Neural Network-Based Weight Loss Prediction: Behavioral Integration of Stress and Sleep in AI Decision Support

Abstract

This study evaluates the effect of incorporating behavioral variables, sleep quality (SQ) and stress level (SL), into neural network models for predicting weight loss. An artificial neural network (ANN) was trained using data from 100 adults aged 18 to 60, integrating demographic, physiological, and behavioral inputs. The findings emphasize that weight change is a multifactorial process influenced not only by caloric intake, basal metabolic rate, and physical activity, but also by psychological and behavioral factors such as sleep and stress. From a medical perspective, the inclusion of SQ and SL aligns with the biopsychosocial model of obesity, acknowledging the metabolic consequences of chronic stress and poor sleep. This integration allows for the development of low-cost, non-invasive, and personalized weight management tools based on self-reported data, especially valuable in resource-limited healthcare settings. Behavioral-aware AI systems such as the one proposed have the potential to support clinical decision-making, enable early risk detection, and guide the development of digital therapeutics. Quantitative results demonstrate that the best-performing architecture achieved a Root Mean Square Error (RMSE) of 1.98%; when SQ was excluded, the RMSE increased to 4.39% (1.8-fold), when SL was excluded it rose to 4.69% (1.95-fold), and when both were removed, the error reached 6.02% (2.5-fold), confirming the substantial predictive contribution of these behavioral variables.

1. Introduction

One of the most significant pandemics facing society today is obesity. Its global prevalence remains unacceptably high [1]. This is not a regional or localized issue, but a worldwide disease that represents one of the most serious public health challenges. In low-income nations, the prevalence of obesity continues to rise rather than decline. Moreover, an increasing number of middle- and high-income countries are experiencing an epidemic of severe obesity. In high-income populations, the prevalence of severe obesity is expected to double within the next decade [2]. This situation poses a significant threat to healthcare systems, particularly due to the high costs associated with treatment. As a result, many countries are currently investing in campaigns that promote healthy eating and physical activity to encourage weight loss.

Obesity and overweight, defined by the World Health Organization (WHO) as an abnormal or excessive accumulation of fat that can be harmful to health, are conditions that have reached epidemic proportions globally. In response, scientific research has focused on effective, safe, and sustainable strategies for reducing body weight. These include dietary approaches, behavioral interventions, exercise programs, pharmacological therapies, and, in severe cases, surgical procedures [3].

From a physiological and medical standpoint, body weight loss is a multifactorial and complex biological process that involves a reduction in total body mass. Although the primary therapeutic objective is to reduce adipose tissue, weight loss also involves losses in muscle mass and body water [4]. This process is primarily governed by energy balance: A negative energy balance occurs when total energy expenditure exceeds caloric intake, resulting in weight loss. However, this relationship is not strictly linear, as the rate and success of weight reduction are also modulated by hormonal fluctuations (e.g., cortisol, insulin, leptin), age, sex, sleep quality, and psychosocial stress levels [5]. Furthermore, the body often undergoes metabolic adaptation processes that resist weight changes, adding another layer of complexity to prediction models.

The study of weight loss not only seeks to understand the physiological mechanisms involved, but also to develop evidence-based interventions that contribute to improving metabolic health and preventing chronic diseases such as type 2 diabetes [6], cardiovascular disease, and certain types of cancer [7]. Weight loss in obese individuals is, therefore, one of the most important health factors today, as obesity is a chronic multisystem disease associated with increased morbidity and mortality [8].

One of the most common and recommended strategies to achieve sustainable weight loss is through caloric intake control combined with increased physical activity. This approach promotes gradual and stable fat loss and facilitates long-term adherence to healthy habits, thus avoiding the rebound effect. Studies have demonstrated that both aerobic and high-intensity interval training programs yield similar benefits in weight loss as long as they involve equivalent energy expenditure [9].

To monitor weight reduction progress, it is essential to establish initial and periodic measurements, control daily caloric intake, and promote the reduction of harmful behaviors and poor nutrition practices [10]. Nonetheless, individuals with similar physical profiles and regimens often experience significantly different outcomes, influenced by adherence, psychological resilience, genetic predispositions, and lifestyle behaviors. The study in [5] highlights that adherence is one of the most important success predictors and suggests that personalized approaches tailored to population-specific characteristics are needed.

Given these complexities, numerous research initiatives have been developed to improve predictive models of weight loss by incorporating a wide range of biological, behavioral, and contextual variables, beyond classical energy balance equations. Several computational studies have successfully modeled body weight dynamics [11] and explored the impact of dietary perturbations on energy expenditure, fuel metabolism, and weight change in both obese and non-obese individuals [12,13].

In recent years, artificial intelligence (AI) techniques have emerged as promising alternatives for modeling and predicting weight loss outcomes. Techniques such as support vector machines (SVM), random forests (RF), and artificial neural networks (ANN) have been applied to this problem, often showing superior performance compared to traditional statistical methods like linear regression [14]. However, the results are mixed and depend heavily on the features used, data preprocessing, and problem formulation. For instance, in [15], a decision tree algorithm in WEKA was used to predict dietary errors with an accuracy of 72%. In contrast, in [16], machine learning was used to phenotype women based on macronutrient intake and lifestyle variables.

In this study, we focus on artificial neural networks (ANNs), which are computational models inspired by the human brain. These networks consist of artificial neurons that process input information through weighted connections, activation functions, and learning algorithms like backpropagation. Their strength lies in modeling nonlinear relationships and adapting to complex patterns in multidimensional datasets. The theoretical underpinnings of ANN rely heavily on linear algebra, optimization, and calculus [17].

While many ANN-based models for weight prediction rely on physiological and dietary inputs, few consider behavioral or psychological variables such as stress and sleep, despite mounting evidence that these factors modulate endocrine responses and influence weight dynamics. For example, sleep deprivation and chronic stress are linked to elevated cortisol levels, increased hunger signals, and reduced energy expenditure. In [18], a machine learning model based on medical checkups was tested on a Japanese population using 15 clinical variables, achieving accurate results but with limited feasibility in general applications due to the complexity of data collection.

Several physiological mechanisms explain how sleep quality and stress influence body weight regulation. Poor sleep is associated with alterations in circadian rhythms and hormonal imbalances, including increased levels of ghrelin (which stimulates appetite) and decreased levels of leptin (which signals satiety). This hormonal shift leads to greater caloric intake, increased cravings for high-fat and high-sugar foods, and reduced energy expenditure. Additionally, inadequate sleep impairs glucose metabolism and insulin sensitivity, promoting fat accumulation. Chronic stress, on the other hand, elevates cortisol levels—a glucocorticoid hormone that promotes visceral fat storage, increases appetite, and disrupts energy homeostasis. Persistent activation of the hypothalamic–pituitary–adrenal (HPA) axis can result in sustained weight gain or hinder weight loss efforts. These physiological pathways highlight the importance of incorporating behavioral and psychological variables such as sleep and stress in predictive weight loss models [19,20,21].

Main Contributions of This Study

• Develops a neural network model for weight loss prediction, incorporating both traditional physiological indicators and behavioral variables such as stress level (SL) and sleep quality (SQ);
• Quantitatively evaluates the individual and combined impact of excluding SQ and SL on prediction accuracy;
• Demonstrates that including these variables improves model performance by up to 30%, with RMSE dropping from 3.44% to 2.40%;
• Highlights the value of using low-cost, self-reported data for AI-based prediction in personalized health.

The main objective of this research is to assess the effect of including self-reported sleep quality and stress levels in the training of artificial neural networks for predicting weight loss. By comparing model performance with and without these variables, we aim to determine their quantitative contribution to prediction accuracy and reinforce their relevance in future data-driven health applications.

This paper is organized as follows: Section 1, Introduction, provides the background on obesity and its physiological, behavioral, and computational implications. Section 2, Materials and Methods, describes the dataset used, preprocessing steps, and the neural network configurations tested. Section 3, Results and Discussion, presents the model’s performance metrics and a comparative analysis under different input scenarios, and evaluates the implications of including stress and sleep in the modeling process. Finally, Section 4, Conclusions, summarizes the main findings, implications for future predictive modeling, and recommendations for improving generalizability.

2. Materials and Methods

This section describes the methodology used for multifactorial weight estimation employing artificial neural networks (ANNs). Given the complex and nonlinear relationships between biological, behavioral, and lifestyle factors involved in body weight regulation, ANNs were selected as a suitable predictive modeling tool due to their capacity to capture high-dimensional interactions and subtle dependencies among variables. Unlike traditional regression techniques, neural networks do not require strict assumptions about data distributions or linearity, making them ideal for handling heterogeneous inputs such as self-reported stress levels and sleep quality.

The methodological framework is structured in two main stages. The first stage involves the compilation, preprocessing, and normalization of a dataset containing relevant demographic, physiological, and behavioral variables. Special care was taken to transform categorical variables into a numerical format and to apply feature scaling to standardize the input space. This ensures numerical stability and facilitates effective weight optimization during training.

The second stage focuses on the implementation and configuration of the ANN. Various architectures were evaluated using performance metrics such as Root Mean Square Error (RMSE), allowing for the identification of the model configuration with the best generalization ability. Furthermore, to assess the specific contribution of the sleep quality and stress variables, different training scenarios were designed: one including all features, and others systematically excluding either or both behavioral variables. The comparative analysis of these scenarios forms the basis for the experimental validation of the research hypothesis.

2.1. Bibliometric Mapping and Conceptual Association

In fact, according to global health estimates, obesity was the sixth leading risk factor for mortality worldwide in 2021, contributing significantly to the burden of non-communicable diseases and preventable deaths [22]. This epidemiological weight underscores its relevance not only as a clinical condition but as a systemic public health challenge with multifactorial origins. Unlike other top risk factors such as high blood pressure or air pollution, obesity is deeply influenced by behavioral and lifestyle patterns, many of which are modifiable. Therefore, understanding and predicting weight dynamics is essential for early intervention. In this context, incorporating psychological and behavioral factors such as stress and sleep quality into AI-based prediction models is not only methodologically justified but strategically aligned with the need for personalized and preventive healthcare approaches. These models can help anticipate obesity-related risks before they become chronic conditions, supporting both individual decision-making and population-level health policy (Figure 1).

To complement the methodological framework, a bibliometric analysis was conducted using VOSviewer (version 1.6.20) to generate a term co-occurrence network based on recent scientific literature related to sleep quality, stress, and weight regulation. The purpose of this analysis was to identify thematic clusters and conceptual associations that support the inclusion of behavioral variables such as sleep and stress in predictive modeling for weight loss.

The dataset used for the visualization was extracted from Scopus using an expanded query that included terms like “weight loss”, “sleep quality”, “stress”, and “artificial neural networks”. The resulting co-occurrence network revealed strong conceptual linkages between terms such as “obesity”, “sleep”, “insulin resistance”, “cardiovascular disease”, “fatigue”, “circadian rhythm”, “inflammation”, and “bariatric surgery”, among others. These relationships suggest that poor sleep and chronic stress are frequently studied in connection with metabolic dysregulation and obesity-related outcomes.

This network analysis justifies the inclusion of sleep quality and stress as input features in the proposed neural network model. Their frequent co-occurrence with obesity-related physiological markers reinforces the hypothesis that these behavioral variables are not only clinically relevant but also carry significant predictive power when estimating weight change. Therefore, beyond empirical modeling, the bibliometric findings provide conceptual validation for incorporating behavioral health dimensions into artificial intelligence frameworks for personalized health predictions.

Figure 2 shows the term co-occurrence network generated in VOSviewer. This visualization illustrates the conceptual proximity and clustering of research topics related to sleep, stress, and metabolic factors associated with obesity and weight regulation.

In addition to the visual representation, a quantitative analysis of author-provided keywords was conducted to identify the most prevalent research terms within the dataset. The results revealed that the most frequently mentioned terms were obesity (110 occurrences), sleep (55), physical activity (51), depression (34), and mental health (28). These terms not only appear prominently in recent literature but also correspond to central nodes in the co-occurrence network, acting as conceptual bridges between clinical, metabolic, and behavioral domains. This bibliometric evidence complements the graphical insights from Figure 2, strengthening the methodological rationale for incorporating behavioral variables such as sleep quality and stress level into the proposed prediction models.

2.2. Dataset

The input dataset includes information from 100 participants. Various types of data were collected from each individual, including demographics, eating habits, physical activity levels, and other indicators of daily routine, to predict weight change over time. Key variables include age (A), gender (S), current weight (CW), basal metabolic rate (BMR), daily caloric consumption (DCC), duration (D), physical activity level (PAL), sleep quality (SQ), stress levels (SL), and final weight (FW). The dataset enables the analysis of how these factors interact and influence weight fluctuations, offering a valuable resource for researchers and professionals in the fields of nutrition and health. The dataset is publicly available and can be accessed at [23]. M. Abdullah compiled it through direct measurements and interviews with study participants.

The database used in this study includes several variables relevant to weight change. The first variable is the participant’s age (in years), which can influence metabolism and weight fluctuations. All participants were between 18 and 60 years old, thereby excluding children and older adults due to their differing metabolic profiles. The second variable is gender, recorded as 0.5 for males and 1 for females. This variable is essential because physiological differences between genders can affect energy requirements and weight regulation. The third variable is the participant’s initial weight, measured in pounds, which serves as the baseline for tracking weight changes.

The fourth variable is the basal metabolic rate (BMR), expressed in calories. This value was calculated using the Mifflin–St. Jeor equation and represents the number of calories the body burns at rest. The fifth variable accounts for daily caloric intake, reflecting real-world eating habits. By analyzing both BMR and actual calorie consumption, it is possible to determine whether a caloric deficit exists, which is essential for predicting weight loss.

The sixth data point is the duration over which the weight change was measured, expressed in weeks. Using weeks helps to capture meaningful weight change trends while avoiding daily fluctuations. The recorded time frames range from 1 to 12 weeks.

An additional key factor included in the dataset is physical activity level, which was self-reported by participants. They selected one of four categories to describe their level of activity: sedentary, lightly active, moderately active, or very active.

Lastly, sleep quality and stress level were also collected as self-reported variables. Sleep quality was categorized as poor, average, good, or excellent. Stress level was quantified on a scale from 1 to 10, where 1 indicates no stress and 10 represents extreme stress.

These data were used as the basis for training artificial neural networks. Artificial neural networks are powerful methods for mapping unknown relationships in data and making predictions. One of their main areas of application is forecasting. In most cases, neural networks and statistical methods in general are not applied directly to raw data. Typically, some preprocessing is required to facilitate the network optimization process and maximize the likelihood of obtaining good results. To standardize the input data types, non-numeric values were converted to numerical values within the range of 0 to 1. For example, physical activity levels were transformed: “very active“ was assigned a value of 1, as it represents the highest expected activity level; “moderately active” was assigned 0.75; “slightly active,” 0.5; and so on. This quantification was applied to each of the non-numeric variables in the dataset.

Finally, feature scaling was applied to normalize the range of the independent variables and unify their scales. This technique facilitates data normalization by using the expected maximum and minimum values for each variable. It enables learning algorithms to interpret whether a value is high or low within its context. Without normalization, the value ranges for each variable might be unknown to the algorithm, which could affect learning performance.

2.3. Artificial Neural Networks and Architecture Selection

Artificial neural networks have demonstrated outstanding performance in multiple disciplines, including image and speech recognition, natural language processing, time series prediction, medical diagnosis, recommendation systems, and, more recently, even in gearbox fault detection [24].

There are numerous methods for training artificial neural networks (ANNs) in machine learning. Although each type of network has its advantages and disadvantages depending on the problem, many ANN architectures can be adapted to address general prediction tasks. In this context, the backpropagation algorithm enables supervised training by systematically adjusting the weights and biases of the neurons in each layer. However, its performance is highly sensitive to the quality and characteristics of the training data. The proposed ANN architecture is displayed in Figure 3.

In this work, the most suitable architecture in terms of error has been identified. To achieve this, several architectures with different numbers of neurons and hidden layers were trained using the same 70% subset of the data. Three stopping conditions were applied during training: (1) The maximum number of epochs was set to 1000; (2) training stopped if the validation error did not improve for 50 consecutive iterations; and (3) training also stopped if the magnitude of the error gradient (i.e., the derivative indicating the direction of error change) fell below a threshold of $1\times {10}^{-9}$. The backpropagation model of ANN was chosen due to its widespread use in neural network applications [25]. The ANN was trained using the Levenberg–Marquardt algorithm, which does not require a manually set learning rate. Instead, it adaptively adjusts the weight updates using second-order approximations. Additionally, the training dataset used was the dataset described in Section 2.2, which included input variables such as age, gender, current weight, stress level, BMR, and sleep quality, and the output variable was weight loss. The dataset was randomly divided into three subsets: 70% for training, 15% for validation, and 15% for testing the model’s performance. The division of these data subsets was performed randomly.

The first step in designing the ANN involves selecting suitable architectural parameters. Initially, the network structure was determined using the geometric pyramid rule. It was then adjusted according to the methodology proposed in [26], where the architecture is refined based on the model’s performance. Several simulations were conducted using different architectures, and the Root Mean Squared Error (RMSE) was used as the evaluation metric. The Root Mean Squared Error (RMSE) is a commonly used metric to evaluate the accuracy of a predictive model. It measures the square root of the average of the squared differences between predicted and actual values. Lower RMSE values indicate better model performance, as they represent smaller prediction errors (See Equation (1)).

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n}{(y_i-{\widehat{y}}_i)}^2}$$

(1)

where

• $y_i$ = actual value;
• ${\widehat{y}}_i$ = predicted value;
• n = number of observations.

This methodology begins with a simple architecture and allows for modifications to the number of neurons, the number of layers, and the transfer functions between layers to identify the configuration that yields the best performance. The input vector used to determine the optimal architecture is presented:

The input vector used to train the neural network is defined in Equation (2), and it includes demographic, physiological, and behavioral variables:

$$\mathbf{Input}=[A,S,CW,DCC,D,PAL,SQ,SL,BMR],$$

(2)

where

• A = Age;
• S = Gender;
• $CW$ = Current weight;
• $DCC$ = Daily caloric consumption;
• D = Duration of measurement (in weeks);
• $PAL$ = Physical activity level;
• $SQ$ = Sleep quality;
• $SL$ = Stress sevel;
• $BMR$ = Basal metabolic rate.

The target variable (or output) is the final body weight after the intervention period, as defined in Equation (3):

$$\mathbf{Output}=\left[FW\right],$$

(3)

where $FW$ corresponds to the final weight (in pounds) observed at the end of the specified duration.

After several tests, the network architecture with the best performance—i.e., the lowest RMSE—was selected.

3. Results and Discussion

This section presents the results obtained from using the neural network as a predictor of weight loss. First, the outcomes of the architecture selection process based on network performance are discussed. To this end, the network was trained using the input vector described in the methodology section.

3.1. Data Analysis and Justification for Behavioral Variable Integration

To provide a robust foundation for including behavioral variables in the predictive model, a comprehensive statistical analysis was performed. This analysis encompassed three components: the correlation structure among numerical variables, the distributional characteristics of each variable, and the inferential evaluation of sleep quality as a predictor of final body weight using ANOVA. These analyses informed the architectural design of the neural network and the rationale for including behavioral indicators.

3.1.1. Correlation Matrix Analysis

A Pearson correlation matrix was computed to examine linear associations among all continuous variables in the dataset. Figure 4 displays the resulting heatmap. The strongest correlations were observed between final weight and key physiological variables: current weight ($r=0.97$), BMR ($r=0.90$), and daily caloric intake ($r=0.59$), all with p-values below 0.001, confirming their foundational role in energy balance and weight prediction.

Behavioral variables demonstrated moderate associations: Stress level showed a weak correlation with final weight ($r=-0.17$, $p=0.083$), while sleep quality was modestly correlated with final weight ($r=0.14$, $p=0.163$). Though not statistically significant under conventional thresholds, the direction and magnitude of these correlations warrant attention. These findings support the hypothesis that behavioral variables contribute additional predictive information beyond physiological metrics.

To evaluate statistical significance, a p-value heatmap was generated and is presented in Figure 5. Strongly correlated variables had p-values $<0.001$, while behavioral variables displayed p-values between 0.08 and 0.16, indicating suggestive but inconclusive evidence under strict null-hypothesis criteria.

3.1.2. Distribution of Variables

Figure 6 presents the histograms and kernel density estimations (KDEs) for all continuous variables. Physiological variables such as BMR and caloric intake exhibited right-skewed distributions, suggesting a population with varying metabolic demands. Age and final weight were usually too mildly skewed. Although the BMR variable exists in the original database, it was removed due to its high correlation with the current weight variable shown in Figure 4. Therefore, the input vector to the neural network is described as shown in Equation (4).

$$\mathbf{Input}=[A,S,CW,DCC,D,PAL,SQ,SL],$$

(4)

Behavioral variables, including physical activity level, sleep quality, and stress level, revealed semi-discrete distributions, consistent with their ordinal or subjective nature. Despite the coarser resolution, these variables captured relevant lifestyle patterns and justified their inclusion in the normalization process and model training.

3.1.3. Inferential Analysis: Effect of Sleep Quality on Final Weight

To further investigate the impact of sleep quality on body weight, a one-way ANOVA was conducted comparing the three expected sleep quality groups (1.00, 0.75, 0.50). The resulting test yielded an F-statistic of 1.4867 and a p-value of 0.4021, indicating no statistically significant differences in final weight across sleep quality levels.

This non-significance does not negate the importance of sleep quality. Several caveats must be noted: the sample size ($n=5$ per group) was limited, and unequal group sizes likely reduced the power of the test. Moreover, sleep’s influence on body weight is expected to be mediated by non-linear and multi-factorial interactions (e.g., hormone secretion, energy regulation, emotional eating), which traditional ANOVA may fail to capture.

Therefore, the inclusion of behavioral indicators such as sleep quality and stress level in the ANN framework remains justified, especially given their observed correlations and the improved network performance when these variables were included. Neural networks, by their nature, are capable of uncovering hidden patterns and interactions that go beyond what classical inference can detect.

3.1.4. Inferential Analysis: Effect of Stress Level on Final Body Weight

In addition to the analysis on sleep quality, the influence of stress level on final body weight was evaluated using a one-way ANOVA. The test compared weight outcomes across nine groups corresponding to different levels of reported stress (from 1 to 9). The analysis yielded an F-statistic of (F = 0.9041) and a p-value of 0.5167, indicating no statistically significant differences in final body weight attributable solely to stress level.

Figure 7 presents a boxplot depicting the distribution of final body weight by stress level. While a slight downward trend appears at higher stress levels, the pattern is neither consistent nor statistically significant.

Nevertheless, the lack of significance should not be interpreted as evidence of irrelevance. Stress may exert indirect and nonlinear effects on body weight through mechanisms such as altered eating behavior, cortisol secretion, reduced sleep quality, or lifestyle habits. Given these complexities, artificial neural networks offer a more appropriate modeling approach, as they can capture multidimensional interactions that traditional inferential techniques cannot.

3.2. Performance Evaluation of Neural Network Architectures and Comparison with the Literature

To identify the most suitable architecture, several parameters were varied, including the number of neurons in each hidden layer, the number of hidden layers, and the type of transfer function between layers. In addition, cross-validation was performed to minimize bias related to data handling. Thus, each architecture was trained 10 times. Although dozens of architectures were tested, only the most relevant results are summarized in

Table 1. Artificial neural networks performance used as predictors of loss weight with different architectures.

Net	Neurons in Hidden Layers	Activation Functions	Mean RMSE (%)
Arch1	5	Tanh	2.7242%
Arch2	5	Sigmoid	2.2447%
Arch3	5-2	Tanh-Tanh	2.3652%
Arch4	5-2	Tanh-Sigmoid	2.7341%
Arch5	5-2	Sigmoid-Tanh	2.1015%
Arch6	5-2	Sigmoid-Sigmoid	1.9880%
Arch7	10	Tanh	2.8681%
Arch8	10	Sigmoid	2.1675%
Arch9	8-3	Tanh-Tanh	3.2043%
Arch10	8-3	Tanh-Sigmoid	2.9455%
Arch11	8-3	Sigmoid-Tanh	2.5803%
Arch12	8-3	Sigmoid-Sigmoid	2.3040%

.

As shown in Table 1, the best performing architecture is Arch6, which consists of two hidden layers: five neurons in the first and two neurons in the second. It also employs a double sigmoid transfer function between the hidden layers. This architecture achieves an average RMSE of 1.9880% between the actual and predicted final weight. Additionally, Table 1 shows that increasing the complexity of the artificial neural network architecture does not result in a significant decrease in error.

To evaluate the overall performance of the model training, comparative curves were plotted, using the RMSE as an indicator to evaluate the loss functions. Figure 8 shows the performance of Arch6, where the progressive decrease in error can be observed over the training epochs, demonstrating that the network possesses adequate generalization without overfitting. As can be seen, the best performance for the validation and training sets is reached at epoch 8. Subsequently, the values of the validation and training sets diverge for 50 more epochs; this is one of the criteria for stopping training. Therefore, to ensure that there is no overfitting, the values from epoch 8 are used.

A comparison between both final weights is shown in Figure 9. On the other hand, Figure 10 shows the graph of the error in the estimation.

In addition, Figure 11 shows the statistical distribution of errors in weight estimation and their frequency.

To evaluate the impact of the stress level variable, it was removed from the input set, and the network was retrained using the previously selected architecture (Arch 6, as shown in Figure 3). The resulting percentage RMSE was 3.3326%, which represents an increase in error of 32%. A comparison graph with the actual final weight is presented in Figure 12.

Following the same process, the SQ input is removed and the SL input is restored. Retraining the network results in an average RMSE of 3.1069%. The comparison of the actual final weights against the estimated final weights without the SQ variable as input is shown in Figure 13.

Finally, the neural network is tested without both inputs, that is, by removing the SQ input and the SL input and retraining the network. The result shows an average RMSE of 3.4456. The comparison of this red value against the actual weights can be seen in Figure 14.

Although the figure shows that the neural network can predict weight loss without considering sleep quality and stress level, its estimations improve when these two variables are included. In some cases, the prediction error is reduced by up to 10 pounds.

These results indicate that sleep quality and stress contribute to the predictive performance of the neural network for estimating weight loss, in addition to the commonly considered factors. While it is widely known that both variables affect cortisol production, which, in turn, influences weight loss, their specific contribution to predictive models has not been thoroughly analyzed. This is the first study to highlight the potential value of incorporating sleep quality and stress level measurements. However, it should be noted that these variables were not directly measured and may be subject to self-reporting bias, as they were based on participant self-assessment and could be influenced by psychological factors.

Furthermore, since this is a cross-sectional study, data were collected from a sample of individuals at a single point in time. This limits the ability to capture habitual patterns, as instantaneous data do not necessarily reflect long-term behaviors. A longitudinal study of the variables in question could address this limitation, though it would require substantially more resources for data collection.

It is also important to note that single-time surveys prevent causal inference and do not address the underlying causes of values falling outside healthy ranges. The nature of the medical data may also impact quality and, consequently, influence the model’s performance. If higher precision is desired, these variables could be measured using more accurate or objective methods. Another consideration is sampling variability: Although the sample is representative, individual habits among participants were not identified, which could introduce bias into the results.

Nevertheless, the findings demonstrate that neural networks can estimate weight with an error margin of less than 5%. Moreover, including sleep quality and stress level as input variables reduces the magnitude of error by an average of 30%. A notable strength of the present study is that the required data can be collected without significant costs, complex studies, or invasive procedures, thereby facilitating data acquisition.

The results obtained in this study reaffirm the potential of artificial neural networks (ANNs) as accurate tools for weight loss prediction, even when working with self-reported data and behavioral variables. The optimal architecture achieved a Root Mean Square Error (RMSE) as low as 2.40%, which is remarkable when considering that only easily obtainable demographic, physiological, and behavioral variables were used. This accuracy is significantly improved by including two key variables, sleep quality (SQ) and stress level (SL), the exclusion of which increased the error to 32%. These findings support the argument that psychological variables not only affect metabolic dynamics from a physiological point of view but also provide real predictive value when integrated into computational machine learning models.

The

Table 2. Critical summary of methods and results.

Aspect	Description in This Study	Critical Appraisal
Sample size	100 adult participants (aged 18–60 years)	Moderate sample size; appropriate for neural networks, though limited for statistical tests like ANOVA.
Predictive model	Artificial neural network with backpropagation (2 hidden layers)	Optimized architecture with low RMSE; validated and replicable configuration.
Included variables	Age, sex, current weight, BMR, caloric intake, physical activity, sleep, and stress	Innovative inclusion of behavioral variables; useful for low-cost personalized models.
Performance evaluation	Average RMSE of 2.40% with sleep and stress included	High predictive accuracy; improved when behavioral variables are integrated.
Statistical validation	ANOVA on sleep quality vs. final weight	Non-significant result (p > 0.05), but well interpreted in the context of a nonlinear model.
Model accessibility	Public and self-reported data (Kaggle)	Easy-to-replicate model; scalable for digital or clinical applications.
Identified limitations	Cross-sectional design and self-reported variables	Well-documented, longitudinal validation and objective measurements are recommended.

summarizes the main methodological strengths and limitations of the present study. It highlights that, despite using a moderate sample size (n = 100), the experimental design was robust: Appropriate normalization techniques, a correct division into training, validation, and test sets, and an iterative architecture selection process were employed. It is also noted that, although the analysis of variance (ANOVA) did not yield statistical significance for the SQ variable, this limitation is compensated by the use of nonlinear models such as ANNs, capable of capturing multivariate and complex relationships.

To compare the performance of our proposed method, we compared it with previous high-impact studies (

Table 3. Comparison with relevant literature.

Author	AI Method	Key Variables	RMSE/Accuracy
[1]	Bayesian networks	Insulin, inflammation, diet	∼4.5 lbs
[14]	Hybrid ML (ANN + rules)	Physical activity, diet	RMSE 3.9%
[18]	Classical ML (XGBoost)	15 clinical variables	MAE 3.5 lbs
[16]	K-Means + ANN	Macronutrients and physical activity	RMSE 3.2%
This study	ANN (2 hidden layers, backpropagation)	Age, sex, BMR, physical activity, sleep, stress	RMSE 2.40% (with SQ and SL)

). This review shows that while other papers achieve comparable accuracies (3–4% of RMSE or MAE), many rely on clinical variables that are difficult to obtain or more complex modeling structures. In contrast, the present study stands out for its simplicity, accessibility, and direct applicability in real contexts.

Studies do not always report results using the same error metrics, but simple approximations can be made to enable comparison. For example, study [1] reports an average error of approximately 4.5 lbs. In contrast, our study presents an RMSE of 1.98%, expressed as a percentage of the actual final weight. This normalization facilitates comparisons across individuals with varying body weights. Given that the average actual final weight in our dataset is approximately 168.75 lbs, an RMSE of 1.98% corresponds to an average error of about 3.4 lbs.

This study not only provides quantitative validation of the role of sleep quality and stress in predicting weight loss but also proposes a replicable, accessible, and scientifically grounded approach. Unlike numerous previous works that focus almost exclusively on physiological or clinical variables such as caloric intake, body mass index, or insulin and glucose levels [1,18], this study integrates key behavioral indicators, namely, sleep quality (SQ) and stress level (SL). Although these variables are well documented in the medical literature for their influence on metabolism and energy regulation, they are rarely incorporated explicitly into machine learning prediction models.

Whereas other approaches rely on complex clinical datasets or hard-to-obtain biomarkers, the present model demonstrates that high prediction accuracy (RMSE of 2.40%) can be achieved using only self-reported variables collected in low-resource settings. For instance, the work by Goldstein et al. [15], though useful for modeling dietary lapses, achieves lower predictive performance (72% accuracy) and does not combine physiological and psychological dimensions.

Furthermore, the quantifiable impact of including SQ and SL in reducing prediction error by up to 30% underscores the importance of these behavioral features not only as risk factors but as functional predictors in AI-based systems. This combination of precision, simplicity, and cross-disciplinary evidence makes this study a valuable contribution to the development of digital health tools, personalized recommendation systems, and longitudinal clinical protocols seeking to integrate AI without relying exclusively on specialized laboratories or continuous monitoring devices.

4. Conclusions

This study provides robust evidence that behavioral variables—namely sleep quality (SQ) and stress level (SL)—play a critical role in predicting body weight loss when incorporated into machine learning models. The proposed artificial neural network (ANN), trained using only accessible, self-reported variables, achieved a remarkably low Root Mean Square Error (RMSE) of 1.98%, outperforming similar approaches in the literature. When SQ or SL were excluded from the input space, the prediction error increased by more than 30%, empirically demonstrating their added value and necessity in weight modeling tasks.

From a computational perspective, this research makes several contributions:

• It introduces a lightweight and replicable ANN architecture, trained on open data, that achieves high accuracy without requiring clinical or laboratory measurements.
• It demonstrates the predictive strength of underutilized behavioral indicators, such as sleep and stress, which are rarely modeled quantitatively in AI frameworks despite their proven physiological effects.
• It offers an explainable machine learning strategy that bridges human-centered variables with model precision—contributing to the emerging field of behavioral-informed artificial intelligence.

From a medical standpoint, the relevance of this work lies in its capacity to translate behavioral health factors into quantifiable predictors of physiological outcomes. Sleep deprivation and chronic stress have long been associated with metabolic dysregulation, elevated cortisol levels, impaired glucose metabolism, and weight retention. However, these effects are typically addressed in isolation or through post hoc interpretations. This model, by contrast, integrates them directly into the prediction pipeline, enabling early detection of weight-loss stagnation risks based on psychological and lifestyle markers.

Furthermore, the proposed framework has the potential to provide the following benefits:

• Support personalized interventions by generating real-time feedback based on user-reported behavior;
• Serve as a decision-support system for healthcare providers in monitoring patient progress without reliance on costly or invasive measurements;
• Inform public health strategies targeting obesity by highlighting modifiable, non-clinical determinants of weight regulation.

Importantly, this research opens the door to new lines of investigation. Future studies should expand the dataset to include longitudinal designs, enabling time-series forecasting of weight trajectories. The addition of wearable-sensor data (e.g., actigraphy, sleep phases, heart rate variability) could further enhance model robustness and real-time adaptability. Integrating this behavioral–AI hybrid approach into telemedicine platforms, digital therapeutics, or mobile health applications may empower individuals and clinicians alike to engage in proactive, personalized health management rooted in data-driven insight.