Machine Learning for Public Health

Francisco Cardozo, Ph.D. Student

Prevention Science and Community Health

Eric C. Brown, Ph.D.

Prevention Science and Community Health

July 30, 2024

Agenda

Part I - Machine Learning

• Key Concepts
• Simulation Examples

Part II - Generative AI

• Key Concepts
• Tools
• Future Directions
• Ethical Dilemmas

Part I - Machine Learning

What is Public Health?

“Public health is the art and science of preventing disease, prolonging life, and promoting health through the organized efforts of society” - (Winslow 1920)

“Public health is primarily concerned with the health of the entire population, rather than the health of individuals. Its features include an emphasis on the promotion of health and the prevention of disease and disability; the collection and use of epidemiological data, population surveillance, and other forms of empirical quantitative assessment; a recognition of the multidimensional nature of the determinants of health; and a focus on the complex interactions of many factors - biological, behavioral, social, and environmental - in developing effective interventions.” - (Childress et al. 2002)

Scoping review of definitions (Azari and Borisch 2023)

The Importance of Prediction in Public Health

• What are the underlying causes of a health problem? For example, what are the causes of youth alcohol initiation?

• What are the risk factors associated with a health problem? Which demographic and behavioral factors increase the likelihood of developing type 2 diabetes among adults?

• Can we predict the future appearance of a health problem? Can we forecast the potential outbreak of seasonal influenza in the upcoming months?

Machine Learning is Good for Prediction

Traditional Statistics

• Focus on inference rather than prediction
• Limited flexibility with complex, non-linear relationships
• Emphasis on hypothesis testing and confidence intervals
• Designed for smaller data sets

Machine Learning

• Prioritizes predictive accuracy over interpretability
• Handles complex, non-linear relationships
• Uses performance metrics (e.g., accuracy, AUC, F1 score)
• Specialized techniques for unstructured data (e.g., NLP, CNNs)

Machine Learning: The Best Tool for Prediction

I suggest these three concepts are at the heart of effective prediction:

• Testing error
• Regularization
• Bias-variance trade-off

Bonus: Machine learning is optimized for prediction.

Understanding Machine Learning Advantages

Testing Error

• Measures the performance of a model on unseen data.
• Helps to evaluate the generalization ability of a model.
• The estimation of this error is key to optimizing model performance.

Regularization

• Technique to prevent overfitting.
• Controls the complexity of the model.
• Adds some bias to the model to reduce variance.

Live Example

#| standalone: true
#| viewerHeight: 500
library(shiny)
library(glmnet)

# Define UI for application
ui <- fluidPage(
    titlePanel("Effect of L1 Penalty (Lasso) on Regression Slope"),
    sidebarLayout(
        sidebarPanel(
            sliderInput("penalty",
                        "Penalty (Lambda):",
                        min = 0,
                        max = 1,
                        value = 0.1,
                        step = 0.01)
        ),
        mainPanel(
           plotOutput("regressionPlot"),
           verbatimTextOutput("modelSummary")
        )
    )
)

# Define server logic
server <- function(input, output) {
    
    output$regressionPlot <- renderPlot({
        # Generate synthetic data
        set.seed(123)
        x <- matrix(rnorm(200), ncol = 2)
        y <- 3 * x[,1] + 2 * x[,2] + rnorm(100)
        
        # Fit model with Lasso penalty (alpha = 1) using both columns of x
        fit <- glmnet(x, y, alpha = 1, lambda = input$penalty)
        
        # Extract coefficients
        intercept <- coef(fit)[1]
        slope1 <- coef(fit)[2]
        slope2 <- coef(fit)[3]
        
        # Calculate predicted values using intercept and slopes
        y_pred <- intercept + x[,1] * slope1 + x[,2] * slope2
        
        # Combine data for plotting
        data <- data.frame(X1 = x[,1], Y = y, Y_Pred = y_pred)
        
        # Plot data and regression line
        par(mfrow = c(1, 2))
        
        # Plot 1: Regression line
        plot(data$X1, data$Y, main = "Lasso Regression Line (X1 vs Y)",
             xlab = "X1", ylab = "Y", pch = 19, col = "red")
        abline(intercept, slope1, col = "blue", lwd = 2)
        
        # Plot 2: Actual vs Predicted values
        plot(data$X1, data$Y, main = "Actual vs Predicted (X1 vs Y)",
             xlab = "X1", ylab = "Y", pch = 19, col = "red")
        points(data$X1, data$Y_Pred, col = "blue", pch = 19)
        
        par(mfrow = c(1, 1))
    })
    
    output$modelSummary <- renderPrint({
        # Generate synthetic data
        set.seed(123)
        x <- matrix(rnorm(200), ncol = 2)
        y <- 3 * x[,1] + 2 * x[,2] + rnorm(100)
        
        # Fit model with Lasso penalty (alpha = 1)
        fit <- glmnet(x, y, alpha = 1, lambda = input$penalty)
        
        # Display coefficients including intercept
        intercept <- coef(fit)[1]
        slope1 <- coef(fit)[2]
        slope2 <- coef(fit)[3]
        
        cat("Intercept:", intercept, "\n")
        cat("Slope for X1:", slope1, "\n")
        cat("Slope for X2:", slope2, "\n")
    })
}

# Run the application 
shinyApp(ui = ui, server = server)

Bias-variance Trade-off

Bias: Error due to overly simplistic assumptions. Variance: Error due to overly complex models. The goal is to minimize the total error.

Examples

Simulation Example

Let’s simulate a data set and apply machine learning techniques to predict the outcome.

• Simulate two data sets:
  • Linear relationship
  • Non-linear relationship
• Fit a linear regression model to the linear data set.
• Fit a linear regression model to the non-linear data set.
• Fit a support vector machine model to the non-linear data set.
• Fit a random forest model to the linear data set.

Simulation Example

Estimate these models in a few cases. Then progressively increase the sample size by one unit. Calculate the training and testing errors each time.

First Scenario

A linear problem - linear regression.

Second Scenario

Non-linear problem - linear regression.

Third Scenario

Non-linear problem - Support Vector Machine.

Fourth Scenario

A linear problem - Random Forest.

Machine Learning in Public Health

• Machine learning is a powerful tool for prediction in public health.
• It prioritizes predictive accuracy over interpretability.
• It handles complex, non-linear relationships.
• It is optimized for prediction.

Key Concepts

• Testing error: Measures the performance of a model on unseen data.
• Regularization: Technique to prevent overfitting.
• Bias-variance trade-off: The goal is to minimize the total error.

Part II - Generative AI

Transformers

What is a Transformer?

• Transformers are a type of neural network architecture that has been used to achieve state-of-the-art performance on a wide range of natural language processing tasks.
• They are based on the self-attention mechanism, which allows the model to focus on different parts of the input sequence when making predictions.
• Transformers have been used to build large language models that can generate human-like text, translate between languages, and perform other natural language processing tasks.

Transformers Applications

• Audio transcription, image captioning, text-to-speech, speech-to-text, translation, text generation, sentiment analysis, summarization, question answering, chatbots, virtual assistants, customer service, content creation, content moderation.

Audio

• Whisper

Audio

Audio Tumaco

Imagen

• Midjourney
• recraft.ai

Text

What are Large Language Models?

• Something that knows how to predict the next word in a sentence or fill in the missing words in a sentence.
• Large language models (LLMs) are a type of artificial intelligence that can generate human-like text.
• They are trained on vast amounts of text data.

What can you do with LLMs?

• Prompt and RAG the models (López Espejel et al. 2023)

Prompt and RAG the models

Prompt and RAG the models

Prompt and RAG the models

groq

nat.dev

h2o

What can you do with LLMs?

• Fine-tune the models
  • Hugging Face Transformers
• Create smaller expert models for narrow tasks

What can you do with LLMs?

• Study, optimize, and debug the models
• Agents (Park et al. 2023; Kapoor et al. 2024)

Learning from Large Language Models

Leader Board

Advance Metric

Future Directions

• Machine learning in public health is getting bigger.
• More research is needed to understand the potential of machine learning in public health.
• The integration of machine learning with traditional statistical approaches is promising.
• We will have agents that can deliver health information, answer questions, and provide support.
• The future of public health is bright!

However

Ethical Dilemmas

• Bias, fairness, security, privacy, transparency, accountability, and trust.
• The potential for misuse of large language models is a major concern.
• How people apply moral and ethical principles to the development and use of large language models is a critical issue.
• How people reason in terms of accountability when interacting with large language models.

Thank you!

Francisco Cardozo

References

Azari, Razieh, and Bettina Borisch. 2023. “What Is Public Health? A Scoping Review.” Archives of Public Health 81 (1). https://doi.org/10.1186/s13690-023-01091-6.

Childress, James F., Ruth R. Faden, Ruth D. Gaare, Lawrence O. Gostin, Jeffrey Kahn, Richard J. Bonnie, Nancy E. Kass, Anna C. Mastroianni, Jonathan D. Moreno, and Phillip Nieburg. 2002. “Public Health Ethics: Mapping the Terrain.” Journal of Law, Medicine &Amp; Ethics 30 (2): 170–78. https://doi.org/10.1111/j.1748-720x.2002.tb00384.x.

Kapoor, Sayash, Benedikt Stroebl, Zachary S. Siegel, Nitya Nadgir, and Arvind Narayanan. 2024. “AI Agents That Matter.” arXiv. https://doi.org/10.48550/ARXIV.2407.01502.

López Espejel, Jessica, El Hassane Ettifouri, Mahaman Sanoussi Yahaya Alassan, El Mehdi Chouham, and Walid Dahhane. 2023. “GPT-3.5, GPT-4, or BARD? Evaluating LLMs Reasoning Ability in Zero-Shot Setting and Performance Boosting Through Prompts.” Natural Language Processing Journal 5 (December): 100032. https://doi.org/10.1016/j.nlp.2023.100032.

Park, Joon Sung, Joseph C. O’Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. 2023. “Generative Agents: Interactive Simulacra of Human Behavior.” arXiv. https://doi.org/10.48550/ARXIV.2304.03442.

Winslow, C.-E. A. 1920. “The Untilled Fields of Public Health.” Science 51 (1306): 23–33. https://doi.org/10.1126/science.51.1306.23.