this data.
 User Profile 64
60

Number of Participants
Health Data Processing
40
28
Nutrient Target Calculation
20 18
10
Optimization Model
Healthy Overweight Diabetic Hypertensive
Participant Category
Personalized Diet Plan
Fig. 6. Distribution of participants across medical categories in the study.
Fig. 4. Workflow pipeline of the personalized diet recommendation system.

TABLE I
2,750 E XAMPLE DAILY M EAL P LAN G ENERATED BY THE S YSTEM

Meal Food Items Calories

2,700
Calories

Breakfast Oatmeal, Banana, Almonds 450

Lunch Brown Rice, Grilled Chicken, Vegetables 750
2,650 Snack Yogurt, Apple 200
Dinner Whole Wheat Roti, Lentils, Salad 600

2,600
TABLE II
Mon Tue Wed Thu Fri Sat Sun C OMPARISON ACROSS STUDY ARMS (8- WEEK AVERAGES ).
Day of Week
Metric Generic Baseline Proposed
Fig. 5. Example weekly calorie distribution generated by the diet recommen- Nutrient deviation (%) 22.4 12.1 5.9
dation system. Weight change (%)* -1.2 -3.8 -5.9
Fasting glucose (mg/dL)** -1.8 -4.2 -7.3
USI (1–5) 3.1 3.8 4.3
Adherence (%) 47 62 74
IX. DATASET AND E XPERIMENTAL P ROTOCOL
* Overweight subgroup average. ** Diabetic subgroup average.
A. Food Database
We curated a database of 320 common items (regional and
international), each annotated with macro/micronutrients
• Weight change (%): relative weight difference at 8
(based on USDA and national food composition
weeks.
tables), typical serving sizes, and categorical tags
• Change in fasting glucose (mg/dL): for the diabetic
(vegan/vegetarian/gluten-free/allergen flags).
group.
B. Participant Cohort • User Satisfaction Index (USI): Likert average from 1 to
5.
120 volunteers (age 18–65; male:female = 66:54), recruited
• Adherence Score: percentage of suggested meals
locally, provided consent. Medical categories: healthy (n=64),
recorded as consumed
overweight/obese (n=28), Type-2 diabetic (n=18), hypertensive
(n=10). All participants were randomized into three arms: XI. R ESULTS
1) Generic diet control (standard RDA-based plan) Table II summarizes principal outcomes.
2) Baseline recommender (rule-based) Nutrient deviation, weight change, and glucose decrease
3) Proposed optimization-based system all demonstrate substantial improvements (p < 0.01) when
Duration: 8 weeks; primary outcomes: nutrient deviation, compared to baseline using paired t-tests. In the suggested arm,
weight change, fasting glucose (when applicable), and user adherence and satisfaction were likewise significantly higher.
satisfaction.
A. Ablation Study
X. E VALUATION M ETRICS To assess the effects of micronutrient penalties and pref-
• Nutrient deviation (%): mean relative deviation across erence weighting (β), we conducted ablations. Removing
macro + selected micros. preference alignment (β = 0) decreased adherence by 9% and
 4.3 A five-point Likert scale was used to gauge user satisfaction
based on survey replies. Because the suggested approach
User Satisfaction (1-5)

offered flexible meal alternatives rather than strict dietary

4 guidelines, participants expressed greater satisfaction with it.
3.8 Overall, the analysis shows that the accuracy and usefulness
of automated diet guidance systems can be greatly increased
by fusing optimisation algorithms with nutritional research.
3.5
XIII. ROBUSTNESS AND S ENSITIVITY
3.1
Sensitivity to activity factor misreporting (±10%) was as-
3 sessed. Calorie mis-specification produced linear effects that
Generic Baseline Proposed were in line with expectations, and nutrient deviation increased
Diet Planning Method somewhat (3

Fig. 7. User satisfaction ratings for each of the assessed dietary advising XIV. C OMPLEXITY AND S CALABILITY
strategies.
In the worst scenario, LP solving time scales approximately
cubic in (m+n); on a commodity server, the typical solve time
80 74 (CVXOPT) for m ≈ 320, n ≈ 20 was 0.2–0.6 s per user. Prac-
62 tical throughput for actual deployments is achieved by using
Percentage / Score

60 warm-starts and a cached problem structure (weekly planning).

47 For production, we advise LP relaxations with deterministic
40 rounding and post-checks for stringent clinical requirements
because MILP situations (integer servings) resulted in longer
22.4 solve times (median 3–8 s).
20 12.1
5.9 5.9 A. Efficiency of Computation
1.23.8
0 The main method used to solve the optimisation problem in
this paper is linear programming. Highly optimised interior-
Nutrient Deviation Weight Change Adherence
point and simplex algorithms are used by contemporary LP
solvers like CVXOPT and Gurobi, allowing for effective
Generic Baseline Proposed calculation even with somewhat large food datasets.
In real-world applications, the number of candidate food
Fig. 8. Performance comparison of the suggested optimization-based system, products that can be recommended is equal to the number of
baseline recommender, and generic diet.
decision variables. The optimisation issue is still computation-
ally tractable with a database size of about 320 food items and
USI by 0.6 points. The average micronutrient shortage rose about 20 nutritional constraints.
from 6% to 18% when micronutrient penalties were excluded. An empirical analysis revealed that, in a typical server setup,
the average time to solve a single user query was between
XII. S YSTEM E VALUATION 0.2 and 0.6 seconds. When used in cloud-based designs, this
performance enables the system to expand to thousands of
We examined the system’s performance across several user
concurrent users with no processing expense.
categories in order to further assess the efficacy of the sug-
Additionally, the system is appropriate for real-time nutri-
gested framework. Nutritional accuracy, improved adherence,
tion recommendation applications because it greatly reduces
and user happiness were the three main focuses of the assess-
computation time for repeated requests by caching commonly
ment.
used optimisation structures and warm-starting the solver.
The percentage difference between the goal nutrient re-
quirements and the recommended nutrient intake was used to
XV. S ECURITY, P RIVACY, AND E THICAL I SSUES
calculate nutritional correctness. Because limited optimisation
was used, the suggested system’s deviations were much fewer Standard procedures for protecting health data include
than those of generic diet programs. explicit consent, role-based access, and encrypted storage.
Meal log data gathered from participants over the course of When more stringent privacy requirements arise, the technol-
the eight-week research period was used to assess adherence. ogy allows on-device or edge processing. We steer clear of
Because the recommendations were tailored to each user’s prescriptive medical advice without doctor supervision and
dietary habits and food choices, users who followed the instead prioritise explainable recommendations (ingredient-
optimised meal plans showed higher adherence rates. level explanations).
 XVI. D ISCUSSION ACKNOWLEDGMENTS
Three practical benefits of the optimization-driven method We thank the volunteers for their participation and the
are as follows: Chandigarh University Institutional Review Board for over-
1) Clinical fidelity: adherence to medical needs is ensured sight of human-subject protocols.
by hard and soft constraints. R EFERENCES
2) Explainability: clinician review is made possible by the
[1] W. Min, X. Kong, Z. Chen, and D. L. Wang, “Food recommendation:
explicit nutrient contribution of each meal. Framework, existing solutions, and challenges,” IEEE Trans. Multime-
3) Adaptability: the framework can be extended with dia, vol. 22, no. 10, pp. 2658–2677, 2020.
predictive dynamics (reinforcement learning) for long- [2] M. Ge, F. Ricci, and D. Massimo, “Health-aware food recommender
system,” in Proc. ACM RecSys, 2015, pp. 333–334.
term behaviour modification and supports incremental [3] D. Elsweiler and M. Harvey, “Towards automatic meal plan recommen-
data (lab findings, wearables). dations for balanced nutrition,” in Proc. ACM RecSys, 2015, pp. 313–
316.
Cohort size and regional concentration are limitations; [4] R. Yera, A. A. Alzahrani, and L. Martinez, “A food recommender system
larger, multi-center trials are required for generalisation. Fur- considering nutritional information and user preferences,” IEEE Access,
thermore, self-reported adherence is still a poor indicator of vol. 7, pp. 96695–96711, 2019.
[5] J. N. Bondevik et al., “A systematic review on food recommender
real intake. systems,” IEEE Intelligent Systems, 2024.
[6] S. Forouzandeh, M. Rostami, K. Berahmand, and R. Sheikhpour,
XVII. D EPLOYMENT C ONSIDERATIONS IN THE R EAL “Health-aware food recommendation system with dual attention in
W ORLD heterogeneous graphs,” Computers in Biology and Medicine, 2024.
[7] L. Yang, C.-K. Hsieh, H. Yang, N. Dell, S. Belongie, and D. Estrin,
The suggested paradigm can be included into clinical nu- “Yum-me: A personalized nutrient-based meal recommender system,”
trition systems, fitness platforms, and mobile health apps for arXiv:1605.07722, 2016.
[8] I. Papastratis et al., “AI nutrition recommendation using a deep gener-
practical use. The backend optimisation engine can function as ative model,” Frontiers in Nutrition, 2024.
a cloud-based service that uses RESTful APIs to gather user [9] S. Abhari, F. Mohamed, and R. J. Couldrey, “A systematic review
data and provide customised meal plans. of nutrition recommendation systems,” Journal of Medical Internet
Research, 2019.
Continuous monitoring of physical activity and calorie [10] I. Orúe-Saiz et al., “Systematic review of nutritional recommendation
expenditure is made possible by integration with wearable systems,” Applied Sciences, vol. 11, 2021.
technology, such as fitness trackers. This data can enhance [11] A. Arefeen et al., “Computational framework for sequential diet recom-
mendation integrating linear optimization and clinical domain knowl-
suggestion accuracy and dynamically modify energy require- edge,” in Proc. IEEE, 2022.
ments. [12] V. O. Pochmann, “Multi-objective bilevel recommender system for food
By offering decision support for individualised food plan- diets,” in Proc. IEEE CEC, 2022.
[13] M. D. Mifflin et al., “A new predictive equation for resting energy expen-
ning, the technology can help dietitians and other medical diture in healthy individuals,” American Journal of Clinical Nutrition,
professionals in clinical settings. Clinicians can examine the vol. 51, no. 2, pp. 241–247, 1990.
nutrient contributions from each suggested food item because [14] Institute of Medicine, Dietary Reference Intakes: The Essential Guide
to Nutrient Requirements. Washington, DC: National Academies Press,
the optimisation model is explainable. 2005.
These deployment scenarios show how computational nu- [15] World Health Organization, “Healthy diet,” Fact Sheet, 2020.
trition systems may support lifestyle management and preven-
tative healthcare on a large scale.

XVIII. U PCOMING P ROJECTS

We intend to:
• Incorporate image-based intake detection (computer
vision-based food recording). Incorporate RL-based
multi-day planning and temporal personalisation
(chrononutrition).
• Extend clinical studies to longer durations and a wider
range of individuals to evaluate long-term results.

XIX. C ONCLUSION
We introduced a solid, clinically conscious paradigm for
personalised food recommendations that combines limited op-
timisation, micronutrient adequacy, and metabolic modelling.
In terms of nutritional correctness, clinical indicators, and user
happiness, experimental results from an 8-week trial show
definite gains over generic menus and baseline recommenders.
For upcoming precision-nutrition applications, the method is
directly deployable, comprehensible, and expandable.