A new neural network architecture is built to stop treating missing medical scans as empty noise.
Researchers behind a new paper propose CAMOS, a model for tracking disease progression from clinical records where tests like PET scans and spinal fluid analysis are missing from roughly half of all visits in ADNI, a widely used Alzheimer's imaging cohort. They first prove a limitation in existing linear state-space models: when a measurement is missing, those models simply mask the input and leave the rest of their internal math untouched, which means they mathematically cannot learn how two missing (or two present) measurements interact. CAMOS fixes this by giving each measurement type a bank of coupled oscillators whose connections switch on or off based on which tests were actually taken, so the model's core transition equations themselves change depending on what data exists at a given visit. To keep that more flexible math from becoming unstable, the authors add a stiffness budget that holds steady across every possible combination of missing and present measurements.
On ADNI, CAMOS beat both uncoupled oscillator models and standard clinical fusion models at same-visit staging, predicting future cognitive decline, and longitudinal forecasting. The more important result came from zero-shot testing on a separate dataset, OASIS-3, where CAMOS was the only model that didn't collapse into just guessing the most common outcome, the usual failure mode when a model meets data patterns it wasn't trained on. That generalization gap matters more than any single benchmark number, since real hospitals rarely have the same missing-data patterns as a research cohort.
It's still one paper on one disease using two overlapping dementia datasets, which is a long way from a system any clinic could plug into its own occasionally-incomplete charts.