Development of Multimodal Machine Learning-Based Prognostic Models for Traumatic Brain Injury Using Clinical Data and Computed Tomography Scans

Abstract

Traumatic brain injury (TBI), a disruption of brain function caused by external forces to the head, is a leading cause of death and disability for trauma patients in the world. In TBI patient care, prognostication—the prediction of disease course and outcome—is essential for planning treatment strategies, patient disposition, and allocation of clinical resources. Nowadays, healthcare professionals manually prognosticate TBI patient outcomes based on clinical and computed tomography (CT) findings. However, accurate and timely prognostication based on multi-modal clinical data requires specialized expertise and can be challenging to inexperienced or early-career clinicians. This thesis has established a machine learning (ML) framework for the timely and standardized prognostication for TBI patients that eliminates the need for manual CT assessments. The developed ML-based prognostic model has demonstrated its ability to predict six-month outcomes for TBI patients, achieving comparable or superior results to conventional models necessitating manual CT interpretations. To facilitate the clinical adoption of this model, a novel weakly supervised anomaly detection (WSAD) algorithm was introduced. This algorithm could efficiently identify TBI-relevant CT slices, which might be used as data for retraining or fine-tuning the prognostic model deployed in a new domain. For the successful deployment of the developed prognostic model in real-world clinical settings, its practical implementation and effectiveness are paramount. To this end, a focus group study was carried out to gather insights from stakeholders involved in TBI care. The comprehensive framework presented in this thesis aims to support clinical decision-making, potentially reducing variability in prognoses that arises from clinician expertise or CT interpretation disparities, thereby contributing to the democratization and standardization of TBI prognostication practices.

Ähnliche Arbeiten