Artificial Intelligence (AI), specifically Natural Language Processing (NLP) models, has played an increasingly vital role in healthcare, from streamlining administrative tasks to supporting clinical decision-making. While OpenAI’s ChatGPT captivated the world in 2022, Chinese AI startup DeepSeek emerged in 2025 as a promising new player, advancing the accessibility and democratisation of AI.
A recent study by Synapxe’s Data Analytics and AI (DNA) team, published by the Singapore Computer Society, compares the effectiveness of applying the two Large Language Models (LLM), DeepSeek and OpenAI, on real-world use cases in healthcare. It aims to provide academic and technical insights, rather than to recommend the use of one AI tool over another.
The study is summarised below, with evaluation focused on four key areas: classification, multilingual responses, medical image interpretation, and synthetic image generation.
1. Adverse Drug Event (ADE) Detection
DeepSeek and OpenAI models were evaluated on their ability to classify ADEs based on clinical narratives, using data from the Active Surveillance System for Adverse Reactions to Medicines and Vaccines (ASAR), a collaborative initiative by Synapxe and the Health Sciences Authority for Singapore’s nationwide ADE monitoring program.
While OpenAI models delivered the highest overall accuracy DeepSeek’s R1 model achieved comparable performance with cost-savings of up to 54 times lower, making it a highly cost-effective solution, especially for resource-constrained environments.
2. Multilingual Conversational AI
To support Singapore’s multilingual population, the study evaluated DeepSeek and OpenAI models in responding to medical inquiries in Chinese, Malay, and Tamil within Synapxe’s digital healthcare companion (more information will be shared later in 2025).
DeepSeek performed best in Chinese, while OpenAI’s GPT models excelled in Malay and Tamil, demonstrating better fluency and coherence.
3. Medical Image Interpretation
Both DeepSeek and OpenAI were tested on identifying masseter muscles in CT scans, a key task for assessing jaw health and frailty. However, neither model achieved reliable results, highlighting the current limitations of LLMs in precise medical image analysis.
4. Synthetic Image Generation
AI-generated images can help enhance training datasets for medical AI models such as FungAI, a joint project by Synapxe and Singapore General Hospital (SGH) that detects a wide range of fungi species, including those that may cause life-threatening diseases.
DeepSeek and OpenAI were tasked with creating realistic microscopy images of Aspergillus fungi. Both models produced results that resembled digital illustrations rather than photorealistic microscopy images. However, DeepSeek’s Janus Pro 7B showed better adherence to the structural characteristics of Aspergillus. The results indicate that AI-generated medical imagery still requires further refinement for real-world applications.
Conclusion
Both AI models offer valuable solutions, but their strengths vary based on specific use cases. For a more detailed explanation of the study’s research methods, the prompts used, and each model’s text and visual responses, download the full study here: DeepSeek in Healthcare: A Brief Exploration of Real-World Use Cases
For more information on the AI projects co-developed by Synapxe, refer to these resources below:
- Active Surveillance System for Adverse Reactions to Medicines and Vaccines (ASAR)
- Clinical Frailty (More information will be published later in 2025)
- FungAI: Leveraging artificial intelligence to identify fungi strains
