Medical artificial intelligence (AI) that reads patient records and extracts critical information works remarkably well in English, but healthcare systems worldwide are discovering a troubling gap: the same technology largely fails in other languages, leaving vast amounts of clinical data unprocessed and inaccessible. A comprehensive review of 46 recent studies reveals that this isn't a simple technical problem to solve, but rather the result of three interconnected resource shortages that vary dramatically depending on which language and healthcare system is involved.

Named entity recognition, or NER, is a foundational natural language processing (NLP) task that teaches AI systems to identify and extract specific medical information from unstructured text. Think of it as teaching a computer to read a doctor's notes and automatically highlight medication names, disease diagnoses, patient symptoms, and treatment procedures. For English medical text, researchers have built specialized AI models like BioBERT, ClinicalBERT, and PubMedBERT that achieve state-of-the-art performance. But when researchers try to apply these same approaches to languages spoken by billions of people, the results deteriorate significantly.

Why Does Medical AI Work Better in English Than Other Languages?

The answer isn't that other languages are inherently harder to process. Rather, the problem stems from what researchers call the "three scarcities framework," which identifies three distinct resource bottlenecks that constrain medical NLP development in low-resource languages.

• Data Scarcity: English-language medical datasets are abundant because major research institutions, hospitals, and pharmaceutical companies have invested decades in digitizing clinical records and publishing research. Other languages lack comparable volumes of annotated training data, meaning AI models have far fewer examples to learn from.
• Model Scarcity: The most powerful pre-trained language models are predominantly built on English text. Adapting these models to other languages requires significant computational resources and expertise that many healthcare systems in lower-income regions cannot afford.
• Infrastructure Scarcity: Even when data and models exist, the computing power, cloud services, and technical expertise needed to deploy and maintain these systems are unevenly distributed globally, creating a digital divide in healthcare technology access.

The interaction of these three scarcities creates a compounding problem. A healthcare system in a country where the primary language is not English faces not just one obstacle, but a combination of all three, making it exponentially harder to build effective medical AI systems.

What Strategies Are Researchers Using to Bridge the Gap?

The review identified several approaches that researchers are testing to overcome these constraints, though the effectiveness of each method depends heavily on the specific scarcity profile of the language and region in question.

• Cross-Lingual Transfer: Researchers train models on high-resource languages like English, then adapt them to work with lower-resource languages by leveraging shared linguistic patterns and structures across languages.
• In-Domain Pre-Training: Instead of relying on general language models, teams build specialized models trained specifically on medical text in the target language, even if the dataset is smaller than ideal.
• Annotation Projection: Researchers use automated tools to transfer annotations from English medical texts to parallel texts in other languages, reducing the manual labeling burden.
• Back-Translation: AI systems translate medical text from a low-resource language into English, process it with established English models, then translate results back, creating a workaround for limited native-language resources.
• Large Language Model-Based Synthetic Data Generation: Researchers use advanced AI models like ChatGPT to generate artificial training examples in target languages, supplementing scarce real-world data.

However, the review found that no single technique works universally. A method that succeeds in one language or healthcare context may fail in another, depending on how data, model, and infrastructure scarcities interact.

How Can Healthcare Systems Decide Which Approach to Use?

The researchers behind this review developed a practitioner-oriented decision roadmap that helps healthcare administrators and AI teams choose appropriate strategies based on their specific resource constraints. Rather than assuming all low-resource settings are identical, the roadmap acknowledges that a hospital system in India faces different constraints than one in Eastern Europe or sub-Saharan Africa, and therefore requires different solutions.

This tailored approach is critical because healthcare systems worldwide generate enormous volumes of textual data in patient records, clinical notes, and medical literature. The value of that data depends entirely on whether language technologies exist to process it. Without effective medical NER systems in non-English languages, hospitals cannot automate information retrieval, support clinical decision-making, or extract insights that could improve patient outcomes.

What Gaps Remain in Current Research?

The review also identified persistent weaknesses in how researchers evaluate medical NER systems across languages. Inconsistent reporting standards and limited use of statistical significance testing make it difficult to compare results across studies or determine whether improvements are meaningful or simply due to chance. This fragmentation in evaluation practices slows progress because researchers cannot easily build on each other's work or identify which approaches are genuinely most effective.

The implications extend far beyond academic research. Billions of people receive healthcare in languages other than English, and their medical data remains largely inaccessible to AI-driven tools that could improve diagnosis, treatment, and clinical outcomes. Closing this gap requires not just better algorithms, but sustained investment in data collection, model development, and infrastructure in regions where healthcare systems currently lack these resources.

What Does This Mean for the Future of Global Healthcare AI?

The three scarcities framework provides a structured way to think about the problem, but solving it will require coordinated effort across multiple stakeholders. Healthcare institutions, technology companies, research organizations, and governments all have roles to play in addressing data, model, and infrastructure scarcities in low-resource languages.

Without deliberate action, the gap between English-language medical AI and AI for other languages will likely widen as English-focused research accelerates. That outcome would entrench existing health inequities, leaving billions of patients in non-English-speaking regions without access to the same AI-powered clinical tools available in wealthier, English-dominant healthcare systems.