Integrating unsupervised and supervised learning techniques to predict traumatic brain injury: A population-based...

Area of Science:

• Computational neuroscience
• Public health informatics
• Machine learning in healthcare

Background:

• Traumatic brain injury (TBI) poses a significant health burden.
• Identifying pre-existing conditions associated with TBI is crucial for prevention.
• Predictive modeling can aid in understanding TBI risk factors and external causes.

Purpose of the Study:

• To identify pre-existing health conditions in patients with TBI.
• To develop predictive models for the first TBI event and its external causes.
• To leverage unsupervised and supervised learning for TBI prediction.

Main Methods:

• Utilized Ontario Health Insurance Plan (OHIP) claims data for 488,107 TBI patients and matched controls.
• Applied Latent Dirichlet Allocation (LDA) for topic modeling of diagnostic codes, identifying 19 relevant topics.
• Employed Random Forest classifiers with topic scores and socio-demographic factors to predict TBI events and external causes.

Main Results:

• Screening identified 314 diagnostic codes significantly associated with TBI.
• LDA model successfully uncovered patterns of pre-morbid conditions.
• High predictive performance achieved for TBI event (AUC=0.85), falls (AUC=0.85), struck by/against (AUC=0.83), and motor vehicle collisions (AUC=0.83).

Conclusions:

• Machine learning models can effectively predict TBI and its external causes.
• The study identified significant pre-existing health conditions and factors contributing to TBI risk.
• This approach demonstrates the feasibility of using large-scale health data for TBI prediction and prevention strategies.