top of page
  • Michael Zhou

Newsletter from The Neural Medwork: Issue 6



Welcome back to the 6th newsletter of The Neural Medwork! This newsletter covers random forest algorithms, a step up from the decision trees covered in our previous newsletter. We then present a paper evaluating GPT-4 as a clinical decision-support tool in Ischemic Stroke Management. Finally, we introduce a novel prompt engineering technique called self-consistency, which you can use to enhance the performance of your GPTs.


Core Concept: Random Forest Algorithms

Welcome back to the sixth edition of The Neural Medwork. Continuing from our last discussion on decision trees, we're diving into Random Forest algorithms, an advanced ensemble learning method in AI. Unlike a singular decision tree, a Random Forest algorithm constructs numerous decision trees during training. Each tree operates on a subset of the data, and the algorithm outputs the most frequent class (or mode) from these trees' results, essentially forming a consensus from multiple decision-making processes. This approach is a type of supervised learning.

Let's consider building an AI algorithm for diagnosing chest pain. A single decision tree would navigate through one algorithmic path to reach a diagnosis. But chest pain diagnosis can vary significantly depending on the specialist's perspective. Here, a Random Forest model becomes advantageous. In this model, the patient's symptoms are evaluated through multiple decision trees, each representing a distinct medical specialist's approach – be it an ER physician, a cardiologist, a family doctor, or an internist. Each tree contributes its diagnosis based on its specific expertise and rules. The Random Forest then aggregates these diverse perspectives to reach a consensus diagnosis.

The Random Forest approach offers several benefits over a single decision tree. It is more resilient to overfitting and can effectively handle large datasets with numerous variables. It also maintains accuracy even when data is missing or incomplete. However, Random Forests come with challenges. They can be computationally intensive, requiring more resources as the number of trees increases. This complexity can also make them less transparent and harder to interpret than simpler models, especially for those new to machine learning.

In summary, Random Forest algorithms provide a powerful and collaborative approach to decision-making in AI. They mirror the multifaceted nature of medical diagnosis, showcasing how varied inputs can lead to more robust and reliable outcomes. As we explore different AI models, understanding these algorithms helps us appreciate their role in enhancing healthcare.


Relevant Research Paper

Title: Evaluating GPT-4 as a Clinical Decision Support Tool in Ischemic Stroke Management

Purpose: The study aimed to assess GPT-4's effectiveness in providing clinical decision support for emergency room neurologists in ischemic stroke management, comparing its recommendations with expert opinions and actual treatment outcomes.


  • Design: Retrospective review of 100 consecutive patients with acute stroke symptoms.

  • Participants: Cases involved patients' history, clinical evaluations, imaging study results, etc.

  • Setting: Emergency department data analysis.

  • Evaluation Criteria: GPT-4 provided scaled treatment recommendations, estimated 90-day mortality probability, and elucidated reasoning, compared against stroke specialist opinions and actual treatment decisions.

Key Findings:

  • Treatment Recommendations: GPT-4's agreement with expert opinion showed an AUC of 0.85 and with actual treatment decisions, an AUC of 0.80.

  • Mortality Prediction: Accurately identified 10 out of 13 patients who died within 90 days among its top 25 high-risk predictions.

  • Explainability: GPT-4 provided reasoning for recommendations, aligning well with routine workflows.

Conclusion: GPT-4 demonstrates potential as a viable clinical decision support tool in ischemic stroke management, with its ability to offer explainable recommendations enhancing its utility. Further prospective validations are suggested for integrating such AI tools into clinical practice.

Haim, A et al. 2024. Evaluating GPT-4 as a Clinical Decision Support Tool in Ischemic Stroke Management. medRxiv. doi:


Tips and Tricks: Enhancing CoT Prompting with Self-Consistency

Self-consistency, as introduced by Wang et al. (2022), represents an advanced leap in prompt engineering for Large Language Models (LLMs). This technique refines the concept of chain-of-thought (CoT) prompting by sampling multiple, diverse reasoning paths. If you have not read our guide on CoT, we strongly recommend you check it out! Instead of relying on a singular, potentially biased chain, self-consistency encourages the generation of various answers through few-shot CoT, aiming to identify the most consistent response among them. This approach significantly enhances LLM performance, especially in tasks requiring arithmetic and commonsense reasoning, by moving beyond the limitations of naive greedy decoding.

What is Self-Consistency: Self-consistency is designed to elevate the accuracy and reliability of LLM outputs by leveraging the model's capacity to explore multiple reasoning avenues. For healthcare professionals, this means enhanced decision-making support from AI, particularly in complex scenarios where nuanced reasoning and high accuracy are critical. Employing self-consistency in medical inquiries helps ensure that the AI's recommendations or analyses are not only based on a broader spectrum of considerations but also align more closely with logical consistency and clinical realities.

Practical Example:

Consider utilizing ChatGPT for evaluating treatment options for a patient with multiple drug allergies. A self-consistency prompt could look like this:

"List three different treatment options for a patient allergic to penicillin and cephalosporins, presenting with community-acquired pneumonia. For each option, provide a brief rationale. Then, evaluate which option appears most consistently recommended across your answers and explain why."

This method allows ChatGPT to critically assess various alternatives, weigh their merits based on the patient's unique context, and converge on the most consistently viable option. It mirrors the analytical depth a clinician would apply, ensuring that the AI's contribution is both thorough and dependable. Through self-consistency, healthcare AI becomes a more reliable collaborator in patient care, supporting professionals with insights grounded in a comprehensive evaluation of possibilities.

Thanks for tuning in,

Sameer & Michael


bottom of page