ADS Capstone Chronicles Revised

36

limited variability within the datasets, especially the restaurant menu items. This finding highlights the need for richer features or external data to capture the nuances in patient scores more effectively. As the scope of the project increases, the variability of menu and individual food items will increase and train the model more effectively. The suitability scoring system provided an interpretable framework for being able to evaluate nutritional content that aligns with dietary guidelines for diabetic patients. For example, foods with higher fiber and lower sugar content were ranked higher , which is consistent with evidence-based recommendations. The strong correlation between the suitability scores and nutritional features like carbohydrates and fiber demonstrates that the scoring system did accurately reflect the key dietary priorities for diabetic patients. Despite its strengths, this project faced several limitations. First, the use of simulated patient data may not fully represent the diversity of real-world diabetic populations. The vastness of simulated data also proved to be too large to properly store throughout the duration of project completion; therefore calling for a stratified sample of the data. Additionally, while the API provided a comprehensive source of nutritional data, the scope of the dataset was influenced by the selection of restaurants and food items queried, which may not fully capture the variety that is available in real-world settings. The negative R-squared values in regression results suggest that the homogeneity within datasets, especially in the menu dataset, made it challenging to capture meaningful variability. Lastly, the predefined scoring weights and thresholds, while following guidelines, may not

perfectly align with the needs of all diabetic individuals. 6.1 Conclusion This study successfully developed a personalized restaurant meal recommendation system for individuals with diabetes based on specialized metrics and general guidelines, which aims to bridge the gap between personalized healthcare and data science. The XGBoost classifier method excelled at predicting binary classifications for food recommendations, minimizing inaccurate recommendations. This project is an extension of past work that has provided food recommendations for people with diabetes, but did not include a personalized element or an emphasis on restaurant meals. Overall, this contribution to the field allows individuals with dietary restrictions, specifically those with diabetes, to make informed decisions about their food choices while dining out. 6.2 Recommend Next Steps/Future Studies To address the aforementioned limitations, several actions can be taken to further improve this study. The most critical area of improvement is in data availability and quality. The obvious restriction of access to an abundance of real, diverse patient data is a top priority to improve the robustness of individualized recommendations. With the appropriate credentials, having access to proper patient data with real-time updates could significantly improve the quality of the input data. Similarly, additional menu data with heterogeneous attributes that are more representative of real-world options can refine the scope of recommendations, therefore improving the practical use of the model. It would also be beneficial to gain access to

238