Researchers develop AI method to diagnose patients from incomplete medical records — handles missing test results and scans without sacrificing accuracy

A team of researchers published a new machine learning technique that processes patient medical data even when some test results, imaging scans, or other diagnostic information is missing. The method reframes diagnosis as a sequence prediction task, using transformer-based AI (the same architecture behind ChatGPT) trained on MIMIC-IV and eICU hospital datasets — two of the largest real-world patient record collections.

The key innovation: the AI learns which data points are actually missing versus simply not recorded, then uses 'contrastive pre-training' (a technique where the model learns by comparing similar and different examples) to understand patient trajectories in a shared conceptual space. This means the model predicts diagnoses not from perfect, complete records, but from the messy reality of sparse clinical data — where a patient might have bloodwork but no recent imaging, or imaging but incomplete test panels.

For hospital IT teams and healthcare AI developers, this solves a major deployment headache: production systems encounter incomplete records constantly due to equipment downtime, patient no-shows, or insurance delays. Previously, models either dropped incomplete patients or required imputation (guessing missing values), both of which degraded accuracy. This work means diagnostic AI can now work reliably on real patient populations without data cleanup, potentially accelerating adoption of AI-assisted diagnosis in understaffed hospitals.