Abstract

Traumatic brain injury (TBI) is a deadly condition, standing as a major cause of death and long-term disability among young people and elders. In the UK it strikes about 110 new patients every day, the daily share of the roughly 40,000 TBI admissions recorded each year. Early outcome forecasts in traumatic brain injury guide life-saving surgical and intensive care choices, give families realistic expectations of recovery, and let hospitals focus scarce critical care and rehabilitation resources on patients who will benefit most. They also provide researchers and policy makers with a common standard for judging treatment quality. It was found that relying only on simple variables such as age, broad demographic information and the initial GCSM0 score failed to deliver reliable accuracy when trying to predict how patients would fare after a traumatic brain injury. In this study we worked with a richly detailed dataset containing many variables to pursue greater predictive accuracy, but choosing the most informative inputs required extensive feature engineering and the use of heuristics such as a genetic algorithm, which explored countless combinations of variables to uncover the sets that produced stronger model performance. By designing stronger features, we outperformed state of the art baselines in the initial admission stage, lifting accuracy for several algorithms; k nearest neighbor rose by 11 percent, random forest by 27 percent, naive Bayes by 30 percent and gradient boosting by 22 percent, with comparable gains seen in rule induction approaches. In future work we intend to explore deep learning approaches, strengthen the explainability of our machine learning models, and carry out a more detailed investigation into how and why specific factors drive predictive accuracy.

Keywords: heuristic feature engineering, prognosis, brain injury

How to Cite:

Alibakhshi, A., (2026) “Optimizing traumatic brain injury prognosis with heuristic feature engineering.”, New Vistas . doi: https://doi.org/10.36828/newvistas.363