Researchers have developed a new artificial intelligence (AI) model designed to better predict patient health outcomes by analyzing electronic health records (EHRs). This breakthrough is important because it not only improves prediction accuracy but also helps doctors understand which medical factors influence those predictions. By making AI decisions more transparent, this approach could increase trust and usefulness in clinical settings.
Key Takeaways
- The new model, called BERT-LER, processes complex patient data from millions of medical records, including detailed lab test results.
- BERT-LER combines prediction accuracy with explainability, highlighting specific medical events that affect its predictions.
- When tested on public benchmarks and asthma severity progression, the model matched or outperformed existing tools, especially on tasks involving laboratory data.
- The explainability method aligns well with known clinical risk factors, potentially aiding doctors in interpreting AI-driven insights.
At the heart of this research is BERT-LER, a model inspired by BERT, a well-known AI language model originally designed to understand human language. Instead of reading text, BERT-LER reads “coded” medical records—structured data that includes diagnoses, treatments, and lab results organized in timelines for each patient. A key innovation is how the model handles laboratory test values. Instead of treating them as simple numbers, the researchers converted these values into categories based on percentiles, preserving important gradations while making the data easier for the model to interpret.
To make the model’s predictions understandable, the team used a technique called Integrated Gradients. This method assigns “attribution scores” to each piece of input data, showing how much each medical event contributed to the final prediction. For example, it can highlight which lab tests or diagnoses had the biggest impact on assessing a patient’s risk of worsening asthma. This level of transparency is crucial for clinical adoption, as doctors need to trust and verify AI recommendations before applying them in real patient care.
The researchers trained BERT-LER on a massive, de-identified dataset covering 75 million patients, which allowed the model to learn patterns across diverse populations and conditions. They then evaluated its performance on the EHRShot benchmark—a public test suite for clinical prediction models—and on a real-world study tracking asthma severity over time. In both cases, BERT-LER demonstrated strong predictive ability, particularly excelling at tasks involving laboratory data, where many previous models struggled.
Beyond improved accuracy, BERT-LER’s ability to provide meaningful explanations about its predictions marks a significant step forward. This combined focus on performance and interpretability addresses a common challenge in healthcare AI, where complex models often act as “black boxes.” By revealing the medical factors driving predictions, BERT-LER could help clinicians better understand patient risks and tailor treatments accordingly.
Looking ahead, the researchers suggest that their model and explainability approach can be adapted to a wide range of diseases and clinical questions. As healthcare systems continue to collect vast amounts of structured EHR data, tools like BERT-LER may become valuable aids in personalized medicine, helping doctors make informed decisions based on both data-driven insights and clinical expertise. Future work will likely explore expanding the model to other therapeutic areas and integrating it into real-world clinical workflows.
Based on research published on arXiv by Jun Ni Du, Lukas Adamek, Maxim Kryukov et al..