top of page
  • Hongjian Zhou

Newsletter from The Neural Medwork: Issue 5



Welcome back to the 5th newsletter of The Neural Medwork!

This newsletter covers Decision Trees in AI, a simple and interpretable method used in healthcare, mirroring physicians' decision-making process. It also introduces a new paper which discusses AI's evolution in healthcare through three epochs, from symbolic AI to generative models, highlighting changing capabilities and risks. Lastly, it explains ChatGPT's temperature setting, which is essential for tailoring responses from accurate medical information to creative brainstorming.


AI Concept: Decision Trees in AI

Decision Trees are a form of supervised learning frequently used for classification and regression tasks. They have been around since the 1960s and remain popular due to their simplicity and interpretability. They are particularly effective for problems with a clear, hierarchical decision-making process. There are three main components of a decision tree:

  1. Root Node: This is where the decision tree starts. It represents the entire dataset and makes the initial, most significant decision to split the data.

  2. Internal Nodes: These nodes test a condition or attribute, leading to further branching based on the outcome. There can be an unlimited number of Internal nodes in decision trees.

  3. Leaf Nodes: The endpoints of the tree, where a prediction or decision is made based on the path followed.

In healthcare, decision trees can be compared to the systematic approach physicians use to make diagnoses. Think about, for instance, when evaluating a patient with a cough, a physician might start at the 'root' by asking about the duration of the cough. Depending on the answer, the 'internal nodes' might involve questions about associated symptoms like fever, weight loss, phlegm, hemoptysis, etc. As you move down the decision you will eventually land in a 'leaf node' that could represent the possible diagnoses, such as pneumonia or lung cancer. A standard and straightforward algorithm that could be turned into a simple decision tree could be the ACLS algorithms we use in cardiac arrest. 

One of the hallmarks of decision trees is that they may be straightforward and can mimic human decision-making, which leads to explainability. However, they have limitations, such as a tendency to overfit data and a lack of flexibility in capturing complex relationships in data without growing excessively large and complex. We will elaborate on some of these concepts in further detail in subsequent newsletters. 


Relevant Research Paper

Title: Three Epochs of Artificial Intelligence in Health Care

Purpose: This special communication written by the team at Google (Michael D. Howell, Greg S. Corrado, and Karen B. DeSalvo) published in JAMA, discusses the evolution of artificial intelligence (AI) in healthcare, categorizing it into three epochs. The first epoch, AI 1.0, includes symbolic AI and probabilistic models, focusing on encoding human knowledge into computational rules. The Decision Tree algorithm we highlighted above falls under this category and allows us to understand the beginnings of AI in healthcare. AI 2.0 marks the era of deep learning, where models learn from labelled examples, significantly advancing healthcare applications. The latest, AI 3.0, involves foundation models and generative AI, capable of performing various tasks without retraining on new datasets and showing new capabilities and risks. Each epoch is characterized by fundamentally different capabilities and risks, offering a framework for understanding AI's evolution in healthcare. The image below summarizes these three epochs and allows us to understand the past, present, and future of AI in healthcare.

Howell MD, Corrado GS, DeSalvo KB. Three Epochs of Artificial Intelligence in Health Care. JAMA. 2024;331(3):242–244. doi:10.1001/jama.2023.25057


Tips and Tricks: Setting the Temperature of ChatGPT

In Large Language Models like ChatGPT, the temperature setting controls the randomness of responses. It ranges from 0 to 1, where 0.1 yields more deterministic and predictable outputs, and 1 leans towards creative and varied responses. The default setting is typically around 0.7.

  • When to Use Lower Temperature: In medical contexts, where accuracy and precision are paramount, setting the temperature closer to 0 can be beneficial. For instance, when seeking specific medical information, diagnosing conditions, or when clarity is crucial, a lower temperature helps ensure focused and reliable answers. 

  • Example: Temperature: 0.2. Please provide a concise, evidence-based explanation of the management guidelines for Type 2 Diabetes, focusing on the latest pharmacological treatments.

  • When to Use Higher Temperature: For brainstorming sessions, exploring creative healthcare solutions, or generating a range of potential hypotheses, a higher temperature can be advantageous. It encourages diverse and inventive ideas, which can be helpful in scenario planning or developing novel approaches to patient care.

  • Example: "Temperature: 0.9. Can you generate some innovative ideas for patient engagement strategies in a primary care setting, focusing on digital technology and community involvement?"

  • Changing the Temperature: The tone of ChatGPT's responses can be adjusted through the temperature parameter. By incorporating this into your prompt, you can tailor the AI's output to your specific needs. For instance, you might specify a temperature setting directly in your query, depending on whether you require straightforward information or more creative input.

  • Experimentation is Key: Finding the optimal temperature setting for your needs might require some trial and error. The context of your inquiry and your desired outcome will guide you in fine-tuning this parameter.

Remember, the temperature setting in ChatGPT is a powerful tool for healthcare professionals, enabling you to customize AI interactions to suit a wide range of clinical and educational scenarios. By mastering this feature, you can make ChatGPT a more effective and versatile aid in your practice.

Thanks for tuning in,

Sameer & Michael


bottom of page