P-MedAgent:Medical diagnostics based on LLM programmable CoT

In the contemporary medical field, artificial intelligence (AI) is emerging as a pivotal force reshaping healthcare, driven by increasingly rich annotated datasets and continuously evolving deep learning models. AI demonstrates significant potential in clinical decision-making processes such as data processing, diagnosis, and drug treatment, promising to fundamentally advance the medical industry. We addressed a general task in medical data analysis: given a set of medical data inputs, the goal is to predict and classify outcomes, such as disease progression and the presence of illness. Traditional approaches typically involve constructing a convolutional neural network (Covnet) trained on the image dataset for this task. However, when training from scratch, setting training parameters, utilizing the training set to minimize the loss function, and continuously adjusting these parameters require substantial human intervention. The effectiveness of data analysis ultimately depends on the constructed network. In this paper, we demonstrate that transforming the data analysis task into a code generation task, by constructing a specialized agent designed for the medical field, called the Programmable Medical Agent (P-MedAgent), can be highly effective. By leveraging techniques such as Programmable CoT and problem decomposition, the agent adjusts input preprocessing or model parameters in real-time based on data format and task requirements, thereby enhancing the model’s generalization capability.We applied our method to five different medical diagnostic tasks, achieving scores of 60.58 SMAPE, 98.75% accuracy, 0.8284 score, 85.84% accuracy, and 91.3% accuracy, respectively. Our results indicate that using our code generation approach to solve these tasks can achieve training levels comparable to human performance.