Building on research by Zador et al (2019) about how existing health conditions and comorbidities can influence the risk of developing sepsis, we've created a detailed dashboard called the Sepsis Insight and Management Platform (SIMPL) that helps identify the sepsis risk level of ICU patients by looking at their specific health conditions and disease interplay. Instead of the usual one-size-fits-all method, we use a statistical technique called Latent Class Analysis to group patients into seven unique categories, recognizing that people often have more than one health issue at the same time. Based on these groups, we then use machine learning techniques to predict the risk of sepsis more accurately, making it easier for doctors and nurses to understand and use this information through a straightforward online tool.

Our project embarked on an ambitious journey to enhance sepsis identification through comprehensive data analysis and visualization tools. Starting with Exploratory Data Analysis (EDA), we delved into patient vital signs from the MIMIC III database, focusing on data cleaning and anomaly detection to ensure the integrity of our dataset. This initial phase was crucial for setting a strong foundation for subsequent analyses. Employing Latent Class Analysis (LCA), we uncovered distinct patient subgroups based on comorbidity profiles, a step that allowed for personalized risk assessments. We then used machine learning, particularly the Random Forest Classifier, to predict sepsis risk categories tailored to the nuances of each subgroup. The culmination of our efforts is an interactive web dashboard designed to provide healthcare professionals with intuitive, subgroup-based sepsis risk information. The EDA process involved meticulous examination of vital signs across 17 "chartevents" tables from the MIMIC III database. Addressing data discrepancies and outliers was a significant part of this stage, ensuring the dataset's reliability for further analysis. Our findings highlighted the importance of considering subgroup-specific vital sign distributions, which could influence model outcomes. LCA played a pivotal role in our methodology, enabling the categorization of patients into seven subgroups, each characterized by unique health profiles. This classification facilitated a more granular understanding of sepsis risk across different patient demographics. In addition to LCA, we applied logistic regression to explore the interplay of comorbidities within subgroups, uncovering significant associations that could inform targeted interventions. The Kruskal-Wallis test further elucidated differences in severity scores among subgroups, reinforcing the value of subgroup-specific analysis in sepsis risk assessment. Our machine learning endeavors extended to the development of subgroup-specific models using Random Forest Classifiers, guided by an emphasis on model interpretability for clinical application. This approach not only enabled precise sepsis risk predictions but also offered insights into feature importance within each subgroup. The final product, an interactive dashboard, integrates our analytical findings with user-friendly visualizations, ensuring healthcare professionals can access and interpret sepsis risk information effectively. This tool represents a significant advancement in data-driven healthcare, offering a more nuanced and comprehensive approach to sepsis risk assessment. Our work underscores the potential of data science in transforming medical diagnostics and patient care, demonstrating the power of combining in-depth data analysis with practical, accessible tools for healthcare professionals.

The development of our dashboard to incorporate sepsis risk assessment hopes to further diagnose sepsis by utilizing and showcasing comorbidity composition for patients. By in- corporating insights from Zador et al (2019) regarding the relationship between pre-existing diseases and sepsis development, we have created a useable dash- board that both include predictions for sepsis risk categorization as well as show doctors patient vitals for professional opinions. Our dashboard goes beyond traditional approaches by providing a more holistic assessment of sepsis risk. Our methodology explored patient vital sign data, utilization of Latent Class Analysis (LCA) to identify distinct comorbidity subgroups, and implementation of machine learning models tailored to each subgroup for predicting sepsis risk categories. The machine learning component, particularly the Random Forest Classifier, proved effective in predicting sepsis risk based on subgroup-specific characteristics compared to the overall population. Furthermore, our dashboard provides healthcare professionals with access to subgroup- based sepsis risk information through an interactive web interface. It presents just-in-time patient information with visualizations of vital sign data overtime, subgroup analysis re- sults, and sepsis risk categorization per patient. Our dashboard equips medical profession- als with insights to support sepsis diagnosis .