Applications of Machine Learning in Prognostication of mTBI: A Systematic Review.

OBJECTIVE: To review the literature regarding the current state and clinical applicability of machine learning (ML) models in prognosticating the outcomes of patients with mild traumatic brain injury (mTBI) in the early clinical presentation. DESIGN: Databases were searched for studies including ML and mTBI from inception to March 10, 2023. Included studies had a primary outcome of predicting post-mTBI prognosis or sequalae. The Prediction model study Risk of Bias for Predictive Models assessment tool (PROBAST) was used for assessing the risk of bias and applicability of included studies. RESULTS: Out of 1235 articles, 10 met the inclusion criteria, including data from 127,929 patients. The most frequently used modeling techniques were Support Vector Machine (SVM) and Artificial Neural Network (NN) and Area Under the Curve (AUC) ranged from 0.66-0.889. Despite promise, several limitations to studies exist such as low sample sizes, database restrictions, inconsistencies in patient presentation definitions and lack of comparison to traditional clinical judgment or tools. CONCLUSION: ML models show potential in early stage mTBI prognostication, but to achieve widespread adoption, future clinical studies prognosticating mTBI using ML need to reduce bias, provide clarity and consistency in defining patient populations targeted, and validate against established benchmarks.

Predicting amputation using machine learning: A systematic review.

Amputation is an irreversible, last-line treatment indicated for a multitude of medical problems. Delaying amputation in favor of limb-sparing treatment may lead to increased risk of morbidity and mortality. This systematic review aims to synthesize the literature on how ML is being applied to predict amputation as an outcome. OVID Embase, OVID Medline, ACM Digital Library, Scopus, Web of Science, and IEEE Xplore were searched from inception to March 5, 2023. 1376 studies were screened; 15 articles were included. In the diabetic population, models ranged from sub-optimal to excellent performance (AUC: 0.6-0.94). In trauma patients, models had strong to excellent performance (AUC: 0.88-0.95). In patients who received amputation secondary to other etiologies (e.g.: burns and peripheral vascular disease), models had similar performance (AUC: 0.81-1.0). Many studies were found to have a high PROBAST risk of bias, most often due to small sample sizes. In conclusion, multiple machine learning models have been successfully developed that have the potential to be superior to traditional modeling techniques and prospective clinical judgment in predicting amputation. Further research is needed to overcome the limitations of current studies and to bring applicability to a clinical setting.