Can DeepSeek transform healthcare in low- and middle-income countries? Equity, governance, and deployment strategies

The global enthusiasm surrounding large language models (LLMs) has given rise to a renewed optimism for addressing health inequities, particularly in low- and middle-income countries (LMICs).1 However, despite significant advances in AI-driven clinical tools, most innovations remain confined to high-resource settings. LMICs continue to grapple with systemic barriers, including fragmented digital infrastructure, persistent clinician shortages, data scarcity, and severe budgetary constraints. Recent studies have confirmed that DeepSeek-R1 performs on par with leading proprietary models. In a 125-case evaluation, DeepSeek-R1 matched GPT-4o in treatment recommendation accuracy (Likert mean score: 4.48 vs. 4.70, p = 0.1522) and significantly outperformed Gemini-2.0 Flash (p = 0.0235). Across a range of diagnostic tasks, DeepSeek demonstrated comparable performance to GPT-4o, thereby substantiating its viability as a cost-effective yet clinically proficient substitute.2,3 In this context, DeepSeek, an open-source, transparent, multilingual, and locally deployable LLM, has emerged as a viable AI solution for under-resourced health systems due to three significant attributes: transparency and explainability, economic accessibility, and local customization and deployment.

Firstly, transparency is paramount in healthcare. DeepSeek’s open-source architecture facilitates comprehensive visibility, thereby enabling full insight into the model weights, data flows, and decision pathways. This auditability is vital for regulators, clinicians, and academic institutions seeking to align AI outputs with national clinical guidelines and ethical frameworks. In contradistinction to the opaque nature of proprietary LLMs, DeepSeek facilitates fine-tuning of the model and fosters trust through verification. However, transparency alone is insufficient. In many LMICs, regulatory ecosystems remain underdeveloped, lacking clear legal, technical, and ethical guidelines for AI oversight. This may result in model deployment not being reviewed, making it difficult to identify and correct biases or delusional outputs. In such settings, the benefits of DeepSeek's explainability are limited unless paired with robust oversight infrastructure.

Secondly, the cost of commercial LLMs—largely driven by cloud-based inference fees—is prohibitive for LMICs. In contrast, DeepSeek can be deployed on local servers or national cloud platforms, reducing recurring expenses and enabling countries to retain control over sensitive health data.1 This further supports both fiscal sustainability and data sovereignty. Nevertheless, local deployment requires foundational digital infrastructure, which many rural and grassroots health facilities in LMICs still lack. Unreliable internet connectivity, inadequate power supply, and insufficient cybersecurity pose substantial barriers. While DeepSeek reduces operational costs in the long term, the initial investments in digital infrastructure and the ongoing maintenance requirements may strain the already overstretched health budgets of LMICs.

Thirdly, in addition to considerations of performance and affordability, changes in the broader policy environment significantly affect AI deployment outcomes. For instance, many Chinese healthcare institutions are competing to localize the deployment of DeepSeek due to recent policy support. Prior to this, China’s policy of prohibiting medical institutions from deploying non-open-source LLMs greatly restricted their use of LLMs. Many LMICs operate within fragmented digital governance ecosystems marked by inconsistent procurement procedures, weak data protection laws, and regulatory uncertainty. Scaling up DeepSeek in LMICs therefore demands more than technical localization. It requires investments in policy support, governance mechanisms, clinical capacity-building, and post-deployment monitoring. Johnson et al. have proposed a three-stage model for equitable AI governance: (1) inclusive model development, (2) workflow-integrated testing, and (3) iterative post-deployment monitoring.4 For instance, in the development stage, fairness metrics should go beyond overall accuracy to address subgroup parity and contextual relevance. During deployment, continuous feedback loops should guide system updates. After deployment, audits must detect hallucinations or demographic drift, while thresholds for safety and user literacy support mechanisms must be enforced.

Taking China’s deployment reality experience as a mirror, as of May 3, 2025, based on the official information released by specific medical institutions, there are approximately 517 institutions that have completed deployment and started to use DeepSeek. The specific distribution is shown in Table 1. As shown in the table, DeepSeek has been deployed mainly in urban tertiary hospitals, tending to cause growing rural-urban divides. This trend underscores a broader concern: without targeted strategies, even low-cost AI models may unintentionally amplify existing inequalities. While localized deployment helps mitigate concerns about cloud data leaks, it shifts data security responsibility onto hospitals—many of which lack the necessary cybersecurity infrastructure. The cost of addressing these vulnerabilities can be substantial, particularly for smaller or rural facilities.