Machine Learning for Subtyping Concussion Using a Clustering Approach.

Area of Science:

• Neurology
• Data Science
• Biostatistics

Background:

• Traditional concussion subtyping methods may be limited by conceptual bias and interdisciplinary expertise.
• A novel approach is needed to overcome the limitations of current concussion classification systems.

Purpose of the Study:

• To investigate the efficacy of an unsupervised machine learning approach for concussion subtyping.
• To determine if a data-driven method can provide a more accurate concussion classification.

Main Methods:

• Retrospective analysis of patient data from a concussion clinic, including PROMIS, DHI, PCS, and ImPACT scores.
• Application of principal component analysis (PCA) and agglomerative clustering for dimensionality reduction and subtype identification.
• Statistical comparison of identified clusters using the Mann-Whitney U test.

Main Results:

• Five distinct concussion subtypes were identified from 275 patient records.
• These subtypes exhibited statistically significant clinical differences (p < 0.05).
• The optimal clustering was validated using Silhouette and Davies-Bouldin scores.

Conclusions:

• Machine learning successfully identified five clinically distinct concussion subtypes.
• These subtypes are best characterized by levels of complexity, from Minimally Complex to Extremely Complex.
• AI-driven concussion classification offers a more accurate reflection of mild traumatic brain injury heterogeneity.