top of page
  • Hongjian Zhou

Newsletter from The Neural Medwork: Issue 13



Welcome to the 13th edition of the Neural Medwork Newsletter. After unpacking the Perceptron in our last discussion, today we're diving into another fascinating and widely used model in machine learning—Support Vector Machines (SVM). We then present a research paper detailing the approach to implementing AI solutions in healthcare. Lastly, we introduce the Advanced Reasoning with Tools technique to help you utilize large language models (LLMs) for complex healthcare use cases.


Core Concept: Support Vector Machine

What is a Support Vector Machine?

Support Vector Machine (SVM) is a robust supervised machine learning algorithm used primarily for classification and regression challenges. At its heart, an SVM model attempts to find the best dividing boundary (or hyperplane) that separates data into classes. This boundary is chosen to maximize the distance between the data points of different categories, known as the margin. The data points that influence where the hyperplane is placed are called support vectors, hence the name.

How Does SVM Work?

  1. Finding the Right Hyperplane: SVM builds a model that assigns new data points to one category or another by finding a hyperplane that best divides a dataset into classes.

  2. Maximizing the Margin: The algorithm looks for the largest possible margin between the data points of different classes. Larger margins are associated with lower generalization errors of the classifier.

  3. Handling Non-linear Data: When data isn't linearly separable, SVM uses a method called the kernel trick to transform data into a higher dimension where a hyperplane can be used to separate classes.

A Clinical Example: Diagnosing Diabetes

To illustrate, let's consider a clinical example relevant to healthcare: developing a model that can predict whether a patient will have diabetes. Imagine we have a dataset containing medical records of patients, including their glucose levels, body mass index (BMI), age, and whether they have diabetes. Our goal is to develop a model that can predict whether a new patient might have diabetes based on these features.

How SVM Works in This Scenario:

  1. Data Preparation:

  • Each patient's record (glucose levels, BMI, age) is plotted in a multidimensional space (each feature represents one dimension).

  • SVM will create a hyperplane that divides the patients into two groups: those with diabetes and those without.

  1. Maximizing the Margin:

  • SVM finds the hyperplane that has the maximum margin between the two classes, using the nearest points (support vectors).

  1. Classification of New Data:

  • When a new patient's data is introduced, SVM uses this hyperplane to determine which side of the plane the new data falls on—predicting whether the patient might have diabetes based on the proximity to the support vectors.

Support Vector Machines offer a powerful tool for medical diagnostics, enhancing our ability to interpret complex datasets and make informed clinical decisions. By utilizing such advanced technologies, we can improve accuracy in diagnostics and potentially increase the effectiveness of our treatment plans.


Relevant Research Paper:  The algorithm journey map: a tangible approach to implementing AI solutions in healthcare

The paper provides a detailed account of integrating the SepsisWatch AI tool into Duke University’s healthcare systems, which illustrates a comprehensive approach to deploying AI solutions in real clinical settings. The purpose of this study is to offer a structured roadmap—the "Algorithm Journey Map"—for the development, integration, and ongoing maintenance of AI tools in healthcare, using SepsisWatch as a case study. 

Development Phase

  • Problem Identification: Early identification of sepsis, a critical health issue, using AI to improve detection rates and patient outcomes.

  • Model Building: Involvement of interdisciplinary teams comprising engineers, clinicians, and data scientists to develop a predictive model based on patient data.

  • Challenges: Addressing data variability, model reliability, and ensuring the model's interpretative capabilities.

Integration Phase

  • System Integration: Detailed steps for embedding the SepsisWatch algorithm into the existing clinical workflows at Duke University Health System.

  • Stakeholder Engagement: Emphasis on involving all relevant stakeholders, including physicians, nurses, IT staff, and administrative personnel to foster acceptance and usability.

  • Training and Adaptation: Training sessions for end-users to ensure smooth adaptation to the AI tool and modifications based on feedback.

Maintenance Phase

  • Ongoing Evaluation: Regular assessment of the tool’s performance and impact on patient care, adjusting strategies based on findings.

  • Updates and Improvements: Continuous updates to the algorithm based on the latest medical research and feedback from users to enhance accuracy and usability.

  • Sustainability Efforts: Efforts to maintain the tool’s relevance and utility over time, including scalability considerations and ethical implications.

Conclusion: This study provides a journey map illustrating the complex processes involved with truly integrating an AI tool into clinical practices. It highlights many lessons in the journey to developing and deploying AI in clinical practice including lessons learned in modelling assumptions, the importance of stakeholder engagement, and the organizational structure required to successfully integrate AI tools into practice.

William Boag, Alifia Hasan, Jee Young Kim, et al. "The algorithm journey map: a tangible approach to implementing AI solutions in healthcare." npj Digital Medicine, vol. 7, no. 87, 2024. DOI: 10.1038/s41746-024-01061-4.


Tips and Tricks: Advanced Reasoning with Tools (ART)

Advanced Reasoning with Tools (ART) is an innovative framework that enhances the capabilities of Large Language Models (LLMs) by incorporating multi-step reasoning and dynamic tool use. Proposed by Paranjape et al., (2023), ART utilizes a frozen LLM to automate the generation of intermediate reasoning steps, integrating external tools as needed throughout the process. This method is particularly suited for ucomplex medical tasks where accuracy and adaptability are paramount.

What is ART: ART strategically combines the chain-of-thought (CoT) prompting approach with selective tool usage to tackle sophisticated challenges. By selecting appropriate demonstrations from a task library and pausing generation to integrate tool outputs, ART enables LLMs to apply detailed, context-specific solutions in a zero-shot fashion. This means the model can handle new tasks effectively without prior direct examples, making it highly adaptable and scalable.

Practical Example:

Consider a situation where a healthcare provider uses an LLM powered by ART to assess a patient's risk of developing complications from multiple medications. The ART framework would:

  1. Select demonstrations that illustrate how to analyze drug interactions.

  2. Pause at key points to incorporate data from updated drug interaction databases.

  3. Resume generating the assessment by synthesizing the information gathered through automated reasoning and external tools.

This process ensures that the AI's output is not only based on deep understanding and logical reasoning but is also informed by the latest medical data, enhancing both the reliability and precision of the consultation.

ART's ability to be extended and refined by adding new tools or updating the task library allows continuous improvement and customization, making it an invaluable asset in the dynamic field of healthcare. This adaptability is crucial for keeping pace with the rapid advancements in medical knowledge and technology, providing healthcare professionals with a powerful tool to enhance patient care and decision-making processes.

Thanks for tuning in,

Sameer & Michael


bottom of page