Understanding social determinants of health (SDoH) is becoming increasingly important in healthcare services and public health research.

Healthy People 2030, a major public health initiative, emphasizes addressing these areas to achieve real health equity.

Traditionally, research has often relied on proxy measures (like using race as a stand-in for the experience of racism) to represent structural factors. While useful, these proxies can miss the complex, relational ways that social determinants interact to shape health outcomes.

Network models offer a new approach: instead of looking at variables one by one, they allow us to model the hidden relationships between different social conditions and populations.

In this project, I used data from the Medical Expenditure Panel Survey (MEPS) to model latent similarity patterns among patients based on SDoH, including food insecurity, access to care, income, and social isolation.

I constructed a patient similarity network using cosine similarity, applied spectral embedding to project patients into a latent space, and used k-means clustering to identify subgroups within the network.

The results revealed subtle but meaningful patterns. Clusters differed in healthcare access barriers, financial instability, and compounded vulnerabilities such as low income paired with food insecurity.

These findings demonstrate how relational network modeling can uncover hidden gradients of disadvantage that traditional feature-by-feature analyses may overlook.

Future work could integrate health outcome data or longitudinal measures to better understand how these latent social structures impact health trajectories over time.

Sources:

[1] Office of Disease Prevention and Health Promotion. Healthy People 2030 Framework. U.S. Department of Health and Human Services.​

[2] Qing, H. (2023). Latent class analysis by regularized spectral clustering. arXiv preprint arXiv:2310.18727. https://doi.org/10.48550/arXiv.2310.18727​
[3] Agency for Healthcare Research and Quality. Medical Expenditure Panel Survey (MEPS), 2022 Full-Year Consolidated Data File.

[4] https://www.geeksforgeeks.org/spectral-embedding/

Background

For my Information Analytics course, I took on the KNIME challenge, where I had full freedom to explore a data science project. KNIME is a data management and analytics platform similar to Alteryx. I planned and executed the entire project independently, exploring different predictive models and evaluating their accuracy in predicting hospital charges. My primary goal was to compare multiple models using a medical cost dataset to determine which one performed the best.

Dataset

I used the Medical Cost Dataset from Kaggle for this project, which includes eight variables like age, BMI, smoker status, and medical charges. My objective was to predict medical charges (target variable) using various predictors (age, BMI, smoker status, etc.).

Model Comparisons

Model 1: Random Forest (Social and Biomedical Variables)

1. R² and Adjusted R²: Both values were high, indicating the model captured 86% of the variance in hospital charges.
2. Mean Absolute Error (MAE): The MAE was 2794.003, representing 21% of the mean charge and 30% of the median charge, indicating potential inaccuracy for predicting lower-cost cases.
3. Root Mean Squared Error (RMSE): The RMSE of 4,406.298, higher than the MAE of 2,794.003, indicates significant outliers in the dataset, as RMSE is more sensitive to large errors and emphasizes the impact of extreme values.
4. Correlation: The correlation between predicted and actual charges was 0.929 (p-value of 0), showing a strong relationship. However, the model underpredicted charges in many cases.

Model 2: Random Forest (Biomedical-Only Variables)

1. R² and Adjusted R²: Both were reported as 1, suggesting a perfect fit, which likely indicates overfitting. Further testing on separate datasets is needed to confirm generalizability.
2. MAE: The MAE was reported as 0, which was unrealistic given discrepancies observed between predicted and actual charges. This raised concerns about the validity of the model’s metrics.
3. Correlation: The biomedical-only model had a correlation of 0.927 (p-value of 0), slightly lower than the social-biomedical model, showing the importance of including social variables.

Model 3: Linear Regression

1. R² and Adjusted R²: These values were lower than Random Forest, explaining 78% of the variance in medical charges.
2. MAE: The MAE was 3770.463, which was higher than in the Random Forest model, representing 28% of the mean charge and 40% of the median charge. This indicates less accuracy in predicting costs, especially in lower-cost instances.
3. Root Mean Squared Error (RMSE):The RMSE was 5,472.896, which is higher than both the MAE and the RMSE of the Random Forest model.
4. Correlation: The model had a correlation of 0.886 (p-value of 0), but it also predicted some charges to be negative, which is unrealistic in a medical cost context.

Model 4: Linear Regression (Biomedical-Only Variables)

1. R² and Adjusted R²: These values remained the same as the social-biomedical model, explaining 78% of the variance.
2. MAE: The MAE was 3,824.172, slightly higher than the social-biomedical linear regression model, indicating less accuracy. This model also underpredicted and overpredicted charges similarly to the combined social-biomedical model.
3. Root Mean Squared Error (RMSE): The RMSE was 5,474.757
4. Correlation: The correlation was 0.886, identical to the social-biomedical model, further suggesting the need for social variables to improve prediction accuracy.

Model 5: K-Nearest Neighbors (KNN)

I experimented with two binning approaches for this model:

1. 5 Bins: The model performed well, with high accuracy:

• Recall: 98.1%
• Precision: 98%
• Overall Accuracy: 99.3% (suggesting simplicity due to fewer bins)

1. 20 Bins: With more bins, accuracy decreased slightly but remained within an acceptable range:

• Recall: 95.5%
• Precision: 95.84%
• Overall Accuracy: 95.5% (which is in line with recommended accuracy ranges for predictive models)

Conclusion

Through this challenge, I gained valuable experience using KNIME and comparing multiple predictive models for hospital charge predictions. I applied both regression and classification models and developed a deeper understanding of how different models perform in healthcare contexts.

Based on performance, I recommend using the Random Forest model with both biomedical and social variables for predicting hospital charges, as it demonstrated the most accurate and reliable results with the given dataset without falling victim to overfitting the training data set.

This project has enhanced my ability to interpret and communicate model results effectively, and I look forward to applying these insights to future healthcare-related predictive analytics projects.

For additional details you can check out the lab report submitted for my course:

TL;DR:

This article explores racial disparities in cancer outcomes in NC using the Mortality-Incidence Ratio (MIR) as a metric to quantify health equity. By comparing MIRs across racial groups with the white MIR and overall MIR as reference points, significant disparities were found, particularly among Black and Native American populations. The analysis underscores the importance of careful benchmark selection in health equity research and highlights the complex factors contributing to these disparities. Addressing these issues requires interdisciplinary research and targeted public health interventions to ensure equitable health outcomes for all

Key Findings:

– Black patients with melanoma had an MIR of 0.43, meaning 43% of those diagnosed in NC died, compared to just 7% of White patients.

– Native American populations faced extreme disparities in ovarian and esophageal cancers, with alarmingly high MIRs.

– Hispanic populations showed fewer disparities compared to the overall and white reference groups, but this finding may be misleading due to the tendency to view this diverse group as a monolith, which can obscure the unique disparities within subgroups.

Introduction

As I embark on my research career, my focus has increasingly centered on health equity—a concept that examines the fairness and justice of health outcomes across different populations. My PhD work aims to develop quantitative methods that better assess and address disparities in healthcare delivery and outcomes. This interest in health equity alongside recent experiences—ranging from my summer internship at the NIH focused on ovarian cancer to my current role in clinical data science within an Oncology clinical trial at Novant Health—have led me to this project.

The Mortality-Incidence Ratio (MIR) offers a powerful metric for this purpose, serving as an indicator of how lethal a disease is relative to its occurrence within a population. By examining the MIR across different racial groups, I aimed to quantify health equity within the realm of cancer care. This analysis compares MIRs using two reference points: the White MIR and the overall MIR, providing insights into how racial disparities manifest in cancer outcomes.

Background

Racial disparities in health outcomes are a critical and well-documented issue in public health, manifesting in various forms across different diseases. These disparities are often driven by a complex interplay of factors, including social determinants of health (such as education, income, and access to healthcare), genetic predispositions, and environmental exposures. The Mortality-Incidence Ratio (MIR) is particularly useful for examining these disparities because it quantifies the severity of a disease by comparing the mortality rate to the incidence rate within a population.

However, it is important to recognize that the MIR is just one piece of the puzzle. Health outcomes, especially in diseases as multifaceted as cancer, are influenced by numerous factors that extend beyond the scope of a single metric. These include healthcare access, the quality of care received, health education, and broader social and environmental determinants, such as healthy food insecurity, access to exercise, and pollution exposure. By understanding these interactions, we can better interpret the disparities revealed through MIR analysis and work toward more equitable health outcomes.

Methods

To investigate racial disparities in cancer outcomes, I conducted an analysis of Mortality-Incidence Ratios (MIRs) across various racial groups using data from the North Carolina Department of Health and Human Services (NC DHHS) for the years 2018-2022. The data was age-adjusted to the standard 2000 population to account for differences in age distribution across racial groups, ensuring that the comparisons were as accurate as possible.

MIRs were calculated by dividing the mortality rate by the incidence rate for each racial group. To provide a comprehensive view of disparities, I used two reference points for comparison: the MIR for the white population and the overall MIR, which represents the aggregated outcomes across all racial groups. This dual approach allowed me to assess how each racial group’s outcomes compared both to the population as a whole and to a specific racial group with historically better health outcomes.

Data

Key findings

The analysis revealed significant disparities in cancer outcomes across racial groups. For instance, from 2018 to 2022, the MIR for melanoma among Black patients in North Carolina was 0.43, indicating that 43% of Black individuals diagnosed with melanoma during this period died from the disease. In contrast, the MIR for melanoma among white patients was markedly lower, at 0.07 (7%).

This disparity likely reflects late-stage diagnosis among Black patients, which can result from several factors:

– Lack of Awareness: There may be a limited understanding among Black patients regarding their risk for melanoma, contributing to delayed diagnosis and treatment.

– Access to Specialized Care: Limited access to dermatological care or healthcare in general can exacerbate the severity of the disease by delaying diagnosis and treatment.

– Physician Training: Many physicians may not receive adequate training on the presentation of melanoma in Black patients, leading to missed or late diagnoses.

Native American populations faced particularly severe disparities in ovarian and esophageal cancers, with MIRs of 1 meaning that all patients diagnosed in 2018 – 2022 with these cancers died. These cancers are already aggressive, but the outcomes were disproportionately worse for Native American patients, likely due to:

– Healthcare Access: Persistent barriers to accessing quality healthcare.

– Intergenerational Trauma: Long-term, intergenerational oppression and its impacts on health.

Interestingly, the data showed fewer disparities for Hispanics when compared to the overall or white reference groups. However, this finding warrants caution. The Hispanic population is often treated as a monolithic group, despite being genetically and culturally diverse. This homogenization can obscure the nuances of disparities within this group. Additionally, underreporting due to immigration status may further distort the data. Notably, the American Cancer Society reports a lifetime cancer mortality risk of 1 in 5 for Hispanic men and 1 in 6 for Hispanic women—figures that are not fully reflected in the North Carolina data.

Discussion

The disparities observed using different reference points underscore the importance of selecting appropriate benchmarks in health equity research. The more pronounced disparities identified using the white MIR suggest that this group benefits from factors—such as better access to care—that improve outcomes, making it a stringent reference point. In contrast, the overall MIR, which averages outcomes across all racial groups, may obscure significant disparities that are critical to understanding and addressing health equity.

Quantifying health equity through metrics like the MIR is essential for identifying where interventions are most needed. However, achieving health equity requires more than just identifying disparities—it demands concerted efforts to address the underlying causes, which are often deeply rooted in social, economic, and political contexts. While health equity may not be a focal point in current political discussions, it is a critical area that must be continuously upheld and prioritized.

Interdisciplinary research is key to advancing our understanding of health disparities. By integrating insights from epidemiology, sociology, economics, and other fields, we can work toward a future where the MIR differential approaches zero across all racial groups. As we look ahead, it will be important to monitor how policy changes, such as the recent expansion of Medicare, impact these disparities—potentially revealing whether the core issue lies in access to care or other systemic factors.

Sources:

Why are so many Black patients dying of skin cancer? | AAMC

Melanoma Among Non-Hispanic Black Americans.

Racial differences in time to treatment for melanoma – PMC

The ongoing racial disparities in melanoma: An analysis of the Surveillance, Epidemiology, and End Results database (1975–2016)).

Cancer statistics for American Indian and Alaska Native individuals, 2022: Including increasing disparities in early onset colorectal cancer

How recognizing diversity among Hispanics could improve health outcomes | AAMC

Cancer Facts & Figures for Hispanics & Latinos 2018-2020