Automatic Sleep Disorder Classification Using Large Language Model Prompting on Sleep Health and Lifestyle Data

Automatic Sleep Disorder Classification Using Large Language Model Prompting on Sleep Health and Lifestyle Data | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Automatic Sleep Disorder Classification Using Large Language Model Prompting on Sleep Health and Lifestyle Data Yutong Zhao This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6124845/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Sleep disorders are a major global health issue, affecting nearly a third of the world’s population. This research explores using the sophisticated language and reasoning abilities of large language models (LLMs) to automatically identify these disorders. We trained LLMs on a dataset of sleep patterns, lifestyle choices, and related health factors, employing three novel prompting approaches to guide their design, training, and evaluation of classifiers. Our results show that a support vector machine classifier, identified through decomposed prompting, achieved an impressive 91.9 percent accuracy (F1-score: 0.919), significantly better than traditional zero-shot and few-shot methods. This work demonstrates a powerful integration of LLM’s understanding and reasoning with automated machine learning, offering a promising new direction for sleep disorder classification in health informatics. Artificial Intelligence and Machine Learning Sleep Disorder Classification Large Language Model Prompt Engineering Health Informatics Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 1 Introduction 1.1 The Gravity of Sleep Disorder Issues Sleep disorders represent a consequential problem, influencing a considerable segment of the population and potentially exerting a substantial influence on general health and the quality of life. As per the statistics provided by the World Health Organization (WHO), roughly 10 percent of the global populace is afflicted by sleep disorders, and this percentage is increasing in the context of rapid urbanization. Sleep disorders can not only result in daytime fatigue and diminished attention but also precipitate chronic ailments like cardiovascular diseases and diabetes. Consequently, the precise identification and classification of sleep disorders hold considerable significance for enhancing public health. 1.2 Dilemmas of Traditional Sleep Disorder Classification Methods In the past, the classification of sleep disorders mainly relied on traditional machine - learning algorithms, such as decision tree algorithms [1], support vector machines (SVM) [2], random forest algorithms [3], etc. These algorithms have played a certain role in sleep disorder research. Researchers used decision tree algorithms to analyze sleep monitoring data, attempting to identify the characteristic patterns of different sleep disorders through a series of conditional judgments and branch decisions. Support vector machines distinguish normal sleep data from sleep disorder data by finding an optimal classification hyperplane. The random forest algorithm improves the accuracy and stability of classification by constructing multiple decision trees and synthesizing their prediction results. However, traditional machine learning algorithms have many difficulties in classifying sleep disorders. Every step of these algorithms requires a large amount of manual operation. In the data pre - processing stage, researchers need to manually handle missing values and outliers in the data, and perform operations such as standardization and normalization to ensure the quality and usability of the data. In the model training and tuning process, researchers need to manually select appropriate algorithms, set model parameters, and optimize the model performance through repeated trials. This manual operation method not only consumes a large amount of human and time costs but is also easily affected by human factors, resulting in certain limitations in the accuracy and reliability of the results. Moreover, the application of traditional machine - learning algorithms highly depends on expert experience. Experts need to select appropriate algorithms, determine the methods of feature engineering, and adjust model parameters based on their professional knowledge and experience. For complex sleep disorder classification problems, the experiences and judgments of different experts may vary, leading to inconsistencies in classification results. These problems of traditional machine - learning algorithms in sleep disorder classification limit the development and application of sleep disorder research, and a new technology and method are needed to break through these dilemmas. 1.3 New Opportunities Brought by Large Language Models Large language models (LLMs), as a cutting - edge technology in the field of natural language processing, have made remarkable progress in recent years. Based on the Transformer [14] architecture, they have learned rich language knowledge through unsupervised pre - training on massive text data. They have demonstrated great potential in numerous fields, providing new ideas and methods for solving complex problems. In the field of medical research [4], [6], [7], [8], [9], the application of large language models has brought new opportunities for sleep disorder classification. Large language models can understand and process natural language, enabling them to directly analyze and interpret the text information in sleep health and lifestyle datasets. They can extract key information from text data such as patients’ sleep logs and descriptions of living habits. Large language models also possess powerful knowledge reasoning capabilities [5]. They can comprehensively analyze and judge the extracted information by combining existing medical knowledge and the diagnostic criteria of sleep disorders, thus achieving the automatic classification of sleep disorders. When faced with complex sleep disorder symptoms, large language models can accurately identify different types of sleep disorders, such as insomnia and sleep apnea, through reasoning and judgment. Moreover, large language models can also discover the potential relationships between sleep disorders and other factors through learning from a large amount of data, providing new bases for the diagnosis and treatment of sleep disorders. 1.4 Research Contributions In the realm of sleep disorder classification research, traditional methodologies have long been heavily reliant on manual processes and expert knowledge. This approach, while valuable, is not only inefficient but also struggles to handle the increasingly complex landscape of sleep health data, presenting significant drawbacks. Prompt engineering, a pivotal technology in the application of large language models, plays a critical role in guiding these models to generate the desired outputs. The efficacy of different prompting strategies varies considerably across tasks. Specifically, the challenge of skillfully employing prompting strategies to empower large language models to accurately extract crucial insights from sleep health and lifestyle data—based on text information alone—and achieve automated sleep disorder classification is a pressing and complex challenge that demands immediate attention. Against this background, this study makes significant and unique contributions in three key areas: • Pioneering Model Application: This research marks a pioneering step in introducing large language models into the field of automated sleep disorder classification. Large language models, with their remarkable capabilities in language understanding and generation, are overcoming the inherent limitations of traditional methods. The core strength of this approach lies in its ability to automatically perform machine learning (AutoML) based solely on textual input, enabling high-precision sleep disorder prediction. This innovation ushers in a new era of efficient and accurate sleep disorder diagnosis. • In-depth Exploration of Prompting Strategies for Automated Classification: This study delves into the application of zero-shot prompting, few-shot prompting, and decomposed prompting techniques within large language models. By systematically examining the novel potential of these prompting strategies in the context of automated sleep disorder classification, the research clearly delineates their advantages and disadvantages in text-based sleep health classification. These in-depth analyses provide clear guidance for subsequent optimization of prompting strategies and facilitate the continuous improvement of LLM performance in text-based automated sleep disorder classification. • Leading Cross-Disciplinary Integration, Automation Innovation, and Application Expansion: Prompt- driven large language models demonstrate robust automation and expansion capabilities. Their integration with diverse technologies, such as sensor technology and wearable devices, further enhances their automation application value in related domains. This not only effectively addresses practical challenges within sleep medicine but, more importantly, accumulates valuable insights for the broader application of LLMs in healthcare. 2 Sleep Health and Lifestyle Dataset 2.1 Source and Composition of the Dataset The sleep health and lifestyle dataset used in this study is sourced from the Kaggle website[10]. The sleep health and lifestyle dataset used this time contains 374 rows and 13 columns of data. Details are as follows: • Personal basic information: Person ID, as the unique identifier for each respondent, helps to accurately identify and track individual data during the data processing and analysis. Gender information can be used to study the differences in the incidence of sleep disorders and sleep patterns between different genders. Age is one of the important factors affecting sleep. As age increases, sleep quality tends to decline, and the incidence of sleep disorders also increases. Occupation reflects the potential impact of work nature, working hours, and work pressure on sleep. • Sleep - related characteristics: Sleep Duration directly reflects an individual’s sleep time. Sufficient sleep duration is crucial for maintaining physical health and normal physiological functions. Quality of Sleep is a subjective evaluation index, measured on a scale from 1 - 10, which reflects aspects such as the depth, continuity, and recovery effect of sleep. • Lifestyle factors Physical Activity Level reflects an individual’s daily exercise volume. Stress Level is one of the important factors affecting sleep. BMI Category is an indicator to measure whether an individual’s weight status is healthy and is closely related to sleep disorders. Daily Steps is a simple and intuitive indicator to measure the physical activity level, through which the daily activity volume of an individual can be understood, and then its impact on sleep can be analyzed. • Health indicators: Blood Pressure is essential for maintaining the normal functions of various organs in the body. Heart Rate reflects the functional state of the heart. These variables are interrelated and jointly reflect the sleep health status and lifestyle characteristics of the respondents. Through in - depth analysis of these variables, the potential relationship between Sleep Disorder and lifestyle can be revealed, providing a scientific basis for the diagnosis, treatment, and prevention of sleep disorders. 2.2 Data Feature Analysis and Visualization 2.2.1 Statistical information of numerical features (Figure 1) The following inferences can be made from the statistical data of these numerical features: • Sleep Duration: The average sleep duration is approximately 7.13 hours, with a relatively small standard deviation, indicating that the overall distribution is relatively concentrated. Most people’s sleep duration is between 6.4 - 7.8 hours, which is within the normal sleep duration range for adults. However, there may still be some individuals with insufficient or excessive sleep duration, which may be related to factors such as living habits and work pressure. • Quality of Sleep score: The average score is 7.31 points, with a standard deviation of 1.20 points, indicating some differences among individuals. The 25 percent - 75 percent quantiles show that most people’s scores are between 6 - 8 points, and the overall sleep quality is above - average. However, a considerable number of respondents have poor sleep quality, with scores below 5 points, which may be comprehensively affected by various factors such as psychological stress and living environment. • Physical Activity Level: The average level is 59.17, with a relatively large standard deviation of 20.83, indicating significant differences in physical activity levels among different individuals. This is closely related to factors such as personal exercise habits and occupational characteristics. For example, there may be a large difference in physical activity levels between people in physically - laborious occupations and those who sit in the office for long hours. • Stress Level: The average stress level is 5.39, with a standard deviation of 1.77, indicating a certain degree of dispersion in the distribution of stress levels. Different individuals face different levels of stress, which may be related to factors such as work nature and life events. Higher stress levels may have a negative impact on health indicators such as sleep quality. • Heart Rate: The average heart rate is 70.17 beats per minute, with a standard deviation of 4.14. The fluctuation range of the heart rate is relatively small, and most people’s heart rates are between 68 - 72 beats per minute. This reflects that the heart rate in this dataset is generally stable, but the heart rates of some individuals may deviate from the normal range, which may be related to physical health conditions, exercise, etc. • Age: The average age is 40.06 years old, with a standard deviation of 13.34, indicating a certain degree of dispersion in the age distribution. People of different age groups may have differences in sleep - related characteristics. For example, older people may have shorter sleep duration or poorer sleep quality. • Daily Steps: The average number of daily steps is 7070.26 steps, with a standard deviation of 3344.52. The differences in the number of steps reflect the different exercise habits and activity levels of different individuals. The number of steps may be related to physical activity levels, sleep quality, etc. Those with more steps may have higher physical activity levels and better sleep quality. 2.2.2 Distribution information of categorical features (Figure 2) From the distribution of categorical features, we can observe the following: • Gender: There are 212 males and 162 females. The number of males is slightly more than that of females. People of different genders may have differences in sleep - related characteristics. For example, due to physiological factors, females may perform differently in terms of sleep quality and stress level. – Occupation: The number of people in each occupation is the same, all 46, reflecting a certain balance in the sample in terms of occupation. Different occupations may show different characteristics in terms of sleep duration, quality, and stress level due to factors such as work pressure and working hours. For example, nurses may have their sleep duration and quality affected and a higher stress level due to the shift - work nature of their jobs. – BMI Category: There are 156 respondents with normal weight, 121 who are overweight, and 97 who are obese. The proportion of respondents with normal weight and overweight is relatively high. People in different BMI categories may have different incidences of sleep disorders. Obese people may be more prone to problems such as sleep apnea. – Sleep Disorder status: There are 236 respondents with normal sleep, 85 with sleep apnea, and 53 with insomnia. Understanding the distribution of different types of sleep disorders helps to conduct targeted research on the causes and preventive measures of sleep disorders. For example, patients with sleep apnea may be related to factors such as obesity and age. 2.2.3 Correlation analysis among features (Figure 3) The Pearson correlation coefficients among various features were calculated (for categorical features, appropri- ate methods were used to analyze their relationships with other features, such as using the chi - square test to analyze the relationship between Gender and Sleep Disorder, etc.). The following conclusions can be drawn from the correlation coefficient matrix: – Sleep Duration and Quality of Sleep: The correlation coefficient between the two is 0.883213, showing a strong positive correlation. This is in line with the general understanding that the longer the sleep duration, the more fully the body and brain can recover, thus improving sleep quality. This result suggests that in measures to improve sleep quality, adjusting sleep duration can be considered. – Physical Activity Level and other features: The correlation coefficients between Physical Activity Level and Sleep Duration, Quality of Sleep are 0.212360 and 0.192896 respectively, showing a weak positive correlation. This indicates that moderate physical activity has a certain positive impact on sleep, but this impact is relatively limited. The correlation coefficient between Physical Activity Level and Stress Level is almost 0, indicating that the linear relationship between the two is not obvious. The correlation coefficient with Heart Rate is 0.136971, showing a weak positive correlation, which may be that a higher physical activity level can cause a certain increase in heart rate. – Stress Level and other features: Stress Level shows a strong negative correlation with Sleep Duration and Quality of Sleep, with correlation coefficients of - 0.811023 and - 0.898752 respectively. This indicates that stress is an important factor affecting sleep. When people are under high stress, the stress hormones secreted by the body will interfere with falling asleep and sleep quality. The correlation coefficient between Stress Level and Heart Rate is 0.670026, showing a positive correlation, indicating that the greater the stress, the higher the heart rate may be, reflecting the impact of stress on physical physiological indicators. – Heart Rate and other features: Heart Rate shows a negative correlation with Sleep Duration and Quality of Sleep, meaning that when the heart rate is high, the sleep duration and quality may be poor. This may be because an abnormal heart rate reflects a certain uncomfortable state of the body, which in turn affects sleep. – Age and other features: Age shows a negative correlation with Sleep Duration and Quality of Sleep, indicating that as age increases, sleep duration may decrease and sleep quality may deteriorate. Age shows a positive correlation with Stress Level, and it is possible that older people face relatively greater life pressure. – Daily Steps and other features: Daily Steps show a positive correlation with Sleep Duration and Quality of Sleep, indicating that more steps may help to extend sleep duration and improve sleep quality. Daily Steps show a strong positive correlation with Physical Activity Level, which is as expected, the more steps, the higher the physical activity level. By conducting statistical analysis and visualizing the distribution of each feature in the dataset, we can gain a deeper understanding of the characteristics and patterns of the data, providing strong data support for the subsequent use of large language models in sleep disorder classification. In subsequent research, we can select appropriate features and models based on these analysis results to improve the accuracy and reliability of sleep disorder classification. 3 Research Methods 3.1 Selection of Large Language Model The selection of a suitable large language model (LLM) is a crucial determinant of accuracy and efficiency in automatic sleep disorder classification research. After careful evaluation of available options, this study employs the Doubao model, a decision grounded in the following key considerations: – Transformer Architecture and Pre-training: The Doubao model is architected upon the Transformer framework [14], a design renowned for its efficacy in natural language processing. Through extensive unsupervised pre-training on massive text corpora, Doubao has acquired a deep and broad understanding of language, encompassing semantics, grammar, pragmatics, and other linguistic facets. Its inherent multi-head attention mechanism enables the model to concurrently process different segments of input text, thereby capturing nuanced semantic relationships and contextual information. Furthermore, the multi-layer neural network architecture facilitates deep feature extraction and information processing across multiple dimensions—lexical, sentential, and textual—allowing for a comprehensive analysis of text content. – Domain-Specific Natural Language Processing Capabilities: The Doubao model demonstrates robust natural language processing capabilities specifically relevant to the medical domain. It exhibits a strong understanding of professional medical terminology within sleep health, accurately interpreting terms such as "apnea-hypopnea index" and "periodic limb movement disorder." Moreover, its excellent generalization performance allows for rapid adaptation to the complex and heterogeneous text data prevalent in sleep health applications. This adaptability extends to diverse data sources, including medical records and log data from sleep monitoring devices, which often exhibit significant variations in format and linguistic style. The Doubao model effectively handles this heterogeneity, ensuring robust performance across varied data inputs. – User-Friendly Interaction and Technical Implementation Advantages: The Doubao model offers unique advantages that streamline the implementation of automated sleep disorder classification. At the user interaction level, certain versions (e.g., the PC version) provide an intuitive interface capable of directly processing common data formats such as CSV files. This user-friendliness significantly simplifies data import and eliminates the need for extensive pre-processing steps like manual data format conversion, saving researchers valuable time and effort. From a technical standpoint, Doubao’s ability to automatically generate and execute code is particularly advantageous. In the automated classification workflow, this feature enables the model to rapidly generate and execute code logic for implementing classification algorithms based on textual analysis. This automation not only dramatically enhances classification efficiency but also mitigates potential syntax errors and logical inconsistencies that may arise during manual code development, providing robust support for efficient and accurate automated sleep disorder classification. In conclusion, the Doubao model’s powerful natural language processing capabilities, high adaptability to sleep health data, and advantageous interactive and technical features make it a compelling and well-justified choice for automated sleep disorder classification in this research. 3.2 Prompt Design Strategies As visually outlined in Figure 4 (detailed prompt examples are provided in the Appendix), this study curate three distinct prompting strategies to investigate and optimize LLM performance in sleep disorder classification: – Zero-shot Prompting: Rule-Guided Baseline Exploration: Zero-shot prompting serves as a baseline strategy, designed to evaluate the model’s inherent capabilities based solely on task descriptions and its pre-existing knowledge. This approach aims to guide the model to generate appropriate responses or complete specific tasks using clear task instructions, without the provision of any task-specific examples. The underlying principle leverages the extensive language understanding and knowledge reserves acquired by LLMs during pre-training on massive datasets. Zero-shot prompting assesses the model’s ability to extract relevant information from its learned knowledge base and apply reasoning and judgment based on semantic cues within the task description. – Few-shot Prompting (Data-Pattern-Driven Optimization): Few-shot prompting represents a data- driven optimization strategy, designed to enhance classification accuracy by guiding the LLM to learn feature-classification outcome relationships from a limited number of provided data samples. This strategy aims to leverage data patterns within a small training set to provide the LLM with more context and information, thereby improving classification accuracy. The key advantage of few-shot prompting lies in its ability to exploit data patterns to enrich the LLM’s understanding and refine its classification performance beyond what is achievable with zero-shot approaches. – Decomposed Prompting (Task Decomposition for Enhanced Classification): Decomposed prompting embodies a task-decomposition strategy, focusing on enhancing task completion efficiency and accuracy by meticulously breaking down complex tasks into a series of more manageable sub-tasks. In the context of natural language processing, this strategy is particularly effective in leveraging the strengths of LLMs to improve both the quality and efficiency of complex task execution. By systematically guiding the LLM through a sequence of well-defined sub-tasks, decomposed prompting aims to facilitate a more structured and accurate approach to sleep disorder classification. Each prompting strategy encompasses two primary sub-tasks. Task 1 involves performing multi-class classifi- cation on data within a specified CSV file and generating a new CSV file containing the classification results. Task 2, consistent across all three prompting strategies, focuses on evaluating the classification outcomes using relevant evaluation metrics and generating corresponding visualizations. 4 Experimental Design and Implementation 4.1 Experimental Environment The construction of the experimental environment is the foundation to ensure the smooth progress of the research. Its hardware and software configurations have a significant impact on the accuracy and reliability of the experimental results. In this sleep disorder classification research, the selected processor for the experiment is Intel (R) Core (TM) i5 - 6400T CPU @ 2.20GHz. This processor has a main frequency of 2201 Mhz, with 4 cores and 4 logical processors. The operating system used is Microsoft Windows 10 Home Chinese Edition, with the version number 10.0.19045 (build 19045). The large language model selected is the Doubao PC version 1.41.6. 4.2 Experimental Procedure 1. Dataset Division and Sample Selection: (a) Randomly select 30 samples from each of the three categories (normal, sleep apnea, and insomnia) in the original dataset, resulting in a total of 90 samples as "prompts 90 examples". Save these 90 samples as Sleep_health_and_lifestyle_dataset_selected_90.csv (training set). (b) Delete the above 90 samples from the original dataset file and save it as Sleep_health_and_lifestyle_dataset_remaining_90.csv as the ground truth. (c) Make a copy of Sleep_health_and_lifestyle_dataset_remaining_90.csv, delete the last column ("Sleep Disorder"),and save it as Sleep_health_and_lifestyle_dataset_remaining_90_without_last_column.csv(test set). 2. Manually upload the above three CSV files to Doubao. This step ensures that the large language model can obtain the required data, providing a data foundation for the subsequent design, training, and evaluation of the classifier. 3. Design, train, and evaluate the classifier according to different prompting strategies. 4.3 Experimental Results 4.3.1 Quantitative Evaluation (Table 1) As detailed in Table 1, we present a quantitative overview of the classification performance achieved by employing different prompting strategies. The metrics provided—accuracy, precision, recall, F1-score, and AUC value—offer a direct numerical comparison of each strategy’s effectiveness in sleep disorder classification. These metrics collectively demonstrate the progressive improvement in model performance as we move from zero-shot prompting to more refined strategies. 4.3.2 Visual Analysis: Confusion Matrices and ROC Curves (Figures 5 and 6) To complement the quantitative analysis and provide a more intuitive understanding of the model’s classification behavior, we present confusion matrices (Figures 5a, 5b, 5c) and Receiver Operating Characteristic (ROC) curves (Figures 6a, 6b, 6c) for each prompting strategy. Confusion matrices visualize the distribution of predicted versus actual classes, offering insights into class-specific accuracy and misclassification patterns. ROC curves, on the other hand, illustrate the trade-off between true positive rate and false positive rate across varying classification thresholds, with curves closer to the upper-left corner signifying superior discriminatory power and higher AUC values. Zero-shot Prompting (Figures 5a and 6a): The confusion matrix for zero-shot prompting (Figure 5a) reveals a considerable degree of misclassification across the three sleep disorder categories. Notably, a significant proportion of normal samples are incorrectly classified as either sleep apnea or insomnia , indicating a limited ability of the zero-shot model to accurately discern normal sleep patterns. Furthermore, some confusion exists between sleep apnea and insomnia classifications. Consistently, the ROC curve for zero-shot prompting (Figure 6a) is positioned furthest from the ideal upper-left corner, exhibiting the lowest AUC value (Table 1) and confirming the suboptimal performance of this strategy. 90-sample Prompting (Figures 5b and 6b): In contrast, the confusion matrix for 90-sample prompting (Figure 5b) demonstrates a marked improvement in classification accuracy. The number of correctly classified normal and sleep apnea samples increases substantially, accompanied by a reduction in misclassifications. However, the classification of insomnia samples still exhibits some misjudgment, suggesting room for further refinement. Correspondingly, the ROC curve (Figure 6b) shifts closer to the upper-left corner compared to zero-shot prompting, and the AUC value improves (Table 1), reflecting the enhanced performance. Decomposed Prompting (Figures 5c and 6c): The decomposed prompting strategy yields the most accurate classification, as evidenced by its confusion matrix (Figure 5c). The matrix shows a significant increase in correctly classified samples across all three categories— normal , sleep apnea , and insomnia —with minimal misclassification instances. This outcome highlights the effectiveness and superiority of decomposed prompting in accurately classifying sleep disorders. The ROC curve for decomposed prompting (Figure 6c) is positioned closest to the upper-left corner and achieves the highest AUC value (Table 1), definitively indicating the best overall classification performance among the evaluated strategies. 5 Analysis of Experimental Results 5.1 Limitations of Zero-shot Prompting The subpar performance of zero-shot prompting in the sleep disorder classification task can be attributed to inherent limitations in data utilization and insufficient model learning capacity. Zero-shot approaches, relying solely on predefined heuristics, lack the ability to leverage training data for in-depth pattern recognition. In the complex domain of sleep disorder classification, diagnoses are rarely determined by single factors but rather by intricate interactions among various clinical and lifestyle variables. These multifaceted relationships often exhibit complex non-linear characteristics. Consequently, zero-shot prompting, devoid of training data exposure, struggles to discern these subtle interdependencies between features and classification outcomes, failing to capture the underlying data patterns essential for accurate sleep disorder identification. 5.2 Performance Gains with 90-sample Prompting: Data-Driven Learning In stark contrast to zero-shot prompting, the 90-sample prompting strategy demonstrates a significant en- hancement in classification performance. By exposing the large language model to a limited set of training examples, the model gains the capacity to discern potential correlations between feature combinations and sleep disorder types. This approach enables the model to learn data patterns and establish a correspondence between input features and desired classification results, effectively mimicking data-driven learning. For instance, by analyzing relationships within the 90-sample training set, the model can begin to understand the influence of factors such as gender, age, and occupation on sleep disorder propensity. Furthermore, the utilization of logistic regression, coupled with essential pre-processing steps like encoding categorical features and standardizing numerical features, contributes to improved data quality and optimized learning. This refined approach leads to more accurate classifications, particularly for common sleep disorder types. However, it is crucial to acknowledge the inherent linearity of the logistic regression model. When confronted with the intricate non-linear relationships that characterize sleep disorders and their multifaceted determinants, the linear nature of logistic regression may limit its ability to fully capture these complexities, ultimately leading to residual classification errors. 5.3 Superiority of Decomposed Prompting: Task Decomposition and Optimized Model Selection The decomposed prompting strategy emerges as the most effective approach in the sleep disorder classification task, achieving the highest performance metrics. This superiority stems from its strategic task decomposition and the optimized selection of a classifier tailored to the data characteristics. Decomposed prompting breaks down the classification task into manageable stages, systematically exploring a suite of classifiers, including logistic regression, decision tree, random forest, gradient boosting, support vector machine (SVM), and K-nearest neighbor (KNN). Rigorous parameter tuning is implemented for each classifier, followed by a comparative evaluation to identify the most suitable model for the specific dataset. This meticulous model selection process reveals the Support Vector Machine (SVM) as the optimal choice for this sleep disorder dataset. SVM’s inherent capability to identify optimal classification hyperplanes in high-dimensional spaces and effectively model complex non-linear relationships proves crucial. The sleep disorder domain is inherently characterized by such non-linearities, evident in the interplay between sleep duration and sleep quality, or stress levels and daily activity volume. SVM’s proficiency in capturing these intricate relationships directly translates into superior classification performance. Specifically, in discerning insomnia from other conditions based on features like stress level and living habits, SVM excels at delineating insomnia samples by identifying the most discriminative classification hyperplane. The task decomposition inherent in decomposed prompting empowers the model to develop a deeper understanding of the task requirements, progressively addressing stages such as data pre-processing, model training, and evaluation. This structured approach not only enhances classification accuracy but also bolsters the overall reliability of the model’s predictions. 5.4 Key Factors for Performance Improvement Through a comparative analysis of the experimental results across different prompting strategies, several key factors for enhancing model performance in sleep disorder classification become apparent. Effective utilization of training data, enhancement of the model’s learning capacity, and the strategic decomposition of tasks coupled with informed classifier selection emerge as pivotal elements. The decomposed prompting strategy demonstrably excels in addressing the complexities of sleep disorder classification, offering valuable methodological insights and a robust framework for future advancements in this domain. 6 Conclusions and Future Work 6.1 Research Conclusions In this rigorous investigation into sleep disorder classification, we meticulously evaluated the application of large language models (LLMs) using a comprehensive sleep health and lifestyle dataset. Through a thoughtfully designed experimental framework and an in-depth, multifaceted analysis, we have convincingly demonstrated the significant potential of LLMs for automated sleep disorder classification. Our findings clearly highlight the pivotal influence of prompting strategies on LLM performance, leading to the following key conclusions: 1. Zero-shot Prompting Sufficiency is Limited: Zero-shot prompting, while straightforward, exhibited relatively weak performance in the complex sleep disorder classification task. This outcome suggests that relying solely on pre-training knowledge and basic heuristics is insufficient for LLMs to effectively address the nuanced complexities inherent in sleep disorder diagnosis without specific examples or in-depth training. 2. 90-sample Prompting Enhances Learning through Data Exposure: The 90-sample prompting strategy, by introducing a limited set of example data points, provided the LLM with more substantial learning information, resulting in a marked improvement in classification performance compared to zero-shot prompting. This improvement unequivocally demonstrates that leveraging data patterns and feature- outcome relationships within a training set significantly enhances the LLM’s comprehension and execution capabilities for sleep disorder classification tasks. However, the inherent linearity of the logistic regression model, employed within the 90-sample prompting approach, presents limitations in accurately capturing and modeling complex non-linear relationships within the data, thereby constraining the model’s ultimate classification potential. 3. Decomposed Prompting Achieves Superior Performance through Task Optimization: The decom- posed prompting strategy yielded the most outstanding results in sleep disorder classification. This approach’s comprehensive exploration of diverse classifiers, coupled with meticulous parameter tuning and rigorous evaluation, highlights the efficacy of task decomposition and judicious classifier selection. Our findings indicate that these techniques effectively harness the inherent strengths of LLMs, leading to enhanced accuracy and reliability in sleep disorder classification. Our research emphasizes that for LLM-based sleep disorder classification, a deep understanding of the data and task is paramount to model performance and directly impacts classification accuracy. Providing detailed examples, relevant reference information, and employing task decomposition strategies are crucial for improving model accuracy. Consequently, future research should prioritize in-depth data and task analysis and focus on optimizing prompting strategies to further enhance model performance. Furthermore, within decomposed prompting methodologies, selecting a classifier model that aligns with task characteristics and implementing effective optimization techniques are central to maximizing classification performance. Given the varying suitability of classifiers across different scenarios, addressing complex sleep disorder classification necessitates a comprehensive consideration of multiple factors, careful selection of the most appropriate classifier, and meticulous parameter adjustments to achieve optimal classification outcomes. 6.2 Limitations and Future Work While this research successfully demonstrates the potential of LLMs in sleep disorder classification, several limitations warrant consideration, particularly concerning data, model selection, experimental design, and prompt engineering. Specifically, the dataset’s diversity remains limited [15], the classifier model selection could benefit from incorporating the advanced feature-learning capabilities of contemporary deep learning architectures, the scope of comparative experiments could be expanded, external validation is absent, and further optimization of prompting strategies is warranted. Future research endeavors can build upon these findings by exploring the following directions to enhance the performance and practical applicability of LLMs in automated sleep disorder classification: – Dataset Expansion and Balancing: Expanding the dataset to encompass a wider range of sleep disorder samples, particularly rare and specialized cases, is a critical next step. Addressing the issue of imbalanced data distribution is also essential. Employing data resampling techniques, such as oversampling (e.g., SMOTE algorithm [16]) and undersampling, can effectively adjust the representation of different sleep disorder types within the dataset. This balanced representation will enable models to learn the char- acteristics of diverse sleep disorders more equitably during training, thereby improving generalization performance, especially for underrepresented groups. – Exploration of Advanced Classifier Models: Future research should prioritize the exploration of more sophisticated classifier architectures and feature engineering methodologies. Deep learning models, with their inherent feature-learning prowess [17], offer promising avenues for improvement. Architectures such as Convolutional Neural Networks (CNNs) [18], Recurrent Neural Networks (RNNs) [19], and their advanced variants like Long Short-Term Memory networks (LSTMs) [20] and Gated Recurrent Units (GRUs) [21] hold significant potential. These models are particularly well-suited for processing time- series data and complex data structures, enabling automatic feature extraction and pattern recognition, and facilitating the exploration of intricate relationships between sleep disorders and factors that are challenging to quantify directly. – Enhanced Experimental Design and External Validation: Future studies should incorporate more comprehensive comparative experiments and rigorous external validation procedures. Systematically comparing the impact of diverse model architectures (including both LLMs and various classification models), prompting strategies, and other relevant factors on classification performance will deepen our understanding of the underlying mechanisms and facilitate the identification of optimal methods and parameter configurations for sleep disorder classification. Furthermore, evaluating model performance on multiple external datasets from diverse sources is crucial for assessing generalization capabilities across varying data distributions and real-world scenarios, and for mitigating the risk of overfitting to the training data. – Refinement of Prompting Strategies: Investigating more advanced and nuanced prompting techniques is essential for further performance gains. Strategies such as Chain-of-Thought Prompting [22] and Contextual Prompting [23] can be explored to guide models toward more in-depth reasoning and analy- sis. Chain-of-Thought Prompting, by encouraging models to articulate their reasoning processes, can enhance model interpretability [24]. Contextual Prompting, by providing models with richer background information and task-specific context, can improve their understanding of task requirements, leading to more accurate sleep disorder classifications. Declarations Acknowledgments My heartfelt thanks go to my parents, whose constant care and support have been absolutely essential to me, both in life and throughout this research project. Their endless encouragement and sacrifices truly made this work possible, and I am deeply grateful. I also want to sincerely thank the data providers for offering the precious sleep health and lifestyle dataset, which was fundamental to this entire study. References Ibomoiye Domor Mienye N, Jere N Survey of Decision Trees: Concepts, Algorithms, and Ap- plications. In IEEE Xplore . https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber= 10562290 Kumari A, Akhtar M, Shah R et al Support matrix machine: A review. https://arxiv.org/abs/2310.19717 Curth A, Jeffares A, van der Schaar M Why do Random Forests Work? Understanding Tree Ensembles as Self-Regularizing Adaptive Smoothers. https://arxiv.org/abs/2402.01502 Kim Y, Xu X, McDuff D et al (2024) Health-LLM: Large Language Models for Health Prediction via Wearable Sensor Data. arXiv preprint arXiv:2401.06885 [cs.LG], [Submitted on 12 Jan 2024 (v1), last revised 27 Apr 2024 (this version, v2)]. Katz DM, Bommarito MJ, Gao S et al (2024) GPT-4 passes the bar exam. Philosophical Trans Royal Soc A 382(2270):20230254 Nori H, Lee YT, Zhang S et al (2023) Can generalist foundation models outcompete special-purpose tuning? Case study in medicine. arXiv preprint arXiv:2311.16452 Saab K, Tu T, Weng W-H et al (2024) Capabilities of Gemini models in medicine. arXiv preprint arXiv:2404.18416 Singhal K, Tu T, Gottweis J et al Towards expert-level medical question answering with large language models. arXiv preprint arXiv:2305.09617, 2023. McDuff D, Schaekermann M, Tu T et al Towards accurate differential diagnosis with large language models. arXiv preprint arXiv:2312.00164, 2023. Sleep Health and Lifestyle Dataset https://www.kaggle.com/datasets/uom190346a/ sleep-health-and-lifestyle-dataset Wang G, Zhao W, Han J et al (2024) MedFound: The First Medical Large Language Model Passing the Physician Qualification Examination. J Artif Intell 5(1):1–12 McDuff D, Xu X, Kim Y et al (2023) Personal Health Large Language Model (PH-LLM): leveraging large language models for personalized health insights. arXiv preprint arXiv:2311.17133 Zhang Y, Maziarka P, Klicpera J et al DiffSBDD: Equivariant diffusion for structure-based drug design. arXiv preprint arXiv:2403.14338, 2024. Vaswani A, Shazeer N, Parmar N, Jones L, Gomez AN, Kaiser L, Polosukhin I (2017) Attention Is All You Need. In Advances in Neural Information Processing Systems 30 (NIPS Liu X, Uchiyama M, Okawa M et al (2000) Prevalence and correlates of insomnia in the Japanese general population: Results from the Japan epidemiological sleep study. Sleep 23(4):497–506 Chawla NV, Bowyer KW, Hall LO, Kegelmeyer WP (2002) SMOTE: Synthetic Minority Over-sampling Technique. J Artif Intell Res 16:321–357 LeCun Y, Bengio Y, Hinton G (2015) Deep Learn Nat 521(7553):436–444 LeCun Y, Bottou L, Bengio Y, Haffner P (1998) Gradient-Based Learning Applied to Document Recognition. Proceedings of the IEEE , 86(11):2278–2324 Graves A, Mohamed A-R, Hinton G (2013) Speech Recognition with Deep Recurrent Neural Networks. In Proceedings of the IEEE international conference on acoustics, speech and signal processing , pages 6645–6649, 2013 Hochreiter S, Schmidhuber J (1997) Long Short-Term Memory. Neural Comput 9(8):1735–1780 Cho K, Van Merriënboer B, Bahdanau D, Bengio Y (2014) On the properties of neural machine translation: Encoder–decoder approaches. arXiv preprint arXiv:1409.1259 Wei J, Wang X, Schuurmans D, Bosma M, Ichtertz R, Joshi S, Zhou D (2022) Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. arXiv preprint arXiv:2201.11903 Zhou X, Huang M, Wang H, Zhang Z (2022) Contextual Prompting for Few-Shot Text Classification. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing , pages 9312–9327 Ribeiro MT, Singh S, Guestrin C, Why Should I (2016) Trust You? Explaining the Predictions of Any Classifier. In Proceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining , pages 1135–1144 Additional Declarations The authors declare no competing interests. Supplementary Files AppendixA.docx AppendixB.docx Appendix.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6124845","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":422066516,"identity":"42c65193-5bdc-41b5-a791-14b80064138c","order_by":0,"name":"Yutong Zhao","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABDElEQVRIiWNgGAWjYBCDBAYG5gMHEipsgGzmBgaGA/gUM8O0sCU++HAmDUgzEq2Fx9hwZtthwloMbuQffFzwqy6Pv73BTJrnzHk5g/uNjUDrGOT5xbDrM7iRzGw8s4+tWOLMgTRpnorbxgbHGJsNZ9xgMJw5OwGXFjZp3h6exIYbCceAttxO3HCMsU2a5wNDgsFtvFokEuffSGyT5m07R6QWnh8GiRuALgR6/wBUyw3cWiTPPDY25m1ISNx45hgj0NfJxpLHEoF+OSOB0y98xxMfPub5U5c473j/B2BU2snxHT588MGHYzby/NLYtSgcABKMbZgSEliVg4B8A4j8g1N+FIyCUTAKRgEDAwAnc2xXQzM46AAAAABJRU5ErkJggg==","orcid":"","institution":"Beijing Guangqumen Middle School","correspondingAuthor":true,"prefix":"","firstName":"Yutong","middleName":"","lastName":"Zhao","suffix":""}],"badges":[],"createdAt":"2025-02-28 03:12:49","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-6124845/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6124845/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":77599656,"identity":"9def6881-435d-4072-bf06-dc0cd76d46e3","added_by":"auto","created_at":"2025-03-03 12:39:08","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":68041,"visible":true,"origin":"","legend":"\u003cp\u003eStatistical chart of numerical features\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-6124845/v1/3cec446f9bb0abbb03ed8b46.png"},{"id":77598320,"identity":"e333d389-2d76-4c5f-8aed-e98fa0ea2257","added_by":"auto","created_at":"2025-03-03 12:31:08","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":110004,"visible":true,"origin":"","legend":"\u003cp\u003eDistribution chart of categorical features\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-6124845/v1/9e1ec1694cc9ce251eb24754.png"},{"id":77598318,"identity":"ee359076-6c40-4397-9208-05eef73e987d","added_by":"auto","created_at":"2025-03-03 12:31:07","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":99663,"visible":true,"origin":"","legend":"\u003cp\u003eHeatmap of Pearson correlation coefficients among features\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-6124845/v1/459758f2473fbd4457f5a4fc.png"},{"id":77598323,"identity":"81358636-fa50-4a3f-88db-c97f2061c33e","added_by":"auto","created_at":"2025-03-03 12:31:08","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":92592,"visible":true,"origin":"","legend":"\u003cp\u003eFlowchart of Prompting Strategies\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-6124845/v1/cf950c1cb5c1cdd4d4ecf0b8.png"},{"id":77599646,"identity":"73cd7afa-0826-4164-94f2-736168536ca3","added_by":"auto","created_at":"2025-03-03 12:39:08","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":69342,"visible":true,"origin":"","legend":"\u003cp\u003eConfusion Matrices for Different Prompting Strategies.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-6124845/v1/62646ba914a4d5be9b3aa2c5.png"},{"id":77598342,"identity":"d0164901-198b-4e44-b32e-c4993f30ff84","added_by":"auto","created_at":"2025-03-03 12:31:08","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":61694,"visible":true,"origin":"","legend":"\u003cp\u003eROC Curves for Different Prompting Strategies.\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-6124845/v1/cd250feb6bce553a52c0ac69.png"},{"id":77601859,"identity":"4e378cd5-0268-4f43-b6b8-592fcf5e5f38","added_by":"auto","created_at":"2025-03-03 13:03:08","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1984188,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6124845/v1/ef5aca06-d479-41c6-9ce5-99225a030a0e.pdf"},{"id":77598324,"identity":"5f2e4acb-fe49-4e6a-9ed1-61bc07cf2166","added_by":"auto","created_at":"2025-03-03 12:31:08","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":34332,"visible":true,"origin":"","legend":"","description":"","filename":"AppendixA.docx","url":"https://assets-eu.researchsquare.com/files/rs-6124845/v1/bd44bbcf1c2433e435226164.docx"},{"id":77598317,"identity":"f79e6713-7113-433b-b743-3cb7e7726117","added_by":"auto","created_at":"2025-03-03 12:31:07","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":18646,"visible":true,"origin":"","legend":"","description":"","filename":"AppendixB.docx","url":"https://assets-eu.researchsquare.com/files/rs-6124845/v1/ae93803e124ce5cb0a705dd6.docx"},{"id":77599644,"identity":"99af2c90-2609-465c-8d48-3b912af9424f","added_by":"auto","created_at":"2025-03-03 12:39:07","extension":"docx","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":14209,"visible":true,"origin":"","legend":"","description":"","filename":"Appendix.docx","url":"https://assets-eu.researchsquare.com/files/rs-6124845/v1/9045d42453c82b0839ef2148.docx"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003eAutomatic Sleep Disorder Classification Using Large Language Model Prompting on Sleep Health and Lifestyle Data\u003c/p\u003e","fulltext":[{"header":"1 Introduction","content":"\u003ch2\u003e1.1 The Gravity of Sleep Disorder Issues\u003c/h2\u003e\n\u003cp\u003eSleep disorders represent a consequential problem, influencing a considerable segment of the population and potentially exerting a substantial influence on general health and the quality of life. As per the statistics provided by the World Health Organization (WHO), roughly 10 percent of the global populace is afflicted by sleep disorders, and this percentage is increasing in the context of rapid urbanization. Sleep disorders can not only result in daytime fatigue and diminished attention but also precipitate chronic ailments like cardiovascular diseases and diabetes. Consequently, the precise identification and classification of sleep disorders hold considerable significance for enhancing public health.\u003c/p\u003e\n\u003ch2\u003e1.2 Dilemmas of Traditional Sleep Disorder Classification Methods\u003c/h2\u003e\n\u003cp\u003eIn the past, the classification of sleep disorders mainly relied on traditional machine - learning algorithms, such as decision tree algorithms [1], support vector machines (SVM) [2], random forest algorithms [3], etc. These algorithms have played a certain role in sleep disorder research. Researchers used decision tree algorithms to analyze sleep monitoring data, attempting to identify the characteristic patterns of different sleep disorders through a series of conditional judgments and branch decisions. Support vector machines distinguish normal sleep data from sleep disorder data by finding an optimal classification hyperplane. The random forest algorithm improves the accuracy and stability of classification by constructing multiple decision trees and synthesizing their prediction results.\u003c/p\u003e\n\u003cp\u003eHowever, traditional machine learning algorithms have many difficulties in classifying sleep disorders. Every step of these algorithms requires a large amount of manual operation. In the data pre - processing stage, researchers need to manually handle missing values and outliers in the data, and perform operations such as standardization and normalization to ensure the quality and usability of the data. In the model training and tuning process, researchers need to manually select appropriate algorithms, set model parameters, and optimize the model performance through repeated trials.\u003c/p\u003e\n\u003cp\u003eThis manual operation method not only consumes a large amount of human and time costs but is also easily affected by human factors, resulting in certain limitations in the accuracy and reliability of the results. Moreover, the application of traditional machine - learning algorithms highly depends on expert experience. Experts need to select appropriate algorithms, determine the methods of feature engineering, and adjust model parameters based on their professional knowledge and experience. For complex sleep disorder classification problems, the experiences and judgments of different experts may vary, leading to inconsistencies in classification results.\u003c/p\u003e\n\u003cp\u003eThese problems of traditional machine - learning algorithms in sleep disorder classification limit the development and application of sleep disorder research, and a new technology and method are needed to break through these dilemmas.\u003c/p\u003e\n\u003ch2\u003e1.3 New Opportunities Brought by Large Language Models\u003c/h2\u003e\n\u003cp\u003eLarge language models (LLMs), as a cutting - edge technology in the field of natural language processing, have made remarkable progress in recent years. Based on the Transformer [14] architecture, they have learned rich language knowledge through unsupervised pre - training on massive text data. They have demonstrated great potential in numerous fields, providing new ideas and methods for solving complex problems.\u003c/p\u003e\n\u003cp\u003eIn the field of medical research [4], [6], [7], [8], [9], the application of large language models has brought new opportunities for sleep disorder classification. Large language models can understand and process natural language, enabling them to directly analyze and interpret the text information in sleep health and lifestyle datasets. They can extract key information from text data such as patients’ sleep logs and descriptions of living habits.\u003c/p\u003e\n\u003cp\u003eLarge language models also possess powerful knowledge reasoning capabilities [5]. They can comprehensively analyze and judge the extracted information by combining existing medical knowledge and the diagnostic criteria of sleep disorders, thus achieving the automatic classification of sleep disorders. When faced with complex sleep disorder symptoms, large language models can accurately identify different types of sleep disorders, such as insomnia and sleep apnea, through reasoning and judgment.\u003c/p\u003e\n\u003cp\u003eMoreover, large language models can also discover the potential relationships between sleep disorders and other factors through learning from a large amount of data, providing new bases for the diagnosis and treatment of sleep disorders.\u003c/p\u003e\n\u003ch2\u003e1.4 Research Contributions\u003c/h2\u003e\n\u003cp\u003eIn the realm of sleep disorder classification research, traditional methodologies have long been heavily reliant on manual processes and expert knowledge. This approach, while valuable, is not only inefficient but also struggles to handle the increasingly complex landscape of sleep health data, presenting significant drawbacks.\u003c/p\u003e\n\u003cp\u003ePrompt engineering, a pivotal technology in the application of large language models, plays a critical role in guiding these models to generate the desired outputs. The efficacy of different prompting strategies varies considerably across tasks. Specifically, the challenge of skillfully employing prompting strategies to empower large language models to accurately extract crucial insights from sleep health and lifestyle data—based on text information alone—and achieve automated sleep disorder classification is a pressing and complex challenge that demands immediate attention. Against this background, this study makes significant and unique contributions in three key areas:\u003c/p\u003e\n\u003cp\u003e• \u003cstrong\u003ePioneering Model Application: \u003c/strong\u003eThis research marks a pioneering step in introducing large language models into the field of automated sleep disorder classification. Large language models, with their remarkable capabilities in language understanding and generation, are overcoming the inherent limitations of traditional methods. The core strength of this approach lies in its ability to automatically perform machine learning (AutoML) based solely on textual input, enabling high-precision sleep disorder prediction. This innovation ushers in a new era of efficient and accurate sleep disorder diagnosis.\u003c/p\u003e\n\u003cp\u003e• \u003cstrong\u003eIn-depth Exploration of Prompting Strategies for Automated Classification: \u003c/strong\u003eThis study delves into the application of zero-shot prompting, few-shot prompting, and decomposed prompting techniques within large language models. By systematically examining the novel potential of these prompting strategies in the context of automated sleep disorder classification, the research clearly delineates their advantages and disadvantages in text-based sleep health classification. These in-depth analyses provide clear guidance for subsequent optimization of prompting strategies and facilitate the continuous improvement of LLM performance in text-based automated sleep disorder classification.\u003c/p\u003e\n\u003cp\u003e• \u003cstrong\u003eLeading Cross-Disciplinary Integration, Automation Innovation, and Application Expansion: \u003c/strong\u003ePrompt- driven large language models demonstrate robust automation and expansion capabilities. Their integration with diverse technologies, such as sensor technology and wearable devices, further enhances their automation application value in related domains. This not only effectively addresses practical challenges within sleep\u003c/p\u003e\n\u003cp\u003emedicine but, more importantly, accumulates valuable insights for the broader application of LLMs in healthcare.\u003c/p\u003e"},{"header":"2 Sleep Health and Lifestyle Dataset","content":"\u003ch2\u003e2.1 Source and Composition of the Dataset\u003c/h2\u003e\n\u003cp\u003eThe sleep health and lifestyle dataset used in this study is sourced from the Kaggle website[10]. The sleep health and lifestyle dataset used this time contains 374 rows and 13 columns of data. Details are as follows:\u003c/p\u003e\n\u003cp\u003e\u0026bull; \u003cstrong\u003ePersonal basic information:\u003c/strong\u003ePerson ID, as the unique identifier for each respondent, helps to accurately identify and track individual data during the data processing and analysis. Gender information can be used to study the differences in the incidence of sleep disorders and sleep patterns between different genders. Age is one of the important factors affecting sleep. As age increases, sleep quality tends to decline, and the incidence of sleep disorders also increases. Occupation reflects the potential impact of work nature, working hours, and work pressure on sleep.\u003c/p\u003e\n\u003cp\u003e\u0026bull; \u003cstrong\u003eSleep - related characteristics:\u003c/strong\u003eSleep Duration directly reflects an individual\u0026rsquo;s sleep time. Sufficient sleep duration is crucial for maintaining physical health and normal physiological functions. Quality of Sleep is a subjective evaluation index, measured on a scale from 1 - 10, which reflects aspects such as the depth, continuity, and recovery effect of sleep.\u003c/p\u003e\n\u003cp\u003e\u0026bull; \u003cstrong\u003eLifestyle factors\u003c/strong\u003ePhysical Activity Level reflects an individual\u0026rsquo;s daily exercise volume. Stress Level is one of the important factors affecting sleep. BMI Category is an indicator to measure whether an individual\u0026rsquo;s weight status is healthy and is closely related to sleep disorders. Daily Steps is a simple and intuitive indicator to measure the physical activity level, through which the daily activity volume of an individual can be understood, and then its impact on sleep can be analyzed.\u003c/p\u003e\n\u003cp\u003e\u0026bull; \u003cstrong\u003eHealth indicators:\u003c/strong\u003eBlood Pressure is essential for maintaining the normal functions of various organs in the body. Heart Rate reflects the functional state of the heart.\u003c/p\u003e\n\u003cp\u003eThese variables are interrelated and jointly reflect the sleep health status and lifestyle characteristics of the respondents. Through in - depth analysis of these variables, the potential relationship between Sleep Disorder and lifestyle can be revealed, providing a scientific basis for the diagnosis, treatment, and prevention of sleep disorders.\u003c/p\u003e\n\u003ch2\u003e2.2 Data Feature Analysis and Visualization\u003c/h2\u003e\n\u003cp\u003e\u003cstrong\u003e2.2.1 \u003c/strong\u003e\u003cstrong\u003eStatistical information of numerical features (Figure 1)\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe following inferences can be made from the statistical data of these numerical features:\u003c/p\u003e\n\u003cp\u003e\u0026bull; \u003cstrong\u003eSleep Duration: \u003c/strong\u003eThe average sleep duration is approximately 7.13 hours, with a relatively small standard deviation, indicating that the overall distribution is relatively concentrated. Most people\u0026rsquo;s sleep duration is between 6.4 - 7.8 hours, which is within the normal sleep duration range for adults. However, there may still be some individuals with insufficient or excessive sleep duration, which may be related to factors such as living habits and work pressure.\u003c/p\u003e\n\u003cp\u003e\u0026bull; \u003cstrong\u003eQuality of Sleep score: \u003c/strong\u003eThe average score is 7.31 points, with a standard deviation of 1.20 points, indicating some differences among individuals. The 25 percent - 75 percent quantiles show that most people\u0026rsquo;s scores are between 6 - 8 points, and the overall sleep quality is above - average. However, a considerable number of respondents have poor sleep quality, with scores below 5 points, which may be comprehensively affected by various factors such as psychological stress and living environment.\u003c/p\u003e\n\u003cp\u003e\u0026bull; \u003cstrong\u003ePhysical Activity Level: \u003c/strong\u003eThe average level is 59.17, with a relatively large standard deviation of 20.83, indicating significant differences in physical activity levels among different individuals. This is closely related to factors such as personal exercise habits and occupational characteristics. For example, there may be a large difference in physical activity levels between people in physically - laborious occupations and those who sit in the office for long hours.\u003c/p\u003e\n\u003cp\u003e\u0026bull; \u003cstrong\u003eStress Level: \u003c/strong\u003eThe average stress level is 5.39, with a standard deviation of 1.77, indicating a certain degree of dispersion in the distribution of stress levels. Different individuals face different levels of stress, which may be related to factors such as work nature and life events. Higher stress levels may have a negative impact on health indicators such as sleep quality.\u003c/p\u003e\n\u003cp\u003e\u0026bull; \u003cstrong\u003eHeart\u003c/strong\u003e\u003cstrong\u003eRate:\u003c/strong\u003eThe average heart rate is 70.17 beats per minute, with a standard deviation of 4.14. The fluctuation range of the heart rate is relatively small, and most people\u0026rsquo;s heart rates are between 68 - 72 beats per minute. This reflects that the heart rate in this dataset is generally stable, but the heart rates of some individuals may deviate from the normal range, which may be related to physical health conditions, exercise, etc.\u003c/p\u003e\n\u003cp\u003e\u0026bull; \u003cstrong\u003eAge: \u003c/strong\u003eThe average age is 40.06 years old, with a standard deviation of 13.34, indicating a certain degree of dispersion in the age distribution. People of different age groups may have differences in sleep - related characteristics. For example, older people may have shorter sleep duration or poorer sleep quality.\u003c/p\u003e\n\u003cp\u003e\u0026bull; \u003cstrong\u003eDaily Steps: \u003c/strong\u003eThe average number of daily steps is 7070.26 steps, with a standard deviation of 3344.52. The differences in the number of steps reflect the different exercise habits and activity levels of different individuals. The number of steps may be related to physical activity levels, sleep quality, etc. Those with more steps may have higher physical activity levels and better sleep quality.\u003c/p\u003e\n\u003ch2\u003e2.2.2 Distribution information of categorical features (Figure 2)\u003c/h2\u003e\n\u003cp\u003eFrom the distribution of categorical features, we can observe the following:\u003c/p\u003e\n\u003cp\u003e\u0026bull; \u003cstrong\u003eGender: \u003c/strong\u003eThere are 212 males and 162 females. The number of males is slightly more than that of females. People of different genders may have differences in sleep - related characteristics. For example, due to physiological factors, females may perform differently in terms of sleep quality and stress level.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026ndash; \u003c/strong\u003e\u003cstrong\u003eOccupation: \u003c/strong\u003eThe number of people in each occupation is the same, all 46, reflecting a certain balance in the sample in terms of occupation. Different occupations may show different characteristics in terms of sleep duration, quality, and stress level due to factors such as work pressure and working hours. For example, nurses may have their sleep duration and quality affected and a higher stress level due to the shift - work nature of their jobs.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026ndash; \u003c/strong\u003e\u003cstrong\u003eBMI Category: \u003c/strong\u003eThere are 156 respondents with normal weight, 121 who are overweight, and 97 who are obese. The proportion of respondents with normal weight and overweight is relatively high. People in different BMI categories may have different incidences of sleep disorders. Obese people may be more prone to problems such as sleep apnea.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026ndash; \u003c/strong\u003e\u003cstrong\u003eSleep Disorder status: \u003c/strong\u003eThere are 236 respondents with normal sleep, 85 with sleep apnea, and 53 with insomnia. Understanding the distribution of different types of sleep disorders helps to conduct targeted research on the causes and preventive measures of sleep disorders. For example, patients with sleep apnea may be related to factors such as obesity and age.\u003c/p\u003e\n\u003ch2\u003e2.2.3 Correlation analysis among features (Figure 3)\u003c/h2\u003e\n\u003cp\u003eThe Pearson correlation coefficients among various features were calculated (for categorical features, appropri- ate methods were used to analyze their relationships with other features, such as using the chi - square test to analyze the relationship between Gender and Sleep Disorder, etc.). The following conclusions can be drawn from the correlation coefficient matrix:\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026ndash; \u003c/strong\u003e\u003cstrong\u003eSleep\u003c/strong\u003e\u003cstrong\u003eDuration\u003c/strong\u003e\u003cstrong\u003eand\u003c/strong\u003e\u003cstrong\u003eQuality\u003c/strong\u003e\u003cstrong\u003eof\u003c/strong\u003e\u003cstrong\u003eSleep:\u003c/strong\u003eThe correlation coefficient between the two is 0.883213, showing a strong positive correlation. This is in line with the general understanding that the longer the sleep duration, the more fully the body and brain can recover, thus improving sleep quality. This result suggests that in measures to improve sleep quality, adjusting sleep duration can be considered.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026ndash; \u003c/strong\u003e\u003cstrong\u003ePhysical Activity Level and other features: \u003c/strong\u003eThe correlation coefficients between Physical Activity Level and Sleep Duration, Quality of Sleep are 0.212360 and 0.192896 respectively, showing a weak positive correlation. This indicates that moderate physical activity has a certain positive impact on sleep, but this impact is relatively limited. The correlation coefficient between Physical Activity Level and Stress Level is almost 0, indicating that the linear relationship between the two is not obvious. The correlation coefficient with Heart Rate is 0.136971, showing a weak positive correlation, which may be that a higher physical activity level can cause a certain increase in heart rate.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026ndash; \u003c/strong\u003e\u003cstrong\u003eStress Level and other features: \u003c/strong\u003eStress Level shows a strong negative correlation with Sleep Duration and Quality of Sleep, with correlation coefficients of - 0.811023 and - 0.898752 respectively. This indicates that stress is an important factor affecting sleep. When people are under high stress, the stress hormones secreted by the body will interfere with falling asleep and sleep quality. The correlation coefficient between Stress Level and Heart Rate is 0.670026, showing a positive correlation, indicating that the greater the stress, the higher the heart rate may be, reflecting the impact of stress on physical physiological indicators.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026ndash; \u003c/strong\u003e\u003cstrong\u003eHeart\u003c/strong\u003e\u003cstrong\u003eRate\u003c/strong\u003e\u003cstrong\u003eand\u003c/strong\u003e\u003cstrong\u003eother\u003c/strong\u003e\u003cstrong\u003efeatures:\u003c/strong\u003eHeart Rate shows a negative correlation with Sleep Duration and Quality of Sleep, meaning that when the heart rate is high, the sleep duration and quality may be poor. This may be because an abnormal heart rate reflects a certain uncomfortable state of the body, which in turn affects sleep.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026ndash; \u003c/strong\u003e\u003cstrong\u003eAge and other features: \u003c/strong\u003eAge shows a negative correlation with Sleep Duration and Quality of Sleep, indicating that as age increases, sleep duration may decrease and sleep quality may deteriorate. Age shows a positive correlation with Stress Level, and it is possible that older people face relatively greater life pressure.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026ndash; \u003c/strong\u003e\u003cstrong\u003eDaily\u003c/strong\u003e\u003cstrong\u003eSteps\u003c/strong\u003e\u003cstrong\u003eand\u003c/strong\u003e\u003cstrong\u003eother\u003c/strong\u003e\u003cstrong\u003efeatures:\u003c/strong\u003eDaily Steps show a positive correlation with Sleep Duration and Quality of Sleep, indicating that more steps may help to extend sleep duration and improve sleep quality. Daily Steps show a strong positive correlation with Physical Activity Level, which is as expected, the more steps, the higher the physical activity level.\u003c/p\u003e\n\u003cp\u003eBy conducting statistical analysis and visualizing the distribution of each feature in the dataset, we can gain a deeper understanding of the characteristics and patterns of the data, providing strong data support for the subsequent use of large language models in sleep disorder classification. In subsequent research, we can select appropriate features and models based on these analysis results to improve the accuracy and reliability of sleep disorder classification.\u003c/p\u003e"},{"header":"3 Research Methods","content":"\u003cp\u003e3.1 Selection of Large Language Model\u003c/p\u003e\n\u003cp\u003eThe selection of a suitable large language model (LLM) is a crucial determinant of accuracy and efficiency in automatic sleep disorder classification research. After careful evaluation of available options, this study employs the Doubao model, a decision grounded in the following key considerations:\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026ndash; \u003c/strong\u003e\u003cstrong\u003eTransformer Architecture and Pre-training: \u003c/strong\u003eThe Doubao model is architected upon the Transformer framework [14], a design renowned for its efficacy in natural language processing. Through extensive unsupervised pre-training on massive text corpora, Doubao has acquired a deep and broad understanding of language, encompassing semantics, grammar, pragmatics, and other linguistic facets. Its inherent multi-head attention mechanism enables the model to concurrently process different segments of input text, thereby capturing nuanced semantic relationships and contextual information. Furthermore, the multi-layer neural network architecture facilitates deep feature extraction and information processing across multiple dimensions\u0026mdash;lexical, sentential, and textual\u0026mdash;allowing for a comprehensive analysis of text content.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026ndash; \u003c/strong\u003e\u003cstrong\u003eDomain-Specific Natural Language Processing Capabilities: \u003c/strong\u003eThe Doubao model demonstrates robust natural language processing capabilities specifically relevant to the medical domain. It exhibits a strong understanding of professional medical terminology within sleep health, accurately interpreting terms such as \u0026quot;apnea-hypopnea index\u0026quot; and \u0026quot;periodic limb movement disorder.\u0026quot; Moreover, its excellent generalization performance allows for rapid adaptation to the complex and heterogeneous text data prevalent in sleep health applications. This adaptability extends to diverse data sources, including medical records and log data from sleep monitoring devices, which often exhibit significant variations in format and linguistic style. The Doubao model effectively handles this heterogeneity, ensuring robust performance across varied data inputs.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026ndash; \u003c/strong\u003e\u003cstrong\u003eUser-Friendly Interaction and Technical Implementation Advantages: \u003c/strong\u003eThe Doubao model offers unique advantages that streamline the implementation of automated sleep disorder classification. At the user interaction level, certain versions (e.g., the PC version) provide an intuitive interface capable of directly processing common data formats such as CSV files. This user-friendliness significantly simplifies data import and eliminates the need for extensive pre-processing steps like manual data format conversion, saving researchers valuable time and effort. From a technical standpoint, Doubao\u0026rsquo;s ability to automatically generate and execute code is particularly advantageous. In the automated classification workflow, this feature enables the model to rapidly generate and execute code logic for implementing classification algorithms based on textual analysis. This automation not only dramatically enhances classification efficiency but also mitigates potential syntax errors and logical inconsistencies that may arise during manual code development, providing robust support for efficient and accurate automated sleep disorder classification.\u003c/p\u003e\n\u003cp\u003eIn conclusion, the Doubao model\u0026rsquo;s powerful natural language processing capabilities, high adaptability to sleep health data, and advantageous interactive and technical features make it a compelling and well-justified choice for automated sleep disorder classification in this research.\u003c/p\u003e\n\u003cp\u003e3.2 Prompt Design Strategies\u003c/p\u003e\n\u003cp\u003eAs visually outlined in Figure 4 (detailed prompt examples are provided in the Appendix), this study curate three distinct prompting strategies to investigate and optimize LLM performance in sleep disorder classification:\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026ndash; \u003c/strong\u003e\u003cstrong\u003eZero-shot Prompting: Rule-Guided Baseline Exploration: \u003c/strong\u003eZero-shot prompting serves as a baseline strategy, designed to evaluate the model\u0026rsquo;s inherent capabilities based solely on task descriptions and its pre-existing knowledge. This approach aims to guide the model to generate appropriate responses or complete specific tasks using clear task instructions, without the provision of any task-specific examples. The underlying principle leverages the extensive language understanding and knowledge reserves acquired by LLMs during pre-training on massive datasets. Zero-shot prompting assesses the model\u0026rsquo;s ability to extract relevant information from its learned knowledge base and apply reasoning and judgment based on semantic cues within the task description.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026ndash; \u003c/strong\u003e\u003cstrong\u003eFew-shot Prompting (Data-Pattern-Driven Optimization): \u003c/strong\u003eFew-shot prompting represents a data- driven optimization strategy, designed to enhance classification accuracy by guiding the LLM to learn feature-classification outcome relationships from a limited number of provided data samples. This strategy aims to leverage data patterns within a small training set to provide the LLM with more context and information, thereby improving classification accuracy. The key advantage of few-shot prompting lies in its ability to exploit data patterns to enrich the LLM\u0026rsquo;s understanding and refine its classification performance beyond what is achievable with zero-shot approaches.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026ndash; \u003c/strong\u003e\u003cstrong\u003eDecomposed\u003c/strong\u003e\u003cstrong\u003ePrompting\u003c/strong\u003e\u003cstrong\u003e(Task\u003c/strong\u003e\u003cstrong\u003eDecomposition\u003c/strong\u003e\u003cstrong\u003efor\u003c/strong\u003e\u003cstrong\u003eEnhanced\u003c/strong\u003e\u003cstrong\u003eClassification): \u003c/strong\u003eDecomposed prompting embodies a task-decomposition strategy, focusing on enhancing task completion efficiency and accuracy by meticulously breaking down complex tasks into a series of more manageable sub-tasks. In the context of natural language processing, this strategy is particularly effective in leveraging the strengths of LLMs to improve both the quality and efficiency of complex task execution. By systematically guiding the LLM through a sequence of well-defined sub-tasks, decomposed prompting aims to facilitate a more structured and accurate approach to sleep disorder classification.\u003c/p\u003e\n\u003cp\u003eEach prompting strategy encompasses two primary sub-tasks. Task 1 involves performing multi-class classifi- cation on data within a specified CSV file and generating a new CSV file containing the classification results. Task 2, consistent across all three prompting strategies, focuses on evaluating the classification outcomes using relevant evaluation metrics and generating corresponding visualizations.\u003c/p\u003e"},{"header":"4 Experimental Design and Implementation","content":"\u003ch2\u003e4.1 Experimental Environment\u003c/h2\u003e\n\u003cp\u003eThe construction of the experimental environment is the foundation to ensure the smooth progress of the research. Its hardware and software configurations have a significant impact on the accuracy and reliability of the experimental results. In this sleep disorder classification research, the selected processor for the experiment is Intel (R) Core (TM) i5 - 6400T CPU @ 2.20GHz. This processor has a main frequency of 2201 Mhz, with 4 cores and 4 logical processors. The operating system used is Microsoft Windows 10 Home Chinese Edition, with the version number 10.0.19045 (build 19045). The large language model selected is the Doubao PC version 1.41.6.\u003c/p\u003e\n\u003ch2\u003e4.2 Experimental Procedure\u003c/h2\u003e\n\u003cp\u003e1. Dataset Division and Sample Selection:\u003c/p\u003e\n\u003cp\u003e(a) Randomly select 30 samples from each of the three categories (normal, sleep apnea, and insomnia) in the original dataset, resulting in a total of 90 samples as \u0026quot;prompts 90 examples\u0026quot;. Save these 90 samples as\u003c/p\u003e\n\u003cp\u003eSleep_health_and_lifestyle_dataset_selected_90.csv (training set).\u003c/p\u003e\n\u003cp\u003e(b) Delete the above 90 samples from the original dataset file and save it as Sleep_health_and_lifestyle_dataset_remaining_90.csv as the ground truth.\u003c/p\u003e\n\u003cp\u003e(c) Make a copy of Sleep_health_and_lifestyle_dataset_remaining_90.csv, delete the last column (\u0026quot;Sleep Disorder\u0026quot;),and save it as Sleep_health_and_lifestyle_dataset_remaining_90_without_last_column.csv(test set).\u003c/p\u003e\n\u003cp\u003e2. Manually upload the above three CSV files to Doubao. This step ensures that the large language model can obtain the required data, providing a data foundation for the subsequent design, training, and evaluation of the classifier.\u003c/p\u003e\n\u003cp\u003e3. Design, train, and evaluate the classifier according to different prompting strategies.\u003c/p\u003e\n\u003ch2\u003e4.3 Experimental Results\u003c/h2\u003e\n\u003cp\u003e\u003cstrong\u003e4.3.1\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003eQuantitative\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;Evaluation (Table\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003e1)\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAs detailed in Table 1, we present a quantitative overview of the classification performance achieved by employing different prompting strategies. The metrics provided\u0026mdash;accuracy, precision, recall, F1-score, and AUC value\u0026mdash;offer a direct numerical comparison of each strategy\u0026rsquo;s effectiveness in sleep disorder classification. These metrics collectively demonstrate the progressive improvement in model performance as we move from zero-shot prompting to more refined strategies.\u003c/p\u003e\n\u003ch2\u003e4.3.2 Visual Analysis: Confusion Matrices and ROC Curves (Figures 5 and 6)\u003c/h2\u003e\n\u003cp\u003eTo complement the quantitative analysis and provide a more intuitive understanding of the model\u0026rsquo;s classification behavior, we present confusion matrices (Figures 5a, 5b, 5c) and Receiver Operating Characteristic (ROC) curves (Figures 6a, 6b, 6c) for each prompting strategy. Confusion matrices visualize the distribution of predicted versus actual classes, offering insights into class-specific accuracy and misclassification patterns. ROC curves, on the other hand, illustrate the trade-off between true positive rate and false positive rate across varying classification thresholds, with curves closer to the upper-left corner signifying superior discriminatory power and higher AUC values.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eZero-shot Prompting (Figures\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003e5a\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;and\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003e6a):\u003c/strong\u003eThe confusion matrix for zero-shot prompting (Figure 5a) reveals a considerable degree of misclassification across the three sleep disorder categories. Notably, a significant proportion of \u003cstrong\u003enormal\u0026nbsp;\u003c/strong\u003esamples are incorrectly classified as either \u003cstrong\u003esleep apnea\u0026nbsp;\u003c/strong\u003eor \u003cstrong\u003einsomnia\u003c/strong\u003e, indicating a limited ability of the zero-shot model to accurately discern normal sleep patterns. Furthermore, some confusion exists between \u003cstrong\u003esleep apnea\u0026nbsp;\u003c/strong\u003eand \u003cstrong\u003einsomnia\u0026nbsp;\u003c/strong\u003eclassifications. Consistently, the ROC curve for zero-shot prompting (Figure 6a) is positioned furthest from the ideal upper-left corner, exhibiting the lowest AUC value (Table 1) and confirming the suboptimal performance of this strategy.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e90-sample Prompting (Figures\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003e5b\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;and\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003e6b):\u003c/strong\u003eIn contrast, the confusion matrix for 90-sample prompting (Figure 5b) demonstrates a marked improvement in classification accuracy. The number of correctly classified \u003cstrong\u003enormal\u0026nbsp;\u003c/strong\u003eand \u003cstrong\u003esleep apnea\u0026nbsp;\u003c/strong\u003esamples increases substantially, accompanied by a reduction in misclassifications. However, the classification of \u003cstrong\u003einsomnia\u0026nbsp;\u003c/strong\u003esamples still exhibits some misjudgment, suggesting room for further refinement. Correspondingly, the ROC curve (Figure 6b) shifts closer to the upper-left corner compared to zero-shot prompting, and the AUC value improves (Table 1), reflecting the enhanced performance.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDecomposed Prompting (Figures\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003e5c\u003c/strong\u003e\u003cstrong\u003eand\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003e6c):\u003c/strong\u003eThe decomposed prompting strategy yields the most accurate classification, as evidenced by its confusion matrix (Figure 5c). The matrix shows a significant increase in correctly classified samples across all three categories\u0026mdash;\u003cstrong\u003enormal\u003c/strong\u003e, \u003cstrong\u003esleep apnea\u003c/strong\u003e, and \u003cstrong\u003einsomnia\u003c/strong\u003e\u0026mdash;with minimal misclassification instances. This outcome highlights the effectiveness and superiority of decomposed prompting in accurately classifying sleep disorders. The ROC curve for decomposed prompting (Figure 6c) is positioned closest to the upper-left corner and achieves the highest AUC value (Table 1), definitively indicating the best overall classification performance among the evaluated strategies.\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAw0AAADgCAYAAABB04ohAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsMAAA7DAcdvqGQAAFsISURBVHhe7d1fbBzHnSfw79jvmxvSTz6vYEyPAhgJoCAZSoFD6mABVo+UIHBgBRx69wABNuT0aNfA7TqkQ9HYM0z/6YlyBxgOSQk2IBwCzeisAMFhyXBoQAY4A2EtyQmJS7APyx4IOsMvyxHj5N11D55fbXVNd08P/4h//P0AA1vV/+tfV3VXFzNKKQUiIiIiIqIYD9kBREREREREJnYaiIiIiIgoETsNRERERESUiJ0GIiIiIiJKxE4DERERERElYqeBiIiIiIgSsdNARERERESJ2GkgIiIiIqJE7DQQEREREVEidhqIiIiIiCgROw1ERERERJSInQYiIiIiIkrETgMRERERESVip4GIiIiIiBKx00BERERERInYaSAiIiIiokTsNBARERERUSJ2GoiIiIiIKBE7DURERERElIidBiIiIiIiSsROww5rt9uoVCrI5/NoNpv24pBWq4ULFy5gYGCg57rUW6vVQqlUQiaTQSaTQalUwurqqr0aPSBMj80ZGhrC0NCQHUwPyMLCAjKZDBYWFuxFu0buK61Wy160Y9rttr4/ZTIZDA0N7ak4Ef3cc/eCVquFgYEBXLhwwV5EtOew09DRbDZ1YybNL21l9OGHH+KDDz5AEAT2opB2u41bt27h2rVr2NjYsBfvmFarhXK5jGKxaC/aFna8mb9isYhLly7Zm2yLVquFQqGAn/3sZ1BKYW5uDktLS/iXf/kXe1V6ALaSHkllc2BgAKVSaU82Xg66hYUFDA0N6XQol8toNpuo1Wqp60fq3+rqKs6fP4/nn38euVyuq0yYv3w+j3K5vC2di2KxiEOHDuH+/ftoNBrY2NjA//pf/8tebde9//77qe65RLQJipRSSjUaDTU6OqqCINBhvu8rAMr3fR0WBIEqFAqq0WjosF5kP2m26WfdrapWq2p0dFQBUK7r2ou3zcrKigKgAKj19XUdVigUFAA1Ojpqb7Jlk5OTynEcO5h2yXakx9zcnAKgCoWCDpufn1fZbFYBUHNzc6H195sgCEJ1zV4m9ZTE+fr6uqpWqzotouovz/PsoC3bT3G2HdbX11WhUAjdp5RVx5rrTk5OKgAqm82qlZWV0Db9mJ+fVwC6jrtXbcd9dCfyK9F+xzcNhjfffBO5XM4ODsnlcnjvvffs4H2pVCqhVqvZwdvuyJEj+v8HBwd12OLiIrLZLK5du7bt5/HJJ58gn8/bwbRLtiM9vvGNbwBGHgKA06dP4//8n/8DAPjJT36yLU9Ud8vFixftoD2p3W5jYmICk5OTePHFF4FOmpRKJXz00UfIZrP2JqjVajuSNvslzrbL+fPnMTQ01HWfMutYMTg4iDfeeAOjo6PY2NjAW2+9Za+S2h/+8Aegc//7Kmi1WpidnbWDib7y2GnoGB4eTl0hHjlyBMPDw3Yw9WlwcBBHjx4FAPzmN7+xFxOlYpbFDz/8MLRsv6jVavumkfKv//qvdpB25MgRlEqlUNjq6irK5XIobDvspzjbDs1mE9euXdMdtbSeeeYZAMC1a9fsRRSh3W5jdHTUDiYidhq2bnV1NfRxZz6fR6VSsVfTarUa8vk8Mp0PQdvttr1KJPs4D+oj0maz+cA/0mq32yiXy7Ef3EkjZGBgAO12G8ViEZnOh4ry//V6HfV6XceXOcZaPpIz08tMh1arFfqQ7sKFC8hkMrhw4ULXsmazqcd1F4tFvR/zg0E77uzrk/Hg5jnUajWUSiW9z3K5rM83arx4rVYLjS+/cOFCV96KGoNur9MP+zqGhoa63hilSY/t1m63dXxUKhXUajUMDAyEPiZOc+5mPgOAS5cu6Xwjadput3W5HBgYCO1jdXVV5wN08p0cz4z7SqWCsbExAMDExAQymYyuQ5rNZugcYEyYIHlhYWFBn1dU49z+AN3+9Wt4eBjZbBZvvvlmV5wBwA9+8AP9/wsLC3jqqaewsbGh01/ydFIa2fWdXa8mxVnU9lH1Zdv6sNf+NZtNtNttXLp0SZ+nvb2dh+xva8z6fmhoCMVicdN5/9133wVi3iqkEfUGCDH1obksk8lgYmICML5RM9dJqleS0jlpmUiTjoi4ht/97nf2Kqm0Wi0cO3YMd+7cAYzrNUVd71beopl1vei3jBM9MPZ4JfoPUd802LLZrBodHVXr6+tqfX1dfyNgjh+V/XiepzzPU77vK8dx9PhsGedvrmuOxVxZWVGO4+iwRqOhstmsymazoW03CwnfNFSr1W357sAeb6s6Y24lHsyx0YVCQf9bviGROAmCIDSOfXR0VHme15VOrutGXpPruqExwXHj5OWYrusqz/NUNptVruuGlklaNhoNPXbY8zw1OTmp5ubmVKPR0OvOz8/r/Y+OjirHcfQ52OPDV1ZWdFihUFCe56n5+Xl93dlsVu9Ldb4XMK/JzG+iWq1GXndUHKWxvr6ustms8jxP531Jh8nJSXv12PToR6PRiDxnczz3ysqKajQa+lxc19XxjU7+S3Pudj6bnJxU1WpVNRoN5bquTi/P81S1Wg2tu76+roIgUI1GQx93cnJS5xcz7wq5NjMPr6ys6LHkcu6qkz9lv2YelGswv+2QazXrKMmTW/kGROoFieNeY91lPdErjexzjqpXo+JM9VFfuq4bKodyPma6zM/P62Obx+lVT6nOeWSNbwkkzmR5v+w4tEl62CS/mvWB8Dr3JLsc2HEqdYqtV72SlM5Jy1Qf6eh16mdZT85hK3EtcWaT8ht1Tr3KQJRGo6Hj1kzbfso40YPUXSpIk8JsV6BCbshmAW50bmRmmN0oFGbjw17XrOxc11XValX/21wv7tz6YVdYtpWVlVAlvRlSiUvFGgSBvhmbHSff97tubhKn5jlK3MkN2f7Iz41opMpN267co26UdpoFQRA6R1gdAaWUruTNuJJzN/ctN0gTIjpmEjcm6ZyYNy3z36pzrtLoUkY+ta9b4nAzN1bP8yI/bDYbBKao9OiXXKuZXxpGx8zMN7KuGQdm4zDtuUscmWkaBEFk2sSVXQCh8ivpAyP/RuUTEdWAicqD6+vrClY5kQaUeU7SydpqepgNG4l/O4+JqOPFpVHaejUuztLUlxIH9raSLqao46Spp3zf78ojk5OTXWUjDXvfUSQdxPz8vM47ZpkRjU6n1mbvR8V0GtLWK3Hp3GtZmnSUTrUdp1Kn2+FpRZU5Kff2Ocl9JSlteonaPm0ZJ3qQODxpCwYHB3H//n28+OKLaHfmhj579iwA4PPPP7dX1x9yCt/3AQAfffRRKNzUbrdRr9cxNjamX5War4tv3Lhhb7Ltjhw5Evr4dCtkmI3jOFhaWoLneVhcXNT7/+CDDzA7Oxu61pGREQBAvV639vYfr+rTvLL/xS9+gWw22/XtiowRvnz5cigcRprlcrmuOPirv/qr0L/znQ997fVsi4uLeliHvJ4HgD/96U/Wmt37+trXvhb699WrVwFrXH8ul8P9+/f1MT7++GNsbGzAcZxQvEp83rx5U2+bRrvdxuzsLAqFgr0IP/3pTwHjvHZCEAR62NPIyAg2Njbg+z5mZmbsVfHtb38b6MRjLpfb9Lmb6SD5x06bJOY4/1wuB8/zACDVlLNJzDwYdT5R9VCaspLG6dOn8fHHH8P3fWSzWczOzsJxnK5hPL3YadRvvWpKW1/+5S9/sbb8knxj1Uuaeuqb3/wm7ty5g4ox/PHUqVOh/fTrO9/5jh3URc7n+9//Pm7duoVqtYrbt2935Y/f/va3CIIgdA0ZYyhOr2FU/dYrdjonLUubjjLlq/2N4eOPPx7693Z4//33AQBPPPFEKLxUKiGbzaJer29pmFKcXmWc6EFip2EbVCoVHDt2DAAwPT1tL44lN++oxqKQjw4bjQY6b4ZCv8XFRXuTPW1xcVGf+/379zEzMxOqCO/cuQPf97uuU35bcefOHTiOYwfrdHiQ83rLGNXl5eUtzcaV5iYlM5/YcSm/8fFxe5NEkiej4lI6WWnOa7OOHj2K27dv6/NfW1tLfQ27fe7ie9/7HpCiEbxVhw4dAgB8+umnOkwasHbDbTMGBwcxPj6Of/u3f9MdoYmJib47DlE2U6+mrS8fffRRAMDdu3dD27fb7ci8YUtTT50+fRq+72NiYgKHDx9GpVLBE0880dXATUMa4PZDgyhKKQRBgGw2i42Nja5Grvjkk0/gum7Xucuv13lud71iSpuO165dg+u61tY745NPPgEAPPbYY/Yi3dn87LPP7EVEBwo7DVs0NDSEDz74AB9//DHGx8cjK5TtYN70Dzr7Rr6d5AO33XThwgX87d/+LX79619jZmZm25789rLdjWG5ie5Hu33u9luqnVIqleC6LqampnT6v/rqq8hms3j55Zft1VOz/yjj4OAgZmZmUK1WAQBvv/12aHm/tlqv9qovc7kcfN/H7OysfqJeq9Vw584dvPPOO/bqkdLUU+Pj4wiCAKVSSXceoj7k7eXJJ58EUh4Tnev71a9+BQB46qmnYo+5trZmB/Vtu+sVU690xDZdQz+SZg8jOujYadgCucm89tprm3ptKJXtj370I3uRJk+Jrly5Yi8CIm7e+12hUECtVtNPQ01bvVZ5ghh3A30QT6xarRbefPNNeJ63LZ0FeVpsz9qCTv5EZ5gEAFy/ft1a4z9mN+mHPKW9deuWvUg7ceKEHbQn7JVzl8aQpM1OkuFDhUIBmUwGrVYLH3300ZbeNMQNqSyVSigUClv6q/ZbqVf7qS+ff/55uK6Ls2fPIpPJYGpqCvPz8zh9+nRomyj91FO5XA4zMzNYWVkBALzwwguh5f3op4Eubzo2Njbw7LPPdp3rd77zHQRBEDkMaWFhoeextrteMaVNx2w2iyAIuq5tJ0h5+eMf/2gv0nq9nSHa79hp2IJ79+7ZQfjzn/9sB8W6fv06stksnn76aXuRNjg4CNd1Ua/Xu6bIrFQqqV5Xb5VMPfgg/PjHP8bGxgaKxWLoptVsNmMb+1HW1ta6zvncuXNAxE1djiPjpndS1Otr+zz7IdNbvvTSS6H9tNttLC8vAwCOHTuGbDaLiYmJ0LW3222cP38+9Thukcvl4LouNjY2uhoG0hg+c+ZMKDwqPXbDZs59O9jXvry8DMdxUjVQt6LVauGpp57C4uIi7t+/r4d2bLXDeu3atdjyKOPcN2sr9Wo/9WWxWMRLL72EtbU1qM4wt7TpkaaeqlQqoWVHjhyB53mbGga52cbo+Pg4XNfV3wGZ8SHfV/zwhz8MdRxarRbeeeedUKdS3nCY2293vWJKm47yrdCrr76ql++U5557DgDw85//3F6EtbU1PTyP6ECzv4ymL60b04Has9oImbmhUCio+fl55fu+nnXB60y9qaxp6GS2FAmzZ2KQ7c1ZQmTqPgAq25n603GcbZlBQWYRcRyna3YNtU1TrsoMGVHXa1s3poSUOCsUCqFp9oIgiJzpRsg1wZqiUSnVNd1k0Jkq0ZwJZd2Y4tGeXUV1pky108jML+ZsIubMGuudaQ0lHavVqqpWq6FZpObm5tRKZ9pQCTPTJerYEuY4jvJ9X01OTirHcULXbk6RKXknm81GXl8aMgNQ1PSDdhonpUc/ZDYRO46jyLqjnWk7TWnP3UxT87zleuy0kRlbzClnpTyPjo7qc5ayb86KYs4iI3WJMmaosc9B9mueb1RZlngoFArK7cxgJVM42unUD3TysO/7+rqCzoxoWWOaUXN9pzN9ppS1uDRKW6/GxVma+lK2lXC3M6OZ7/tqbm4udD7meYo09ZTfmT1J8pjUW1FTEqch+SuKXI+dJ5SVjyVOhexTlkk8yznb29szAKapV+LSudeyNOkoZVn20Wg0VLVa1efrdaar7pfEw/z8vJqcnNR53JxaW+pzz/O66oJ+mHnR3EfaMk70IEXXQF9xZuPE/JkVoZBKxHEcVa1WdSUW1WgzbzKjo6ORN1bzZ1aOKysrunGYzWbV5OTklisNqZSSrlEaU5u90dn7l18SqYjNm4E9RV/cOUddk3m89fV1NTk5qfftdBraJntbMx2ilsWdU9S5qE5eyHYarRKv0uCoVquR+S/qGOZ5mdcUlbeU9fcnHMfpagD0SxqJcj52g0SlSI80oq7dvn5T1DFtvc496pi+7/edNnIuZvqYDUmTNODcTgczar9RYa7rxp6XNPjMusfedjPm5uZUEATKN/7mDBLynjQupV7slUZp61U7zkSa+nJ0dDQ2XrKdxr99nmZ8JdVTqnMfMbfPbqGTrow8aecd+9zlZzIb4PZyMw3tvBmVr+xzSKpX7Pgzj5u0TKRJRzM/SRz7nQ6b3YFKS+JL6mTT3Nycjq+s8fdeNiMqflVEmiaVcaIHKaO+zKBERLTNisUi6vU6dquardVq+Pzzz/W0wqbV1VU8++yzD/xD0r1gdXUVly5dipymt9VqYXJyEn/3d3+36WFBO6VUKmFgYCDyvImIdhq/aSAiOoBarRbK5TK++93v2ouAztSRJ0+etIO/El544YXY7zpyuRwGBgZipyrdTb/85S9x+/bt2O9JiIh2Et80EBHtgHa7jWPHjiEIAqysrMQ2UnfK6uoqvvWtb8FxHJw7d05P24nODDAfffQRfvnLX/Y9Q9FBMDAwAADwPA/f+9739BS4n376KX7zm9/gZz/72QNPr7Ta7TZeffVVvP7661/JtCOi3cNOAxHRNms2m/ovBAvf97f0B682o9ls4t1338XS0pKeBrVQKOCFF17AmTNnvrKNzlarhYsXL2JpaUnPZuQ4DkZHR/H8889vaTraB6HdbuP69et4+umn9/y5EtHBwU4DEREREREl4jcNRERERESUiJ0GIiIiIiJKxE4DERERERElYqeBiIiIiIgSsdNARERERESJ2GkgIiIiIqJE7DQQEREREVEidhqIiIiIiCgROw1ERERERJSInQYiIiIiIkrETgMRERERESVip4GIiIiIiBKx00BERERERInYaSAiIiIiokTsNBARERERUSJ2GoiIiIiIKBE7DURERERElIidBiIiIiIiSsROAxERERERJWKngYiIiIiIErHTQEREREREidhpICIiIiKiROw0EBERERFRInYaiIiIiIgoUUYppezA7fLf//t/x2uvvWYHExEREdED8M477+Dv//7v7WCivu1op2G/GxoaAgDcvn3bXkQALly4gOeffx65XM5eREREREQHyJ4antRut1EulzEwMIBMJoOhoSFcunRJh3/VDQ0NoVgs2sGxKpWKjst8Po+FhQV7lVjNZhNDQ0PIZDLIZDIol8totVqhdd544w1cvHgRq6uroXAiIiIiOlj2TKeh3W7j8OHDuH37Nu7cuQOlFN577z189NFHeOSRR7oarNstqlNy+/btPfOWoVKp4M6dO3ZwrAsXLuDGjRt45ZVX4HkegiDA97//fTSbTXvVLs1mEz/84Q8xNDQE3/dRKBQwOzuLQqHQlQ6vv/46XnjhBXYciIiIiA6wPdNp+B//439gY2MDi4uLerjLkSNHUKvV4Hmevfq2qtVqXY3hvaTZbOJ3v/udHRyr1Wrh0KFDWFxcxPj4OGZmZjA3NwcAuHnzpr16l3fffRcfffQRZmZmMD4+jtu3b2N0dBQbGxu4ePFiaN3BwUH84z/+I5599lm02+3QMiIiIiI6GPZMp+GTTz6xg7SXX37ZDto2q6urkW8Z9op2u43/9t/+G375y1/ai2J97Wtfw4svvhgK+8Y3vqGXJWm323jmmWdw5MiRUPibb74JdDoktlKphCAI8P7779uLiIiIiOgA2DOdhh/96EcAgGKx2PXEOpfL4Tvf+U4oDAAWFhb0uPuBgQGUy+WubVutVtd3EjJEZ2FhAU899RQ2NjZQr9eRyWT0NwPNZlNvZ1tdXUWpVNLj/UulUtewn1qthlKppK+nXC7rbwvsdZOcP38er732GgYHB+1FsaLW/eMf/4hCodDVmbANDg6iVCrZwfrtz3/6T//JXgQAcF0Xly9ftoOJiIiI6ADYM52GM2fOoFAo4M6dOzh8+DAuXboUWv7GG2+E/l2r1fBP//RPuHbtGpRSeOuttzA7O4u/+Zu/0eu0Wi0UCgUcOXIE9+/fRxAE2NjYwMjICFqtFk6fPo379+8DnUavUgqLi4tYXV3Fn//8Z8zOzmJjY8M46pediW9961t46qmnoJRCEAQIggAjIyP6Q+PV1VXcu3cP165dQ7vdxquvvoof/OAHmJ+fx/379/HDH/4wtM84CwsLcBwHp0+fthelJp2mn/3sZ1hcXLQXpyadsWeeecZeBAD4zne+gyAI+G0DERER0UGk9pD19XXleZ4CoAAox3FUtVq1V1Pr6+sqm82qIAhC4a7rKgCq0Wjof7uuG1pncnJSAVArKys6DEDXesrYn8lxHOV5XigsCAIFQGWzWbW+vq7DAahCoRBaV44v5xgnCALlum7X/qLOM06j0dBxKediXnc/qtWqchzHDtZ831cAlO/79iIiIiIi2uf2zJsGdIbGzMzMoNFowHVdBEGAsbExDA0NhZ5gf/zxx9jY2IDjOHqIUCaTQb1eBzof+7ZaLdTrdZw4ccI4wpdvLJRSXWP201hYWEAQBF3b5nI5/aHwhx9+GFpmDxXq9U2BKJfL8H2/a/t+DA8PQymF+fl5/RbnhRdesFfrqd1uY2pqCr/+9a/tRdqhQ4fsICIiIiI6IPZUp0EMDw9jcXER8/PzcBwHd+7cwVNPPaWHyPzhD38AvnwFEPkbHx/HZ599Zu116+S4f/3Xf20vwre//W0AwL179+xFfatUKvjRj37U1TnZrNOnT+P27ds6LvsdQvTqq69ieno68Xwee+wxO4iIiIiIDog902mwv2FAp7H78ccfo1AoYGNjA9evXw8tj5rJ50GQzsNOmZiYwE9+8pPQW5RMJgMA+oPtSqVib9bT9PQ0AOAvf/mLvShWrVZDNpuN/DiaiIiIiL4a9kyn4fPPP498Aj44OIjXXntNrwMA3/zmNwGgqxOBzlCaWq2GRx99FADwwQcf2KsAncZwv2QITtLfTHjyySftoL75vh/5AwDHceD7/qaOI28DJG56qdVqWF5e7voIPUqav/9ARERERPvTnuk0AMBbb71lBwEA/vznPwNGo/3YsWPIZrOYmJgIvaFot9s4f/48jh49ilwup8fx228xKpUKnnjiiVBYGk8//TSy2SyuXbvW9Zbj7t27cBwHw8PDofDNGB8fj/wBQD6fx/j4eOg4q6urXecT5ebNmxgdHdXTp6ITZ1FTwEqHYWZmJhQunbI4m+nMEBEREdHetqc6DdeuXUOxWNRTl6LTeC2Xy/A8Tw+RkQ+mAeAnP/kJ8vk8isUiDh8+jG9/+9u6Ufzee+/pdcrlMiqVCorFIu7evds1Pn9tbU3/bQZ0Gse3bt0COo1ydI77q1/9CgAwOjqqG+qXLl1CrVYLfSgsDfF2ux362xHyluKPf/yjDtuK1dVVfOtb30KhUNBhFy5cwMDAACqVij72pUuX8MEHH3T9kbjz589jZGQk1LGSOH/88cdRqVRCv2PHjuHo0aOhfQDAjRs3tq3TRERERER7jD2d0m6Zm5tTSik1Pz+vRkdHQ9Ouxk3jOT8/rwqFgl5P9mFaWVkJrRO1r2q1qpevrKx0TVVqT5HaaDT0PgGo0dHR0FSmMv2ovb0d1s/0qSpmytUgCFQ2mw1N7To/P68cx9HHKRQKyvf90PStQqaAnZ+f19va52n+7ClkRTabjYxbIiIiItr/MurLxijRpi0sLOCll17Cxx9/vKUpYomIiIhob2KngbZsaGgI7733XteQLyIiIiI6GNhpoC25cOECvve97+H06dP2IiIiIiI6INhpoE27dOkSvvvd7/INAxEREdEBx04DEREREREl2lNTrhIRERER0d7DTgMRERERESVip4GIiIiIiBKx00BERERERInYaSAiIiIiokTsNBARERERUSJ2GoiIiIiIKBE7DURERERElIidBiIiIiIiSsROAxERERERJWKngYiIiIiIErHTQEREREREidhpICIiIiKiROw0EBERERFRInYaiIiIiIgoETsNRERERESUiJ0GIiIiIiJKxE4DERERERElYqeBiIiIiIgSsdNARERERESJ2GkgIiIiIqJE7DQQEREREVEidhqIiIiIiCgROw10oCml7CAior4opTZVl2x2u6+6uHiLCyfa6w5K3mWn4StAMmvc76BaWFjAQw89hIWFBXvRAxUV3wc53ok2wy4n9m+3tFotDA4OYmpqyl6UaLPb7RQ7PqN+e0Gr1cLU1BQGBwfRbDZ1eLPZxPnz5/HQQ5trttjXGveLk7QsSdr9HyRprtOOl7j46bVc2OslrWuyt4nbLs06cWq1Gk6dOoVTp07Zi/adzZU+2jcqlQoefvhhPPTQQ/r38MMP4+GHH8bhw4dx/vx5rKys2JsdGJlMxg56YJRSqFQq+PrXv46HH34YX//61zE1NYUgCHD+/Hl7ddpnqtUqqtWqHUybJPWSXU89/PDDOHXqFObm5vq+WW+XTCazqbpks9vthKT4lXtBEAT2Zg9Uu93GrVu3cO3aNWxsbOjw1dVVfPrpp5idnQ2tn5Z9HzSv3fzFdUiq1SoOHz4c6sSkMT8/j6NHj+Lhhx/GI488gvPnz6PRaKBarfa9r/3k/PnzaLVadrBmp4ekyenTpxPXe/jhh1GpVELrrK+vY25uDqdOndLp+PWvf13n57m5udi4bjabkecRtX6v84izsLCAe/fuoV6v24v2J0VfCaOjowqAmpycVEoptb6+rubm5lQmk1HZbFYtLy/bm+wrV69eVY1Gww7eVa7rqoGBAX1eQRAo3/dVJpNRUUVPlm+3ndrvV53jOMpxHDuYtmBlZUUBUADU+vq6DhsaGlIA1OjoqL0J9cGMX7G+vq4uXLigAKhsNqtWVlZC2+wG3/cVgK463XXdyLozjfX1dX0fjKoP5+fnVTabDYWtrKyoubk5lc1mI88nydtvv60AqLm5OaU6x69Wq2pgYKDvfe0nQRAoAMrzPHtRF0kPx3F0ebetr68r13VVoVDoWuf3v/+9chxHDQwMqGq1qpf3E9dBEOj6xXEcFQSBvYpSRv7JZDKqWq3ai3sCoFzXtYP3nc2VPtp3pBK2K0sJ38+Nny+++EI5jpNYMTxo8/PzCoCan5+3F6lqtRp54/M8ryt9tsNO7ferrNFoKHQaX1FpTJsn8WpaX1/XDbfN3LDpP0TFrzIacHuhY7YTnQaVcB8USeFR5xNnfX1dwXhIZ1pZWVHZbDb1vvYbiSsYHf84ci/s1cHwfb8rbX7/+9+rbDarBgYGYo+TNq6lPrc7jbbJycmu80gLB6TTEP0ujr4ynnzySQBAEARYXV3V4TIMIG4MX5rxfWZ43Lpx4bLM/P+o9ZRS+jVk3HKbuV7cfk3mOvYvzh/+8Acg5vilUgmO44TCqtUqZmdnI/ebdK69zidpv0ixPSLW6fWL2/YgmZ6eRqlUAgC888479uJISfEk4taJCksTbv+/vTxqmSluvbhwe9l2GRwcxNGjRwEAv/nNb3S4eZyoYyadp6nXer3Cey2P0u+2Setuh2eeeQYAcO3aNXuRluYc0l5X0jq7ZXx83A7alH/913+1g7QjR47ousOWNl7SriP/jVo37bH60W638fbbb2N0dBQA8P7779urhDz22GMAgMcff9xe1NMrr7yCjY0N/PKXv8Tg4KC9GOjE9SuvvIJPP/3UXhQyPDwM13WxsbGROOT02rVreP755+3gTcVl3Ppx4fayqOUPAjsNFNJqtVCpVPT4zampKTz00EOhj/lWVlYwNjamxw+OjY2h0Wjo5e12G7VaDUePHkWlUkGr1dLrHz58WH9DUavV8PWvfx0PPfQQTp06hXa7Hdr+1KlTqFQqWFlZwalTp/DQQw/h8OHDmJ+f18c6ffo05ubmAAD/5b/8Fz0esd1u49KlS/ocRK1Ww9jYGE6dOoX19XWcP39en5d5DaLRaOgxqfaYRnv8pcl1XQDAf/2v/xUrKytdBdy8aVQqFfzN3/wNAOBnP/uZHi/ZKy3sdDh8+DB83++5XzE/P49jx47hYWO87fr6ul6OTmfSPIZ5/RIH5nhQM04qlUpk+H63urqKtbU1vPvuu8hms6jX65FjYIVSCrVaTce1fBxrx3Wj0egqVzLG3Ix7OZY9Hlesrq7i/PnzGBwcxPr6ui47MiGAOc764YcfxtGjR0NlSqQ9HztfSZ6ww7fbwsKCLssrKys4fPgwBgcH9VhqKd+PPPJI7HWmSRuzzjBVq1X9vdLRo0dx6tSpUD6I2w4xZc/8nkA+Bpayv7CwoOvKnf4eKpvN2kFQ1vdZdl0jkvIMItJkcHAQ58+f13X/blldXU0sw/0aHh7GwMAA3nzzTVSr1a76/wc/+EHo3+jUtXZ+te9Jcelg5tde9w3E5D+7PtqM69evo1Qq4e/+7u8AAG+//faOpG2z2US9Xkc2m43tgInnn38e9+7ds4O7vPTSSwCAV1991V4EdMrzyZMnuzooaetTU1zdXTG+4bDrDft+PzY2tjvfo9qvHuhginstOzs7qwCogYEBpTrDagqFgkLnVVq5XFYDAwOqWCwqpZRaXl5WMMZpmuMB//mf/1mpzqs+z/MUOq+6y+WyajQa+lWk4ziqWq2qCxcuqEajocrlsoLxKrfRaOjz9TxPDQ0NKd/3ValUUui89jTH3Ua9Op6fn+8av7qysqLXLRQKqlwuq/n5eT2W1X41Ka8sze1ljGSacb/lcll/vzA5Oan+/d//3V5Fs4+lUqRFNptVo6Ojan19PTRe1zy3qP2qzjcgQ0NDevzm3NycPo6QISHmMSStzXG6kn5Rr14nJyf3xHCH7WQO95J8nvR6fXJyMhTXZt4W//zP/xx6jS7pZg4blPQ183kQBHrYjvxb8jOM8pfJZJTv+3pMu5x/EATKcZyuvN/rfMxyb+ctpZQqFAo6j/QLEcNn/v3f/105jqPznow1R6csF4tFVSwWVSaTUY1GQ62vr4fOwTxfM/56pY1ZF5n5W4Y9SFmTus2Mr6jtVGesu/mtU6PRUAMDAyqbzaq1tTWlOmVfrlfyW8OoKzcbtyomfr/44gs99KdcLoeWqU4+L5fLuh6Q83j77bf1Or3yjOrkYXPcuMSRfT1RdbraxuFJb7/9tvriiy/0z/O8rmOZ4s4nSbVa1fW/67pqbW1NffHFF/ZqSiml1tbWVDabDeVXSX9zjL3rupH1dqFQ0Ov0um+kqfs3yxwmLOd/9epVezVN8khUHWLyreFJk5OT23bOpnw+rxAz5NR13a77ftr6VHXKnX1/lTgyydBmc93f//73Kp/Ph8qWtFnihmbtlM2XPtpX7Juh6lRq0rgwxwnbFXkQBDpjOo7T1UAKOh8+mRlYKgN7TKc0suxCaRcS2d7++EkqC7MhGlehx1VIsl+T7NfchzTSTFLB2vuM4/u+juNsNqt8348s5HHnGpcW0qA3b7ayj6gwc7+yrXkzUsYNWeJArtWME6kkoyo/s3Egoira/UzyusSdxIcZZpL4N+Mw6DT0JQ9LethpXygUQjefuHwe1ZCSMIl7+a+cT1SnW8LSno9cu10fSH7YLDs+gyDQZdGuD2B9RCnX4Pt+13nJtUveTZM2wtxOdfYfVYfYaWNvJ/nH/i5DOh32MWDVleudsfJbaSxJ/Ir5+XmdX+z4VZ14ikpPcz9p84zrupFxa4f1k9f7IfuN+tnHMsWdTy9m5w+dshJVT7iu25Wmck+yO6b29nJPNeO+133D3odd929GtVoN5RO5f0TlHSFl0M43Nt/qNMj52mV8q6LKoeqcpx0m4WYaqYj6VETtNy4/2+u6rttVZ8hxesXdduPwpK+YpaUl5PN5ZDIZjI2NwXEczM/PR77i+8Y3vgEAyOVyGBwcxMLCAoIgwJEjR0Lr5XI5jI6OYmNjAx9++GFo2de+9rXQv2Xs4l/91V+FwuP8+Mc/Dr0O/Id/+Aegcx1bYb9itM8TAP70pz/ZQTpO0hofH8edO3fgeR42NjYwMTGBw4cP9/23I+y0GBwcxP379/Hiiy+i3W6jUqng7NmzAIDPP//c2jrs448/xsbGBhzH0VNCZjIZPSXczZs3gZj92GmPTlxOT08jCALUajUdLq/7o7bZry5evAjP85DL5YDOtclQtKjxu1evXgU6wxVELpfD/fv3dVx9+OGH2NjY0N8Xidu3b+P+/fuhsH5J3Mt/h4eHoZTCkSNH0Gq1UC6X8fbbbwMA/vKXvwB9nM+RI0fgeR5mZ2dD0yu+//77OHfunP73ZpXLZWQyGTiOg6WlJXieh8XFxa6ym8/ndZhc5wcffIDZ2dlQ/h4ZGQEAnc/TpE2cb37zm7hz5w4qlYoefmEPJ4gieeSJJ54IhZdKJT3UzZ6q0qwr7WvfComX73//+7h16xaq1Spu377ddYzf/va3CIIgFJfmNLLNZjN1nllcXNRxW6vVMDQ0BMTUtTvJ9310HppCKYXJyUl7lW1x+vRpfPzxx/B9H9lsFrOzs3AcJzRsr9VqoV6v48SJE6Ft33jjDV1WAeAXv/gFstmsrnvEiy++CAC4fPlyKBwR9420df9mTE1NYXp6Wv/7zJkzyGazCIKg7/tdWpv5FiKJfG9oDzm9evWqHr5kSlOfblW73Ua9XsfY2FgozSYmJgAAN27csDfZUew0fMWcO3cOa2trurK8fft26vHm8nHvX//1X9uL8O1vfxsAUo0d3IrBwUEUCoXQHN47RSpn8yb+5z//GQBw6NAhHdZLLpfDzMwMVlZW9MdW3//+97dlDG2lUsGxY8eAzse5aZgfaUf95GNAuUbzIzJpINk3Lqlsp6am9DpXr17VHZmDQL61ee6550Lhco2zs7Nd43ftBmCUnS4ztna7jXK5jJMnT+L48ePwPC+0vJ/zefnllwEg1Oi6fPkyzpw5Y6y1OYuLizpP3r9/HzMzM10N2jh37tzpahiaP6RMmzinT5+G7/v6IUClUsETTzwR6oBE+eSTTwDjA1CTfOj92Wef2Yt2hFIKQRAgm81iY2OjqyMjPvnkE7iu2xWH8hseHu4rzywsLCCfz2N5eRnvvfeevXhXRH3cul0GBwcxPj6Of/u3f9NlbWJiQncc0qb3nTt3uibQgNFRTvM3NtLW/f1qNpu4f/8+nn76aR02ODjY92QRacn9ZycazD/96U8B437aarWwtLQU207qVZ9ulXxU32g0utJLKYXFxUV7kx3FTgP1TSqe3ZK24bBVL7/8MhzHweTkJNrtNtrtNv7pn/4JhUIh8s2MqdlshmajQqdyX1xc1JXKu+++G1rer6GhIXzwwQf4+OOPMT4+HtkQSdKr0VQqleC6LqampvS6r776KrLZrG4smuRtw/vvv68r2l7xtJ+8//772NjYwMjISOiJz9jYGABgY2MD169ftzfbU1qtFg4fPgwAWFtbQ6lUinzLllYul4Pnebh27Rqazab+WNDuVO6Gu3fv2kHbanx8HEEQoFQq6c6DXebjJM2u8yDlcjn86le/AgA89dRTsee/trZmB23KhQsX8Ld/+7f49a9/jZmZmT3zFjKXy/Xs8PXr0qVLoX8PDg5iZmZGz84jT6T7cefOHTtoU3rV/f2anp7GxsYGHnnkkVDdKH+Mr16vx+YtpGz8m/XU8ePHgW3MlyZ5QyLn/P777+uOhG2769MkvWaAelDYaaDU5Mnz7373O3uRZr+e3glra2soFAp28LbL5XL46U9/ij/96U945JFH8Mgjj2BoaCh1z/5f/uVf7CAAwOuvvw5s8ZV8rVbDnTt38Nprr/XdifrmN78JdGa6sMnTdCGv1QuFAjKZDFqtFj766KPIRqG8bXj77bdx8eLFbRmispe8/fbb+q8S2z950v7zn/88tI3EU9TreYlnKVe//e1vrTW+vCltxxspcfHiRWxsbOg8GKXf85EO5PT0NH7xi190vYnZDYVCAbVarevND4zGXJq06cV8iwgAL7zwgr1KiBzzj3/8o71I2+7Gay/y1mRjYwPPPvtsV5x95zvfQRAEXemOTty1Wq1UeabVauHNN9+E53l7prOwkz766CM7COjUk+bb8kcffRToDKmLInlR3jLENb5lmGSSfur+tFZXV1Gv1/UsgfZP7tV2J8pk5znb7373u9DQ4KeffloPfYrKl6Z2ZybFtAYHB/HKK68AAN566y1cu3Yt9s1pmvp0q+QN4JUrV+xFQI943QnsNFBqUlCvXbvW9aTi7t27cBxn22949rj6VquFIAjwj//4j6HwnbCwsID33nsvNEyinyESP//5zyMrQ7mmqIZ3WlHDAWToVC/Hjh1DNpvFxMREqMJpt9s4f/68HibRarXw1FNPYXFxEffv39evQpNu+PLEaXZ2dkdf+T9oly5dguM4evyw7R/+4R/0Tcy88crUii+99FIoL7TbbSwvLwOdcgUAb775ZleDYHJyUpcp6ZDbw8X6edpml1tElLG05yPkbYOMi7aX74Yf//jH2NjYQLFYDF2z+QYwTdrEqXSmRBbyfUevISLSobI7l+g8DNnuoQ1pjY+Pw3VdBEGAYrEYig/5VuOHP/xhqIHWarXwzjvvIJfLpcozUcNwourHg+LatWtdcSHkuwJ0yk+hUMCdO3e6GoAy7A2docWIaCRKPkwzFDRt3d+Pt956C5OTk7H3hf/5P/8n0Bm+adc/w8PDcBwHd+7ciY2r1dVVLC0theqVwcFB/Ybs7NmzXfs1vfrqq7GN/jhy77p27VrkNKsi6rh2fZpEvmMx92PHw+DgIFzXRb1eR7lcDpWZSqWyY282YtlfRtPBs25Mxxk1O4bJXDfqq3yZDqxQKOgZGObm5lTWmH5QwhAxK4bs25wJwJx9SfYpsxJks1m9bhAEqlAodM1aYs4iUK1W9foSbp6D7NeOBzkvc+YhmdlAZrZwXVf/Rches0zIcZzO9LJyrEajoa/BPL6sPzo6qubn55XfmWUpLi3MdJD1zRklzOlr7f0qY5YIOUfXdbtmP5H4KxQK+vpl+kd7JgdT1Axb+1mjM71dr2sqdKY5zFp/gVTS0HEc5fu+mpycVI7jRJaXbDar85jkHSHlRNK8Wq2G0n1yclKtrKyowJiu0c6n5uxjjc7UyHLec3NzOv3TnI8pblagfkl+TbMvs4zZs8Gsd6ZclX25rqtn8okq90lpYx5HtvU7sydJ/Eqcm7PFRW2njDTwPE/PaCPpYK4n6WrGg8xYZe8zraT4XTemgJQ8JmSGHlkm52bmr155Zr0z85PU6dVqNXRfmutMpauMazfr43Xjr4LbM9OkYdandlz3EnV/6EWu1fd9nT+Dzkxg9v1S0hXGFLtS35rMsqqMe6K5XtJ9Q6Ws+9OSNLfzkknSXeLdLqtyL3McJ5TngiBQfmf2QXu2RSHXYsezMmYF20xeUUaet8/XlLY+jasL5Pxle9/3dXzINa2vr6uVzhTPEu66rk67B42dhgPOj5liLq6CsNeLypSNTsNX1hkdHQ0VTKnwzZ+K2XfU+TUaDV3IRkdHQ42xycnJrsp+3WggyE3bPoekY9lhcs1SUOVGav/i4lB14mhlZUWtrKwoz/NCBV4aCzappFzXDVW09nkJqbDkxhx0pou0Gz32fsW8MZ+34zhdN0RpRJhpnXQ+YisV9V4TlT+i2OvAyh+Tk5M6D9jlRVQ7UxYiotFmrpPtzM9t5nXP89TKykrk+Zrnsb6+rstGoVBQKysr+iZl54805yMkr2yFfd7yixJXlk3rnca4Ge9RDYCktIk6joSbdYzd6IrbTszNzem4jaoT7G2T6q+07G2jzstsnNjL/U4nAJ38EHXsXnkmKv8WOp05aXja5+e6bs983UtU3KXZR9RxEZMnbXNzc7rha95D7DwmVlZWQvVx1Lmtr6+H8mvUeva5RtXTver+NOw4jTpOXPzZeUfuk2Y8SbmIiitTEARd2xYKBeXHTG+elnTwkqSpT+14sq/fvo+rzt9g8q1O0MrKiu4MSvnZyvVtVkZ9mcmI9pRms4mRkRH4vr/pGR22qlwu48UXX4x87bqwsIB33nkn9fcN+1GtVsPnn38eOSRndXUVzz77bNfQmFarhdHRUdy+fTsUTgdbrVbD//2//xdvvPGGvYiIiA4IftNAFGFhYQG1Wi2ywwAA//k//+euebUPklZnzunvfve79iKgM2XkyZMn7WBMTk4+kO9NaO9ot9uYmpo6UN+wEBFRN3YaaE+S2UWSZmraSf/v//0//SHlpUuX0Gw29e/ChQv43//7f+/aG5AH4S9/+YueTaVSqYSu/9KlSzh//jxef/11tNttDAwMIJPJYGhoCH/6058O1DSrFK3ZbCKTyWBgYADHjh3DuXPntvRhPxER7X0cnkR7TrFY1DOxiN3IprVaDVeuXAmdy+joKJ555pmvRMO42Wzi3XffxdLSkp4esFAo4IUXXsCZM2f0jBJDQ0P6r16//vrrsTNN0MHRarX0VIqvvPLKge5AExHRl9hpICIiIiKiRByeREREREREidhpICIiIiKiROw0EBERERFRInYaiIiIiIgoETsNRERERESUiJ0GIiIiIiJKxE4DERERERElYqeBiIiIiIgSsdNARERERESJ2GkgIiIiIqJE7DQQEREREVEidhqIiIiIiCgROw1ERERERJSInQYiIiIiIkrETgMRERERESVip4GIiIiIiBKx00BERERERInYaSAiIiIiokTsNBARERERUSJ2GoiIiIiIKBE7DURERERElIidBiIiIiIiSpRRSik7cCsymYwdREREREQPyDY37YiAneg0EBERERHRwcLhSURERERElIidBiIiIiIiSsROAxERERERJWKnoU+tVguZTAa1Ws1eRDEqlQo/kCciIiLax3a905DJZBJ/zWbT3mTXtFotOI5jB29auVy2gwCjkR31q1Qq9up7jn1d4+PjnMlhE4rFYupyUKvVuvJKJpNBsVgMrVcul/WyXh3ffD7ftU6z2dx3+dHWT7za4uLVjpd8Ph9aborbh1nu91K8mnnLLttR8vl8KC6i4loevvSKK3SOH7WOeZyo5btls/nLjBMzntPEp33PsMstIuoIO//tV2b8tFote3GsrcQ3jDi3w0U/dS3RvqF2URAEynVdO1h5nqcAqEajYS/adUEQKACqWq3ai/rSaDQir93kOI7yPE//W+LFDNtrqtXqnj6//cJ1XR2PkueSykNUnHueF8qnnueF8lxSPpa8Zi63y2u1WlUAlO/7Omyv6zdehe/7iWXPdd2e+0nah+/7ynEc/W/HcfZEvEoaCzP+olSr1a54CIIgdG12vNvXbpJ17eWu66ogCPS/o9bZDZvNX1Le7DRPE5+SRhIfccd1HCcUZweBWU4ajUYoHpJsJb5V57gAIuNZ9VnXEu0nu9ppaDQaXQVprzdEtqvT4DhO350GCUtbMe4GxDSKKD25+Znsm5ApCILI/GDuQ/KtuV5cY61arSrXdbvyeVSed1039rz2mn7jVUhjP6pxoFI+AOi1D3uZ3VjfLY7jhNK9V8Ms6vp83w/V557nddURdl4TUk/anQ77OHbDeTdsNn9JAzbq3O3rVDHxaZdjO92qB/BhTrVa7bruXp1atQ3xLSS97W36qWuJ9pvdvysZpLDFFS5pyMDqVJgNeakQpCBLwY7aLonc5OUJolTA5rFknaiKI+m4Zri9zBTVaZA4kOPJOnZnS85TflH7cTtP6+x1JA7t6zIbMrJc0so+HoxGgN2wNG9ysm7UjdXcV1QFf1BFNTSiGiRJGlZDNuqmFXVzU8YTSTMN40Q1APeqzcSrLE+KB7NesusBlWIfcR2EuP09KHLeUfkj7lqiSB0qouIiKh/5vq/rdDvv2nql44Owmfwlad9POtvxadf9dtlX1pNxOz33KzeigxDVkTBtR3wLSVt7WT91LdF+s+vfNJhOnjwJAFhaWrIXIZ/P4+zZs1BKodFoYGJiArVaLfSdwZUrV3D8+HG9TbPZxMjICIIgCG3Xa1xus9nEjRs3oJTC1NQURkZG7FUwNTUFfHk3gOu6mJ6e1st6HVe2cV0XSimMj4/rbXtZW1sDADz66KPI5/MIggCtVgvLy8twXRcwvr1oNBpQSiEIAszOzuoxrMViEfV6HWtrazh58iSUUqhWq3qd48ePd11XpVLB2NgY0BmTrb7scCIIAhSLReRyOSil4DgOPM+DUgqlUkkfS5TLZczOzob202g0UK/XQ2NDM5kMfN/X5+Y4zoEah5uk1Wohl8vZwUAnb6Vx9epVnD17Vv/77t27seO+b926pf+/WCxGlr84S0tLeO655+zgPWkz8To9PQ3HcXDv3j09PtnOg1NTU1BKwfd9jIyMdC3vtY979+7Ffit18+ZNO+iB+fTTTwEgMs6Wl5ftoEitVgtra2sYHh7W/waAxx57zFozXO83m03cvXsXpVIptE6cmzdvwvM8O/iB2kz+mpqaguu6uHr1qs4bSfcnOz4BoFQqwfd9TExMIJPJYHp6GouLi6FtlpaWoJSC53lwHCfxGPvF2toaHn/8cTtY3xOjbEd895K2riXaj/ZMp6FcLiMIAlSr1a6Kt9lsIp/P6xvI8PAwHMfBlStXkMvlEAQB0Lm5lUolKKUwPDyM6elpeJ6n9zc8PAzP8zA7OxtbqQhpnA8PD6NardqLce7cOd3YP3HihF4fnUbCZo+bpFKpIAgCve+1tTU4joO1tTXMzMxgcXER4+PjuHjxIlzX1RVdLpeD7/u6Yb64uKg7GHLeR48eBQCcPXtWx7McA52PmX3fBzqdHiH7jbsu81gAMDMzA8/zdIcKnfiB0UiRG+yZM2eAzk3RcRz4vh+6GVK82dnZ1A0uUavVcPbs2a7yF6fZbOLkyZN93VD3G7PDa3ZwzYaGXL988F+v10MfMqfZx0F1/fp1nDt3zg7uaXp6GjMzM3ZwrMuXL/e1/l7QarUQBAHW1tb0gxp5eBP3IXxcfI6Pj+uOp5nf0KnHpUzPzMzoh0hftY9ztzO+ib6q9kSnoVarYXZ2Fp7nRTZ0bt68iXq9rp8MZDIZXfhN5lsGdCpP+0mEPBX97LPPQuGm4eFhBEGATGdWjlKp1HVehw4dCv1bOi7YwnGjzM7O6muemJiA53ldN0d5QyOWlpa6Gn7SAJeGOTpvb7bqySefBDZxXVFPVu/duxf6t7nPfD6Pu3fvhpZTtFqt1vdTV3lbZefzJP027Par6elp/YDAfAAQp1qt4saNG6GwfvdxUFy+fFnXPWmVy+W+8lWlUsGVK1fs4H3jypUrutyVSiW4rovLly/bqwEJ8ZnJZPTbBPl3HHmIlPZt0UGzHfFN9FW1652GVquFsbExOI6TeKOQoTzmz+409KtpTZOYMaZsk1e50nnYracyMtRHfklxdFAMDw/rV8jo5JGojthBlcvlut7cSGcvzVP95eXlrg70448/3lVepFN29OhRXL9+PdRBlU7d2NhYZOeyWCzuu7y41XgVvfJh1NAbm7mPQ4cOhR46wBjGI53y3SDXYcdZEARd+SuKbGc+wJD/Nx9eoLOuPPyYnZ3VwxEzmQxmZ2dj62H5dz/pt1O2K3/ZD3xEVHyi08ky32xLxyHu6TkiHnrtR1EPkmSonx1HSeLWjYvvXnrVtUT72a53GpK+Y5DX94cOHep65Wouj+M4TtcTBLMSHx4e7uqI5HI5NJtN1Go1zMzM6M6DfMOQRq/j7rR8Pt/1FHOnKq2duq6pqSndiHU630n08+3Hfvbcc8915Xfzm5VeooYmnTlzpmus782bN/UNVobWyE8asdVqtesGWC6XMTU11ffNdLdtJl5d1418ip20zaeffooTJ07of/fah6SVOe5dxj5vd7nqx3BnGKg5DlvOMU09Eje0w/O8rqfc9Xpdd0TsOtnrjMNXne+kRLPZxPLy8p6pF/rNX7lcDo7j6IcjJvvtMRLi0+6ooBPHdoPadO/evVQdv73s7NmzXe2GGzduRMYdtjG+e+lV1xLta/aX0Q+SzNJjz6ShIuZGdqwpSoMg0DMnxM30IrMbmDMsOCnmP7dnDfF9Xx876lgyi5K9fdJxzRlB7BkgRNTsSbaodeQczfiyZ5pwrRmNoq7LnrVErtOeCcr8t7lfOZ59LHu/ytqPnfZfRWZ69TPzhsw2E8We3QURM3+IqPygOvuww6rValfYXtVvvEpZluuTbeLirdFodOXdNPuwZ1xx+pyhaKfIbDPCdd2e9adwYv4ugH39Zv0aJaq+iJohSHXObzf1m7/s2Xzse48pLj6lXjbzk/1vU7UzpfJBYN5XJe6SbEd8C9k2Kp77qWuJ9pPkEraDpPAm/eyKzTGmjZObiBRc+dk3NHt5mhtxozPFqnk8CbeP5RnTk8KofOx17eOay+2KSW4C5i+qwjGX23ElNyz5mfFix6N9rnITN8OU1WmI2q+y0jXqWPZ+7fOUG64ZZv6+Ssy4s/OILLPD3ZjpAYUZ/0nrSbqY+dZOO/nZDbq9rt94tcuHGW/2MrscCnu9qLg36xK7vthNacq73VFtxDTqhRkfSeupiE5D0r0jKl4ftH7zl3099jYqRXza9wwzHuz922m135nXZosKt+NjM/Ft14VRcZq2riXaTzLqy4JF1FOlUsHExAQeRJapVCpdww5arRYuXry478bSExEREe13u/5NA5GtXC5HfgCay+V6foRKRERERNuPnQbac5aWlkJ/mEzk8/mutw9EREREtPPYaaBUyuUyJiYmgM4c4FEzdmyXtbU15PP5rulw7ZkyiIiIiOjB4DcNRERERESUiG8aiIiIiIgoETsNRERERESUiJ0GIiIiIiJKxE7DA9RqtZDJZFCr1exFXznFYhHFYtEOJiIiIqI9aNc7Dc1mMzRDTrPZtFcJrbNfG5qtVguO49jBqRWLxa7ZhJLibC+p1Wpd57i4uIjFxcVQGIWZaW7HX5RKpRLKF1GdU3N5pVIJLavVal15KxNR5syZrcrlcmjZftBvvNrxEjVzmBn3drwKcx82eaAgv6i02y3m9adNb/Na8vm8vbgrr9qiZk/LWOnVax+7pd/8ZV9HVNqby6PyV5p9iHK5rNeLysv7jZlX0lyPHVdRzDhKikt0jm+vY9cZ9nKifcv+E9EPUqPRCP0Zd/m3+SfX7XU8z0v88+57WRAECoCqVqv2olSitnccpytsr3FdN5Sm1JvrusrzPKWMdE+Kw2q1GionUdvY+QSA8n1f/1uOZ/I8ryu/mds4jhO53V612Xi1/y3xrJRSvu8rx3H0v+04Ulbc2/sMgiC0ve/7XcfYLfa5mvEXx84Tdp1tx5dcr6hWq11pEhVH9r/NfeyWzeavrZTbNPsww/dCPG0Xs6zZbYUoafKNnV/t+Dd5nte1vNFohPK/HCNuH0T7ya7WHlE3ILvAuq4bqiDjKsT9QM59s5VH1PYSZlaEe4nc0PZjeu0WufmZ7HJh8zyvKw84jqPzin2zVFaDMAiCyJuteR52A1KlvFHvFZuJ16gOgF1v2fnbjiff97uOYe43Ku7sfe4WMw+pFOkdVUdVq9VQ3rOXq4jGn833/VA69NrHbthM/tpquVUp9qGMdLHvt/uZna9URNm09co3Ek9m/o5KA9U5vuu6Xfu0968iOtJE+9WuDk9aW1vD448/Hgo7fvw46vU60HllX6/X8eSTT+rluVwOruvi5s2bxlbR5LVlrVbreq1uv/62l5XL5dCwKHklbL56Nl+FyvCCZrMZu04c81WofZ695HI5AEAQBID13YTsV16R93plKtdprifrmPFlKpfLKBaLobiSa6hUKhgbGwMAjIyMhLaVY4mkOBfmsqghDwfF1atX4bpuKOy5557T5SLK8ePHEQSBjrNms4l8Po9SqQQAuHv3blecHT16VK+by+V0XhLNZjN0Hvfu3esaYjc8PAwAuHXrVih8L9pMvAZBgEOHDoXCTpw4of/QoJQPiQcAOs6l3F2+fBknTpzQywHg3LlzuHHjBmCUYdFqteC6bmifu6HZbCIIAp1P0LlOx3Fi0zuXy8FxHF3uAeDKlSu4cuUK0Lk2AHjsscf0cgA4efKkjo+o675x44a+D6TZx27YTP7aarlFin2gEzeu62JmZkaH7XdXrlzByZMnQ2Fnz56N/SOgafLN9evX4ThOqEyeOXMGQRB03cunpqYi49OMd5HP57vaOkT70a52GtC5GURptVr47LPPAACPPvqovTh2O1EulzE9PQ2lFJaXlzE7O6uXFYtF5PN5KKV0Y1sq3Hw+jyAIsLS0pLf3fR8TExMoFouYmpqCUgqO4+DixYtAp5KWxtTIyIhex3XdrkaWrVgs4vHHH9fnMjs729VYTiIVmeM4aBnfTVy5cgXHjx/X69VqNYyNjaHzdgnVahVjY2P6WNKgv3z5MpaXl6GUgud5mJqaQj6fx9LSEr58UPMfcVUsFjE7O4t6va7jqtFo6GsYHx9Ho9EAADQaDb293fFIinPRbDYxMjKi93Py5MnYzsV+12q1uhqSQhoKtlKppOMsk8lgenq665uRuMbLp59+agcBnUbQ2bNnQ2FRN090OhR73WbiFQCWl5ftIF1vRHWkhDzYiOp4ICY9Wq0WTp482ZV2u0HyRVScRcWJWFtbA4xx+FNTU10dgaiHPrKdrdVqYW1tbUv7eBA2k7+2o9z22ketVkMQBDh79qxOk0zKB1p7WdRDRyTUUSIp30R10oTZUS4Wi7Gdkyj1eh1nzpyxg4n2nV3tNJw7dw71ej30xFtuRnGVbz+kITMzM9P1BEj2L0/G7t69C3QqD8dxkM/ndcUrT7jMm18+n9cV0/DwcKhxLOtMTU0BxtNIm9wMx8fHAeMtyuXLl+1VY8lT/enpaeRyOd2YyeVyKJVKUEpheHgYU1NT8H1fb1cqleC6rm6YS4P+5MmT+umJPMFaWlrS8WXG1eLiIlzXheu6Oq6Gh4d7XoMcSyTFucTxzZs34TiOjtuXX34Z6MS3xN9X3fj4uG7A2g2N5557DjDyC4yboP3kTczOzoaemslNz9yHNIaiGsUHged5mJ2dDTX6bty4EdtR2IpisQjHcRAEQVfHer+pVqv6/81GmtRxExMToYbd0tJSbGPt+vXrOHfunP73Zvaxl21HuU3ah9xTr1y5oh8aOY6zI3l4L9uufFOr1XD27NnUbZRarQbf91OvT7SX7WqnYXx8HJ7nYWxsTD/9mJ2d7Wrgb8bx48f1k5dKpRJ68rK4uIiZmRk93EUa2qbtKODSwI17Cnvr1i3dQJBfvV6PPB+TGV/1eh3VarXrlaj5lqHVakU+7ZSnyGYFGvXkpl8nTpzoeQ1RouJc3jbBeLoLY924p+RfRZlMJvRGyGx4Dg8Po1qtYnZ2VucdGUJiP8FF50bneV4oLJfLodFooF6v632MjIwAxpCJg0YeOMjwOilz9rCI7bC4uKjf8MFqKO4nlUpFv62UJ+Dm28DFxUXdaJU4DYKgawiXuHz5ctdT2n73sZdtR7lN2geA0IMddDoQSHigdVBtNd+0Wi0sLy933W+TXLlyhQ+26MDY1U4DOjdlefohjUJ5Qi/DksyGIzpPpqWQm98DZIzpIeUpOwDdeRDSWZChMLv5xMVxHH395i9JtVoNrdtPBbZfSaUrNzn5b9xT8v0sl8t1vV6XzlFUAx+dcuB5nu5MSR4yG2tSJuTnOE7o7ZNpeXk51PEUw8PDoX24rhs67l62mXiF0ZhXneF3MN50HTp0qKuDLMeQt2WO43Q9OLh7927sw5GoN6O7QcqWHWdBEETmDXTWnZiY0PEzPj7eNdQQnTpc4lTeSkQ1rOTYUfkr7T4elM3kr+0ot2n2YYsa8rvf5PN5/dZbyHDBqPwikvLN448/3jXETdofR48exfXr10MdOGk7jI2NRb6tMN+eEx0Eu95pMDmOA8/zdAUrrxPN19vy1FxuyGanQymlC6g8pZNwGJXoyMgIqtXqjhdmu/Fge+yxxxBEjL/c7ieMUoHK0yWRpoLdjKQG0Vb4vq/fsoyNjaFarcbejPezqI8nl5eXE+PUzkPoDK2xb6pCOtdxjSx7aFKUcrmMer0e+THgXrSZeDW1Wi2MjIyEhhpIHJnDl2T4iORN86NnsbS0lPh0M5fLbctbv60YjvjoWa4z7s2S/YAHxrC2qDzabDZ1WY5iD02K0msfD8pm8ldUnPRbbnvtQyYXiVovLh33g6iPnm/cuJH6LWBUvon66FmGxuZyOYyPj4faG/LAoFqtdnU2+v3ugWhfsKdT2g0yhVzUlGT2FH+e50WuZ7OnupMp6GRKNXuKNXOf9vRocg7mVICu64b2L+vY08Wa/46ajtB13a7p3OKm6Iva3ha3jsRx0rSF9nLZxpx+zo4bmXJO4saOK/N8qsb86/ax7P3a+6lWq6nS/aAwpw6MmgbQ5nfmAjfzqP1vlXKK3jRxjX0633u/8SokfqOm9LSnZLSnu1RWWatGTBVpipq6c7dIHSBcawrsKLDqQT9iylkVM8e9zXGcxPRJs48Hqd/8tR3lNs0+HMfpujf1KuP7gTldaj/lJinf2G0HOy5NcfdbN+JvE/m+3xVGtN+kK2E7RBqcvSpWuXEhpmMRxfM8XYnY20mFIZWwnIfT+UNp8nNdN3RsqRzM9aQSl2NJBS7bC/NcZD1hxgNiKj17naj1ko6hrHhERMNefp7nha5D1jX/LdfmdjpG5vnZFaMs8zwv8lj2fu3zlM6GGWauf1CZ+cwuH7LMDI9Ks6hl9g3OFnXDU1b+sfPWftJPvJrrJjHzcVT8mvnebvjZ6bbX8rR5fna6S56w6+Wk64mLB1uj0ejaVqTdx27oJ3+piPTfTLlN2ocwz8tOr/3MvG6bHS7/7pVvku5npqhOg92OkF9cXibaTzLqy4JEW2ROCXoQh8zEkdflD2Ko161bt7qGzMi3DXY4EREREW2fPfVNA1GckydPRnYMHnvssQP5MTQRERHRXsJOA+158vG7/YF4q9XC9PT0V+rNDhEREdFu4PAkIiIiIiJKxDcNRERERESUiJ0GIiIiIiJKxE4DERERERElYqeBiIiIiIgSsdNwQGUyGVQqFTv4wCmXy8jn83YwEREREW2jXes0VCoVZDKZxB9tTlLclcvlrniWn/yhtL2q1Wp1dYRmZmawtrYWCjsoisWiTptms2kv7mKXqTiZTEb/UT5bmn2Y69jpsR/0G6+1Wi0UJ61Wy14FAJDP52M7sL320Ww2Q8v3Urya525PexwnzbVIPopLg177aLVaenlcvO+GfvNXmjKHPsttVF3eKz73q3w+r6/LLldR7LiKkxTfwsyDZtmwy3OafEC0L9h/IvpBcxyn60/aNxoN/afXfd8PLaN0esWdvdx13a6wvcb3/T19ftvJdV1dLoIgUABUo9GwV9N831eO44T+bRfvarWqy5XruqFlKuU+7HUcx9lXadJvvEqc2f8OgkCHmfWVGTei1z6CIAilhyzfC/Fqn7sZf3EAqGq1Gvq3fS2O4+g4i4r/Xvuw087Ol7ul3/xln3dUmetVbqPyk33cXvG5X5n1j5RDs2zatiO+hed5kfHYaDRC28l6RAfBrufkqE6DkBtLUqVL0aIqM1PUcqko9yK5EdrnfBDJzc/keV7iDcxuFKiEBr3rupH7SrMPuzzajcq9bDPxal+/Smg4e54X2XDttQ87zmV50nk9KI7jhM6vV8PMbpSphDwi+7Lr9zT78DyvKw2i8u+DtJn8FXXOUflFJeSJqHxnplua+NyPqtVq13XFlU2xHfGtjI5AVDmww+TeZYcT7Ue7NjwpjStXrgAApqenQ+Hm69+o16zma0H7Na09TMBeXiwWUSwWI187mkN7zNeNsm6z2Qydm/2q1H4tar+ylOXFYhHNZrPr3JK2NV+H2tul5TgO0LkeGN9FyHnJfu1Xr3Ya5PN5lMvl0HqyTlz8SLrAuE4ZctBsNvW5TUxMhK5fjiWS0k+Yy+SYe8nVq1fhum4o7LnnnkO9Xg+FCYnHxx57LBR+8uRJ3LhxIxQWJ80+JP3Nv8BdKpWAThrtdf3GKwAEQYBDhw6Fwk6cOIGlpaVQWJJe+5A4NOVyOeRyOTv4gWo2mwiCAEePHtVhw8PDcBwHt27dCq0r7t692zVUSLZPm0fS7GN2dhbHjx8PreN5HpaXl0NhD1K/+StNmUvj+PHjCIJA17HNZhP5fF7nqzTxuR9duXIFJ0+eDIWdPXs2tmxuV3zXajXMzs6i0WhEllE77LPPPoPneV3hRPvRnu40yA3KrHTz+TzOnj0LpRQajQYmJiZ0Y0Yag41GA0op+L6PsbExXVnUajWMjY2h84YF1WoVY2NjoQZtvV7H2toaTp48qdeZnZ1FsVjE8ePHoZSC67q6I2M2aEdGRjA1NaXXkXB0OgSXL1/Wx/Z9HyMjI6GG+I0bN6CUwtTUFEZGRvS2cl1BEIS2letqNpsYGRnRyzd74wyCAOhUetKYlspUrsU+lqSBNMzz+TyCIMDS0hKmp6f1+U5MTKBYLOr4cRwHFy9eBDpxMzY2BnQ6DBJHQRCgWCxieHgYSikAgO/7UEpheHhYH0skpZ/EVavVguM4qFar+tziOhe7pdVqxd5gkm7yN2/etIP6/t4jaR/37t0L5WlT1HZ7zWbjNao8mfkujX73sbS0hOeee84OfqA+/fRTIKIRhJjrEXGNZNlfGkn7iGv8oRNvu2Wz+Suq7PRTbkulkq5jM5kMpqensbi4GFonKT73q7W1NTz++ON2MIIg0Hkkylbje2pqCq7r4urVqz3vHc1mE9PT05iZmbEXEe1Le7rTYLOfoEinQt5IXLx4Ea7r6iehZ86cATo9fXQKu+/7en+lUgmu62JiYgIAsLi4qJ8USSUiT2TOnj2rj5vL5fTy4eFhNBoNAECj0dDHnpqaAoynsxMTE6E3JuPj43AcR68H45jDw8OoVqs6/Pr16/B9X9+Q5LquX78OdM7NXL6ZCkoqPYkfaaSjc65ra2solUqYnp4OPTUZHh6G53m6Yb62tgbHcZDP5/WN68knnwQ6cSLxk8/ndcU+Pj4eeVzf91Gv12NvAHIskZR+kgfkCamk5fj4OACgWq1uKt72glwup/OxGVdLS0tdTxjjbMc+DiLJ22aj78aNG7Gdpyj97qPZbOLkyZOhNzr7hXR0zEaUlLmoRn6U7djHfrCdZU7uJ4joIHxV4rOX7YjvVquFIAiwtramHyLKg6moN+4jIyOo1+up90+01+2rTsPNmzdRr9dDQ0ukACPiSU8ul9NPpaWw28MEzp49CxivLtEp7FslN/x79+7pxoJdQZ87d04/bRweHkYQBMh0huWUSiXdsL1x44Z+ipTJZPTN4e7du/q6pGHeD3Ofs7Oz8H1fN6LFiRMnQv+u1+tdT3fkpiQNc8Q8neyXXJO53zSi0s9+omamt+M4uHfvXmj5frO4uAjHceA4Tqhs2OmXZDv2cdDMzMzAdV2MjIzoOKnX613DIpL0u4/9/GRSHnjMzs7qa5W3iGk7Qduxj/1iu8pcJpPB0tKSfuiSMYZdfpXis5ftiu8rV67o+7M8fLx8+XJonbW1NajOqIPAGD5GtJ/t6U5DszOm1hwn6rquHr4iv35eLe5lSil4nqc7D+a3CTIsx/xttWFh79PuMBxEpVIJjuPotzSSx+zO5G7K5XJdb1ek05N0k5eblOo8/YLxJiWtpH0cOnSoa0iNnOdmOq0P2mbjdXFxUceJvFV8+eWX7dUSpd1HsVjccrneLvKQw46zIAi6vicwlUqlUL3iOE7oDW8aSfuQBxJRDwLiOmIPwmbzV1KZS6NcLofe/qpOx8FspCbF536Vz+dx9+7dUJgMoUx6aLXV+I6SdDzpqBAdBHu60yBvAWQIz6FDh7pevcJ47ZrL5SLHtFYqFV2oZSiTSFPJbIbZmHr00UeBzodyprt37+oOUbPz4fPMzAxUp/Mg153L5SI/1CqXy/q8o8Zp7gTHcbqeqKS5MW7GTu13enpav2UZGRmB7/uRH6PulqiPJ5eXl0Od5yTNZhNjY2OhIW79itqHxJE5zEaGOWx3Gu2ErcZrq9XS+WWz9UXSPsrlMqamprrCd8twxEfPkvbmx9FJZJ77rTTKovYR9dFzvV5P7MzstK3mr6gyl4bdUUEnfuwGtYiKz/0o6qPnGzdupO44bia+c7kcHMfpupej80F1nHw+v6ceTBFtmj2d0oMWNeWqOe951PRo5jRoQRB0zYtt7q9areop/WSaOXsKSfMY9jRrsk9zHXuKOzlfczt7PzIftOzHnkvbnsrQ9329fdR0o9VqVe9Lpn+zr9OOC5O9vyhR68h5mvu1p6yz0zRqasW4+LHTJmq/QRDocPtY9n7t9GtYc2jvVebUgf1M2Sd5wS43JjuObEn7sKdvNKd23A82G69R+dNm1wu2pH14ntcVj2YZ3y1SlwjXdSPP3yZxmxQfUfWCKWkfdv1p1pe7abP5K6nMibhyK/nKjEf736pHfO5X5r0nasrbOFuJbykTcfduWzVialii/SpdCdsBUtHF/aIKq3CMPwxkF0YpwPKzG81mg9os+Cpiv/a+5MZkhinjmOY1RZ2/fc1mJdNoNHTlYu5bSIUvP/u6zPOSCjHq5i7L4s5DpTiWHS9mxWuGu67bFd/2NUr6mQ0qWWafvyyTbexj2fu1z9P3/a5rM9ffa8zrsdNIlkl4muuw48Peb5p9KCsPJd1096p+4jWuPJqi8lRcvRLFrlPk1ysdHpSkMinlW+oIc92kvGFfs1nHpN2HmZ+j6tvd0k/+kvWS0rpXuVUR9xYz/6WNz/3KvG6bHb5d8W3f18zl9j3Wvn8S7WcZ9WVBoi1odqYhNWdPov5UKhVMTEzgQWTHSqXS9Wq+2Wzi5s2bXeFEREREtMe/aSDabsViUU9Za3r00Uc55pSIiIgoBjsN9JVSr9cj/xDPyZMn99TH0ERERER7CTsNW1Sr1fRfbzb/wjOlVy6X9R/Yy2QykbOBbBfVmaJX5uiW30GZtpeIiIhoJ/CbBiIiIiIiSsQ3DURERERElIidBiIiIiIiSsROAxERERERJWKngYiIiIiIEm37h9APPcR+CBEREdFu+eKLL+wgoi3b9k4DEREREREdLHwtQEREREREidhpICIiIiKiROw0EBERERFRInYaiIiIiIgo0f8HIDOs3lFOfX8AAAAASUVORK5CYII=\"\u003e\u003c/p\u003e"},{"header":"5\tAnalysis of Experimental Results","content":"\u003ch2\u003e5.1 Limitations of Zero-shot Prompting\u003c/h2\u003e\n\u003cp\u003eThe subpar performance of zero-shot prompting in the sleep disorder classification task can be attributed to inherent limitations in data utilization and insufficient model learning capacity. Zero-shot approaches, relying solely on predefined heuristics, lack the ability to leverage training data for in-depth pattern recognition. In the complex domain of sleep disorder classification, diagnoses are rarely determined by single factors but rather by intricate interactions among various clinical and lifestyle variables. These multifaceted relationships often exhibit complex non-linear characteristics. Consequently, zero-shot prompting, devoid of training data exposure, struggles to discern these subtle interdependencies between features and classification outcomes, failing to capture the underlying data patterns essential for accurate sleep disorder identification.\u003c/p\u003e\n\u003ch2\u003e5.2 Performance Gains with 90-sample Prompting: Data-Driven Learning\u003c/h2\u003e\n\u003cp\u003eIn stark contrast to zero-shot prompting, the 90-sample prompting strategy demonstrates a significant en- hancement in classification performance. By exposing the large language model to a limited set of training examples, the model gains the capacity to discern potential correlations between feature combinations and sleep disorder types. This approach enables the model to learn data patterns and establish a correspondence between input features and desired classification results, effectively mimicking data-driven learning.\u0026nbsp;For instance, by analyzing relationships within the 90-sample training set, the model can begin to understand\u0026nbsp;the influence of factors such as gender, age, and occupation on sleep disorder propensity. Furthermore, the utilization of logistic regression, coupled with essential pre-processing steps like encoding categorical features and standardizing numerical features, contributes to improved data quality and optimized learning. This refined approach leads to more accurate classifications, particularly for common sleep disorder types.\u0026nbsp;However, it\u0026nbsp;is crucial to acknowledge the inherent linearity of the logistic regression model. When confronted with the intricate non-linear relationships that characterize sleep disorders and their multifaceted determinants, the linear\u0026nbsp;nature\u0026nbsp;of\u0026nbsp;logistic\u0026nbsp;regression\u0026nbsp;may\u0026nbsp;limit\u0026nbsp;its\u0026nbsp;ability\u0026nbsp;to\u0026nbsp;fully\u0026nbsp;capture\u0026nbsp;these\u0026nbsp;complexities,\u0026nbsp;ultimately\u0026nbsp;leading to residual classification errors.\u003c/p\u003e\n\u003ch2\u003e5.3 Superiority of Decomposed Prompting: Task Decomposition and Optimized Model Selection\u003c/h2\u003e\n\u003cp\u003eThe decomposed prompting strategy emerges as the most effective approach in the sleep disorder classification task, achieving the highest performance metrics. This superiority stems from its strategic task decomposition and the optimized selection of a classifier tailored to the data characteristics. Decomposed prompting breaks down the classification task into manageable stages, systematically exploring a suite of classifiers, including logistic regression, decision tree, random forest, gradient boosting, support vector machine (SVM), and K-nearest neighbor (KNN). Rigorous parameter tuning is implemented for each classifier, followed by a comparative evaluation to identify the most suitable model for the specific dataset. This meticulous model selection process reveals the Support Vector Machine (SVM) as the optimal choice for this sleep disorder dataset. SVM\u0026rsquo;s inherent capability to identify optimal classification hyperplanes in high-dimensional spaces and effectively model complex non-linear relationships proves crucial. The sleep disorder domain is inherently characterized by such non-linearities, evident in the interplay between sleep duration and sleep quality, or stress levels and daily activity volume. SVM\u0026rsquo;s proficiency in capturing these intricate relationships directly translates into superior classification performance. Specifically, in discerning insomnia from other conditions based on features like stress level and living habits, SVM excels at delineating insomnia samples by identifying the most discriminative classification hyperplane. The task decomposition inherent in decomposed prompting empowers the model to develop a deeper understanding of the task requirements, progressively addressing stages such as data pre-processing, model training, and evaluation. This structured approach not only enhances classification accuracy but also bolsters the overall reliability of the model\u0026rsquo;s predictions.\u003c/p\u003e\n\u003ch2\u003e5.4 Key\u0026nbsp;Factors\u0026nbsp;for\u0026nbsp;Performance\u0026nbsp;Improvement\u003c/h2\u003e\n\u003cp\u003eThrough a comparative analysis of the experimental results across different prompting strategies, several key factors for enhancing model performance in sleep disorder classification become apparent. Effective utilization of training data, enhancement of the model\u0026rsquo;s learning capacity, and the strategic decomposition of tasks coupled with informed classifier selection emerge as pivotal elements. The decomposed prompting strategy demonstrably excels in addressing the complexities of sleep disorder classification, offering valuable methodological insights and a robust framework for future advancements in this domain.\u003c/p\u003e"},{"header":"6 Conclusions and Future Work","content":"\u003ch2\u003e6.1 Research Conclusions\u003c/h2\u003e\n\u003cp\u003eIn this rigorous investigation into sleep disorder classification, we meticulously evaluated the application of large language models (LLMs) using a comprehensive sleep health and lifestyle dataset. Through a thoughtfully designed experimental framework and an in-depth, multifaceted analysis, we have convincingly demonstrated the significant potential of LLMs for automated sleep disorder classification. Our findings clearly highlight the pivotal influence of prompting strategies on LLM performance, leading to the following key conclusions:\u003c/p\u003e\n\u003cp\u003e1. \u003cstrong\u003eZero-shot Prompting Sufficiency is Limited: \u003c/strong\u003eZero-shot prompting, while straightforward, exhibited relatively weak performance in the complex sleep disorder classification task. This outcome suggests that relying solely on pre-training knowledge and basic heuristics is insufficient for LLMs to effectively address the nuanced complexities inherent in sleep disorder diagnosis without specific examples or in-depth training.\u003c/p\u003e\n\u003cp\u003e2. \u003cstrong\u003e90-sample\u003c/strong\u003e\u003cstrong\u003ePrompting\u003c/strong\u003e\u003cstrong\u003eEnhances\u003c/strong\u003e\u003cstrong\u003eLearning\u003c/strong\u003e\u003cstrong\u003ethrough\u003c/strong\u003e\u003cstrong\u003eData\u003c/strong\u003e\u003cstrong\u003eExposure:\u003c/strong\u003eThe 90-sample prompting strategy, by introducing a limited set of example data points, provided the LLM with more substantial learning information, resulting in a marked improvement in classification performance compared to zero-shot prompting. This improvement unequivocally demonstrates that leveraging data patterns and feature- outcome relationships within a training set significantly enhances the LLM\u0026rsquo;s comprehension and execution capabilities for sleep disorder classification tasks. However, the inherent linearity of the logistic regression model, employed within the 90-sample prompting approach, presents limitations in accurately capturing and modeling complex non-linear relationships within the data, thereby constraining the model\u0026rsquo;s ultimate classification potential.\u003c/p\u003e\n\u003cp\u003e3. \u003cstrong\u003eDecomposed Prompting Achieves Superior Performance through Task Optimization: \u003c/strong\u003eThe decom- posed prompting strategy yielded the most outstanding results in sleep disorder classification. This approach\u0026rsquo;s comprehensive exploration of diverse classifiers, coupled with meticulous parameter tuning and rigorous evaluation, highlights the efficacy of task decomposition and judicious classifier selection. Our findings indicate that these techniques effectively harness the inherent strengths of LLMs, leading to enhanced accuracy and reliability in sleep disorder classification.\u003c/p\u003e\n\u003cp\u003eOur research emphasizes that for LLM-based sleep disorder classification, a deep understanding of the data and task is paramount to model performance and directly impacts classification accuracy. Providing detailed examples, relevant reference information, and employing task decomposition strategies are crucial for improving model accuracy. Consequently, future research should prioritize in-depth data and task analysis and focus on optimizing prompting strategies to further enhance model performance. Furthermore, within decomposed prompting methodologies, selecting a classifier model that aligns with task characteristics and implementing effective optimization techniques are central to maximizing classification performance. Given the varying suitability of classifiers across different scenarios, addressing complex sleep disorder classification necessitates a comprehensive consideration of multiple factors, careful selection of the most appropriate classifier, and meticulous parameter adjustments to achieve optimal classification outcomes.\u003c/p\u003e\n\u003ch2\u003e6.2 Limitations and Future Work\u003c/h2\u003e\n\u003cp\u003eWhile this research successfully demonstrates the potential of LLMs in sleep disorder classification, several limitations warrant consideration, particularly concerning data, model selection, experimental design, and prompt engineering. Specifically, the dataset\u0026rsquo;s diversity remains limited [15], the classifier model selection could benefit from incorporating the advanced feature-learning capabilities of contemporary deep learning architectures, the scope of comparative experiments could be expanded, external validation is absent, and further optimization of prompting strategies is warranted. Future research endeavors can build upon these findings by exploring the following directions to enhance the performance and practical applicability of LLMs in automated sleep disorder classification:\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026ndash; \u003c/strong\u003e\u003cstrong\u003eDataset Expansion and Balancing: \u003c/strong\u003eExpanding the dataset to encompass a wider range of sleep disorder samples, particularly rare and specialized cases, is a critical next step. Addressing the issue of imbalanced data distribution is also essential. Employing data resampling techniques, such as oversampling (e.g., SMOTE algorithm [16]) and undersampling, can effectively adjust the representation of different sleep disorder types within the dataset. This balanced representation will enable models to learn the char- acteristics of diverse sleep disorders more equitably during training, thereby improving generalization performance, especially for underrepresented groups.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026ndash; \u003c/strong\u003e\u003cstrong\u003eExploration of Advanced Classifier Models: \u003c/strong\u003eFuture research should prioritize the exploration of more sophisticated classifier architectures and feature engineering methodologies. Deep learning models, with their inherent feature-learning prowess [17], offer promising avenues for improvement. Architectures such as Convolutional Neural Networks (CNNs) [18], Recurrent Neural Networks (RNNs) [19], and their advanced variants like Long Short-Term Memory networks (LSTMs) [20] and Gated Recurrent Units (GRUs) [21] hold significant potential. These models are particularly well-suited for processing time- series data and complex data structures, enabling automatic feature extraction and pattern recognition, and facilitating the exploration of intricate relationships between sleep disorders and factors that are challenging to quantify directly.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026ndash; \u003c/strong\u003e\u003cstrong\u003eEnhanced Experimental Design and External Validation: \u003c/strong\u003eFuture studies should incorporate more comprehensive comparative experiments and rigorous external validation procedures. Systematically comparing the impact of diverse model architectures (including both LLMs and various classification models), prompting strategies, and other relevant factors on classification performance will deepen our understanding of the underlying mechanisms and facilitate the identification of optimal methods and parameter configurations for sleep disorder classification. Furthermore, evaluating model performance on multiple external datasets from diverse sources is crucial for assessing generalization capabilities across varying data distributions and real-world scenarios, and for mitigating the risk of overfitting to the training data.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026ndash; \u003c/strong\u003e\u003cstrong\u003eRefinement of Prompting Strategies: \u003c/strong\u003eInvestigating more advanced and nuanced prompting techniques is essential for further performance gains. Strategies such as Chain-of-Thought Prompting [22] and Contextual Prompting [23] can be explored to guide models toward more in-depth reasoning and analy- sis. Chain-of-Thought Prompting, by encouraging models to articulate their reasoning processes, can enhance model interpretability [24]. Contextual Prompting, by providing models with richer background information and task-specific context, can improve their understanding of task requirements, leading to more accurate sleep disorder classifications.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eAcknowledgments\u003c/p\u003e\n\u003cp\u003eMy\u0026nbsp;heartfelt\u0026nbsp;thanks\u0026nbsp;go\u0026nbsp;to\u0026nbsp;my\u0026nbsp;parents,\u0026nbsp;whose\u0026nbsp;constant\u0026nbsp;care\u0026nbsp;and\u0026nbsp;support\u0026nbsp;have\u0026nbsp;been\u0026nbsp;absolutely\u0026nbsp;essential\u0026nbsp;to me,\u003c/p\u003e\n\u003cp\u003eboth in life and throughout this research project. Their endless encouragement and sacrifices truly made this work possible, and I am deeply grateful. I also want to sincerely thank the data providers for offering the precious sleep health and lifestyle dataset, which was fundamental to this entire study.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eIbomoiye Domor Mienye N, Jere N Survey of Decision Trees: Concepts, Algorithms, and Ap- plications. In \u003cem\u003eIEEE Xplore\u003c/em\u003e. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://ieeexplore.ieee.org/stamp/stamp.jsp?tp=\u0026amp;arnumber=\u003c/span\u003e\u003cspan address=\"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=\u0026amp;arnumber=\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e 10562290\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKumari A, Akhtar M, Shah R et al Support matrix machine: A review. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://arxiv.org/abs/2310.19717\u003c/span\u003e\u003cspan address=\"https://arxiv.org/abs/2310.19717\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCurth A, Jeffares A, van der Schaar M Why do Random Forests Work? Understanding Tree Ensembles as Self-Regularizing Adaptive Smoothers. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://arxiv.org/abs/2402.01502\u003c/span\u003e\u003cspan address=\"https://arxiv.org/abs/2402.01502\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKim Y, Xu X, McDuff D et al (2024) Health-LLM: Large Language Models for Health Prediction via Wearable Sensor Data. arXiv preprint arXiv:2401.06885 [cs.LG], [Submitted on 12 Jan 2024 (v1), last revised 27 Apr 2024 (this version, v2)].\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKatz DM, Bommarito MJ, Gao S et al (2024) GPT-4 passes the bar exam. Philosophical Trans Royal Soc A 382(2270):20230254\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNori H, Lee YT, Zhang S et al (2023) Can generalist foundation models outcompete special-purpose tuning? Case study in medicine. arXiv preprint arXiv:2311.16452\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSaab K, Tu T, Weng W-H et al (2024) Capabilities of Gemini models in medicine. arXiv preprint arXiv:2404.18416\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSinghal K, Tu T, Gottweis J et al Towards expert-level medical question answering with large language models. arXiv preprint arXiv:2305.09617, 2023.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMcDuff D, Schaekermann M, Tu T et al Towards accurate differential diagnosis with large language models. arXiv preprint arXiv:2312.00164, 2023.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSleep Health and Lifestyle Dataset \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.kaggle.com/datasets/uom190346a/ sleep-health-and-lifestyle-dataset\u003c/span\u003e\u003cspan address=\"https://www.kaggle.com/datasets/uom190346a/ sleep-health-and-lifestyle-dataset\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang G, Zhao W, Han J et al (2024) MedFound: The First Medical Large Language Model Passing the Physician Qualification Examination. J Artif Intell 5(1):1\u0026ndash;12\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMcDuff D, Xu X, Kim Y et al (2023) Personal Health Large Language Model (PH-LLM): leveraging large language models for personalized health insights. arXiv preprint arXiv:2311.17133\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang Y, Maziarka P, Klicpera J et al DiffSBDD: Equivariant diffusion for structure-based drug design. arXiv preprint arXiv:2403.14338, 2024.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVaswani A, Shazeer N, Parmar N, Jones L, Gomez AN, Kaiser L, Polosukhin I (2017) Attention Is All You Need. In \u003cem\u003eAdvances in Neural Information Processing Systems 30 (NIPS\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu X, Uchiyama M, Okawa M et al (2000) Prevalence and correlates of insomnia in the Japanese general population: Results from the Japan epidemiological sleep study. Sleep 23(4):497\u0026ndash;506\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChawla NV, Bowyer KW, Hall LO, Kegelmeyer WP (2002) SMOTE: Synthetic Minority Over-sampling Technique. J Artif Intell Res 16:321\u0026ndash;357\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLeCun Y, Bengio Y, Hinton G (2015) Deep Learn Nat 521(7553):436\u0026ndash;444\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLeCun Y, Bottou L, Bengio Y, Haffner P (1998) Gradient-Based Learning Applied to Document Recognition. \u003cem\u003eProceedings of the IEEE\u003c/em\u003e, 86(11):2278\u0026ndash;2324\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGraves A, Mohamed A-R, Hinton G (2013) Speech Recognition with Deep Recurrent Neural Networks. In \u003cem\u003eProceedings of the IEEE international conference on acoustics, speech and signal processing\u003c/em\u003e, pages 6645\u0026ndash;6649, 2013\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHochreiter S, Schmidhuber J (1997) Long Short-Term Memory. Neural Comput 9(8):1735\u0026ndash;1780\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCho K, Van Merri\u0026euml;nboer B, Bahdanau D, Bengio Y (2014) On the properties of neural machine translation: Encoder\u0026ndash;decoder approaches. arXiv preprint arXiv:1409.1259\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWei J, Wang X, Schuurmans D, Bosma M, Ichtertz R, Joshi S, Zhou D (2022) Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. arXiv preprint arXiv:2201.11903\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhou X, Huang M, Wang H, Zhang Z (2022) Contextual Prompting for Few-Shot Text Classification. In \u003cem\u003eProceedings of the 2022 Conference on Empirical Methods in Natural Language Processing\u003c/em\u003e, pages 9312\u0026ndash;9327\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRibeiro MT, Singh S, Guestrin C, Why Should I (2016) Trust You? Explaining the Predictions of Any Classifier. In \u003cem\u003eProceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining\u003c/em\u003e, pages 1135\u0026ndash;1144\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"Beijing Guangqumen Middle School ","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Sleep Disorder Classification, Large Language Model, Prompt Engineering, Health Informatics","lastPublishedDoi":"10.21203/rs.3.rs-6124845/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6124845/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eSleep disorders are a major global health issue, affecting nearly a third of the world\u0026rsquo;s population. This research explores using the sophisticated language and reasoning abilities of large language models (LLMs) to automatically identify these disorders. We trained LLMs on a dataset of sleep patterns, lifestyle choices, and related health factors, employing three novel prompting approaches to guide their design, training, and evaluation of classifiers. Our results show that a support vector machine classifier, identified through decomposed prompting, achieved an impressive 91.9 percent accuracy (F1-score: 0.919), significantly better than traditional zero-shot and few-shot methods. This work demonstrates a powerful integration of LLM\u0026rsquo;s understanding and reasoning with automated machine learning, offering a promising new direction for sleep disorder classification in health informatics.\u003c/p\u003e","manuscriptTitle":"Automatic Sleep Disorder Classification Using Large Language Model Prompting on Sleep Health and Lifestyle Data","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-03-03 12:31:03","doi":"10.21203/rs.3.rs-6124845/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"4b8a360a-00d8-44b1-9df8-28265e77815c","owner":[],"postedDate":"March 3rd, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":44989151,"name":"Artificial Intelligence and Machine Learning"}],"tags":[],"updatedAt":"2025-03-03T12:31:03+00:00","versionOfRecord":[],"versionCreatedAt":"2025-03-03 12:31:03","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6124845","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6124845","identity":"rs-6124845","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}