Quantifying impairment and disease severity using AI models trained on healthy subjects

Area of Science:

• Artificial Intelligence in Medicine
• Medical Diagnostics
• Digital Health

Background:

• Assessing disease severity and impairment automatically is crucial for data-driven medicine.
• Current clinical evaluations can be time-consuming and infrequent, limiting patient monitoring and treatment adaptation.

Purpose of the Study:

• To introduce a novel AI framework, the COnfidence-Based chaRacterization of Anomalies (COBRA) score, for quantifying disease severity.
• To demonstrate the COBRA score's effectiveness in evaluating upper-body impairment in stroke patients and knee osteoarthritis severity.

Main Methods:

• Developed an AI framework using models trained exclusively on healthy individuals.
• The COBRA score quantifies patient deviation from healthy norms by measuring model confidence.
• Applied the COBRA score to data from wearable sensors, video, and MRI scans.

Main Results:

• The COBRA score strongly correlated with the Fugl-Meyer Assessment for stroke patients using wearable sensors (ρ = 0.814) and video (ρ = 0.736).
• The COBRA score also showed significant correlation with clinical assessments for knee osteoarthritis severity from MRI scans (ρ = 0.644).
• The COBRA score is computed automatically in under one minute, significantly faster than traditional methods.

Conclusions:

• The COBRA score offers a rapid, automated method for assessing impairment and disease severity.
• This AI-driven approach demonstrates generalizability across different conditions and data modalities.
• The COBRA score has the potential to enhance patient monitoring and personalize rehabilitation protocols.