Contextual Errors Keep Medical AI From Reaching Its Full Potential in Clinical Settings

Why Medical AI Performance Falls Short in Clinical Practice

Thousands of medical AI models have been developed across academia and industry, yet very few successfully transition into real-world clinical settings. Marinka Zitnik, associate professor of biomedical informatics at Harvard Medical School, and her colleagues have identified a fundamental problem: contextual errors that severely limit real-world performance 1

1

. In a paper published February 3 in Nature Medicine, the researchers explain that while medical AI models may produce responses that appear useful and correct, they often fail to account for the specific context in which they're being used 2

2

.

"This is not a minor fluke," said Zitnik, who is also associate faculty at the Kempner Institute for the Study of Natural and Artificial Intelligence at Harvard University. "It is a broad limitation of all the types of medical AI models that we are developing in the field" 1

1

. The gap between AI performance on standardized tests and actual patient-care settings represents a critical challenge for AI implementation in healthcare.

How Medical Specialty Context Creates AI Blind Spots

The first major source of contextual errors involves medical specialty boundaries. Patients with complex symptoms often span multiple specialties, but AI models in clinical settings struggle to adapt. When a patient arrives at the emergency department with neurological symptoms and breathing problems, they might see both a neurologist and a pulmonologist. Each specialist focuses on their organ system based on their training and experience. An AI model trained mostly on one medical specialty may provide answers based on data from the wrong specialty or miss that the combination of symptoms points to a multisystem disease 2

2

.

Zitnik emphasizes the need to develop AI models trained in multiple specialties that can switch between contexts in real time to focus on whatever information is most relevant. This capability would enable actionable recommendations that account for the full complexity of patient presentations rather than narrow specialty-specific responses.

Geographic Location Shapes Treatment Guidelines and Availability

Geographic location in AI represents another critical contextual gap. If a model gives the same answer to the same question in different locations, that answer is likely incorrect because each place has specific conditions and constraints. For a patient susceptible to a disease that could lead to organ dysfunction or failure, whether they're in South Africa, the United States, or Sweden makes a significant difference in terms of disease prevalence and what treatments and procedures are approved and available 1

1

.

Zitnik's lab is actively working on models that can incorporate geographic information to produce location-specific responses, which could have major implications for global health. This approach addresses health inequities by ensuring AI recommendations align with local treatment guidelines and healthcare infrastructure.

Socioeconomic Factors in AI Determine Clinical Relevance

The third category of contextual errors involves socioeconomic factors that affect patient behavior. Consider a patient who arrives at the emergency department with severe symptoms after previously being referred to an oncologist but never making an appointment. A typical response might be to remind the patient to schedule the oncology appointment, but this recommendation lacks clinical relevance if it ignores the underlying barriers preventing the patient from accessing care 2

2

.

Integrating Context Into AI Models Through Three Key Steps

Contextual errors happen because important information for making clinical decisions is not contained in the training datasets used to develop medical AI models. The models then generate recommendations that seem reasonable and sensible but are not actually relevant or actionable for patients 1

1

.

Zitnik proposes three steps to improve AI performance. First, incorporate contextual information into training datasets. Second, create enhanced computational benchmarks to test models after training. Third, build context awareness into the architectural design of the models themselves. These steps will help ensure that errors are detected before models are implemented in actual patient-care settings and that AI can adapt recommendations in real time based on specific contextual information 2

2

.

The research highlights that addressing contextual errors is essential for building trust in AI and closing the gap between promising test results and successful deployment in hospitals and doctors' offices. As the field moves forward, integrating context into AI models will determine whether these tools can deliver on their potential to analyze vast amounts of information and recognize subtle patterns in ways that genuinely improve patient care.