Machine learning algorithms can identify patients at risk for cervical spondylotic myelopathy—the most common cause of spinal cord dysfunction in older adults—up to 30 months before clinical diagnosis using electronic health records from 2 million patients. Researchers compared various AI architectures, from simple clinically-guided networks to large foundation models like clmbr-t-5k-CSM, finding that while complex models performed better on internal datasets, simpler domain-informed approaches showed superior external validation across different healthcare systems.
This represents a significant advancement in addressing CSM's notorious diagnostic delays, which often span years due to subtle symptom onset that mimics normal aging. Early identification could fundamentally change patient outcomes, as CSM typically progresses irreversibly without timely surgical intervention. The finding that clinically-informed models generalize better than foundation models challenges the current AI trend toward increasingly complex architectures. For healthcare systems, this suggests that targeted, interpretable algorithms may prove more reliable for real-world deployment than sophisticated black-box models. The 30-month prediction window provides substantial lead time for preventive interventions, though the study's reliance on retrospective claims data means prospective validation remains essential before clinical implementation.