Computational Gene Mapping to Analyze Continuous Automated Real-Time Vital Signs Monitoring Data.

摘要

This project explored machine learning 'gene mapping' algorithms (MLA) as possible analytic platforms for advanced, forward-deployable patient care instrumentation. A patient database was developed to (1) identify physiologic data collected electronically by continuous automated critical care patient monitoring, (2) derive clinically useful vital signs (VS) 'features' of potential use in predicting patient functional outcome after severe brain trauma, (3) define outcomes, (4) train and test machine learning algorithms, and (5) cross-validate and finalize results. The computer-based experimental work developed in three stages: (1) feature selection by conventional univariate methodology, (2) feature selection using logistic regression, and (3) feature selection using other procedures that could modulate the limitations of the first two approaches. Stage 1 results suggested that long-term patient outcomes may be predictable using MLA and VS features gathered within the first 12 hours of critical care. Stage 2 results were much stronger but showed signs of 'overfitting,' an inherent pitfall of feature selection using logistic regression. Stage 3 results confirmed Stage 1 and 2 results with much better correlations, particularly for early (<6 weeks post discharge) and late (3-6 months) patient functional outcomes after severe brain trauma and using only the first 12 hours of continuous monitoring data.