Referral Notes:

  • The AI-generated COnfidence-Based chaRacterization of Anomalies (COBRA) score quantifies stroke motor impairment in under one minute using data from wearables and video.
  • It strongly correlates with the Fugl-Meyer Assessment, the gold-standard for stroke sensorimotor evaluation, which takes upward of an hour to complete.
  • COBRA is currently intended for research use only, but has the potential to support continuous and standardized monitoring of patient recovery.
  • The AI framework may be ideal for applications with limited patient data, as the model was trained on healthy subjects and detects deviations from the norm.

The evaluation of upper extremity impairment is key to guiding and adapting rehabilitation protocols that support recovery in patients with stroke. Current assessment tools, including the gold standard, the Fugl-Meyer Assessment (FMA), are time-consuming and require in-person administration by a trained clinician.

To solve this problem, Heidi Schambra, MD, an associate professor of neurology and rehabilitation medicine at NYU Langone Health; Carlos Fernandez-Granda, PhD, an associate professor of mathematics and data science at New York University; and their research teams developed a new framework. Based on artificial intelligence (AI) and data-driven, it automatically quantifies impairment. Their work was published in npj Digital Medicine.

“Think about having a system that effortlessly automates patient assessment and monitoring,” says Dr. Schambra. “It could provide real-time, continuous feedback, saving time and providing clinicians with actionable information.”

Anomaly Detection

The COnfidence-Based chaRacterization of Anomalies (COBRA) score was designed based on anomaly detection. The approach incorporates AI models trained exclusively on wearable sensor and video data from healthy individuals.

“Anomaly detection involves training a model to recognize patterns of movement that are normal or expected. If the model encounters an unfamiliar pattern, it will raise an alert.”

Heidi Schambra, MD

“Anomaly detection involves training a model to recognize patterns of movement that are normal or expected,” explains Dr. Schambra. “If the model encounters an unfamiliar pattern, it will raise an alert.”

The anomaly-detection framework is tailored to a specific medical condition, like stroke. This is achieved by training AI models to predict, for example, the upper body motions required for a healthy individual to raise a drinking glass with their arm. When the models are presented with data from a patient with arm impairments due to stroke, the average model confidence drops in proportion to severity, allowing for quantification.

“When presented with data from patients with impairments, the model quantifies their deviation from the healthy population,” says Dr. Schambra.

Moving Beyond the Gold Standard

According to Dr. Schambra, the FMA scale remains the gold standard, but it’s far from perfect. In fact, the tool is the current “rate-limiting step” in the evaluation of upper body impairment after stroke.

“The FMA takes specialized training in order to administer,” says Dr. Schambra. “It can take up to an hour to complete, which few busy therapists have the time for. After three hours of rehab each day, patients are also very tired. We need something faster and less burdensome for everyone.”

“We need something faster and less burdensome for everyone.”

In their recent publication, the research teams showed that the COBRA score, computed automatically in under one minute, strongly correlates with the FMA. They compared both assessments across several movements and activities, including arranging objects on a table or shelf, putting on eyeglasses, brushing hair and teeth, and eating. Although the system isn’t ready for full-scale clinical deployment, it’s an exciting first step.

“This technology is a way to support therapists in their delivery of rehabilitation training,” says Dr. Schambra. “They could use the COBRA scores to better focus on problem areas and to track what’s working and what’s not. It will enable us to deliver therapy in a more standardized fashion.”

Future Applications

For next steps, the researchers will prioritize refining the framework, moving beyond proof of concept. The team plans to expand their testing of the COBRA framework in hundreds of stroke patients to ensure its generalizability.

“Our results suggest that fine-tuned annotations describing clinically relevant attributes can be useful, even if they are only available for healthy individuals.”

Moreover, the framework could be suitable for applications where it is difficult to obtain large-scale databases of patients with different degrees of impairment or severity, since it only requires data from a healthy cohort of moderate size, Dr. Schambra explains. For instance, the team is applying the framework to quantify knee osteoarthritis severity using MRI scans. “Our results suggest that fine-tuned annotations describing clinically relevant attributes can be useful, even if they are only available for healthy individuals,” notes Dr. Schambra.