AI Detects Spinal Cord Disease 30 Months Early

AI Detects Spinal Cord Disease 30 Months Early

Cervical spondylotic myelopathy (CSM) is the leading cause of spinal cord dysfunction in older adults. Yet, diagnosis often comes far too late — sometimes years after symptoms begin. A groundbreaking study from Washington University in St. Louis (WashU) now shows that clinically informed AI models can predict CSM up to 30 months before doctors make a formal diagnosis. Moreover, these specialized models outperform larger, general-purpose foundation models in real-world clinical settings.

What Is Cervical Spondylotic Myelopathy?

A Progressive and Often Missed Condition

CSM occurs when arthritis in the neck compresses the spinal cord. It develops slowly, causing neck pain, muscle weakness, and difficulty walking. Because symptoms tend to appear gradually, clinicians frequently miss the condition until it reaches an advanced stage. By then, treatment options shrink significantly. Early intervention, therefore, is not just helpful — it is essential for better patient outcomes.

How AI Is Changing Early Diagnosis

Using Electronic Health Records to Predict Risk

A multidisciplinary team of surgeon-scientists, computer scientists, and researchers at WashU developed AI models that analyze structured electronic health record (EHR) data to identify CSM risk. The team trained these models using two datasets — a large external database of approximately two million patients from the Merative MarketScan claims database, and a smaller dataset from a St. Louis-based health system.

Their goal was clear. “We wanted to know if we could use the information within the electronic health record to try to identify these patients early enough and at a clinically relevant interval where we could potentially intervene at an earlier stage to lead to better outcomes,” said co-senior author Dr. Greenberg.

Furthermore, the study tested multiple modeling strategies. These ranged from simple, clinically guided models to large-scale foundation models pretrained on massive clinical datasets.

Clinically Informed vs. Foundation Models

Bigger Is Not Always Better

The research team evaluated both large foundation models — often referred to as “out-of-the-box” systems — and smaller, domain-informed models built with clinical knowledge. During internal testing on the larger dataset, foundation models performed slightly better overall. However, during external validation across a separate health system, clinically guided models proved more reliable.

This is a critical finding. A model must generalize well across different care settings to be useful in real clinical practice. Clinically informed models, despite their smaller size, delivered stronger and more consistent results across varied patient populations. As a result, the study highlights a key insight: clinical domain knowledge can compensate for smaller model size when real-world generalization matters.

The findings appear in the journal npj Digital Medicine, published in January 2026.

Why Early Detection Matters

Opening a Window for Treatment

Currently, many CSM patients receive a diagnosis only after significant neurological damage has already occurred. Earlier identification opens a critical treatment window. Clinicians can then recommend surgery, physical therapy, or lifestyle modifications before the condition worsens.

The WashU AI approach offers a 30-month prediction horizon. In practical terms, this means doctors could identify at-risk patients two and a half years before symptoms become disabling. Consequently, patients gain access to timely care that can meaningfully preserve their quality of life.

In addition, the AI models use data that already exists in hospital systems — EHR records — so there is no need for costly new diagnostic tests. This makes the approach both scalable and practical for healthcare systems worldwide.

What This Means for Patients and Clinicians

A New Standard for Spinal Cord Care

This research represents a major step forward in AI-assisted spinal cord care. Clinicians now have a tool that scans existing patient data to flag high-risk individuals far earlier than traditional methods allow. Additionally, the study underlines why domain-specific AI models — built with clinical insight — often surpass general-purpose models in healthcare applications.

Going forward, wider adoption of such tools could reduce diagnostic delays, cut healthcare costs, and — most importantly — improve lives. Patients who might have faced paralysis or severe disability now have a realistic chance of receiving care before irreversible damage occurs.

WashU’s work sets a compelling benchmark for how AI and clinical expertise together can transform patient outcomes in spinal cord disease management.